fs/btrfs/inode.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/kernel.h>
   7 #include <linux/bio.h>
   8 #include <linux/buffer_head.h>
   9 #include <linux/file.h>
  10 #include <linux/fs.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/highmem.h>
  13 #include <linux/time.h>
  14 #include <linux/init.h>
  15 #include <linux/string.h>
  16 #include <linux/backing-dev.h>
  17 #include <linux/writeback.h>
  18 #include <linux/compat.h>
  19 #include <linux/xattr.h>
  20 #include <linux/posix_acl.h>
  21 #include <linux/falloc.h>
  22 #include <linux/slab.h>
  23 #include <linux/ratelimit.h>
  24 #include <linux/btrfs.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/posix_acl_xattr.h>
  27 #include <linux/uio.h>
  28 #include <linux/magic.h>
  29 #include <linux/iversion.h>
  30 #include <linux/swap.h>
  31 #include <asm/unaligned.h>
  32 #include "ctree.h"
  33 #include "disk-io.h"
  34 #include "transaction.h"
  35 #include "btrfs_inode.h"
  36 #include "print-tree.h"
  37 #include "ordered-data.h"
  38 #include "xattr.h"
  39 #include "tree-log.h"
  40 #include "volumes.h"
  41 #include "compression.h"
  42 #include "locking.h"
  43 #include "free-space-cache.h"
  44 #include "inode-map.h"
  45 #include "backref.h"
  46 #include "props.h"
  47 #include "qgroup.h"
  48 #include "dedupe.h"
  49
  50 struct btrfs_iget_args {
  51         struct btrfs_key *location;
  52         struct btrfs_root *root;
  53 };
  54
  55 struct btrfs_dio_data {
  56         u64 reserve;
  57         u64 unsubmitted_oe_range_start;
  58         u64 unsubmitted_oe_range_end;
  59         int overwrite;
  60 };
  61
  62 static const struct inode_operations btrfs_dir_inode_operations;
  63 static const struct inode_operations btrfs_symlink_inode_operations;
  64 static const struct inode_operations btrfs_dir_ro_inode_operations;
  65 static const struct inode_operations btrfs_special_inode_operations;
  66 static const struct inode_operations btrfs_file_inode_operations;
  67 static const struct address_space_operations btrfs_aops;
  68 static const struct file_operations btrfs_dir_file_operations;
  69 static const struct extent_io_ops btrfs_extent_io_ops;
  70
  71 static struct kmem_cache *btrfs_inode_cachep;
  72 struct kmem_cache *btrfs_trans_handle_cachep;
  73 struct kmem_cache *btrfs_path_cachep;
  74 struct kmem_cache *btrfs_free_space_cachep;
  75
  76 #define S_SHIFT 12
  77 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  78         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  79         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  80         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  81         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  82         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  83         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
  84         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
  85 };
  86
  87 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
  88 static int btrfs_truncate(struct inode *inode, bool skip_writeback);
  89 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
  90 static noinline int cow_file_range(struct inode *inode,
  91                                    struct page *locked_page,
  92                                    u64 start, u64 end, u64 delalloc_end,
  93                                    int *page_started, unsigned long *nr_written,
  94                                    int unlock, struct btrfs_dedupe_hash *hash);
  95 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
  96                                        u64 orig_start, u64 block_start,
  97                                        u64 block_len, u64 orig_block_len,
  98                                        u64 ram_bytes, int compress_type,
  99                                        int type);
 100
 101 static void __endio_write_update_ordered(struct inode *inode,
 102                                          const u64 offset, const u64 bytes,
 103                                          const bool uptodate);
 104
 105 /*
 106  * Cleanup all submitted ordered extents in specified range to handle errors
 107  * from the btrfs_run_delalloc_range() callback.
 108  *
 109  * NOTE: caller must ensure that when an error happens, it can not call
 110  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 111  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 112  * to be released, which we want to happen only when finishing the ordered
 113  * extent (btrfs_finish_ordered_io()).
 114  */
 115 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 116                                                  struct page *locked_page,
 117                                                  u64 offset, u64 bytes)
 118 {
 119         unsigned long index = offset >> PAGE_SHIFT;
 120         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 121         u64 page_start = page_offset(locked_page);
 122         u64 page_end = page_start + PAGE_SIZE - 1;
 123
 124         struct page *page;
 125
 126         while (index <= end_index) {
 127                 page = find_get_page(inode->i_mapping, index);
 128                 index++;
 129                 if (!page)
 130                         continue;
 131                 ClearPagePrivate2(page);
 132                 put_page(page);
 133         }
 134
 135         /*
 136          * In case this page belongs to the delalloc range being instantiated
 137          * then skip it, since the first page of a range is going to be
 138          * properly cleaned up by the caller of run_delalloc_range
 139          */
 140         if (page_start >= offset && page_end <= (offset + bytes - 1)) {
 141                 offset += PAGE_SIZE;
 142                 bytes -= PAGE_SIZE;
 143         }
 144
 145         return __endio_write_update_ordered(inode, offset, bytes, false);
 146 }
 147
 148 static int btrfs_dirty_inode(struct inode *inode);
 149
 150 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 151 void btrfs_test_inode_set_ops(struct inode *inode)
 152 {
 153         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 154 }
 155 #endif
 156
 157 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 158                                      struct inode *inode,  struct inode *dir,
 159                                      const struct qstr *qstr)
 160 {
 161         int err;
 162
 163         err = btrfs_init_acl(trans, inode, dir);
 164         if (!err)
 165                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 166         return err;
 167 }
 168
 169 /*
 170  * this does all the hard work for inserting an inline extent into
 171  * the btree.  The caller should have done a btrfs_drop_extents so that
 172  * no overlapping inline items exist in the btree
 173  */
 174 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 175                                 struct btrfs_path *path, int extent_inserted,
 176                                 struct btrfs_root *root, struct inode *inode,
 177                                 u64 start, size_t size, size_t compressed_size,
 178                                 int compress_type,
 179                                 struct page **compressed_pages)
 180 {
 181         struct extent_buffer *leaf;
 182         struct page *page = NULL;
 183         char *kaddr;
 184         unsigned long ptr;
 185         struct btrfs_file_extent_item *ei;
 186         int ret;
 187         size_t cur_size = size;
 188         unsigned long offset;
 189
 190         if (compressed_size && compressed_pages)
 191                 cur_size = compressed_size;
 192
 193         inode_add_bytes(inode, size);
 194
 195         if (!extent_inserted) {
 196                 struct btrfs_key key;
 197                 size_t datasize;
 198
 199                 key.objectid = btrfs_ino(BTRFS_I(inode));
 200                 key.offset = start;
 201                 key.type = BTRFS_EXTENT_DATA_KEY;
 202
 203                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 204                 path->leave_spinning = 1;
 205                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 206                                               datasize);
 207                 if (ret)
 208                         goto fail;
 209         }
 210         leaf = path->nodes[0];
 211         ei = btrfs_item_ptr(leaf, path->slots[0],
 212                             struct btrfs_file_extent_item);
 213         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 214         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 215         btrfs_set_file_extent_encryption(leaf, ei, 0);
 216         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 217         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 218         ptr = btrfs_file_extent_inline_start(ei);
 219
 220         if (compress_type != BTRFS_COMPRESS_NONE) {
 221                 struct page *cpage;
 222                 int i = 0;
 223                 while (compressed_size > 0) {
 224                         cpage = compressed_pages[i];
 225                         cur_size = min_t(unsigned long, compressed_size,
 226                                        PAGE_SIZE);
 227
 228                         kaddr = kmap_atomic(cpage);
 229                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 230                         kunmap_atomic(kaddr);
 231
 232                         i++;
 233                         ptr += cur_size;
 234                         compressed_size -= cur_size;
 235                 }
 236                 btrfs_set_file_extent_compression(leaf, ei,
 237                                                   compress_type);
 238         } else {
 239                 page = find_get_page(inode->i_mapping,
 240                                      start >> PAGE_SHIFT);
 241                 btrfs_set_file_extent_compression(leaf, ei, 0);
 242                 kaddr = kmap_atomic(page);
 243                 offset = offset_in_page(start);
 244                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 245                 kunmap_atomic(kaddr);
 246                 put_page(page);
 247         }
 248         btrfs_mark_buffer_dirty(leaf);
 249         btrfs_release_path(path);
 250
 251         /*
 252          * we're an inline extent, so nobody can
 253          * extend the file past i_size without locking
 254          * a page we already have locked.
 255          *
 256          * We must do any isize and inode updates
 257          * before we unlock the pages.  Otherwise we
 258          * could end up racing with unlink.
 259          */
 260         BTRFS_I(inode)->disk_i_size = inode->i_size;
 261         ret = btrfs_update_inode(trans, root, inode);
 262
 263 fail:
 264         return ret;
 265 }
 266
 267
 268 /*
 269  * conditionally insert an inline extent into the file.  This
 270  * does the checks required to make sure the data is small enough
 271  * to fit as an inline extent.
 272  */
 273 static noinline int cow_file_range_inline(struct inode *inode, u64 start,
 274                                           u64 end, size_t compressed_size,
 275                                           int compress_type,
 276                                           struct page **compressed_pages)
 277 {
 278         struct btrfs_root *root = BTRFS_I(inode)->root;
 279         struct btrfs_fs_info *fs_info = root->fs_info;
 280         struct btrfs_trans_handle *trans;
 281         u64 isize = i_size_read(inode);
 282         u64 actual_end = min(end + 1, isize);
 283         u64 inline_len = actual_end - start;
 284         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 285         u64 data_len = inline_len;
 286         int ret;
 287         struct btrfs_path *path;
 288         int extent_inserted = 0;
 289         u32 extent_item_size;
 290
 291         if (compressed_size)
 292                 data_len = compressed_size;
 293
 294         if (start > 0 ||
 295             actual_end > fs_info->sectorsize ||
 296             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 297             (!compressed_size &&
 298             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 299             end + 1 < isize ||
 300             data_len > fs_info->max_inline) {
 301                 return 1;
 302         }
 303
 304         path = btrfs_alloc_path();
 305         if (!path)
 306                 return -ENOMEM;
 307
 308         trans = btrfs_join_transaction(root);
 309         if (IS_ERR(trans)) {
 310                 btrfs_free_path(path);
 311                 return PTR_ERR(trans);
 312         }
 313         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 314
 315         if (compressed_size && compressed_pages)
 316                 extent_item_size = btrfs_file_extent_calc_inline_size(
 317                    compressed_size);
 318         else
 319                 extent_item_size = btrfs_file_extent_calc_inline_size(
 320                     inline_len);
 321
 322         ret = __btrfs_drop_extents(trans, root, inode, path,
 323                                    start, aligned_end, NULL,
 324                                    1, 1, extent_item_size, &extent_inserted);
 325         if (ret) {
 326                 btrfs_abort_transaction(trans, ret);
 327                 goto out;
 328         }
 329
 330         if (isize > actual_end)
 331                 inline_len = min_t(u64, isize, actual_end);
 332         ret = insert_inline_extent(trans, path, extent_inserted,
 333                                    root, inode, start,
 334                                    inline_len, compressed_size,
 335                                    compress_type, compressed_pages);
 336         if (ret && ret != -ENOSPC) {
 337                 btrfs_abort_transaction(trans, ret);
 338                 goto out;
 339         } else if (ret == -ENOSPC) {
 340                 ret = 1;
 341                 goto out;
 342         }
 343
 344         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 345         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 346 out:
 347         /*
 348          * Don't forget to free the reserved space, as for inlined extent
 349          * it won't count as data extent, free them directly here.
 350          * And at reserve time, it's always aligned to page size, so
 351          * just free one page here.
 352          */
 353         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 354         btrfs_free_path(path);
 355         btrfs_end_transaction(trans);
 356         return ret;
 357 }
 358
 359 struct async_extent {
 360         u64 start;
 361         u64 ram_size;
 362         u64 compressed_size;
 363         struct page **pages;
 364         unsigned long nr_pages;
 365         int compress_type;
 366         struct list_head list;
 367 };
 368
 369 struct async_cow {
 370         struct inode *inode;
 371         struct btrfs_fs_info *fs_info;
 372         struct page *locked_page;
 373         u64 start;
 374         u64 end;
 375         unsigned int write_flags;
 376         struct list_head extents;
 377         struct btrfs_work work;
 378 };
 379
 380 static noinline int add_async_extent(struct async_cow *cow,
 381                                      u64 start, u64 ram_size,
 382                                      u64 compressed_size,
 383                                      struct page **pages,
 384                                      unsigned long nr_pages,
 385                                      int compress_type)
 386 {
 387         struct async_extent *async_extent;
 388
 389         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 390         BUG_ON(!async_extent); /* -ENOMEM */
 391         async_extent->start = start;
 392         async_extent->ram_size = ram_size;
 393         async_extent->compressed_size = compressed_size;
 394         async_extent->pages = pages;
 395         async_extent->nr_pages = nr_pages;
 396         async_extent->compress_type = compress_type;
 397         list_add_tail(&async_extent->list, &cow->extents);
 398         return 0;
 399 }
 400
 401 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 402 {
 403         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 404
 405         /* force compress */
 406         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 407                 return 1;
 408         /* defrag ioctl */
 409         if (BTRFS_I(inode)->defrag_compress)
 410                 return 1;
 411         /* bad compression ratios */
 412         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 413                 return 0;
 414         if (btrfs_test_opt(fs_info, COMPRESS) ||
 415             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 416             BTRFS_I(inode)->prop_compress)
 417                 return btrfs_compress_heuristic(inode, start, end);
 418         return 0;
 419 }
 420
 421 static inline void inode_should_defrag(struct btrfs_inode *inode,
 422                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 423 {
 424         /* If this is a small write inside eof, kick off a defrag */
 425         if (num_bytes < small_write &&
 426             (start > 0 || end + 1 < inode->disk_i_size))
 427                 btrfs_add_inode_defrag(NULL, inode);
 428 }
 429
 430 /*
 431  * we create compressed extents in two phases.  The first
 432  * phase compresses a range of pages that have already been
 433  * locked (both pages and state bits are locked).
 434  *
 435  * This is done inside an ordered work queue, and the compression
 436  * is spread across many cpus.  The actual IO submission is step
 437  * two, and the ordered work queue takes care of making sure that
 438  * happens in the same order things were put onto the queue by
 439  * writepages and friends.
 440  *
 441  * If this code finds it can't get good compression, it puts an
 442  * entry onto the work queue to write the uncompressed bytes.  This
 443  * makes sure that both compressed inodes and uncompressed inodes
 444  * are written in the same order that the flusher thread sent them
 445  * down.
 446  */
 447 static noinline void compress_file_range(struct inode *inode,
 448                                         struct page *locked_page,
 449                                         u64 start, u64 end,
 450                                         struct async_cow *async_cow,
 451                                         int *num_added)
 452 {
 453         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 454         u64 blocksize = fs_info->sectorsize;
 455         u64 actual_end;
 456         u64 isize = i_size_read(inode);
 457         int ret = 0;
 458         struct page **pages = NULL;
 459         unsigned long nr_pages;
 460         unsigned long total_compressed = 0;
 461         unsigned long total_in = 0;
 462         int i;
 463         int will_compress;
 464         int compress_type = fs_info->compress_type;
 465         int redirty = 0;
 466
 467         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 468                         SZ_16K);
 469
 470         actual_end = min_t(u64, isize, end + 1);
 471 again:
 472         will_compress = 0;
 473         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 474         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 475         nr_pages = min_t(unsigned long, nr_pages,
 476                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 477
 478         /*
 479          * we don't want to send crud past the end of i_size through
 480          * compression, that's just a waste of CPU time.  So, if the
 481          * end of the file is before the start of our current
 482          * requested range of bytes, we bail out to the uncompressed
 483          * cleanup code that can deal with all of this.
 484          *
 485          * It isn't really the fastest way to fix things, but this is a
 486          * very uncommon corner.
 487          */
 488         if (actual_end <= start)
 489                 goto cleanup_and_bail_uncompressed;
 490
 491         total_compressed = actual_end - start;
 492
 493         /*
 494          * skip compression for a small file range(<=blocksize) that
 495          * isn't an inline extent, since it doesn't save disk space at all.
 496          */
 497         if (total_compressed <= blocksize &&
 498            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 499                 goto cleanup_and_bail_uncompressed;
 500
 501         total_compressed = min_t(unsigned long, total_compressed,
 502                         BTRFS_MAX_UNCOMPRESSED);
 503         total_in = 0;
 504         ret = 0;
 505
 506         /*
 507          * we do compression for mount -o compress and when the
 508          * inode has not been flagged as nocompress.  This flag can
 509          * change at any time if we discover bad compression ratios.
 510          */
 511         if (inode_need_compress(inode, start, end)) {
 512                 WARN_ON(pages);
 513                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 514                 if (!pages) {
 515                         /* just bail out to the uncompressed code */
 516                         nr_pages = 0;
 517                         goto cont;
 518                 }
 519
 520                 if (BTRFS_I(inode)->defrag_compress)
 521                         compress_type = BTRFS_I(inode)->defrag_compress;
 522                 else if (BTRFS_I(inode)->prop_compress)
 523                         compress_type = BTRFS_I(inode)->prop_compress;
 524
 525                 /*
 526                  * we need to call clear_page_dirty_for_io on each
 527                  * page in the range.  Otherwise applications with the file
 528                  * mmap'd can wander in and change the page contents while
 529                  * we are compressing them.
 530                  *
 531                  * If the compression fails for any reason, we set the pages
 532                  * dirty again later on.
 533                  *
 534                  * Note that the remaining part is redirtied, the start pointer
 535                  * has moved, the end is the original one.
 536                  */
 537                 if (!redirty) {
 538                         extent_range_clear_dirty_for_io(inode, start, end);
 539                         redirty = 1;
 540                 }
 541
 542                 /* Compression level is applied here and only here */
 543                 ret = btrfs_compress_pages(
 544                         compress_type | (fs_info->compress_level << 4),
 545                                            inode->i_mapping, start,
 546                                            pages,
 547                                            &nr_pages,
 548                                            &total_in,
 549                                            &total_compressed);
 550
 551                 if (!ret) {
 552                         unsigned long offset = offset_in_page(total_compressed);
 553                         struct page *page = pages[nr_pages - 1];
 554                         char *kaddr;
 555
 556                         /* zero the tail end of the last page, we might be
 557                          * sending it down to disk
 558                          */
 559                         if (offset) {
 560                                 kaddr = kmap_atomic(page);
 561                                 memset(kaddr + offset, 0,
 562                                        PAGE_SIZE - offset);
 563                                 kunmap_atomic(kaddr);
 564                         }
 565                         will_compress = 1;
 566                 }
 567         }
 568 cont:
 569         if (start == 0) {
 570                 /* lets try to make an inline extent */
 571                 if (ret || total_in < actual_end) {
 572                         /* we didn't compress the entire range, try
 573                          * to make an uncompressed inline extent.
 574                          */
 575                         ret = cow_file_range_inline(inode, start, end, 0,
 576                                                     BTRFS_COMPRESS_NONE, NULL);
 577                 } else {
 578                         /* try making a compressed inline extent */
 579                         ret = cow_file_range_inline(inode, start, end,
 580                                                     total_compressed,
 581                                                     compress_type, pages);
 582                 }
 583                 if (ret <= 0) {
 584                         unsigned long clear_flags = EXTENT_DELALLOC |
 585                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 586                                 EXTENT_DO_ACCOUNTING;
 587                         unsigned long page_error_op;
 588
 589                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 590
 591                         /*
 592                          * inline extent creation worked or returned error,
 593                          * we don't need to create any more async work items.
 594                          * Unlock and free up our temp pages.
 595                          *
 596                          * We use DO_ACCOUNTING here because we need the
 597                          * delalloc_release_metadata to be done _after_ we drop
 598                          * our outstanding extent for clearing delalloc for this
 599                          * range.
 600                          */
 601                         extent_clear_unlock_delalloc(inode, start, end, end,
 602                                                      NULL, clear_flags,
 603                                                      PAGE_UNLOCK |
 604                                                      PAGE_CLEAR_DIRTY |
 605                                                      PAGE_SET_WRITEBACK |
 606                                                      page_error_op |
 607                                                      PAGE_END_WRITEBACK);
 608                         goto free_pages_out;
 609                 }
 610         }
 611
 612         if (will_compress) {
 613                 /*
 614                  * we aren't doing an inline extent round the compressed size
 615                  * up to a block size boundary so the allocator does sane
 616                  * things
 617                  */
 618                 total_compressed = ALIGN(total_compressed, blocksize);
 619
 620                 /*
 621                  * one last check to make sure the compression is really a
 622                  * win, compare the page count read with the blocks on disk,
 623                  * compression must free at least one sector size
 624                  */
 625                 total_in = ALIGN(total_in, PAGE_SIZE);
 626                 if (total_compressed + blocksize <= total_in) {
 627                         *num_added += 1;
 628
 629                         /*
 630                          * The async work queues will take care of doing actual
 631                          * allocation on disk for these compressed pages, and
 632                          * will submit them to the elevator.
 633                          */
 634                         add_async_extent(async_cow, start, total_in,
 635                                         total_compressed, pages, nr_pages,
 636                                         compress_type);
 637
 638                         if (start + total_in < end) {
 639                                 start += total_in;
 640                                 pages = NULL;
 641                                 cond_resched();
 642                                 goto again;
 643                         }
 644                         return;
 645                 }
 646         }
 647         if (pages) {
 648                 /*
 649                  * the compression code ran but failed to make things smaller,
 650                  * free any pages it allocated and our page pointer array
 651                  */
 652                 for (i = 0; i < nr_pages; i++) {
 653                         WARN_ON(pages[i]->mapping);
 654                         put_page(pages[i]);
 655                 }
 656                 kfree(pages);
 657                 pages = NULL;
 658                 total_compressed = 0;
 659                 nr_pages = 0;
 660
 661                 /* flag the file so we don't compress in the future */
 662                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 663                     !(BTRFS_I(inode)->prop_compress)) {
 664                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 665                 }
 666         }
 667 cleanup_and_bail_uncompressed:
 668         /*
 669          * No compression, but we still need to write the pages in the file
 670          * we've been given so far.  redirty the locked page if it corresponds
 671          * to our extent and set things up for the async work queue to run
 672          * cow_file_range to do the normal delalloc dance.
 673          */
 674         if (page_offset(locked_page) >= start &&
 675             page_offset(locked_page) <= end)
 676                 __set_page_dirty_nobuffers(locked_page);
 677                 /* unlocked later on in the async handlers */
 678
 679         if (redirty)
 680                 extent_range_redirty_for_io(inode, start, end);
 681         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 682                          BTRFS_COMPRESS_NONE);
 683         *num_added += 1;
 684
 685         return;
 686
 687 free_pages_out:
 688         for (i = 0; i < nr_pages; i++) {
 689                 WARN_ON(pages[i]->mapping);
 690                 put_page(pages[i]);
 691         }
 692         kfree(pages);
 693 }
 694
 695 static void free_async_extent_pages(struct async_extent *async_extent)
 696 {
 697         int i;
 698
 699         if (!async_extent->pages)
 700                 return;
 701
 702         for (i = 0; i < async_extent->nr_pages; i++) {
 703                 WARN_ON(async_extent->pages[i]->mapping);
 704                 put_page(async_extent->pages[i]);
 705         }
 706         kfree(async_extent->pages);
 707         async_extent->nr_pages = 0;
 708         async_extent->pages = NULL;
 709 }
 710
 711 /*
 712  * phase two of compressed writeback.  This is the ordered portion
 713  * of the code, which only gets called in the order the work was
 714  * queued.  We walk all the async extents created by compress_file_range
 715  * and send them down to the disk.
 716  */
 717 static noinline void submit_compressed_extents(struct inode *inode,
 718                                               struct async_cow *async_cow)
 719 {
 720         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 721         struct async_extent *async_extent;
 722         u64 alloc_hint = 0;
 723         struct btrfs_key ins;
 724         struct extent_map *em;
 725         struct btrfs_root *root = BTRFS_I(inode)->root;
 726         struct extent_io_tree *io_tree;
 727         int ret = 0;
 728
 729 again:
 730         while (!list_empty(&async_cow->extents)) {
 731                 async_extent = list_entry(async_cow->extents.next,
 732                                           struct async_extent, list);
 733                 list_del(&async_extent->list);
 734
 735                 io_tree = &BTRFS_I(inode)->io_tree;
 736
 737 retry:
 738                 /* did the compression code fall back to uncompressed IO? */
 739                 if (!async_extent->pages) {
 740                         int page_started = 0;
 741                         unsigned long nr_written = 0;
 742
 743                         lock_extent(io_tree, async_extent->start,
 744                                          async_extent->start +
 745                                          async_extent->ram_size - 1);
 746
 747                         /* allocate blocks */
 748                         ret = cow_file_range(inode, async_cow->locked_page,
 749                                              async_extent->start,
 750                                              async_extent->start +
 751                                              async_extent->ram_size - 1,
 752                                              async_extent->start +
 753                                              async_extent->ram_size - 1,
 754                                              &page_started, &nr_written, 0,
 755                                              NULL);
 756
 757                         /* JDM XXX */
 758
 759                         /*
 760                          * if page_started, cow_file_range inserted an
 761                          * inline extent and took care of all the unlocking
 762                          * and IO for us.  Otherwise, we need to submit
 763                          * all those pages down to the drive.
 764                          */
 765                         if (!page_started && !ret)
 766                                 extent_write_locked_range(inode,
 767                                                   async_extent->start,
 768                                                   async_extent->start +
 769                                                   async_extent->ram_size - 1,
 770                                                   WB_SYNC_ALL);
 771                         else if (ret)
 772                                 unlock_page(async_cow->locked_page);
 773                         kfree(async_extent);
 774                         cond_resched();
 775                         continue;
 776                 }
 777
 778                 lock_extent(io_tree, async_extent->start,
 779                             async_extent->start + async_extent->ram_size - 1);
 780
 781                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 782                                            async_extent->compressed_size,
 783                                            async_extent->compressed_size,
 784                                            0, alloc_hint, &ins, 1, 1);
 785                 if (ret) {
 786                         free_async_extent_pages(async_extent);
 787
 788                         if (ret == -ENOSPC) {
 789                                 unlock_extent(io_tree, async_extent->start,
 790                                               async_extent->start +
 791                                               async_extent->ram_size - 1);
 792
 793                                 /*
 794                                  * we need to redirty the pages if we decide to
 795                                  * fallback to uncompressed IO, otherwise we
 796                                  * will not submit these pages down to lower
 797                                  * layers.
 798                                  */
 799                                 extent_range_redirty_for_io(inode,
 800                                                 async_extent->start,
 801                                                 async_extent->start +
 802                                                 async_extent->ram_size - 1);
 803
 804                                 goto retry;
 805                         }
 806                         goto out_free;
 807                 }
 808                 /*
 809                  * here we're doing allocation and writeback of the
 810                  * compressed pages
 811                  */
 812                 em = create_io_em(inode, async_extent->start,
 813                                   async_extent->ram_size, /* len */
 814                                   async_extent->start, /* orig_start */
 815                                   ins.objectid, /* block_start */
 816                                   ins.offset, /* block_len */
 817                                   ins.offset, /* orig_block_len */
 818                                   async_extent->ram_size, /* ram_bytes */
 819                                   async_extent->compress_type,
 820                                   BTRFS_ORDERED_COMPRESSED);
 821                 if (IS_ERR(em))
 822                         /* ret value is not necessary due to void function */
 823                         goto out_free_reserve;
 824                 free_extent_map(em);
 825
 826                 ret = btrfs_add_ordered_extent_compress(inode,
 827                                                 async_extent->start,
 828                                                 ins.objectid,
 829                                                 async_extent->ram_size,
 830                                                 ins.offset,
 831                                                 BTRFS_ORDERED_COMPRESSED,
 832                                                 async_extent->compress_type);
 833                 if (ret) {
 834                         btrfs_drop_extent_cache(BTRFS_I(inode),
 835                                                 async_extent->start,
 836                                                 async_extent->start +
 837                                                 async_extent->ram_size - 1, 0);
 838                         goto out_free_reserve;
 839                 }
 840                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 841
 842                 /*
 843                  * clear dirty, set writeback and unlock the pages.
 844                  */
 845                 extent_clear_unlock_delalloc(inode, async_extent->start,
 846                                 async_extent->start +
 847                                 async_extent->ram_size - 1,
 848                                 async_extent->start +
 849                                 async_extent->ram_size - 1,
 850                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 851                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 852                                 PAGE_SET_WRITEBACK);
 853                 if (btrfs_submit_compressed_write(inode,
 854                                     async_extent->start,
 855                                     async_extent->ram_size,
 856                                     ins.objectid,
 857                                     ins.offset, async_extent->pages,
 858                                     async_extent->nr_pages,
 859                                     async_cow->write_flags)) {
 860                         struct page *p = async_extent->pages[0];
 861                         const u64 start = async_extent->start;
 862                         const u64 end = start + async_extent->ram_size - 1;
 863
 864                         p->mapping = inode->i_mapping;
 865                         btrfs_writepage_endio_finish_ordered(p, start, end, 0);
 866
 867                         p->mapping = NULL;
 868                         extent_clear_unlock_delalloc(inode, start, end, end,
 869                                                      NULL, 0,
 870                                                      PAGE_END_WRITEBACK |
 871                                                      PAGE_SET_ERROR);
 872                         free_async_extent_pages(async_extent);
 873                 }
 874                 alloc_hint = ins.objectid + ins.offset;
 875                 kfree(async_extent);
 876                 cond_resched();
 877         }
 878         return;
 879 out_free_reserve:
 880         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 881         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 882 out_free:
 883         extent_clear_unlock_delalloc(inode, async_extent->start,
 884                                      async_extent->start +
 885                                      async_extent->ram_size - 1,
 886                                      async_extent->start +
 887                                      async_extent->ram_size - 1,
 888                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 889                                      EXTENT_DELALLOC_NEW |
 890                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 891                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 892                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 893                                      PAGE_SET_ERROR);
 894         free_async_extent_pages(async_extent);
 895         kfree(async_extent);
 896         goto again;
 897 }
 898
 899 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 900                                       u64 num_bytes)
 901 {
 902         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 903         struct extent_map *em;
 904         u64 alloc_hint = 0;
 905
 906         read_lock(&em_tree->lock);
 907         em = search_extent_mapping(em_tree, start, num_bytes);
 908         if (em) {
 909                 /*
 910                  * if block start isn't an actual block number then find the
 911                  * first block in this inode and use that as a hint.  If that
 912                  * block is also bogus then just don't worry about it.
 913                  */
 914                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 915                         free_extent_map(em);
 916                         em = search_extent_mapping(em_tree, 0, 0);
 917                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 918                                 alloc_hint = em->block_start;
 919                         if (em)
 920                                 free_extent_map(em);
 921                 } else {
 922                         alloc_hint = em->block_start;
 923                         free_extent_map(em);
 924                 }
 925         }
 926         read_unlock(&em_tree->lock);
 927
 928         return alloc_hint;
 929 }
 930
 931 /*
 932  * when extent_io.c finds a delayed allocation range in the file,
 933  * the call backs end up in this code.  The basic idea is to
 934  * allocate extents on disk for the range, and create ordered data structs
 935  * in ram to track those extents.
 936  *
 937  * locked_page is the page that writepage had locked already.  We use
 938  * it to make sure we don't do extra locks or unlocks.
 939  *
 940  * *page_started is set to one if we unlock locked_page and do everything
 941  * required to start IO on it.  It may be clean and already done with
 942  * IO when we return.
 943  */
 944 static noinline int cow_file_range(struct inode *inode,
 945                                    struct page *locked_page,
 946                                    u64 start, u64 end, u64 delalloc_end,
 947                                    int *page_started, unsigned long *nr_written,
 948                                    int unlock, struct btrfs_dedupe_hash *hash)
 949 {
 950         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 951         struct btrfs_root *root = BTRFS_I(inode)->root;
 952         u64 alloc_hint = 0;
 953         u64 num_bytes;
 954         unsigned long ram_size;
 955         u64 cur_alloc_size = 0;
 956         u64 blocksize = fs_info->sectorsize;
 957         struct btrfs_key ins;
 958         struct extent_map *em;
 959         unsigned clear_bits;
 960         unsigned long page_ops;
 961         bool extent_reserved = false;
 962         int ret = 0;
 963
 964         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 965                 WARN_ON_ONCE(1);
 966                 ret = -EINVAL;
 967                 goto out_unlock;
 968         }
 969
 970         num_bytes = ALIGN(end - start + 1, blocksize);
 971         num_bytes = max(blocksize,  num_bytes);
 972         ASSERT(num_bytes <= btrfs_super_total_bytes(fs_info->super_copy));
 973
 974         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 975
 976         if (start == 0) {
 977                 /* lets try to make an inline extent */
 978                 ret = cow_file_range_inline(inode, start, end, 0,
 979                                             BTRFS_COMPRESS_NONE, NULL);
 980                 if (ret == 0) {
 981                         /*
 982                          * We use DO_ACCOUNTING here because we need the
 983                          * delalloc_release_metadata to be run _after_ we drop
 984                          * our outstanding extent for clearing delalloc for this
 985                          * range.
 986                          */
 987                         extent_clear_unlock_delalloc(inode, start, end,
 988                                      delalloc_end, NULL,
 989                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 990                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 991                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
 992                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 993                                      PAGE_END_WRITEBACK);
 994                         *nr_written = *nr_written +
 995                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 996                         *page_started = 1;
 997                         goto out;
 998                 } else if (ret < 0) {
 999                         goto out_unlock;
1000                 }
1001         }
1002
1003         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1004         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1005                         start + num_bytes - 1, 0);
1006
1007         while (num_bytes > 0) {
1008                 cur_alloc_size = num_bytes;
1009                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1010                                            fs_info->sectorsize, 0, alloc_hint,
1011                                            &ins, 1, 1);
1012                 if (ret < 0)
1013                         goto out_unlock;
1014                 cur_alloc_size = ins.offset;
1015                 extent_reserved = true;
1016
1017                 ram_size = ins.offset;
1018                 em = create_io_em(inode, start, ins.offset, /* len */
1019                                   start, /* orig_start */
1020                                   ins.objectid, /* block_start */
1021                                   ins.offset, /* block_len */
1022                                   ins.offset, /* orig_block_len */
1023                                   ram_size, /* ram_bytes */
1024                                   BTRFS_COMPRESS_NONE, /* compress_type */
1025                                   BTRFS_ORDERED_REGULAR /* type */);
1026                 if (IS_ERR(em)) {
1027                         ret = PTR_ERR(em);
1028                         goto out_reserve;
1029                 }
1030                 free_extent_map(em);
1031
1032                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1033                                                ram_size, cur_alloc_size, 0);
1034                 if (ret)
1035                         goto out_drop_extent_cache;
1036
1037                 if (root->root_key.objectid ==
1038                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1039                         ret = btrfs_reloc_clone_csums(inode, start,
1040                                                       cur_alloc_size);
1041                         /*
1042                          * Only drop cache here, and process as normal.
1043                          *
1044                          * We must not allow extent_clear_unlock_delalloc()
1045                          * at out_unlock label to free meta of this ordered
1046                          * extent, as its meta should be freed by
1047                          * btrfs_finish_ordered_io().
1048                          *
1049                          * So we must continue until @start is increased to
1050                          * skip current ordered extent.
1051                          */
1052                         if (ret)
1053                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1054                                                 start + ram_size - 1, 0);
1055                 }
1056
1057                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1058
1059                 /* we're not doing compressed IO, don't unlock the first
1060                  * page (which the caller expects to stay locked), don't
1061                  * clear any dirty bits and don't set any writeback bits
1062                  *
1063                  * Do set the Private2 bit so we know this page was properly
1064                  * setup for writepage
1065                  */
1066                 page_ops = unlock ? PAGE_UNLOCK : 0;
1067                 page_ops |= PAGE_SET_PRIVATE2;
1068
1069                 extent_clear_unlock_delalloc(inode, start,
1070                                              start + ram_size - 1,
1071                                              delalloc_end, locked_page,
1072                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1073                                              page_ops);
1074                 if (num_bytes < cur_alloc_size)
1075                         num_bytes = 0;
1076                 else
1077                         num_bytes -= cur_alloc_size;
1078                 alloc_hint = ins.objectid + ins.offset;
1079                 start += cur_alloc_size;
1080                 extent_reserved = false;
1081
1082                 /*
1083                  * btrfs_reloc_clone_csums() error, since start is increased
1084                  * extent_clear_unlock_delalloc() at out_unlock label won't
1085                  * free metadata of current ordered extent, we're OK to exit.
1086                  */
1087                 if (ret)
1088                         goto out_unlock;
1089         }
1090 out:
1091         return ret;
1092
1093 out_drop_extent_cache:
1094         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1095 out_reserve:
1096         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1097         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1098 out_unlock:
1099         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1100                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1101         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1102                 PAGE_END_WRITEBACK;
1103         /*
1104          * If we reserved an extent for our delalloc range (or a subrange) and
1105          * failed to create the respective ordered extent, then it means that
1106          * when we reserved the extent we decremented the extent's size from
1107          * the data space_info's bytes_may_use counter and incremented the
1108          * space_info's bytes_reserved counter by the same amount. We must make
1109          * sure extent_clear_unlock_delalloc() does not try to decrement again
1110          * the data space_info's bytes_may_use counter, therefore we do not pass
1111          * it the flag EXTENT_CLEAR_DATA_RESV.
1112          */
1113         if (extent_reserved) {
1114                 extent_clear_unlock_delalloc(inode, start,
1115                                              start + cur_alloc_size,
1116                                              start + cur_alloc_size,
1117                                              locked_page,
1118                                              clear_bits,
1119                                              page_ops);
1120                 start += cur_alloc_size;
1121                 if (start >= end)
1122                         goto out;
1123         }
1124         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1125                                      locked_page,
1126                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1127                                      page_ops);
1128         goto out;
1129 }
1130
1131 /*
1132  * work queue call back to started compression on a file and pages
1133  */
1134 static noinline void async_cow_start(struct btrfs_work *work)
1135 {
1136         struct async_cow *async_cow;
1137         int num_added = 0;
1138         async_cow = container_of(work, struct async_cow, work);
1139
1140         compress_file_range(async_cow->inode, async_cow->locked_page,
1141                             async_cow->start, async_cow->end, async_cow,
1142                             &num_added);
1143         if (num_added == 0) {
1144                 btrfs_add_delayed_iput(async_cow->inode);
1145                 async_cow->inode = NULL;
1146         }
1147 }
1148
1149 /*
1150  * work queue call back to submit previously compressed pages
1151  */
1152 static noinline void async_cow_submit(struct btrfs_work *work)
1153 {
1154         struct btrfs_fs_info *fs_info;
1155         struct async_cow *async_cow;
1156         unsigned long nr_pages;
1157
1158         async_cow = container_of(work, struct async_cow, work);
1159
1160         fs_info = async_cow->fs_info;
1161         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1162                 PAGE_SHIFT;
1163
1164         /* atomic_sub_return implies a barrier */
1165         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1166             5 * SZ_1M)
1167                 cond_wake_up_nomb(&fs_info->async_submit_wait);
1168
1169         if (async_cow->inode)
1170                 submit_compressed_extents(async_cow->inode, async_cow);
1171 }
1172
1173 static noinline void async_cow_free(struct btrfs_work *work)
1174 {
1175         struct async_cow *async_cow;
1176         async_cow = container_of(work, struct async_cow, work);
1177         if (async_cow->inode)
1178                 btrfs_add_delayed_iput(async_cow->inode);
1179         kfree(async_cow);
1180 }
1181
1182 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1183                                 u64 start, u64 end, int *page_started,
1184                                 unsigned long *nr_written,
1185                                 unsigned int write_flags)
1186 {
1187         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1188         struct async_cow *async_cow;
1189         unsigned long nr_pages;
1190         u64 cur_end;
1191
1192         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1193                          1, 0, NULL);
1194         while (start < end) {
1195                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1196                 BUG_ON(!async_cow); /* -ENOMEM */
1197                 async_cow->inode = igrab(inode);
1198                 async_cow->fs_info = fs_info;
1199                 async_cow->locked_page = locked_page;
1200                 async_cow->start = start;
1201                 async_cow->write_flags = write_flags;
1202
1203                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1204                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1205                         cur_end = end;
1206                 else
1207                         cur_end = min(end, start + SZ_512K - 1);
1208
1209                 async_cow->end = cur_end;
1210                 INIT_LIST_HEAD(&async_cow->extents);
1211
1212                 btrfs_init_work(&async_cow->work,
1213                                 btrfs_delalloc_helper,
1214                                 async_cow_start, async_cow_submit,
1215                                 async_cow_free);
1216
1217                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1218                         PAGE_SHIFT;
1219                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1220
1221                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1222
1223                 *nr_written += nr_pages;
1224                 start = cur_end + 1;
1225         }
1226         *page_started = 1;
1227         return 0;
1228 }
1229
1230 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1231                                         u64 bytenr, u64 num_bytes)
1232 {
1233         int ret;
1234         struct btrfs_ordered_sum *sums;
1235         LIST_HEAD(list);
1236
1237         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1238                                        bytenr + num_bytes - 1, &list, 0);
1239         if (ret == 0 && list_empty(&list))
1240                 return 0;
1241
1242         while (!list_empty(&list)) {
1243                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1244                 list_del(&sums->list);
1245                 kfree(sums);
1246         }
1247         if (ret < 0)
1248                 return ret;
1249         return 1;
1250 }
1251
1252 /*
1253  * when nowcow writeback call back.  This checks for snapshots or COW copies
1254  * of the extents that exist in the file, and COWs the file as required.
1255  *
1256  * If no cow copies or snapshots exist, we write directly to the existing
1257  * blocks on disk
1258  */
1259 static noinline int run_delalloc_nocow(struct inode *inode,
1260                                        struct page *locked_page,
1261                               u64 start, u64 end, int *page_started, int force,
1262                               unsigned long *nr_written)
1263 {
1264         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1265         struct btrfs_root *root = BTRFS_I(inode)->root;
1266         struct extent_buffer *leaf;
1267         struct btrfs_path *path;
1268         struct btrfs_file_extent_item *fi;
1269         struct btrfs_key found_key;
1270         struct extent_map *em;
1271         u64 cow_start;
1272         u64 cur_offset;
1273         u64 extent_end;
1274         u64 extent_offset;
1275         u64 disk_bytenr;
1276         u64 num_bytes;
1277         u64 disk_num_bytes;
1278         u64 ram_bytes;
1279         int extent_type;
1280         int ret;
1281         int type;
1282         int nocow;
1283         int check_prev = 1;
1284         bool nolock;
1285         u64 ino = btrfs_ino(BTRFS_I(inode));
1286
1287         path = btrfs_alloc_path();
1288         if (!path) {
1289                 extent_clear_unlock_delalloc(inode, start, end, end,
1290                                              locked_page,
1291                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1292                                              EXTENT_DO_ACCOUNTING |
1293                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1294                                              PAGE_CLEAR_DIRTY |
1295                                              PAGE_SET_WRITEBACK |
1296                                              PAGE_END_WRITEBACK);
1297                 return -ENOMEM;
1298         }
1299
1300         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1301
1302         cow_start = (u64)-1;
1303         cur_offset = start;
1304         while (1) {
1305                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1306                                                cur_offset, 0);
1307                 if (ret < 0)
1308                         goto error;
1309                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1310                         leaf = path->nodes[0];
1311                         btrfs_item_key_to_cpu(leaf, &found_key,
1312                                               path->slots[0] - 1);
1313                         if (found_key.objectid == ino &&
1314                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1315                                 path->slots[0]--;
1316                 }
1317                 check_prev = 0;
1318 next_slot:
1319                 leaf = path->nodes[0];
1320                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1321                         ret = btrfs_next_leaf(root, path);
1322                         if (ret < 0) {
1323                                 if (cow_start != (u64)-1)
1324                                         cur_offset = cow_start;
1325                                 goto error;
1326                         }
1327                         if (ret > 0)
1328                                 break;
1329                         leaf = path->nodes[0];
1330                 }
1331
1332                 nocow = 0;
1333                 disk_bytenr = 0;
1334                 num_bytes = 0;
1335                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1336
1337                 if (found_key.objectid > ino)
1338                         break;
1339                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1340                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1341                         path->slots[0]++;
1342                         goto next_slot;
1343                 }
1344                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1345                     found_key.offset > end)
1346                         break;
1347
1348                 if (found_key.offset > cur_offset) {
1349                         extent_end = found_key.offset;
1350                         extent_type = 0;
1351                         goto out_check;
1352                 }
1353
1354                 fi = btrfs_item_ptr(leaf, path->slots[0],
1355                                     struct btrfs_file_extent_item);
1356                 extent_type = btrfs_file_extent_type(leaf, fi);
1357
1358                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1359                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1360                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1361                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1362                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1363                         extent_end = found_key.offset +
1364                                 btrfs_file_extent_num_bytes(leaf, fi);
1365                         disk_num_bytes =
1366                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1367                         if (extent_end <= start) {
1368                                 path->slots[0]++;
1369                                 goto next_slot;
1370                         }
1371                         if (disk_bytenr == 0)
1372                                 goto out_check;
1373                         if (btrfs_file_extent_compression(leaf, fi) ||
1374                             btrfs_file_extent_encryption(leaf, fi) ||
1375                             btrfs_file_extent_other_encoding(leaf, fi))
1376                                 goto out_check;
1377                         /*
1378                          * Do the same check as in btrfs_cross_ref_exist but
1379                          * without the unnecessary search.
1380                          */
1381                         if (!nolock &&
1382                             btrfs_file_extent_generation(leaf, fi) <=
1383                             btrfs_root_last_snapshot(&root->root_item))
1384                                 goto out_check;
1385                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1386                                 goto out_check;
1387                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1388                                 goto out_check;
1389                         ret = btrfs_cross_ref_exist(root, ino,
1390                                                     found_key.offset -
1391                                                     extent_offset, disk_bytenr);
1392                         if (ret) {
1393                                 /*
1394                                  * ret could be -EIO if the above fails to read
1395                                  * metadata.
1396                                  */
1397                                 if (ret < 0) {
1398                                         if (cow_start != (u64)-1)
1399                                                 cur_offset = cow_start;
1400                                         goto error;
1401                                 }
1402
1403                                 WARN_ON_ONCE(nolock);
1404                                 goto out_check;
1405                         }
1406                         disk_bytenr += extent_offset;
1407                         disk_bytenr += cur_offset - found_key.offset;
1408                         num_bytes = min(end + 1, extent_end) - cur_offset;
1409                         /*
1410                          * if there are pending snapshots for this root,
1411                          * we fall into common COW way.
1412                          */
1413                         if (!nolock && atomic_read(&root->snapshot_force_cow))
1414                                 goto out_check;
1415                         /*
1416                          * force cow if csum exists in the range.
1417                          * this ensure that csum for a given extent are
1418                          * either valid or do not exist.
1419                          */
1420                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1421                                                   num_bytes);
1422                         if (ret) {
1423                                 /*
1424                                  * ret could be -EIO if the above fails to read
1425                                  * metadata.
1426                                  */
1427                                 if (ret < 0) {
1428                                         if (cow_start != (u64)-1)
1429                                                 cur_offset = cow_start;
1430                                         goto error;
1431                                 }
1432                                 WARN_ON_ONCE(nolock);
1433                                 goto out_check;
1434                         }
1435                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr))
1436                                 goto out_check;
1437                         nocow = 1;
1438                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1439                         extent_end = found_key.offset +
1440                                 btrfs_file_extent_ram_bytes(leaf, fi);
1441                         extent_end = ALIGN(extent_end,
1442                                            fs_info->sectorsize);
1443                 } else {
1444                         BUG_ON(1);
1445                 }
1446 out_check:
1447                 if (extent_end <= start) {
1448                         path->slots[0]++;
1449                         if (nocow)
1450                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1451                         goto next_slot;
1452                 }
1453                 if (!nocow) {
1454                         if (cow_start == (u64)-1)
1455                                 cow_start = cur_offset;
1456                         cur_offset = extent_end;
1457                         if (cur_offset > end)
1458                                 break;
1459                         path->slots[0]++;
1460                         goto next_slot;
1461                 }
1462
1463                 btrfs_release_path(path);
1464                 if (cow_start != (u64)-1) {
1465                         ret = cow_file_range(inode, locked_page,
1466                                              cow_start, found_key.offset - 1,
1467                                              end, page_started, nr_written, 1,
1468                                              NULL);
1469                         if (ret) {
1470                                 if (nocow)
1471                                         btrfs_dec_nocow_writers(fs_info,
1472                                                                 disk_bytenr);
1473                                 goto error;
1474                         }
1475                         cow_start = (u64)-1;
1476                 }
1477
1478                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1479                         u64 orig_start = found_key.offset - extent_offset;
1480
1481                         em = create_io_em(inode, cur_offset, num_bytes,
1482                                           orig_start,
1483                                           disk_bytenr, /* block_start */
1484                                           num_bytes, /* block_len */
1485                                           disk_num_bytes, /* orig_block_len */
1486                                           ram_bytes, BTRFS_COMPRESS_NONE,
1487                                           BTRFS_ORDERED_PREALLOC);
1488                         if (IS_ERR(em)) {
1489                                 if (nocow)
1490                                         btrfs_dec_nocow_writers(fs_info,
1491                                                                 disk_bytenr);
1492                                 ret = PTR_ERR(em);
1493                                 goto error;
1494                         }
1495                         free_extent_map(em);
1496                 }
1497
1498                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1499                         type = BTRFS_ORDERED_PREALLOC;
1500                 } else {
1501                         type = BTRFS_ORDERED_NOCOW;
1502                 }
1503
1504                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1505                                                num_bytes, num_bytes, type);
1506                 if (nocow)
1507                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1508                 BUG_ON(ret); /* -ENOMEM */
1509
1510                 if (root->root_key.objectid ==
1511                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1512                         /*
1513                          * Error handled later, as we must prevent
1514                          * extent_clear_unlock_delalloc() in error handler
1515                          * from freeing metadata of created ordered extent.
1516                          */
1517                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1518                                                       num_bytes);
1519
1520                 extent_clear_unlock_delalloc(inode, cur_offset,
1521                                              cur_offset + num_bytes - 1, end,
1522                                              locked_page, EXTENT_LOCKED |
1523                                              EXTENT_DELALLOC |
1524                                              EXTENT_CLEAR_DATA_RESV,
1525                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1526
1527                 cur_offset = extent_end;
1528
1529                 /*
1530                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1531                  * handler, as metadata for created ordered extent will only
1532                  * be freed by btrfs_finish_ordered_io().
1533                  */
1534                 if (ret)
1535                         goto error;
1536                 if (cur_offset > end)
1537                         break;
1538         }
1539         btrfs_release_path(path);
1540
1541         if (cur_offset <= end && cow_start == (u64)-1)
1542                 cow_start = cur_offset;
1543
1544         if (cow_start != (u64)-1) {
1545                 cur_offset = end;
1546                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1547                                      page_started, nr_written, 1, NULL);
1548                 if (ret)
1549                         goto error;
1550         }
1551
1552 error:
1553         if (ret && cur_offset < end)
1554                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1555                                              locked_page, EXTENT_LOCKED |
1556                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1557                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1558                                              PAGE_CLEAR_DIRTY |
1559                                              PAGE_SET_WRITEBACK |
1560                                              PAGE_END_WRITEBACK);
1561         btrfs_free_path(path);
1562         return ret;
1563 }
1564
1565 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1566 {
1567
1568         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1569             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1570                 return 0;
1571
1572         /*
1573          * @defrag_bytes is a hint value, no spinlock held here,
1574          * if is not zero, it means the file is defragging.
1575          * Force cow if given extent needs to be defragged.
1576          */
1577         if (BTRFS_I(inode)->defrag_bytes &&
1578             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1579                            EXTENT_DEFRAG, 0, NULL))
1580                 return 1;
1581
1582         return 0;
1583 }
1584
1585 /*
1586  * Function to process delayed allocation (create CoW) for ranges which are
1587  * being touched for the first time.
1588  */
1589 int btrfs_run_delalloc_range(void *private_data, struct page *locked_page,
1590                 u64 start, u64 end, int *page_started, unsigned long *nr_written,
1591                 struct writeback_control *wbc)
1592 {
1593         struct inode *inode = private_data;
1594         int ret;
1595         int force_cow = need_force_cow(inode, start, end);
1596         unsigned int write_flags = wbc_to_write_flags(wbc);
1597
1598         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1599                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1600                                          page_started, 1, nr_written);
1601         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1602                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1603                                          page_started, 0, nr_written);
1604         } else if (!inode_need_compress(inode, start, end)) {
1605                 ret = cow_file_range(inode, locked_page, start, end, end,
1606                                       page_started, nr_written, 1, NULL);
1607         } else {
1608                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1609                         &BTRFS_I(inode)->runtime_flags);
1610                 ret = cow_file_range_async(inode, locked_page, start, end,
1611                                            page_started, nr_written,
1612                                            write_flags);
1613         }
1614         if (ret)
1615                 btrfs_cleanup_ordered_extents(inode, locked_page, start,
1616                                               end - start + 1);
1617         return ret;
1618 }
1619
1620 void btrfs_split_delalloc_extent(struct inode *inode,
1621                                  struct extent_state *orig, u64 split)
1622 {
1623         u64 size;
1624
1625         /* not delalloc, ignore it */
1626         if (!(orig->state & EXTENT_DELALLOC))
1627                 return;
1628
1629         size = orig->end - orig->start + 1;
1630         if (size > BTRFS_MAX_EXTENT_SIZE) {
1631                 u32 num_extents;
1632                 u64 new_size;
1633
1634                 /*
1635                  * See the explanation in btrfs_merge_delalloc_extent, the same
1636                  * applies here, just in reverse.
1637                  */
1638                 new_size = orig->end - split + 1;
1639                 num_extents = count_max_extents(new_size);
1640                 new_size = split - orig->start;
1641                 num_extents += count_max_extents(new_size);
1642                 if (count_max_extents(size) >= num_extents)
1643                         return;
1644         }
1645
1646         spin_lock(&BTRFS_I(inode)->lock);
1647         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1648         spin_unlock(&BTRFS_I(inode)->lock);
1649 }
1650
1651 /*
1652  * Handle merged delayed allocation extents so we can keep track of new extents
1653  * that are just merged onto old extents, such as when we are doing sequential
1654  * writes, so we can properly account for the metadata space we'll need.
1655  */
1656 void btrfs_merge_delalloc_extent(struct inode *inode, struct extent_state *new,
1657                                  struct extent_state *other)
1658 {
1659         u64 new_size, old_size;
1660         u32 num_extents;
1661
1662         /* not delalloc, ignore it */
1663         if (!(other->state & EXTENT_DELALLOC))
1664                 return;
1665
1666         if (new->start > other->start)
1667                 new_size = new->end - other->start + 1;
1668         else
1669                 new_size = other->end - new->start + 1;
1670
1671         /* we're not bigger than the max, unreserve the space and go */
1672         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1673                 spin_lock(&BTRFS_I(inode)->lock);
1674                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1675                 spin_unlock(&BTRFS_I(inode)->lock);
1676                 return;
1677         }
1678
1679         /*
1680          * We have to add up either side to figure out how many extents were
1681          * accounted for before we merged into one big extent.  If the number of
1682          * extents we accounted for is <= the amount we need for the new range
1683          * then we can return, otherwise drop.  Think of it like this
1684          *
1685          * [ 4k][MAX_SIZE]
1686          *
1687          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1688          * need 2 outstanding extents, on one side we have 1 and the other side
1689          * we have 1 so they are == and we can return.  But in this case
1690          *
1691          * [MAX_SIZE+4k][MAX_SIZE+4k]
1692          *
1693          * Each range on their own accounts for 2 extents, but merged together
1694          * they are only 3 extents worth of accounting, so we need to drop in
1695          * this case.
1696          */
1697         old_size = other->end - other->start + 1;
1698         num_extents = count_max_extents(old_size);
1699         old_size = new->end - new->start + 1;
1700         num_extents += count_max_extents(old_size);
1701         if (count_max_extents(new_size) >= num_extents)
1702                 return;
1703
1704         spin_lock(&BTRFS_I(inode)->lock);
1705         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1706         spin_unlock(&BTRFS_I(inode)->lock);
1707 }
1708
1709 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1710                                       struct inode *inode)
1711 {
1712         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1713
1714         spin_lock(&root->delalloc_lock);
1715         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1716                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1717                               &root->delalloc_inodes);
1718                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1719                         &BTRFS_I(inode)->runtime_flags);
1720                 root->nr_delalloc_inodes++;
1721                 if (root->nr_delalloc_inodes == 1) {
1722                         spin_lock(&fs_info->delalloc_root_lock);
1723                         BUG_ON(!list_empty(&root->delalloc_root));
1724                         list_add_tail(&root->delalloc_root,
1725                                       &fs_info->delalloc_roots);
1726                         spin_unlock(&fs_info->delalloc_root_lock);
1727                 }
1728         }
1729         spin_unlock(&root->delalloc_lock);
1730 }
1731
1732
1733 void __btrfs_del_delalloc_inode(struct btrfs_root *root,
1734                                 struct btrfs_inode *inode)
1735 {
1736         struct btrfs_fs_info *fs_info = root->fs_info;
1737
1738         if (!list_empty(&inode->delalloc_inodes)) {
1739                 list_del_init(&inode->delalloc_inodes);
1740                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1741                           &inode->runtime_flags);
1742                 root->nr_delalloc_inodes--;
1743                 if (!root->nr_delalloc_inodes) {
1744                         ASSERT(list_empty(&root->delalloc_inodes));
1745                         spin_lock(&fs_info->delalloc_root_lock);
1746                         BUG_ON(list_empty(&root->delalloc_root));
1747                         list_del_init(&root->delalloc_root);
1748                         spin_unlock(&fs_info->delalloc_root_lock);
1749                 }
1750         }
1751 }
1752
1753 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1754                                      struct btrfs_inode *inode)
1755 {
1756         spin_lock(&root->delalloc_lock);
1757         __btrfs_del_delalloc_inode(root, inode);
1758         spin_unlock(&root->delalloc_lock);
1759 }
1760
1761 /*
1762  * Properly track delayed allocation bytes in the inode and to maintain the
1763  * list of inodes that have pending delalloc work to be done.
1764  */
1765 void btrfs_set_delalloc_extent(struct inode *inode, struct extent_state *state,
1766                                unsigned *bits)
1767 {
1768         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1769
1770         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1771                 WARN_ON(1);
1772         /*
1773          * set_bit and clear bit hooks normally require _irqsave/restore
1774          * but in this case, we are only testing for the DELALLOC
1775          * bit, which is only set or cleared with irqs on
1776          */
1777         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1778                 struct btrfs_root *root = BTRFS_I(inode)->root;
1779                 u64 len = state->end + 1 - state->start;
1780                 u32 num_extents = count_max_extents(len);
1781                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1782
1783                 spin_lock(&BTRFS_I(inode)->lock);
1784                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1785                 spin_unlock(&BTRFS_I(inode)->lock);
1786
1787                 /* For sanity tests */
1788                 if (btrfs_is_testing(fs_info))
1789                         return;
1790
1791                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1792                                          fs_info->delalloc_batch);
1793                 spin_lock(&BTRFS_I(inode)->lock);
1794                 BTRFS_I(inode)->delalloc_bytes += len;
1795                 if (*bits & EXTENT_DEFRAG)
1796                         BTRFS_I(inode)->defrag_bytes += len;
1797                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1798                                          &BTRFS_I(inode)->runtime_flags))
1799                         btrfs_add_delalloc_inodes(root, inode);
1800                 spin_unlock(&BTRFS_I(inode)->lock);
1801         }
1802
1803         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1804             (*bits & EXTENT_DELALLOC_NEW)) {
1805                 spin_lock(&BTRFS_I(inode)->lock);
1806                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1807                         state->start;
1808                 spin_unlock(&BTRFS_I(inode)->lock);
1809         }
1810 }
1811
1812 /*
1813  * Once a range is no longer delalloc this function ensures that proper
1814  * accounting happens.
1815  */
1816 void btrfs_clear_delalloc_extent(struct inode *vfs_inode,
1817                                  struct extent_state *state, unsigned *bits)
1818 {
1819         struct btrfs_inode *inode = BTRFS_I(vfs_inode);
1820         struct btrfs_fs_info *fs_info = btrfs_sb(vfs_inode->i_sb);
1821         u64 len = state->end + 1 - state->start;
1822         u32 num_extents = count_max_extents(len);
1823
1824         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1825                 spin_lock(&inode->lock);
1826                 inode->defrag_bytes -= len;
1827                 spin_unlock(&inode->lock);
1828         }
1829
1830         /*
1831          * set_bit and clear bit hooks normally require _irqsave/restore
1832          * but in this case, we are only testing for the DELALLOC
1833          * bit, which is only set or cleared with irqs on
1834          */
1835         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1836                 struct btrfs_root *root = inode->root;
1837                 bool do_list = !btrfs_is_free_space_inode(inode);
1838
1839                 spin_lock(&inode->lock);
1840                 btrfs_mod_outstanding_extents(inode, -num_extents);
1841                 spin_unlock(&inode->lock);
1842
1843                 /*
1844                  * We don't reserve metadata space for space cache inodes so we
1845                  * don't need to call delalloc_release_metadata if there is an
1846                  * error.
1847                  */
1848                 if (*bits & EXTENT_CLEAR_META_RESV &&
1849                     root != fs_info->tree_root)
1850                         btrfs_delalloc_release_metadata(inode, len, false);
1851
1852                 /* For sanity tests. */
1853                 if (btrfs_is_testing(fs_info))
1854                         return;
1855
1856                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1857                     do_list && !(state->state & EXTENT_NORESERVE) &&
1858                     (*bits & EXTENT_CLEAR_DATA_RESV))
1859                         btrfs_free_reserved_data_space_noquota(
1860                                         &inode->vfs_inode,
1861                                         state->start, len);
1862
1863                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1864                                          fs_info->delalloc_batch);
1865                 spin_lock(&inode->lock);
1866                 inode->delalloc_bytes -= len;
1867                 if (do_list && inode->delalloc_bytes == 0 &&
1868                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1869                                         &inode->runtime_flags))
1870                         btrfs_del_delalloc_inode(root, inode);
1871                 spin_unlock(&inode->lock);
1872         }
1873
1874         if ((state->state & EXTENT_DELALLOC_NEW) &&
1875             (*bits & EXTENT_DELALLOC_NEW)) {
1876                 spin_lock(&inode->lock);
1877                 ASSERT(inode->new_delalloc_bytes >= len);
1878                 inode->new_delalloc_bytes -= len;
1879                 spin_unlock(&inode->lock);
1880         }
1881 }
1882
1883 /*
1884  * btrfs_bio_fits_in_stripe - Checks whether the size of the given bio will fit
1885  * in a chunk's stripe. This function ensures that bios do not span a
1886  * stripe/chunk
1887  *
1888  * @page - The page we are about to add to the bio
1889  * @size - size we want to add to the bio
1890  * @bio - bio we want to ensure is smaller than a stripe
1891  * @bio_flags - flags of the bio
1892  *
1893  * return 1 if page cannot be added to the bio
1894  * return 0 if page can be added to the bio
1895  * return error otherwise
1896  */
1897 int btrfs_bio_fits_in_stripe(struct page *page, size_t size, struct bio *bio,
1898                              unsigned long bio_flags)
1899 {
1900         struct inode *inode = page->mapping->host;
1901         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1902         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1903         u64 length = 0;
1904         u64 map_length;
1905         int ret;
1906
1907         if (bio_flags & EXTENT_BIO_COMPRESSED)
1908                 return 0;
1909
1910         length = bio->bi_iter.bi_size;
1911         map_length = length;
1912         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1913                               NULL, 0);
1914         if (ret < 0)
1915                 return ret;
1916         if (map_length < length + size)
1917                 return 1;
1918         return 0;
1919 }
1920
1921 /*
1922  * in order to insert checksums into the metadata in large chunks,
1923  * we wait until bio submission time.   All the pages in the bio are
1924  * checksummed and sums are attached onto the ordered extent record.
1925  *
1926  * At IO completion time the cums attached on the ordered extent record
1927  * are inserted into the btree
1928  */
1929 static blk_status_t btrfs_submit_bio_start(void *private_data, struct bio *bio,
1930                                     u64 bio_offset)
1931 {
1932         struct inode *inode = private_data;
1933         blk_status_t ret = 0;
1934
1935         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1936         BUG_ON(ret); /* -ENOMEM */
1937         return 0;
1938 }
1939
1940 /*
1941  * extent_io.c submission hook. This does the right thing for csum calculation
1942  * on write, or reading the csums from the tree before a read.
1943  *
1944  * Rules about async/sync submit,
1945  * a) read:                             sync submit
1946  *
1947  * b) write without checksum:           sync submit
1948  *
1949  * c) write with checksum:
1950  *    c-1) if bio is issued by fsync:   sync submit
1951  *         (sync_writers != 0)
1952  *
1953  *    c-2) if root is reloc root:       sync submit
1954  *         (only in case of buffered IO)
1955  *
1956  *    c-3) otherwise:                   async submit
1957  */
1958 static blk_status_t btrfs_submit_bio_hook(void *private_data, struct bio *bio,
1959                                  int mirror_num, unsigned long bio_flags,
1960                                  u64 bio_offset)
1961 {
1962         struct inode *inode = private_data;
1963         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1964         struct btrfs_root *root = BTRFS_I(inode)->root;
1965         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1966         blk_status_t ret = 0;
1967         int skip_sum;
1968         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1969
1970         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1971
1972         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
1973                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1974
1975         if (bio_op(bio) != REQ_OP_WRITE) {
1976                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
1977                 if (ret)
1978                         goto out;
1979
1980                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1981                         ret = btrfs_submit_compressed_read(inode, bio,
1982                                                            mirror_num,
1983                                                            bio_flags);
1984                         goto out;
1985                 } else if (!skip_sum) {
1986                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
1987                         if (ret)
1988                                 goto out;
1989                 }
1990                 goto mapit;
1991         } else if (async && !skip_sum) {
1992                 /* csum items have already been cloned */
1993                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1994                         goto mapit;
1995                 /* we're doing a write, do the async checksumming */
1996                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
1997                                           bio_offset, inode,
1998                                           btrfs_submit_bio_start);
1999                 goto out;
2000         } else if (!skip_sum) {
2001                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2002                 if (ret)
2003                         goto out;
2004         }
2005
2006 mapit:
2007         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
2008
2009 out:
2010         if (ret) {
2011                 bio->bi_status = ret;
2012                 bio_endio(bio);
2013         }
2014         return ret;
2015 }
2016
2017 /*
2018  * given a list of ordered sums record them in the inode.  This happens
2019  * at IO completion time based on sums calculated at bio submission time.
2020  */
2021 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2022                              struct inode *inode, struct list_head *list)
2023 {
2024         struct btrfs_ordered_sum *sum;
2025         int ret;
2026
2027         list_for_each_entry(sum, list, list) {
2028                 trans->adding_csums = true;
2029                 ret = btrfs_csum_file_blocks(trans,
2030                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2031                 trans->adding_csums = false;
2032                 if (ret)
2033                         return ret;
2034         }
2035         return 0;
2036 }
2037
2038 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2039                               unsigned int extra_bits,
2040                               struct extent_state **cached_state, int dedupe)
2041 {
2042         WARN_ON(PAGE_ALIGNED(end));
2043         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2044                                    extra_bits, cached_state);
2045 }
2046
2047 /* see btrfs_writepage_start_hook for details on why this is required */
2048 struct btrfs_writepage_fixup {
2049         struct page *page;
2050         struct btrfs_work work;
2051 };
2052
2053 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2054 {
2055         struct btrfs_writepage_fixup *fixup;
2056         struct btrfs_ordered_extent *ordered;
2057         struct extent_state *cached_state = NULL;
2058         struct extent_changeset *data_reserved = NULL;
2059         struct page *page;
2060         struct inode *inode;
2061         u64 page_start;
2062         u64 page_end;
2063         int ret;
2064
2065         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2066         page = fixup->page;
2067 again:
2068         lock_page(page);
2069         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2070                 ClearPageChecked(page);
2071                 goto out_page;
2072         }
2073
2074         inode = page->mapping->host;
2075         page_start = page_offset(page);
2076         page_end = page_offset(page) + PAGE_SIZE - 1;
2077
2078         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2079                          &cached_state);
2080
2081         /* already ordered? We're done */
2082         if (PagePrivate2(page))
2083                 goto out;
2084
2085         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2086                                         PAGE_SIZE);
2087         if (ordered) {
2088                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2089                                      page_end, &cached_state);
2090                 unlock_page(page);
2091                 btrfs_start_ordered_extent(inode, ordered, 1);
2092                 btrfs_put_ordered_extent(ordered);
2093                 goto again;
2094         }
2095
2096         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2097                                            PAGE_SIZE);
2098         if (ret) {
2099                 mapping_set_error(page->mapping, ret);
2100                 end_extent_writepage(page, ret, page_start, page_end);
2101                 ClearPageChecked(page);
2102                 goto out;
2103          }
2104
2105         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2106                                         &cached_state, 0);
2107         if (ret) {
2108                 mapping_set_error(page->mapping, ret);
2109                 end_extent_writepage(page, ret, page_start, page_end);
2110                 ClearPageChecked(page);
2111                 goto out;
2112         }
2113
2114         ClearPageChecked(page);
2115         set_page_dirty(page);
2116         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, false);
2117 out:
2118         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2119                              &cached_state);
2120 out_page:
2121         unlock_page(page);
2122         put_page(page);
2123         kfree(fixup);
2124         extent_changeset_free(data_reserved);
2125 }
2126
2127 /*
2128  * There are a few paths in the higher layers of the kernel that directly
2129  * set the page dirty bit without asking the filesystem if it is a
2130  * good idea.  This causes problems because we want to make sure COW
2131  * properly happens and the data=ordered rules are followed.
2132  *
2133  * In our case any range that doesn't have the ORDERED bit set
2134  * hasn't been properly setup for IO.  We kick off an async process
2135  * to fix it up.  The async helper will wait for ordered extents, set
2136  * the delalloc bit and make it safe to write the page.
2137  */
2138 int btrfs_writepage_cow_fixup(struct page *page, u64 start, u64 end)
2139 {
2140         struct inode *inode = page->mapping->host;
2141         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2142         struct btrfs_writepage_fixup *fixup;
2143
2144         /* this page is properly in the ordered list */
2145         if (TestClearPagePrivate2(page))
2146                 return 0;
2147
2148         if (PageChecked(page))
2149                 return -EAGAIN;
2150
2151         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2152         if (!fixup)
2153                 return -EAGAIN;
2154
2155         SetPageChecked(page);
2156         get_page(page);
2157         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2158                         btrfs_writepage_fixup_worker, NULL, NULL);
2159         fixup->page = page;
2160         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2161         return -EBUSY;
2162 }
2163
2164 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2165                                        struct inode *inode, u64 file_pos,
2166                                        u64 disk_bytenr, u64 disk_num_bytes,
2167                                        u64 num_bytes, u64 ram_bytes,
2168                                        u8 compression, u8 encryption,
2169                                        u16 other_encoding, int extent_type)
2170 {
2171         struct btrfs_root *root = BTRFS_I(inode)->root;
2172         struct btrfs_file_extent_item *fi;
2173         struct btrfs_path *path;
2174         struct extent_buffer *leaf;
2175         struct btrfs_key ins;
2176         u64 qg_released;
2177         int extent_inserted = 0;
2178         int ret;
2179
2180         path = btrfs_alloc_path();
2181         if (!path)
2182                 return -ENOMEM;
2183
2184         /*
2185          * we may be replacing one extent in the tree with another.
2186          * The new extent is pinned in the extent map, and we don't want
2187          * to drop it from the cache until it is completely in the btree.
2188          *
2189          * So, tell btrfs_drop_extents to leave this extent in the cache.
2190          * the caller is expected to unpin it and allow it to be merged
2191          * with the others.
2192          */
2193         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2194                                    file_pos + num_bytes, NULL, 0,
2195                                    1, sizeof(*fi), &extent_inserted);
2196         if (ret)
2197                 goto out;
2198
2199         if (!extent_inserted) {
2200                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2201                 ins.offset = file_pos;
2202                 ins.type = BTRFS_EXTENT_DATA_KEY;
2203
2204                 path->leave_spinning = 1;
2205                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2206                                               sizeof(*fi));
2207                 if (ret)
2208                         goto out;
2209         }
2210         leaf = path->nodes[0];
2211         fi = btrfs_item_ptr(leaf, path->slots[0],
2212                             struct btrfs_file_extent_item);
2213         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2214         btrfs_set_file_extent_type(leaf, fi, extent_type);
2215         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2216         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2217         btrfs_set_file_extent_offset(leaf, fi, 0);
2218         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2219         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2220         btrfs_set_file_extent_compression(leaf, fi, compression);
2221         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2222         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2223
2224         btrfs_mark_buffer_dirty(leaf);
2225         btrfs_release_path(path);
2226
2227         inode_add_bytes(inode, num_bytes);
2228
2229         ins.objectid = disk_bytenr;
2230         ins.offset = disk_num_bytes;
2231         ins.type = BTRFS_EXTENT_ITEM_KEY;
2232
2233         /*
2234          * Release the reserved range from inode dirty range map, as it is
2235          * already moved into delayed_ref_head
2236          */
2237         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2238         if (ret < 0)
2239                 goto out;
2240         qg_released = ret;
2241         ret = btrfs_alloc_reserved_file_extent(trans, root,
2242                                                btrfs_ino(BTRFS_I(inode)),
2243                                                file_pos, qg_released, &ins);
2244 out:
2245         btrfs_free_path(path);
2246
2247         return ret;
2248 }
2249
2250 /* snapshot-aware defrag */
2251 struct sa_defrag_extent_backref {
2252         struct rb_node node;
2253         struct old_sa_defrag_extent *old;
2254         u64 root_id;
2255         u64 inum;
2256         u64 file_pos;
2257         u64 extent_offset;
2258         u64 num_bytes;
2259         u64 generation;
2260 };
2261
2262 struct old_sa_defrag_extent {
2263         struct list_head list;
2264         struct new_sa_defrag_extent *new;
2265
2266         u64 extent_offset;
2267         u64 bytenr;
2268         u64 offset;
2269         u64 len;
2270         int count;
2271 };
2272
2273 struct new_sa_defrag_extent {
2274         struct rb_root root;
2275         struct list_head head;
2276         struct btrfs_path *path;
2277         struct inode *inode;
2278         u64 file_pos;
2279         u64 len;
2280         u64 bytenr;
2281         u64 disk_len;
2282         u8 compress_type;
2283 };
2284
2285 static int backref_comp(struct sa_defrag_extent_backref *b1,
2286                         struct sa_defrag_extent_backref *b2)
2287 {
2288         if (b1->root_id < b2->root_id)
2289                 return -1;
2290         else if (b1->root_id > b2->root_id)
2291                 return 1;
2292
2293         if (b1->inum < b2->inum)
2294                 return -1;
2295         else if (b1->inum > b2->inum)
2296                 return 1;
2297
2298         if (b1->file_pos < b2->file_pos)
2299                 return -1;
2300         else if (b1->file_pos > b2->file_pos)
2301                 return 1;
2302
2303         /*
2304          * [------------------------------] ===> (a range of space)
2305          *     |<--->|   |<---->| =============> (fs/file tree A)
2306          * |<---------------------------->| ===> (fs/file tree B)
2307          *
2308          * A range of space can refer to two file extents in one tree while
2309          * refer to only one file extent in another tree.
2310          *
2311          * So we may process a disk offset more than one time(two extents in A)
2312          * and locate at the same extent(one extent in B), then insert two same
2313          * backrefs(both refer to the extent in B).
2314          */
2315         return 0;
2316 }
2317
2318 static void backref_insert(struct rb_root *root,
2319                            struct sa_defrag_extent_backref *backref)
2320 {
2321         struct rb_node **p = &root->rb_node;
2322         struct rb_node *parent = NULL;
2323         struct sa_defrag_extent_backref *entry;
2324         int ret;
2325
2326         while (*p) {
2327                 parent = *p;
2328                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2329
2330                 ret = backref_comp(backref, entry);
2331                 if (ret < 0)
2332                         p = &(*p)->rb_left;
2333                 else
2334                         p = &(*p)->rb_right;
2335         }
2336
2337         rb_link_node(&backref->node, parent, p);
2338         rb_insert_color(&backref->node, root);
2339 }
2340
2341 /*
2342  * Note the backref might has changed, and in this case we just return 0.
2343  */
2344 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2345                                        void *ctx)
2346 {
2347         struct btrfs_file_extent_item *extent;
2348         struct old_sa_defrag_extent *old = ctx;
2349         struct new_sa_defrag_extent *new = old->new;
2350         struct btrfs_path *path = new->path;
2351         struct btrfs_key key;
2352         struct btrfs_root *root;
2353         struct sa_defrag_extent_backref *backref;
2354         struct extent_buffer *leaf;
2355         struct inode *inode = new->inode;
2356         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2357         int slot;
2358         int ret;
2359         u64 extent_offset;
2360         u64 num_bytes;
2361
2362         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2363             inum == btrfs_ino(BTRFS_I(inode)))
2364                 return 0;
2365
2366         key.objectid = root_id;
2367         key.type = BTRFS_ROOT_ITEM_KEY;
2368         key.offset = (u64)-1;
2369
2370         root = btrfs_read_fs_root_no_name(fs_info, &key);
2371         if (IS_ERR(root)) {
2372                 if (PTR_ERR(root) == -ENOENT)
2373                         return 0;
2374                 WARN_ON(1);
2375                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2376                          inum, offset, root_id);
2377                 return PTR_ERR(root);
2378         }
2379
2380         key.objectid = inum;
2381         key.type = BTRFS_EXTENT_DATA_KEY;
2382         if (offset > (u64)-1 << 32)
2383                 key.offset = 0;
2384         else
2385                 key.offset = offset;
2386
2387         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2388         if (WARN_ON(ret < 0))
2389                 return ret;
2390         ret = 0;
2391
2392         while (1) {
2393                 cond_resched();
2394
2395                 leaf = path->nodes[0];
2396                 slot = path->slots[0];
2397
2398                 if (slot >= btrfs_header_nritems(leaf)) {
2399                         ret = btrfs_next_leaf(root, path);
2400                         if (ret < 0) {
2401                                 goto out;
2402                         } else if (ret > 0) {
2403                                 ret = 0;
2404                                 goto out;
2405                         }
2406                         continue;
2407                 }
2408
2409                 path->slots[0]++;
2410
2411                 btrfs_item_key_to_cpu(leaf, &key, slot);
2412
2413                 if (key.objectid > inum)
2414                         goto out;
2415
2416                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2417                         continue;
2418
2419                 extent = btrfs_item_ptr(leaf, slot,
2420                                         struct btrfs_file_extent_item);
2421
2422                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2423                         continue;
2424
2425                 /*
2426                  * 'offset' refers to the exact key.offset,
2427                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2428                  * (key.offset - extent_offset).
2429                  */
2430                 if (key.offset != offset)
2431                         continue;
2432
2433                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2434                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2435
2436                 if (extent_offset >= old->extent_offset + old->offset +
2437                     old->len || extent_offset + num_bytes <=
2438                     old->extent_offset + old->offset)
2439                         continue;
2440                 break;
2441         }
2442
2443         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2444         if (!backref) {
2445                 ret = -ENOENT;
2446                 goto out;
2447         }
2448
2449         backref->root_id = root_id;
2450         backref->inum = inum;
2451         backref->file_pos = offset;
2452         backref->num_bytes = num_bytes;
2453         backref->extent_offset = extent_offset;
2454         backref->generation = btrfs_file_extent_generation(leaf, extent);
2455         backref->old = old;
2456         backref_insert(&new->root, backref);
2457         old->count++;
2458 out:
2459         btrfs_release_path(path);
2460         WARN_ON(ret);
2461         return ret;
2462 }
2463
2464 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2465                                    struct new_sa_defrag_extent *new)
2466 {
2467         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2468         struct old_sa_defrag_extent *old, *tmp;
2469         int ret;
2470
2471         new->path = path;
2472
2473         list_for_each_entry_safe(old, tmp, &new->head, list) {
2474                 ret = iterate_inodes_from_logical(old->bytenr +
2475                                                   old->extent_offset, fs_info,
2476                                                   path, record_one_backref,
2477                                                   old, false);
2478                 if (ret < 0 && ret != -ENOENT)
2479                         return false;
2480
2481                 /* no backref to be processed for this extent */
2482                 if (!old->count) {
2483                         list_del(&old->list);
2484                         kfree(old);
2485                 }
2486         }
2487
2488         if (list_empty(&new->head))
2489                 return false;
2490
2491         return true;
2492 }
2493
2494 static int relink_is_mergable(struct extent_buffer *leaf,
2495                               struct btrfs_file_extent_item *fi,
2496                               struct new_sa_defrag_extent *new)
2497 {
2498         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2499                 return 0;
2500
2501         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2502                 return 0;
2503
2504         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2505                 return 0;
2506
2507         if (btrfs_file_extent_encryption(leaf, fi) ||
2508             btrfs_file_extent_other_encoding(leaf, fi))
2509                 return 0;
2510
2511         return 1;
2512 }
2513
2514 /*
2515  * Note the backref might has changed, and in this case we just return 0.
2516  */
2517 static noinline int relink_extent_backref(struct btrfs_path *path,
2518                                  struct sa_defrag_extent_backref *prev,
2519                                  struct sa_defrag_extent_backref *backref)
2520 {
2521         struct btrfs_file_extent_item *extent;
2522         struct btrfs_file_extent_item *item;
2523         struct btrfs_ordered_extent *ordered;
2524         struct btrfs_trans_handle *trans;
2525         struct btrfs_root *root;
2526         struct btrfs_key key;
2527         struct extent_buffer *leaf;
2528         struct old_sa_defrag_extent *old = backref->old;
2529         struct new_sa_defrag_extent *new = old->new;
2530         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2531         struct inode *inode;
2532         struct extent_state *cached = NULL;
2533         int ret = 0;
2534         u64 start;
2535         u64 len;
2536         u64 lock_start;
2537         u64 lock_end;
2538         bool merge = false;
2539         int index;
2540
2541         if (prev && prev->root_id == backref->root_id &&
2542             prev->inum == backref->inum &&
2543             prev->file_pos + prev->num_bytes == backref->file_pos)
2544                 merge = true;
2545
2546         /* step 1: get root */
2547         key.objectid = backref->root_id;
2548         key.type = BTRFS_ROOT_ITEM_KEY;
2549         key.offset = (u64)-1;
2550
2551         index = srcu_read_lock(&fs_info->subvol_srcu);
2552
2553         root = btrfs_read_fs_root_no_name(fs_info, &key);
2554         if (IS_ERR(root)) {
2555                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2556                 if (PTR_ERR(root) == -ENOENT)
2557                         return 0;
2558                 return PTR_ERR(root);
2559         }
2560
2561         if (btrfs_root_readonly(root)) {
2562                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2563                 return 0;
2564         }
2565
2566         /* step 2: get inode */
2567         key.objectid = backref->inum;
2568         key.type = BTRFS_INODE_ITEM_KEY;
2569         key.offset = 0;
2570
2571         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2572         if (IS_ERR(inode)) {
2573                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2574                 return 0;
2575         }
2576
2577         srcu_read_unlock(&fs_info->subvol_srcu, index);
2578
2579         /* step 3: relink backref */
2580         lock_start = backref->file_pos;
2581         lock_end = backref->file_pos + backref->num_bytes - 1;
2582         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2583                          &cached);
2584
2585         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2586         if (ordered) {
2587                 btrfs_put_ordered_extent(ordered);
2588                 goto out_unlock;
2589         }
2590
2591         trans = btrfs_join_transaction(root);
2592         if (IS_ERR(trans)) {
2593                 ret = PTR_ERR(trans);
2594                 goto out_unlock;
2595         }
2596
2597         key.objectid = backref->inum;
2598         key.type = BTRFS_EXTENT_DATA_KEY;
2599         key.offset = backref->file_pos;
2600
2601         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2602         if (ret < 0) {
2603                 goto out_free_path;
2604         } else if (ret > 0) {
2605                 ret = 0;
2606                 goto out_free_path;
2607         }
2608
2609         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2610                                 struct btrfs_file_extent_item);
2611
2612         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2613             backref->generation)
2614                 goto out_free_path;
2615
2616         btrfs_release_path(path);
2617
2618         start = backref->file_pos;
2619         if (backref->extent_offset < old->extent_offset + old->offset)
2620                 start += old->extent_offset + old->offset -
2621                          backref->extent_offset;
2622
2623         len = min(backref->extent_offset + backref->num_bytes,
2624                   old->extent_offset + old->offset + old->len);
2625         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2626
2627         ret = btrfs_drop_extents(trans, root, inode, start,
2628                                  start + len, 1);
2629         if (ret)
2630                 goto out_free_path;
2631 again:
2632         key.objectid = btrfs_ino(BTRFS_I(inode));
2633         key.type = BTRFS_EXTENT_DATA_KEY;
2634         key.offset = start;
2635
2636         path->leave_spinning = 1;
2637         if (merge) {
2638                 struct btrfs_file_extent_item *fi;
2639                 u64 extent_len;
2640                 struct btrfs_key found_key;
2641
2642                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2643                 if (ret < 0)
2644                         goto out_free_path;
2645
2646                 path->slots[0]--;
2647                 leaf = path->nodes[0];
2648                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2649
2650                 fi = btrfs_item_ptr(leaf, path->slots[0],
2651                                     struct btrfs_file_extent_item);
2652                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2653
2654                 if (extent_len + found_key.offset == start &&
2655                     relink_is_mergable(leaf, fi, new)) {
2656                         btrfs_set_file_extent_num_bytes(leaf, fi,
2657                                                         extent_len + len);
2658                         btrfs_mark_buffer_dirty(leaf);
2659                         inode_add_bytes(inode, len);
2660
2661                         ret = 1;
2662                         goto out_free_path;
2663                 } else {
2664                         merge = false;
2665                         btrfs_release_path(path);
2666                         goto again;
2667                 }
2668         }
2669
2670         ret = btrfs_insert_empty_item(trans, root, path, &key,
2671                                         sizeof(*extent));
2672         if (ret) {
2673                 btrfs_abort_transaction(trans, ret);
2674                 goto out_free_path;
2675         }
2676
2677         leaf = path->nodes[0];
2678         item = btrfs_item_ptr(leaf, path->slots[0],
2679                                 struct btrfs_file_extent_item);
2680         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2681         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2682         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2683         btrfs_set_file_extent_num_bytes(leaf, item, len);
2684         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2685         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2686         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2687         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2688         btrfs_set_file_extent_encryption(leaf, item, 0);
2689         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2690
2691         btrfs_mark_buffer_dirty(leaf);
2692         inode_add_bytes(inode, len);
2693         btrfs_release_path(path);
2694
2695         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2696                         new->disk_len, 0,
2697                         backref->root_id, backref->inum,
2698                         new->file_pos); /* start - extent_offset */
2699         if (ret) {
2700                 btrfs_abort_transaction(trans, ret);
2701                 goto out_free_path;
2702         }
2703
2704         ret = 1;
2705 out_free_path:
2706         btrfs_release_path(path);
2707         path->leave_spinning = 0;
2708         btrfs_end_transaction(trans);
2709 out_unlock:
2710         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2711                              &cached);
2712         iput(inode);
2713         return ret;
2714 }
2715
2716 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2717 {
2718         struct old_sa_defrag_extent *old, *tmp;
2719
2720         if (!new)
2721                 return;
2722
2723         list_for_each_entry_safe(old, tmp, &new->head, list) {
2724                 kfree(old);
2725         }
2726         kfree(new);
2727 }
2728
2729 static void relink_file_extents(struct new_sa_defrag_extent *new)
2730 {
2731         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2732         struct btrfs_path *path;
2733         struct sa_defrag_extent_backref *backref;
2734         struct sa_defrag_extent_backref *prev = NULL;
2735         struct rb_node *node;
2736         int ret;
2737
2738         path = btrfs_alloc_path();
2739         if (!path)
2740                 return;
2741
2742         if (!record_extent_backrefs(path, new)) {
2743                 btrfs_free_path(path);
2744                 goto out;
2745         }
2746         btrfs_release_path(path);
2747
2748         while (1) {
2749                 node = rb_first(&new->root);
2750                 if (!node)
2751                         break;
2752                 rb_erase(node, &new->root);
2753
2754                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2755
2756                 ret = relink_extent_backref(path, prev, backref);
2757                 WARN_ON(ret < 0);
2758
2759                 kfree(prev);
2760
2761                 if (ret == 1)
2762                         prev = backref;
2763                 else
2764                         prev = NULL;
2765                 cond_resched();
2766         }
2767         kfree(prev);
2768
2769         btrfs_free_path(path);
2770 out:
2771         free_sa_defrag_extent(new);
2772
2773         atomic_dec(&fs_info->defrag_running);
2774         wake_up(&fs_info->transaction_wait);
2775 }
2776
2777 static struct new_sa_defrag_extent *
2778 record_old_file_extents(struct inode *inode,
2779                         struct btrfs_ordered_extent *ordered)
2780 {
2781         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2782         struct btrfs_root *root = BTRFS_I(inode)->root;
2783         struct btrfs_path *path;
2784         struct btrfs_key key;
2785         struct old_sa_defrag_extent *old;
2786         struct new_sa_defrag_extent *new;
2787         int ret;
2788
2789         new = kmalloc(sizeof(*new), GFP_NOFS);
2790         if (!new)
2791                 return NULL;
2792
2793         new->inode = inode;
2794         new->file_pos = ordered->file_offset;
2795         new->len = ordered->len;
2796         new->bytenr = ordered->start;
2797         new->disk_len = ordered->disk_len;
2798         new->compress_type = ordered->compress_type;
2799         new->root = RB_ROOT;
2800         INIT_LIST_HEAD(&new->head);
2801
2802         path = btrfs_alloc_path();
2803         if (!path)
2804                 goto out_kfree;
2805
2806         key.objectid = btrfs_ino(BTRFS_I(inode));
2807         key.type = BTRFS_EXTENT_DATA_KEY;
2808         key.offset = new->file_pos;
2809
2810         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2811         if (ret < 0)
2812                 goto out_free_path;
2813         if (ret > 0 && path->slots[0] > 0)
2814                 path->slots[0]--;
2815
2816         /* find out all the old extents for the file range */
2817         while (1) {
2818                 struct btrfs_file_extent_item *extent;
2819                 struct extent_buffer *l;
2820                 int slot;
2821                 u64 num_bytes;
2822                 u64 offset;
2823                 u64 end;
2824                 u64 disk_bytenr;
2825                 u64 extent_offset;
2826
2827                 l = path->nodes[0];
2828                 slot = path->slots[0];
2829
2830                 if (slot >= btrfs_header_nritems(l)) {
2831                         ret = btrfs_next_leaf(root, path);
2832                         if (ret < 0)
2833                                 goto out_free_path;
2834                         else if (ret > 0)
2835                                 break;
2836                         continue;
2837                 }
2838
2839                 btrfs_item_key_to_cpu(l, &key, slot);
2840
2841                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2842                         break;
2843                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2844                         break;
2845                 if (key.offset >= new->file_pos + new->len)
2846                         break;
2847
2848                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2849
2850                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2851                 if (key.offset + num_bytes < new->file_pos)
2852                         goto next;
2853
2854                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2855                 if (!disk_bytenr)
2856                         goto next;
2857
2858                 extent_offset = btrfs_file_extent_offset(l, extent);
2859
2860                 old = kmalloc(sizeof(*old), GFP_NOFS);
2861                 if (!old)
2862                         goto out_free_path;
2863
2864                 offset = max(new->file_pos, key.offset);
2865                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2866
2867                 old->bytenr = disk_bytenr;
2868                 old->extent_offset = extent_offset;
2869                 old->offset = offset - key.offset;
2870                 old->len = end - offset;
2871                 old->new = new;
2872                 old->count = 0;
2873                 list_add_tail(&old->list, &new->head);
2874 next:
2875                 path->slots[0]++;
2876                 cond_resched();
2877         }
2878
2879         btrfs_free_path(path);
2880         atomic_inc(&fs_info->defrag_running);
2881
2882         return new;
2883
2884 out_free_path:
2885         btrfs_free_path(path);
2886 out_kfree:
2887         free_sa_defrag_extent(new);
2888         return NULL;
2889 }
2890
2891 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2892                                          u64 start, u64 len)
2893 {
2894         struct btrfs_block_group_cache *cache;
2895
2896         cache = btrfs_lookup_block_group(fs_info, start);
2897         ASSERT(cache);
2898
2899         spin_lock(&cache->lock);
2900         cache->delalloc_bytes -= len;
2901         spin_unlock(&cache->lock);
2902
2903         btrfs_put_block_group(cache);
2904 }
2905
2906 /* as ordered data IO finishes, this gets called so we can finish
2907  * an ordered extent if the range of bytes in the file it covers are
2908  * fully written.
2909  */
2910 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2911 {
2912         struct inode *inode = ordered_extent->inode;
2913         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2914         struct btrfs_root *root = BTRFS_I(inode)->root;
2915         struct btrfs_trans_handle *trans = NULL;
2916         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2917         struct extent_state *cached_state = NULL;
2918         struct new_sa_defrag_extent *new = NULL;
2919         int compress_type = 0;
2920         int ret = 0;
2921         u64 logical_len = ordered_extent->len;
2922         bool nolock;
2923         bool truncated = false;
2924         bool range_locked = false;
2925         bool clear_new_delalloc_bytes = false;
2926         bool clear_reserved_extent = true;
2927
2928         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2929             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2930             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2931                 clear_new_delalloc_bytes = true;
2932
2933         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2934
2935         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2936                 ret = -EIO;
2937                 goto out;
2938         }
2939
2940         btrfs_free_io_failure_record(BTRFS_I(inode),
2941                         ordered_extent->file_offset,
2942                         ordered_extent->file_offset +
2943                         ordered_extent->len - 1);
2944
2945         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2946                 truncated = true;
2947                 logical_len = ordered_extent->truncated_len;
2948                 /* Truncated the entire extent, don't bother adding */
2949                 if (!logical_len)
2950                         goto out;
2951         }
2952
2953         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2954                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2955
2956                 /*
2957                  * For mwrite(mmap + memset to write) case, we still reserve
2958                  * space for NOCOW range.
2959                  * As NOCOW won't cause a new delayed ref, just free the space
2960                  */
2961                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
2962                                        ordered_extent->len);
2963                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2964                 if (nolock)
2965                         trans = btrfs_join_transaction_nolock(root);
2966                 else
2967                         trans = btrfs_join_transaction(root);
2968                 if (IS_ERR(trans)) {
2969                         ret = PTR_ERR(trans);
2970                         trans = NULL;
2971                         goto out;
2972                 }
2973                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
2974                 ret = btrfs_update_inode_fallback(trans, root, inode);
2975                 if (ret) /* -ENOMEM or corruption */
2976                         btrfs_abort_transaction(trans, ret);
2977                 goto out;
2978         }
2979
2980         range_locked = true;
2981         lock_extent_bits(io_tree, ordered_extent->file_offset,
2982                          ordered_extent->file_offset + ordered_extent->len - 1,
2983                          &cached_state);
2984
2985         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2986                         ordered_extent->file_offset + ordered_extent->len - 1,
2987                         EXTENT_DEFRAG, 0, cached_state);
2988         if (ret) {
2989                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2990                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2991                         /* the inode is shared */
2992                         new = record_old_file_extents(inode, ordered_extent);
2993
2994                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2995                         ordered_extent->file_offset + ordered_extent->len - 1,
2996                         EXTENT_DEFRAG, 0, 0, &cached_state);
2997         }
2998
2999         if (nolock)
3000                 trans = btrfs_join_transaction_nolock(root);
3001         else
3002                 trans = btrfs_join_transaction(root);
3003         if (IS_ERR(trans)) {
3004                 ret = PTR_ERR(trans);
3005                 trans = NULL;
3006                 goto out;
3007         }
3008
3009         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3010
3011         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3012                 compress_type = ordered_extent->compress_type;
3013         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3014                 BUG_ON(compress_type);
3015                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3016                                        ordered_extent->len);
3017                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3018                                                 ordered_extent->file_offset,
3019                                                 ordered_extent->file_offset +
3020                                                 logical_len);
3021         } else {
3022                 BUG_ON(root == fs_info->tree_root);
3023                 ret = insert_reserved_file_extent(trans, inode,
3024                                                 ordered_extent->file_offset,
3025                                                 ordered_extent->start,
3026                                                 ordered_extent->disk_len,
3027                                                 logical_len, logical_len,
3028                                                 compress_type, 0, 0,
3029                                                 BTRFS_FILE_EXTENT_REG);
3030                 if (!ret) {
3031                         clear_reserved_extent = false;
3032                         btrfs_release_delalloc_bytes(fs_info,
3033                                                      ordered_extent->start,
3034                                                      ordered_extent->disk_len);
3035                 }
3036         }
3037         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3038                            ordered_extent->file_offset, ordered_extent->len,
3039                            trans->transid);
3040         if (ret < 0) {
3041                 btrfs_abort_transaction(trans, ret);
3042                 goto out;
3043         }
3044
3045         ret = add_pending_csums(trans, inode, &ordered_extent->list);
3046         if (ret) {
3047                 btrfs_abort_transaction(trans, ret);
3048                 goto out;
3049         }
3050
3051         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3052         ret = btrfs_update_inode_fallback(trans, root, inode);
3053         if (ret) { /* -ENOMEM or corruption */
3054                 btrfs_abort_transaction(trans, ret);
3055                 goto out;
3056         }
3057         ret = 0;
3058 out:
3059         if (range_locked || clear_new_delalloc_bytes) {
3060                 unsigned int clear_bits = 0;
3061
3062                 if (range_locked)
3063                         clear_bits |= EXTENT_LOCKED;
3064                 if (clear_new_delalloc_bytes)
3065                         clear_bits |= EXTENT_DELALLOC_NEW;
3066                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3067                                  ordered_extent->file_offset,
3068                                  ordered_extent->file_offset +
3069                                  ordered_extent->len - 1,
3070                                  clear_bits,
3071                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3072                                  0, &cached_state);
3073         }
3074
3075         if (trans)
3076                 btrfs_end_transaction(trans);
3077
3078         if (ret || truncated) {
3079                 u64 start, end;
3080
3081                 if (truncated)
3082                         start = ordered_extent->file_offset + logical_len;
3083                 else
3084                         start = ordered_extent->file_offset;
3085                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3086                 clear_extent_uptodate(io_tree, start, end, NULL);
3087
3088                 /* Drop the cache for the part of the extent we didn't write. */
3089                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3090
3091                 /*
3092                  * If the ordered extent had an IOERR or something else went
3093                  * wrong we need to return the space for this ordered extent
3094                  * back to the allocator.  We only free the extent in the
3095                  * truncated case if we didn't write out the extent at all.
3096                  *
3097                  * If we made it past insert_reserved_file_extent before we
3098                  * errored out then we don't need to do this as the accounting
3099                  * has already been done.
3100                  */
3101                 if ((ret || !logical_len) &&
3102                     clear_reserved_extent &&
3103                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3104                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3105                         btrfs_free_reserved_extent(fs_info,
3106                                                    ordered_extent->start,
3107                                                    ordered_extent->disk_len, 1);
3108         }
3109
3110
3111         /*
3112          * This needs to be done to make sure anybody waiting knows we are done
3113          * updating everything for this ordered extent.
3114          */
3115         btrfs_remove_ordered_extent(inode, ordered_extent);
3116
3117         /* for snapshot-aware defrag */
3118         if (new) {
3119                 if (ret) {
3120                         free_sa_defrag_extent(new);
3121                         atomic_dec(&fs_info->defrag_running);
3122                 } else {
3123                         relink_file_extents(new);
3124                 }
3125         }
3126
3127         /* once for us */
3128         btrfs_put_ordered_extent(ordered_extent);
3129         /* once for the tree */
3130         btrfs_put_ordered_extent(ordered_extent);
3131
3132         return ret;
3133 }
3134
3135 static void finish_ordered_fn(struct btrfs_work *work)
3136 {
3137         struct btrfs_ordered_extent *ordered_extent;
3138         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3139         btrfs_finish_ordered_io(ordered_extent);
3140 }
3141
3142 void btrfs_writepage_endio_finish_ordered(struct page *page, u64 start,
3143                                           u64 end, int uptodate)
3144 {
3145         struct inode *inode = page->mapping->host;
3146         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3147         struct btrfs_ordered_extent *ordered_extent = NULL;
3148         struct btrfs_workqueue *wq;
3149         btrfs_work_func_t func;
3150
3151         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3152
3153         ClearPagePrivate2(page);
3154         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3155                                             end - start + 1, uptodate))
3156                 return;
3157
3158         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
3159                 wq = fs_info->endio_freespace_worker;
3160                 func = btrfs_freespace_write_helper;
3161         } else {
3162                 wq = fs_info->endio_write_workers;
3163                 func = btrfs_endio_write_helper;
3164         }
3165
3166         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3167                         NULL);
3168         btrfs_queue_work(wq, &ordered_extent->work);
3169 }
3170
3171 static int __readpage_endio_check(struct inode *inode,
3172                                   struct btrfs_io_bio *io_bio,
3173                                   int icsum, struct page *page,
3174                                   int pgoff, u64 start, size_t len)
3175 {
3176         char *kaddr;
3177         u32 csum_expected;
3178         u32 csum = ~(u32)0;
3179
3180         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3181
3182         kaddr = kmap_atomic(page);
3183         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3184         btrfs_csum_final(csum, (u8 *)&csum);
3185         if (csum != csum_expected)
3186                 goto zeroit;
3187
3188         kunmap_atomic(kaddr);
3189         return 0;
3190 zeroit:
3191         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3192                                     io_bio->mirror_num);
3193         memset(kaddr + pgoff, 1, len);
3194         flush_dcache_page(page);
3195         kunmap_atomic(kaddr);
3196         return -EIO;
3197 }
3198
3199 /*
3200  * when reads are done, we need to check csums to verify the data is correct
3201  * if there's a match, we allow the bio to finish.  If not, the code in
3202  * extent_io.c will try to find good copies for us.
3203  */
3204 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3205                                       u64 phy_offset, struct page *page,
3206                                       u64 start, u64 end, int mirror)
3207 {
3208         size_t offset = start - page_offset(page);
3209         struct inode *inode = page->mapping->host;
3210         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3211         struct btrfs_root *root = BTRFS_I(inode)->root;
3212
3213         if (PageChecked(page)) {
3214                 ClearPageChecked(page);
3215                 return 0;
3216         }
3217
3218         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3219                 return 0;
3220
3221         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3222             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3223                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3224                 return 0;
3225         }
3226
3227         phy_offset >>= inode->i_sb->s_blocksize_bits;
3228         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3229                                       start, (size_t)(end - start + 1));
3230 }
3231
3232 /*
3233  * btrfs_add_delayed_iput - perform a delayed iput on @inode
3234  *
3235  * @inode: The inode we want to perform iput on
3236  *
3237  * This function uses the generic vfs_inode::i_count to track whether we should
3238  * just decrement it (in case it's > 1) or if this is the last iput then link
3239  * the inode to the delayed iput machinery. Delayed iputs are processed at
3240  * transaction commit time/superblock commit/cleaner kthread.
3241  */
3242 void btrfs_add_delayed_iput(struct inode *inode)
3243 {
3244         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3245         struct btrfs_inode *binode = BTRFS_I(inode);
3246
3247         if (atomic_add_unless(&inode->i_count, -1, 1))
3248                 return;
3249
3250         spin_lock(&fs_info->delayed_iput_lock);
3251         ASSERT(list_empty(&binode->delayed_iput));
3252         list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3253         spin_unlock(&fs_info->delayed_iput_lock);
3254         if (!test_bit(BTRFS_FS_CLEANER_RUNNING, &fs_info->flags))
3255                 wake_up_process(fs_info->cleaner_kthread);
3256 }
3257
3258 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3259 {
3260
3261         spin_lock(&fs_info->delayed_iput_lock);
3262         while (!list_empty(&fs_info->delayed_iputs)) {
3263                 struct btrfs_inode *inode;
3264
3265                 inode = list_first_entry(&fs_info->delayed_iputs,
3266                                 struct btrfs_inode, delayed_iput);
3267                 list_del_init(&inode->delayed_iput);
3268                 spin_unlock(&fs_info->delayed_iput_lock);
3269                 iput(&inode->vfs_inode);
3270                 spin_lock(&fs_info->delayed_iput_lock);
3271         }
3272         spin_unlock(&fs_info->delayed_iput_lock);
3273 }
3274
3275 /*
3276  * This creates an orphan entry for the given inode in case something goes wrong
3277  * in the middle of an unlink.
3278  */
3279 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3280                      struct btrfs_inode *inode)
3281 {
3282         int ret;
3283
3284         ret = btrfs_insert_orphan_item(trans, inode->root, btrfs_ino(inode));
3285         if (ret && ret != -EEXIST) {
3286                 btrfs_abort_transaction(trans, ret);
3287                 return ret;
3288         }
3289
3290         return 0;
3291 }
3292
3293 /*
3294  * We have done the delete so we can go ahead and remove the orphan item for
3295  * this particular inode.
3296  */
3297 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3298                             struct btrfs_inode *inode)
3299 {
3300         return btrfs_del_orphan_item(trans, inode->root, btrfs_ino(inode));
3301 }
3302
3303 /*
3304  * this cleans up any orphans that may be left on the list from the last use
3305  * of this root.
3306  */
3307 int btrfs_orphan_cleanup(struct btrfs_root *root)
3308 {
3309         struct btrfs_fs_info *fs_info = root->fs_info;
3310         struct btrfs_path *path;
3311         struct extent_buffer *leaf;
3312         struct btrfs_key key, found_key;
3313         struct btrfs_trans_handle *trans;
3314         struct inode *inode;
3315         u64 last_objectid = 0;
3316         int ret = 0, nr_unlink = 0;
3317
3318         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3319                 return 0;
3320
3321         path = btrfs_alloc_path();
3322         if (!path) {
3323                 ret = -ENOMEM;
3324                 goto out;
3325         }
3326         path->reada = READA_BACK;
3327
3328         key.objectid = BTRFS_ORPHAN_OBJECTID;
3329         key.type = BTRFS_ORPHAN_ITEM_KEY;
3330         key.offset = (u64)-1;
3331
3332         while (1) {
3333                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3334                 if (ret < 0)
3335                         goto out;
3336
3337                 /*
3338                  * if ret == 0 means we found what we were searching for, which
3339                  * is weird, but possible, so only screw with path if we didn't
3340                  * find the key and see if we have stuff that matches
3341                  */
3342                 if (ret > 0) {
3343                         ret = 0;
3344                         if (path->slots[0] == 0)
3345                                 break;
3346                         path->slots[0]--;
3347                 }
3348
3349                 /* pull out the item */
3350                 leaf = path->nodes[0];
3351                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3352
3353                 /* make sure the item matches what we want */
3354                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3355                         break;
3356                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3357                         break;
3358
3359                 /* release the path since we're done with it */
3360                 btrfs_release_path(path);
3361
3362                 /*
3363                  * this is where we are basically btrfs_lookup, without the
3364                  * crossing root thing.  we store the inode number in the
3365                  * offset of the orphan item.
3366                  */
3367
3368                 if (found_key.offset == last_objectid) {
3369                         btrfs_err(fs_info,
3370                                   "Error removing orphan entry, stopping orphan cleanup");
3371                         ret = -EINVAL;
3372                         goto out;
3373                 }
3374
3375                 last_objectid = found_key.offset;
3376
3377                 found_key.objectid = found_key.offset;
3378                 found_key.type = BTRFS_INODE_ITEM_KEY;
3379                 found_key.offset = 0;
3380                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3381                 ret = PTR_ERR_OR_ZERO(inode);
3382                 if (ret && ret != -ENOENT)
3383                         goto out;
3384
3385                 if (ret == -ENOENT && root == fs_info->tree_root) {
3386                         struct btrfs_root *dead_root;
3387                         struct btrfs_fs_info *fs_info = root->fs_info;
3388                         int is_dead_root = 0;
3389
3390                         /*
3391                          * this is an orphan in the tree root. Currently these
3392                          * could come from 2 sources:
3393                          *  a) a snapshot deletion in progress
3394                          *  b) a free space cache inode
3395                          * We need to distinguish those two, as the snapshot
3396                          * orphan must not get deleted.
3397                          * find_dead_roots already ran before us, so if this
3398                          * is a snapshot deletion, we should find the root
3399                          * in the dead_roots list
3400                          */
3401                         spin_lock(&fs_info->trans_lock);
3402                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3403                                             root_list) {
3404                                 if (dead_root->root_key.objectid ==
3405                                     found_key.objectid) {
3406                                         is_dead_root = 1;
3407                                         break;
3408                                 }
3409                         }
3410                         spin_unlock(&fs_info->trans_lock);
3411                         if (is_dead_root) {
3412                                 /* prevent this orphan from being found again */
3413                                 key.offset = found_key.objectid - 1;
3414                                 continue;
3415                         }
3416
3417                 }
3418
3419                 /*
3420                  * If we have an inode with links, there are a couple of
3421                  * possibilities. Old kernels (before v3.12) used to create an
3422                  * orphan item for truncate indicating that there were possibly
3423                  * extent items past i_size that needed to be deleted. In v3.12,
3424                  * truncate was changed to update i_size in sync with the extent
3425                  * items, but the (useless) orphan item was still created. Since
3426                  * v4.18, we don't create the orphan item for truncate at all.
3427                  *
3428                  * So, this item could mean that we need to do a truncate, but
3429                  * only if this filesystem was last used on a pre-v3.12 kernel
3430                  * and was not cleanly unmounted. The odds of that are quite
3431                  * slim, and it's a pain to do the truncate now, so just delete
3432                  * the orphan item.
3433                  *
3434                  * It's also possible that this orphan item was supposed to be
3435                  * deleted but wasn't. The inode number may have been reused,
3436                  * but either way, we can delete the orphan item.
3437                  */
3438                 if (ret == -ENOENT || inode->i_nlink) {
3439                         if (!ret)
3440                                 iput(inode);
3441                         trans = btrfs_start_transaction(root, 1);
3442                         if (IS_ERR(trans)) {
3443                                 ret = PTR_ERR(trans);
3444                                 goto out;
3445                         }
3446                         btrfs_debug(fs_info, "auto deleting %Lu",
3447                                     found_key.objectid);
3448                         ret = btrfs_del_orphan_item(trans, root,
3449                                                     found_key.objectid);
3450                         btrfs_end_transaction(trans);
3451                         if (ret)
3452                                 goto out;
3453                         continue;
3454                 }
3455
3456                 nr_unlink++;
3457
3458                 /* this will do delete_inode and everything for us */
3459                 iput(inode);
3460         }
3461         /* release the path since we're done with it */
3462         btrfs_release_path(path);
3463
3464         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3465
3466         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3467                 trans = btrfs_join_transaction(root);
3468                 if (!IS_ERR(trans))
3469                         btrfs_end_transaction(trans);
3470         }
3471
3472         if (nr_unlink)
3473                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3474
3475 out:
3476         if (ret)
3477                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3478         btrfs_free_path(path);
3479         return ret;
3480 }
3481
3482 /*
3483  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3484  * don't find any xattrs, we know there can't be any acls.
3485  *
3486  * slot is the slot the inode is in, objectid is the objectid of the inode
3487  */
3488 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3489                                           int slot, u64 objectid,
3490                                           int *first_xattr_slot)
3491 {
3492         u32 nritems = btrfs_header_nritems(leaf);
3493         struct btrfs_key found_key;
3494         static u64 xattr_access = 0;
3495         static u64 xattr_default = 0;
3496         int scanned = 0;
3497
3498         if (!xattr_access) {
3499                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3500                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3501                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3502                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3503         }
3504
3505         slot++;
3506         *first_xattr_slot = -1;
3507         while (slot < nritems) {
3508                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3509
3510                 /* we found a different objectid, there must not be acls */
3511                 if (found_key.objectid != objectid)
3512                         return 0;
3513
3514                 /* we found an xattr, assume we've got an acl */
3515                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3516                         if (*first_xattr_slot == -1)
3517                                 *first_xattr_slot = slot;
3518                         if (found_key.offset == xattr_access ||
3519                             found_key.offset == xattr_default)
3520                                 return 1;
3521                 }
3522
3523                 /*
3524                  * we found a key greater than an xattr key, there can't
3525                  * be any acls later on
3526                  */
3527                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3528                         return 0;
3529
3530                 slot++;
3531                 scanned++;
3532
3533                 /*
3534                  * it goes inode, inode backrefs, xattrs, extents,
3535                  * so if there are a ton of hard links to an inode there can
3536                  * be a lot of backrefs.  Don't waste time searching too hard,
3537                  * this is just an optimization
3538                  */
3539                 if (scanned >= 8)
3540                         break;
3541         }
3542         /* we hit the end of the leaf before we found an xattr or
3543          * something larger than an xattr.  We have to assume the inode
3544          * has acls
3545          */
3546         if (*first_xattr_slot == -1)
3547                 *first_xattr_slot = slot;
3548         return 1;
3549 }
3550
3551 /*
3552  * read an inode from the btree into the in-memory inode
3553  */
3554 static int btrfs_read_locked_inode(struct inode *inode,
3555                                    struct btrfs_path *in_path)
3556 {
3557         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3558         struct btrfs_path *path = in_path;
3559         struct extent_buffer *leaf;
3560         struct btrfs_inode_item *inode_item;
3561         struct btrfs_root *root = BTRFS_I(inode)->root;
3562         struct btrfs_key location;
3563         unsigned long ptr;
3564         int maybe_acls;
3565         u32 rdev;
3566         int ret;
3567         bool filled = false;
3568         int first_xattr_slot;
3569
3570         ret = btrfs_fill_inode(inode, &rdev);
3571         if (!ret)
3572                 filled = true;
3573
3574         if (!path) {
3575                 path = btrfs_alloc_path();
3576                 if (!path)
3577                         return -ENOMEM;
3578         }
3579
3580         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3581
3582         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3583         if (ret) {
3584                 if (path != in_path)
3585                         btrfs_free_path(path);
3586                 return ret;
3587         }
3588
3589         leaf = path->nodes[0];
3590
3591         if (filled)
3592                 goto cache_index;
3593
3594         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3595                                     struct btrfs_inode_item);
3596         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3597         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3598         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3599         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3600         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3601
3602         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3603         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3604
3605         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3606         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3607
3608         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3609         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3610
3611         BTRFS_I(inode)->i_otime.tv_sec =
3612                 btrfs_timespec_sec(leaf, &inode_item->otime);
3613         BTRFS_I(inode)->i_otime.tv_nsec =
3614                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3615
3616         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3617         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3618         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3619
3620         inode_set_iversion_queried(inode,
3621                                    btrfs_inode_sequence(leaf, inode_item));
3622         inode->i_generation = BTRFS_I(inode)->generation;
3623         inode->i_rdev = 0;
3624         rdev = btrfs_inode_rdev(leaf, inode_item);
3625
3626         BTRFS_I(inode)->index_cnt = (u64)-1;
3627         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3628
3629 cache_index:
3630         /*
3631          * If we were modified in the current generation and evicted from memory
3632          * and then re-read we need to do a full sync since we don't have any
3633          * idea about which extents were modified before we were evicted from
3634          * cache.
3635          *
3636          * This is required for both inode re-read from disk and delayed inode
3637          * in delayed_nodes_tree.
3638          */
3639         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3640                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3641                         &BTRFS_I(inode)->runtime_flags);
3642
3643         /*
3644          * We don't persist the id of the transaction where an unlink operation
3645          * against the inode was last made. So here we assume the inode might
3646          * have been evicted, and therefore the exact value of last_unlink_trans
3647          * lost, and set it to last_trans to avoid metadata inconsistencies
3648          * between the inode and its parent if the inode is fsync'ed and the log
3649          * replayed. For example, in the scenario:
3650          *
3651          * touch mydir/foo
3652          * ln mydir/foo mydir/bar
3653          * sync
3654          * unlink mydir/bar
3655          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3656          * xfs_io -c fsync mydir/foo
3657          * <power failure>
3658          * mount fs, triggers fsync log replay
3659          *
3660          * We must make sure that when we fsync our inode foo we also log its
3661          * parent inode, otherwise after log replay the parent still has the
3662          * dentry with the "bar" name but our inode foo has a link count of 1
3663          * and doesn't have an inode ref with the name "bar" anymore.
3664          *
3665          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3666          * but it guarantees correctness at the expense of occasional full
3667          * transaction commits on fsync if our inode is a directory, or if our
3668          * inode is not a directory, logging its parent unnecessarily.
3669          */
3670         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3671         /*
3672          * Similar reasoning for last_link_trans, needs to be set otherwise
3673          * for a case like the following:
3674          *
3675          * mkdir A
3676          * touch foo
3677          * ln foo A/bar
3678          * echo 2 > /proc/sys/vm/drop_caches
3679          * fsync foo
3680          * <power failure>
3681          *
3682          * Would result in link bar and directory A not existing after the power
3683          * failure.
3684          */
3685         BTRFS_I(inode)->last_link_trans = BTRFS_I(inode)->last_trans;
3686
3687         path->slots[0]++;
3688         if (inode->i_nlink != 1 ||
3689             path->slots[0] >= btrfs_header_nritems(leaf))
3690                 goto cache_acl;
3691
3692         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3693         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3694                 goto cache_acl;
3695
3696         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3697         if (location.type == BTRFS_INODE_REF_KEY) {
3698                 struct btrfs_inode_ref *ref;
3699
3700                 ref = (struct btrfs_inode_ref *)ptr;
3701                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3702         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3703                 struct btrfs_inode_extref *extref;
3704
3705                 extref = (struct btrfs_inode_extref *)ptr;
3706                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3707                                                                      extref);
3708         }
3709 cache_acl:
3710         /*
3711          * try to precache a NULL acl entry for files that don't have
3712          * any xattrs or acls
3713          */
3714         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3715                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3716         if (first_xattr_slot != -1) {
3717                 path->slots[0] = first_xattr_slot;
3718                 ret = btrfs_load_inode_props(inode, path);
3719                 if (ret)
3720                         btrfs_err(fs_info,
3721                                   "error loading props for ino %llu (root %llu): %d",
3722                                   btrfs_ino(BTRFS_I(inode)),
3723                                   root->root_key.objectid, ret);
3724         }
3725         if (path != in_path)
3726                 btrfs_free_path(path);
3727
3728         if (!maybe_acls)
3729                 cache_no_acl(inode);
3730
3731         switch (inode->i_mode & S_IFMT) {
3732         case S_IFREG:
3733                 inode->i_mapping->a_ops = &btrfs_aops;
3734                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3735                 inode->i_fop = &btrfs_file_operations;
3736                 inode->i_op = &btrfs_file_inode_operations;
3737                 break;
3738         case S_IFDIR:
3739                 inode->i_fop = &btrfs_dir_file_operations;
3740                 inode->i_op = &btrfs_dir_inode_operations;
3741                 break;
3742         case S_IFLNK:
3743                 inode->i_op = &btrfs_symlink_inode_operations;
3744                 inode_nohighmem(inode);
3745                 inode->i_mapping->a_ops = &btrfs_aops;
3746                 break;
3747         default:
3748                 inode->i_op = &btrfs_special_inode_operations;
3749                 init_special_inode(inode, inode->i_mode, rdev);
3750                 break;
3751         }
3752
3753         btrfs_sync_inode_flags_to_i_flags(inode);
3754         return 0;
3755 }
3756
3757 /*
3758  * given a leaf and an inode, copy the inode fields into the leaf
3759  */
3760 static void fill_inode_item(struct btrfs_trans_handle *trans,
3761                             struct extent_buffer *leaf,
3762                             struct btrfs_inode_item *item,
3763                             struct inode *inode)
3764 {
3765         struct btrfs_map_token token;
3766
3767         btrfs_init_map_token(&token);
3768
3769         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3770         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3771         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3772                                    &token);
3773         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3774         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3775
3776         btrfs_set_token_timespec_sec(leaf, &item->atime,
3777                                      inode->i_atime.tv_sec, &token);
3778         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3779                                       inode->i_atime.tv_nsec, &token);
3780
3781         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3782                                      inode->i_mtime.tv_sec, &token);
3783         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3784                                       inode->i_mtime.tv_nsec, &token);
3785
3786         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3787                                      inode->i_ctime.tv_sec, &token);
3788         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3789                                       inode->i_ctime.tv_nsec, &token);
3790
3791         btrfs_set_token_timespec_sec(leaf, &item->otime,
3792                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3793         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3794                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3795
3796         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3797                                      &token);
3798         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3799                                          &token);
3800         btrfs_set_token_inode_sequence(leaf, item, inode_peek_iversion(inode),
3801                                        &token);
3802         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3803         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3804         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3805         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3806 }
3807
3808 /*
3809  * copy everything in the in-memory inode into the btree.
3810  */
3811 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3812                                 struct btrfs_root *root, struct inode *inode)
3813 {
3814         struct btrfs_inode_item *inode_item;
3815         struct btrfs_path *path;
3816         struct extent_buffer *leaf;
3817         int ret;
3818
3819         path = btrfs_alloc_path();
3820         if (!path)
3821                 return -ENOMEM;
3822
3823         path->leave_spinning = 1;
3824         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3825                                  1);
3826         if (ret) {
3827                 if (ret > 0)
3828                         ret = -ENOENT;
3829                 goto failed;
3830         }
3831
3832         leaf = path->nodes[0];
3833         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3834                                     struct btrfs_inode_item);
3835
3836         fill_inode_item(trans, leaf, inode_item, inode);
3837         btrfs_mark_buffer_dirty(leaf);
3838         btrfs_set_inode_last_trans(trans, inode);
3839         ret = 0;
3840 failed:
3841         btrfs_free_path(path);
3842         return ret;
3843 }
3844
3845 /*
3846  * copy everything in the in-memory inode into the btree.
3847  */
3848 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3849                                 struct btrfs_root *root, struct inode *inode)
3850 {
3851         struct btrfs_fs_info *fs_info = root->fs_info;
3852         int ret;
3853
3854         /*
3855          * If the inode is a free space inode, we can deadlock during commit
3856          * if we put it into the delayed code.
3857          *
3858          * The data relocation inode should also be directly updated
3859          * without delay
3860          */
3861         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
3862             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3863             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
3864                 btrfs_update_root_times(trans, root);
3865
3866                 ret = btrfs_delayed_update_inode(trans, root, inode);
3867                 if (!ret)
3868                         btrfs_set_inode_last_trans(trans, inode);
3869                 return ret;
3870         }
3871
3872         return btrfs_update_inode_item(trans, root, inode);
3873 }
3874
3875 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3876                                          struct btrfs_root *root,
3877                                          struct inode *inode)
3878 {
3879         int ret;
3880
3881         ret = btrfs_update_inode(trans, root, inode);
3882         if (ret == -ENOSPC)
3883                 return btrfs_update_inode_item(trans, root, inode);
3884         return ret;
3885 }
3886
3887 /*
3888  * unlink helper that gets used here in inode.c and in the tree logging
3889  * recovery code.  It remove a link in a directory with a given name, and
3890  * also drops the back refs in the inode to the directory
3891  */
3892 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3893                                 struct btrfs_root *root,
3894                                 struct btrfs_inode *dir,
3895                                 struct btrfs_inode *inode,
3896                                 const char *name, int name_len)
3897 {
3898         struct btrfs_fs_info *fs_info = root->fs_info;
3899         struct btrfs_path *path;
3900         int ret = 0;
3901         struct extent_buffer *leaf;
3902         struct btrfs_dir_item *di;
3903         struct btrfs_key key;
3904         u64 index;
3905         u64 ino = btrfs_ino(inode);
3906         u64 dir_ino = btrfs_ino(dir);
3907
3908         path = btrfs_alloc_path();
3909         if (!path) {
3910                 ret = -ENOMEM;
3911                 goto out;
3912         }
3913
3914         path->leave_spinning = 1;
3915         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3916                                     name, name_len, -1);
3917         if (IS_ERR_OR_NULL(di)) {
3918                 ret = di ? PTR_ERR(di) : -ENOENT;
3919                 goto err;
3920         }
3921         leaf = path->nodes[0];
3922         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3923         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3924         if (ret)
3925                 goto err;
3926         btrfs_release_path(path);
3927
3928         /*
3929          * If we don't have dir index, we have to get it by looking up
3930          * the inode ref, since we get the inode ref, remove it directly,
3931          * it is unnecessary to do delayed deletion.
3932          *
3933          * But if we have dir index, needn't search inode ref to get it.
3934          * Since the inode ref is close to the inode item, it is better
3935          * that we delay to delete it, and just do this deletion when
3936          * we update the inode item.
3937          */
3938         if (inode->dir_index) {
3939                 ret = btrfs_delayed_delete_inode_ref(inode);
3940                 if (!ret) {
3941                         index = inode->dir_index;
3942                         goto skip_backref;
3943                 }
3944         }
3945
3946         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3947                                   dir_ino, &index);
3948         if (ret) {
3949                 btrfs_info(fs_info,
3950                         "failed to delete reference to %.*s, inode %llu parent %llu",
3951                         name_len, name, ino, dir_ino);
3952                 btrfs_abort_transaction(trans, ret);
3953                 goto err;
3954         }
3955 skip_backref:
3956         ret = btrfs_delete_delayed_dir_index(trans, dir, index);
3957         if (ret) {
3958                 btrfs_abort_transaction(trans, ret);
3959                 goto err;
3960         }
3961
3962         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
3963                         dir_ino);
3964         if (ret != 0 && ret != -ENOENT) {
3965                 btrfs_abort_transaction(trans, ret);
3966                 goto err;
3967         }
3968
3969         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
3970                         index);
3971         if (ret == -ENOENT)
3972                 ret = 0;
3973         else if (ret)
3974                 btrfs_abort_transaction(trans, ret);
3975 err:
3976         btrfs_free_path(path);
3977         if (ret)
3978                 goto out;
3979
3980         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
3981         inode_inc_iversion(&inode->vfs_inode);
3982         inode_inc_iversion(&dir->vfs_inode);
3983         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
3984                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
3985         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
3986 out:
3987         return ret;
3988 }
3989
3990 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3991                        struct btrfs_root *root,
3992                        struct btrfs_inode *dir, struct btrfs_inode *inode,
3993                        const char *name, int name_len)
3994 {
3995         int ret;
3996         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
3997         if (!ret) {
3998                 drop_nlink(&inode->vfs_inode);
3999                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4000         }
4001         return ret;
4002 }
4003
4004 /*
4005  * helper to start transaction for unlink and rmdir.
4006  *
4007  * unlink and rmdir are special in btrfs, they do not always free space, so
4008  * if we cannot make our reservations the normal way try and see if there is
4009  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4010  * allow the unlink to occur.
4011  */
4012 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4013 {
4014         struct btrfs_root *root = BTRFS_I(dir)->root;
4015
4016         /*
4017          * 1 for the possible orphan item
4018          * 1 for the dir item
4019          * 1 for the dir index
4020          * 1 for the inode ref
4021          * 1 for the inode
4022          */
4023         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4024 }
4025
4026 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4027 {
4028         struct btrfs_root *root = BTRFS_I(dir)->root;
4029         struct btrfs_trans_handle *trans;
4030         struct inode *inode = d_inode(dentry);
4031         int ret;
4032
4033         trans = __unlink_start_trans(dir);
4034         if (IS_ERR(trans))
4035                 return PTR_ERR(trans);
4036
4037         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4038                         0);
4039
4040         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4041                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4042                         dentry->d_name.len);
4043         if (ret)
4044                 goto out;
4045
4046         if (inode->i_nlink == 0) {
4047                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4048                 if (ret)
4049                         goto out;
4050         }
4051
4052 out:
4053         btrfs_end_transaction(trans);
4054         btrfs_btree_balance_dirty(root->fs_info);
4055         return ret;
4056 }
4057
4058 static int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4059                                struct inode *dir, u64 objectid,
4060                                const char *name, int name_len)
4061 {
4062         struct btrfs_root *root = BTRFS_I(dir)->root;
4063         struct btrfs_path *path;
4064         struct extent_buffer *leaf;
4065         struct btrfs_dir_item *di;
4066         struct btrfs_key key;
4067         u64 index;
4068         int ret;
4069         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4070
4071         path = btrfs_alloc_path();
4072         if (!path)
4073                 return -ENOMEM;
4074
4075         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4076                                    name, name_len, -1);
4077         if (IS_ERR_OR_NULL(di)) {
4078                 ret = di ? PTR_ERR(di) : -ENOENT;
4079                 goto out;
4080         }
4081
4082         leaf = path->nodes[0];
4083         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4084         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4085         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4086         if (ret) {
4087                 btrfs_abort_transaction(trans, ret);
4088                 goto out;
4089         }
4090         btrfs_release_path(path);
4091
4092         ret = btrfs_del_root_ref(trans, objectid, root->root_key.objectid,
4093                                  dir_ino, &index, name, name_len);
4094         if (ret < 0) {
4095                 if (ret != -ENOENT) {
4096                         btrfs_abort_transaction(trans, ret);
4097                         goto out;
4098                 }
4099                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4100                                                  name, name_len);
4101                 if (IS_ERR_OR_NULL(di)) {
4102                         if (!di)
4103                                 ret = -ENOENT;
4104                         else
4105                                 ret = PTR_ERR(di);
4106                         btrfs_abort_transaction(trans, ret);
4107                         goto out;
4108                 }
4109
4110                 leaf = path->nodes[0];
4111                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4112                 index = key.offset;
4113         }
4114         btrfs_release_path(path);
4115
4116         ret = btrfs_delete_delayed_dir_index(trans, BTRFS_I(dir), index);
4117         if (ret) {
4118                 btrfs_abort_transaction(trans, ret);
4119                 goto out;
4120         }
4121
4122         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4123         inode_inc_iversion(dir);
4124         dir->i_mtime = dir->i_ctime = current_time(dir);
4125         ret = btrfs_update_inode_fallback(trans, root, dir);
4126         if (ret)
4127                 btrfs_abort_transaction(trans, ret);
4128 out:
4129         btrfs_free_path(path);
4130         return ret;
4131 }
4132
4133 /*
4134  * Helper to check if the subvolume references other subvolumes or if it's
4135  * default.
4136  */
4137 static noinline int may_destroy_subvol(struct btrfs_root *root)
4138 {
4139         struct btrfs_fs_info *fs_info = root->fs_info;
4140         struct btrfs_path *path;
4141         struct btrfs_dir_item *di;
4142         struct btrfs_key key;
4143         u64 dir_id;
4144         int ret;
4145
4146         path = btrfs_alloc_path();
4147         if (!path)
4148                 return -ENOMEM;
4149
4150         /* Make sure this root isn't set as the default subvol */
4151         dir_id = btrfs_super_root_dir(fs_info->super_copy);
4152         di = btrfs_lookup_dir_item(NULL, fs_info->tree_root, path,
4153                                    dir_id, "default", 7, 0);
4154         if (di && !IS_ERR(di)) {
4155                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
4156                 if (key.objectid == root->root_key.objectid) {
4157                         ret = -EPERM;
4158                         btrfs_err(fs_info,
4159                                   "deleting default subvolume %llu is not allowed",
4160                                   key.objectid);
4161                         goto out;
4162                 }
4163                 btrfs_release_path(path);
4164         }
4165
4166         key.objectid = root->root_key.objectid;
4167         key.type = BTRFS_ROOT_REF_KEY;
4168         key.offset = (u64)-1;
4169
4170         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
4171         if (ret < 0)
4172                 goto out;
4173         BUG_ON(ret == 0);
4174
4175         ret = 0;
4176         if (path->slots[0] > 0) {
4177                 path->slots[0]--;
4178                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
4179                 if (key.objectid == root->root_key.objectid &&
4180                     key.type == BTRFS_ROOT_REF_KEY)
4181                         ret = -ENOTEMPTY;
4182         }
4183 out:
4184         btrfs_free_path(path);
4185         return ret;
4186 }
4187
4188 /* Delete all dentries for inodes belonging to the root */
4189 static void btrfs_prune_dentries(struct btrfs_root *root)
4190 {
4191         struct btrfs_fs_info *fs_info = root->fs_info;
4192         struct rb_node *node;
4193         struct rb_node *prev;
4194         struct btrfs_inode *entry;
4195         struct inode *inode;
4196         u64 objectid = 0;
4197
4198         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
4199                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
4200
4201         spin_lock(&root->inode_lock);
4202 again:
4203         node = root->inode_tree.rb_node;
4204         prev = NULL;
4205         while (node) {
4206                 prev = node;
4207                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4208
4209                 if (objectid < btrfs_ino(entry))
4210                         node = node->rb_left;
4211                 else if (objectid > btrfs_ino(entry))
4212                         node = node->rb_right;
4213                 else
4214                         break;
4215         }
4216         if (!node) {
4217                 while (prev) {
4218                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
4219                         if (objectid <= btrfs_ino(entry)) {
4220                                 node = prev;
4221                                 break;
4222                         }
4223                         prev = rb_next(prev);
4224                 }
4225         }
4226         while (node) {
4227                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4228                 objectid = btrfs_ino(entry) + 1;
4229                 inode = igrab(&entry->vfs_inode);
4230                 if (inode) {
4231                         spin_unlock(&root->inode_lock);
4232                         if (atomic_read(&inode->i_count) > 1)
4233                                 d_prune_aliases(inode);
4234                         /*
4235                          * btrfs_drop_inode will have it removed from the inode
4236                          * cache when its usage count hits zero.
4237                          */
4238                         iput(inode);
4239                         cond_resched();
4240                         spin_lock(&root->inode_lock);
4241                         goto again;
4242                 }
4243
4244                 if (cond_resched_lock(&root->inode_lock))
4245                         goto again;
4246
4247                 node = rb_next(node);
4248         }
4249         spin_unlock(&root->inode_lock);
4250 }
4251
4252 int btrfs_delete_subvolume(struct inode *dir, struct dentry *dentry)
4253 {
4254         struct btrfs_fs_info *fs_info = btrfs_sb(dentry->d_sb);
4255         struct btrfs_root *root = BTRFS_I(dir)->root;
4256         struct inode *inode = d_inode(dentry);
4257         struct btrfs_root *dest = BTRFS_I(inode)->root;
4258         struct btrfs_trans_handle *trans;
4259         struct btrfs_block_rsv block_rsv;
4260         u64 root_flags;
4261         int ret;
4262         int err;
4263
4264         /*
4265          * Don't allow to delete a subvolume with send in progress. This is
4266          * inside the inode lock so the error handling that has to drop the bit
4267          * again is not run concurrently.
4268          */
4269         spin_lock(&dest->root_item_lock);
4270         if (dest->send_in_progress) {
4271                 spin_unlock(&dest->root_item_lock);
4272                 btrfs_warn(fs_info,
4273                            "attempt to delete subvolume %llu during send",
4274                            dest->root_key.objectid);
4275                 return -EPERM;
4276         }
4277         root_flags = btrfs_root_flags(&dest->root_item);
4278         btrfs_set_root_flags(&dest->root_item,
4279                              root_flags | BTRFS_ROOT_SUBVOL_DEAD);
4280         spin_unlock(&dest->root_item_lock);
4281
4282         down_write(&fs_info->subvol_sem);
4283
4284         err = may_destroy_subvol(dest);
4285         if (err)
4286                 goto out_up_write;
4287
4288         btrfs_init_block_rsv(&block_rsv, BTRFS_BLOCK_RSV_TEMP);
4289         /*
4290          * One for dir inode,
4291          * two for dir entries,
4292          * two for root ref/backref.
4293          */
4294         err = btrfs_subvolume_reserve_metadata(root, &block_rsv, 5, true);
4295         if (err)
4296                 goto out_up_write;
4297
4298         trans = btrfs_start_transaction(root, 0);
4299         if (IS_ERR(trans)) {
4300                 err = PTR_ERR(trans);
4301                 goto out_release;
4302         }
4303         trans->block_rsv = &block_rsv;
4304         trans->bytes_reserved = block_rsv.size;
4305
4306         btrfs_record_snapshot_destroy(trans, BTRFS_I(dir));
4307
4308         ret = btrfs_unlink_subvol(trans, dir, dest->root_key.objectid,
4309                                   dentry->d_name.name, dentry->d_name.len);
4310         if (ret) {
4311                 err = ret;
4312                 btrfs_abort_transaction(trans, ret);
4313                 goto out_end_trans;
4314         }
4315
4316         btrfs_record_root_in_trans(trans, dest);
4317
4318         memset(&dest->root_item.drop_progress, 0,
4319                 sizeof(dest->root_item.drop_progress));
4320         dest->root_item.drop_level = 0;
4321         btrfs_set_root_refs(&dest->root_item, 0);
4322
4323         if (!test_and_set_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &dest->state)) {
4324                 ret = btrfs_insert_orphan_item(trans,
4325                                         fs_info->tree_root,
4326                                         dest->root_key.objectid);
4327                 if (ret) {
4328                         btrfs_abort_transaction(trans, ret);
4329                         err = ret;
4330                         goto out_end_trans;
4331                 }
4332         }
4333
4334         ret = btrfs_uuid_tree_remove(trans, dest->root_item.uuid,
4335                                   BTRFS_UUID_KEY_SUBVOL,
4336                                   dest->root_key.objectid);
4337         if (ret && ret != -ENOENT) {
4338                 btrfs_abort_transaction(trans, ret);
4339                 err = ret;
4340                 goto out_end_trans;
4341         }
4342         if (!btrfs_is_empty_uuid(dest->root_item.received_uuid)) {
4343                 ret = btrfs_uuid_tree_remove(trans,
4344                                           dest->root_item.received_uuid,
4345                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
4346                                           dest->root_key.objectid);
4347                 if (ret && ret != -ENOENT) {
4348                         btrfs_abort_transaction(trans, ret);
4349                         err = ret;
4350                         goto out_end_trans;
4351                 }
4352         }
4353
4354 out_end_trans:
4355         trans->block_rsv = NULL;
4356         trans->bytes_reserved = 0;
4357         ret = btrfs_end_transaction(trans);
4358         if (ret && !err)
4359                 err = ret;
4360         inode->i_flags |= S_DEAD;
4361 out_release:
4362         btrfs_subvolume_release_metadata(fs_info, &block_rsv);
4363 out_up_write:
4364         up_write(&fs_info->subvol_sem);
4365         if (err) {
4366                 spin_lock(&dest->root_item_lock);
4367                 root_flags = btrfs_root_flags(&dest->root_item);
4368                 btrfs_set_root_flags(&dest->root_item,
4369                                 root_flags & ~BTRFS_ROOT_SUBVOL_DEAD);
4370                 spin_unlock(&dest->root_item_lock);
4371         } else {
4372                 d_invalidate(dentry);
4373                 btrfs_prune_dentries(dest);
4374                 ASSERT(dest->send_in_progress == 0);
4375
4376                 /* the last ref */
4377                 if (dest->ino_cache_inode) {
4378                         iput(dest->ino_cache_inode);
4379                         dest->ino_cache_inode = NULL;
4380                 }
4381         }
4382
4383         return err;
4384 }
4385
4386 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4387 {
4388         struct inode *inode = d_inode(dentry);
4389         int err = 0;
4390         struct btrfs_root *root = BTRFS_I(dir)->root;
4391         struct btrfs_trans_handle *trans;
4392         u64 last_unlink_trans;
4393
4394         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4395                 return -ENOTEMPTY;
4396         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4397                 return btrfs_delete_subvolume(dir, dentry);
4398
4399         trans = __unlink_start_trans(dir);
4400         if (IS_ERR(trans))
4401                 return PTR_ERR(trans);
4402
4403         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4404                 err = btrfs_unlink_subvol(trans, dir,
4405                                           BTRFS_I(inode)->location.objectid,
4406                                           dentry->d_name.name,
4407                                           dentry->d_name.len);
4408                 goto out;
4409         }
4410
4411         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4412         if (err)
4413                 goto out;
4414
4415         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4416
4417         /* now the directory is empty */
4418         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4419                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4420                         dentry->d_name.len);
4421         if (!err) {
4422                 btrfs_i_size_write(BTRFS_I(inode), 0);
4423                 /*
4424                  * Propagate the last_unlink_trans value of the deleted dir to
4425                  * its parent directory. This is to prevent an unrecoverable
4426                  * log tree in the case we do something like this:
4427                  * 1) create dir foo
4428                  * 2) create snapshot under dir foo
4429                  * 3) delete the snapshot
4430                  * 4) rmdir foo
4431                  * 5) mkdir foo
4432                  * 6) fsync foo or some file inside foo
4433                  */
4434                 if (last_unlink_trans >= trans->transid)
4435                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4436         }
4437 out:
4438         btrfs_end_transaction(trans);
4439         btrfs_btree_balance_dirty(root->fs_info);
4440
4441         return err;
4442 }
4443
4444 /*
4445  * Return this if we need to call truncate_block for the last bit of the
4446  * truncate.
4447  */
4448 #define NEED_TRUNCATE_BLOCK 1
4449
4450 /*
4451  * this can truncate away extent items, csum items and directory items.
4452  * It starts at a high offset and removes keys until it can't find
4453  * any higher than new_size
4454  *
4455  * csum items that cross the new i_size are truncated to the new size
4456  * as well.
4457  *
4458  * min_type is the minimum key type to truncate down to.  If set to 0, this
4459  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4460  */
4461 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4462                                struct btrfs_root *root,
4463                                struct inode *inode,
4464                                u64 new_size, u32 min_type)
4465 {
4466         struct btrfs_fs_info *fs_info = root->fs_info;
4467         struct btrfs_path *path;
4468         struct extent_buffer *leaf;
4469         struct btrfs_file_extent_item *fi;
4470         struct btrfs_key key;
4471         struct btrfs_key found_key;
4472         u64 extent_start = 0;
4473         u64 extent_num_bytes = 0;
4474         u64 extent_offset = 0;
4475         u64 item_end = 0;
4476         u64 last_size = new_size;
4477         u32 found_type = (u8)-1;
4478         int found_extent;
4479         int del_item;
4480         int pending_del_nr = 0;
4481         int pending_del_slot = 0;
4482         int extent_type = -1;
4483         int ret;
4484         u64 ino = btrfs_ino(BTRFS_I(inode));
4485         u64 bytes_deleted = 0;
4486         bool be_nice = false;
4487         bool should_throttle = false;
4488
4489         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4490
4491         /*
4492          * for non-free space inodes and ref cows, we want to back off from
4493          * time to time
4494          */
4495         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4496             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4497                 be_nice = true;
4498
4499         path = btrfs_alloc_path();
4500         if (!path)
4501                 return -ENOMEM;
4502         path->reada = READA_BACK;
4503
4504         /*
4505          * We want to drop from the next block forward in case this new size is
4506          * not block aligned since we will be keeping the last block of the
4507          * extent just the way it is.
4508          */
4509         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4510             root == fs_info->tree_root)
4511                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4512                                         fs_info->sectorsize),
4513                                         (u64)-1, 0);
4514
4515         /*
4516          * This function is also used to drop the items in the log tree before
4517          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4518          * it is used to drop the logged items. So we shouldn't kill the delayed
4519          * items.
4520          */
4521         if (min_type == 0 && root == BTRFS_I(inode)->root)
4522                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4523
4524         key.objectid = ino;
4525         key.offset = (u64)-1;
4526         key.type = (u8)-1;
4527
4528 search_again:
4529         /*
4530          * with a 16K leaf size and 128MB extents, you can actually queue
4531          * up a huge file in a single leaf.  Most of the time that
4532          * bytes_deleted is > 0, it will be huge by the time we get here
4533          */
4534         if (be_nice && bytes_deleted > SZ_32M &&
4535             btrfs_should_end_transaction(trans)) {
4536                 ret = -EAGAIN;
4537                 goto out;
4538         }
4539
4540         path->leave_spinning = 1;
4541         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4542         if (ret < 0)
4543                 goto out;
4544
4545         if (ret > 0) {
4546                 ret = 0;
4547                 /* there are no items in the tree for us to truncate, we're
4548                  * done
4549                  */
4550                 if (path->slots[0] == 0)
4551                         goto out;
4552                 path->slots[0]--;
4553         }
4554
4555         while (1) {
4556                 fi = NULL;
4557                 leaf = path->nodes[0];
4558                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4559                 found_type = found_key.type;
4560
4561                 if (found_key.objectid != ino)
4562                         break;
4563
4564                 if (found_type < min_type)
4565                         break;
4566
4567                 item_end = found_key.offset;
4568                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4569                         fi = btrfs_item_ptr(leaf, path->slots[0],
4570                                             struct btrfs_file_extent_item);
4571                         extent_type = btrfs_file_extent_type(leaf, fi);
4572                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4573                                 item_end +=
4574                                     btrfs_file_extent_num_bytes(leaf, fi);
4575
4576                                 trace_btrfs_truncate_show_fi_regular(
4577                                         BTRFS_I(inode), leaf, fi,
4578                                         found_key.offset);
4579                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4580                                 item_end += btrfs_file_extent_ram_bytes(leaf,
4581                                                                         fi);
4582
4583                                 trace_btrfs_truncate_show_fi_inline(
4584                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4585                                         found_key.offset);
4586                         }
4587                         item_end--;
4588                 }
4589                 if (found_type > min_type) {
4590                         del_item = 1;
4591                 } else {
4592                         if (item_end < new_size)
4593                                 break;
4594                         if (found_key.offset >= new_size)
4595                                 del_item = 1;
4596                         else
4597                                 del_item = 0;
4598                 }
4599                 found_extent = 0;
4600                 /* FIXME, shrink the extent if the ref count is only 1 */
4601                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4602                         goto delete;
4603
4604                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4605                         u64 num_dec;
4606                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4607                         if (!del_item) {
4608                                 u64 orig_num_bytes =
4609                                         btrfs_file_extent_num_bytes(leaf, fi);
4610                                 extent_num_bytes = ALIGN(new_size -
4611                                                 found_key.offset,
4612                                                 fs_info->sectorsize);
4613                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4614                                                          extent_num_bytes);
4615                                 num_dec = (orig_num_bytes -
4616                                            extent_num_bytes);
4617                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4618                                              &root->state) &&
4619                                     extent_start != 0)
4620                                         inode_sub_bytes(inode, num_dec);
4621                                 btrfs_mark_buffer_dirty(leaf);
4622                         } else {
4623                                 extent_num_bytes =
4624                                         btrfs_file_extent_disk_num_bytes(leaf,
4625                                                                          fi);
4626                                 extent_offset = found_key.offset -
4627                                         btrfs_file_extent_offset(leaf, fi);
4628
4629                                 /* FIXME blocksize != 4096 */
4630                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4631                                 if (extent_start != 0) {
4632                                         found_extent = 1;
4633                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4634                                                      &root->state))
4635                                                 inode_sub_bytes(inode, num_dec);
4636                                 }
4637                         }
4638                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4639                         /*
4640                          * we can't truncate inline items that have had
4641                          * special encodings
4642                          */
4643                         if (!del_item &&
4644                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4645                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4646                             btrfs_file_extent_compression(leaf, fi) == 0) {
4647                                 u32 size = (u32)(new_size - found_key.offset);
4648
4649                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4650                                 size = btrfs_file_extent_calc_inline_size(size);
4651                                 btrfs_truncate_item(root->fs_info, path, size, 1);
4652                         } else if (!del_item) {
4653                                 /*
4654                                  * We have to bail so the last_size is set to
4655                                  * just before this extent.
4656                                  */
4657                                 ret = NEED_TRUNCATE_BLOCK;
4658                                 break;
4659                         }
4660
4661                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4662                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4663                 }
4664 delete:
4665                 if (del_item)
4666                         last_size = found_key.offset;
4667                 else
4668                         last_size = new_size;
4669                 if (del_item) {
4670                         if (!pending_del_nr) {
4671                                 /* no pending yet, add ourselves */
4672                                 pending_del_slot = path->slots[0];
4673                                 pending_del_nr = 1;
4674                         } else if (pending_del_nr &&
4675                                    path->slots[0] + 1 == pending_del_slot) {
4676                                 /* hop on the pending chunk */
4677                                 pending_del_nr++;
4678                                 pending_del_slot = path->slots[0];
4679                         } else {
4680                                 BUG();
4681                         }
4682                 } else {
4683                         break;
4684                 }
4685                 should_throttle = false;
4686
4687                 if (found_extent &&
4688                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4689                      root == fs_info->tree_root)) {
4690                         btrfs_set_path_blocking(path);
4691                         bytes_deleted += extent_num_bytes;
4692                         ret = btrfs_free_extent(trans, root, extent_start,
4693                                                 extent_num_bytes, 0,
4694                                                 btrfs_header_owner(leaf),
4695                                                 ino, extent_offset);
4696                         if (ret) {
4697                                 btrfs_abort_transaction(trans, ret);
4698                                 break;
4699                         }
4700                         if (be_nice) {
4701                                 if (btrfs_should_throttle_delayed_refs(trans))
4702                                         should_throttle = true;
4703                         }
4704                 }
4705
4706                 if (found_type == BTRFS_INODE_ITEM_KEY)
4707                         break;
4708
4709                 if (path->slots[0] == 0 ||
4710                     path->slots[0] != pending_del_slot ||
4711                     should_throttle) {
4712                         if (pending_del_nr) {
4713                                 ret = btrfs_del_items(trans, root, path,
4714                                                 pending_del_slot,
4715                                                 pending_del_nr);
4716                                 if (ret) {
4717                                         btrfs_abort_transaction(trans, ret);
4718                                         break;
4719                                 }
4720                                 pending_del_nr = 0;
4721                         }
4722                         btrfs_release_path(path);
4723
4724                         /*
4725                          * We can generate a lot of delayed refs, so we need to
4726                          * throttle every once and a while and make sure we're
4727                          * adding enough space to keep up with the work we are
4728                          * generating.  Since we hold a transaction here we
4729                          * can't flush, and we don't want to FLUSH_LIMIT because
4730                          * we could have generated too many delayed refs to
4731                          * actually allocate, so just bail if we're short and
4732                          * let the normal reservation dance happen higher up.
4733                          */
4734                         if (should_throttle) {
4735                                 ret = btrfs_delayed_refs_rsv_refill(fs_info,
4736                                                         BTRFS_RESERVE_NO_FLUSH);
4737                                 if (ret) {
4738                                         ret = -EAGAIN;
4739                                         break;
4740                                 }
4741                         }
4742                         goto search_again;
4743                 } else {
4744                         path->slots[0]--;
4745                 }
4746         }
4747 out:
4748         if (ret >= 0 && pending_del_nr) {
4749                 int err;
4750
4751                 err = btrfs_del_items(trans, root, path, pending_del_slot,
4752                                       pending_del_nr);
4753                 if (err) {
4754                         btrfs_abort_transaction(trans, err);
4755                         ret = err;
4756                 }
4757         }
4758         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4759                 ASSERT(last_size >= new_size);
4760                 if (!ret && last_size > new_size)
4761                         last_size = new_size;
4762                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4763         }
4764
4765         btrfs_free_path(path);
4766         return ret;
4767 }
4768
4769 /*
4770  * btrfs_truncate_block - read, zero a chunk and write a block
4771  * @inode - inode that we're zeroing
4772  * @from - the offset to start zeroing
4773  * @len - the length to zero, 0 to zero the entire range respective to the
4774  *      offset
4775  * @front - zero up to the offset instead of from the offset on
4776  *
4777  * This will find the block for the "from" offset and cow the block and zero the
4778  * part we want to zero.  This is used with truncate and hole punching.
4779  */
4780 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4781                         int front)
4782 {
4783         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4784         struct address_space *mapping = inode->i_mapping;
4785         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4786         struct btrfs_ordered_extent *ordered;
4787         struct extent_state *cached_state = NULL;
4788         struct extent_changeset *data_reserved = NULL;
4789         char *kaddr;
4790         u32 blocksize = fs_info->sectorsize;
4791         pgoff_t index = from >> PAGE_SHIFT;
4792         unsigned offset = from & (blocksize - 1);
4793         struct page *page;
4794         gfp_t mask = btrfs_alloc_write_mask(mapping);
4795         int ret = 0;
4796         u64 block_start;
4797         u64 block_end;
4798
4799         if (IS_ALIGNED(offset, blocksize) &&
4800             (!len || IS_ALIGNED(len, blocksize)))
4801                 goto out;
4802
4803         block_start = round_down(from, blocksize);
4804         block_end = block_start + blocksize - 1;
4805
4806         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4807                                            block_start, blocksize);
4808         if (ret)
4809                 goto out;
4810
4811 again:
4812         page = find_or_create_page(mapping, index, mask);
4813         if (!page) {
4814                 btrfs_delalloc_release_space(inode, data_reserved,
4815                                              block_start, blocksize, true);
4816                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize, true);
4817                 ret = -ENOMEM;
4818                 goto out;
4819         }
4820
4821         if (!PageUptodate(page)) {
4822                 ret = btrfs_readpage(NULL, page);
4823                 lock_page(page);
4824                 if (page->mapping != mapping) {
4825                         unlock_page(page);
4826                         put_page(page);
4827                         goto again;
4828                 }
4829                 if (!PageUptodate(page)) {
4830                         ret = -EIO;
4831                         goto out_unlock;
4832                 }
4833         }
4834         wait_on_page_writeback(page);
4835
4836         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4837         set_page_extent_mapped(page);
4838
4839         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4840         if (ordered) {
4841                 unlock_extent_cached(io_tree, block_start, block_end,
4842                                      &cached_state);
4843                 unlock_page(page);
4844                 put_page(page);
4845                 btrfs_start_ordered_extent(inode, ordered, 1);
4846                 btrfs_put_ordered_extent(ordered);
4847                 goto again;
4848         }
4849
4850         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4851                           EXTENT_DIRTY | EXTENT_DELALLOC |
4852                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4853                           0, 0, &cached_state);
4854
4855         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
4856                                         &cached_state, 0);
4857         if (ret) {
4858                 unlock_extent_cached(io_tree, block_start, block_end,
4859                                      &cached_state);
4860                 goto out_unlock;
4861         }
4862
4863         if (offset != blocksize) {
4864                 if (!len)
4865                         len = blocksize - offset;
4866                 kaddr = kmap(page);
4867                 if (front)
4868                         memset(kaddr + (block_start - page_offset(page)),
4869                                 0, offset);
4870                 else
4871                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4872                                 0, len);
4873                 flush_dcache_page(page);
4874                 kunmap(page);
4875         }
4876         ClearPageChecked(page);
4877         set_page_dirty(page);
4878         unlock_extent_cached(io_tree, block_start, block_end, &cached_state);
4879
4880 out_unlock:
4881         if (ret)
4882                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
4883                                              blocksize, true);
4884         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize, (ret != 0));
4885         unlock_page(page);
4886         put_page(page);
4887 out:
4888         extent_changeset_free(data_reserved);
4889         return ret;
4890 }
4891
4892 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4893                              u64 offset, u64 len)
4894 {
4895         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4896         struct btrfs_trans_handle *trans;
4897         int ret;
4898
4899         /*
4900          * Still need to make sure the inode looks like it's been updated so
4901          * that any holes get logged if we fsync.
4902          */
4903         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4904                 BTRFS_I(inode)->last_trans = fs_info->generation;
4905                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4906                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4907                 return 0;
4908         }
4909
4910         /*
4911          * 1 - for the one we're dropping
4912          * 1 - for the one we're adding
4913          * 1 - for updating the inode.
4914          */
4915         trans = btrfs_start_transaction(root, 3);
4916         if (IS_ERR(trans))
4917                 return PTR_ERR(trans);
4918
4919         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4920         if (ret) {
4921                 btrfs_abort_transaction(trans, ret);
4922                 btrfs_end_transaction(trans);
4923                 return ret;
4924         }
4925
4926         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4927                         offset, 0, 0, len, 0, len, 0, 0, 0);
4928         if (ret)
4929                 btrfs_abort_transaction(trans, ret);
4930         else
4931                 btrfs_update_inode(trans, root, inode);
4932         btrfs_end_transaction(trans);
4933         return ret;
4934 }
4935
4936 /*
4937  * This function puts in dummy file extents for the area we're creating a hole
4938  * for.  So if we are truncating this file to a larger size we need to insert
4939  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4940  * the range between oldsize and size
4941  */
4942 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4943 {
4944         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4945         struct btrfs_root *root = BTRFS_I(inode)->root;
4946         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4947         struct extent_map *em = NULL;
4948         struct extent_state *cached_state = NULL;
4949         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4950         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
4951         u64 block_end = ALIGN(size, fs_info->sectorsize);
4952         u64 last_byte;
4953         u64 cur_offset;
4954         u64 hole_size;
4955         int err = 0;
4956
4957         /*
4958          * If our size started in the middle of a block we need to zero out the
4959          * rest of the block before we expand the i_size, otherwise we could
4960          * expose stale data.
4961          */
4962         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4963         if (err)
4964                 return err;
4965
4966         if (size <= hole_start)
4967                 return 0;
4968
4969         while (1) {
4970                 struct btrfs_ordered_extent *ordered;
4971
4972                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4973                                  &cached_state);
4974                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), hole_start,
4975                                                      block_end - hole_start);
4976                 if (!ordered)
4977                         break;
4978                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4979                                      &cached_state);
4980                 btrfs_start_ordered_extent(inode, ordered, 1);
4981                 btrfs_put_ordered_extent(ordered);
4982         }
4983
4984         cur_offset = hole_start;
4985         while (1) {
4986                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
4987                                 block_end - cur_offset, 0);
4988                 if (IS_ERR(em)) {
4989                         err = PTR_ERR(em);
4990                         em = NULL;
4991                         break;
4992                 }
4993                 last_byte = min(extent_map_end(em), block_end);
4994                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
4995                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4996                         struct extent_map *hole_em;
4997                         hole_size = last_byte - cur_offset;
4998
4999                         err = maybe_insert_hole(root, inode, cur_offset,
5000                                                 hole_size);
5001                         if (err)
5002                                 break;
5003                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
5004                                                 cur_offset + hole_size - 1, 0);
5005                         hole_em = alloc_extent_map();
5006                         if (!hole_em) {
5007                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5008                                         &BTRFS_I(inode)->runtime_flags);
5009                                 goto next;
5010                         }
5011                         hole_em->start = cur_offset;
5012                         hole_em->len = hole_size;
5013                         hole_em->orig_start = cur_offset;
5014
5015                         hole_em->block_start = EXTENT_MAP_HOLE;
5016                         hole_em->block_len = 0;
5017                         hole_em->orig_block_len = 0;
5018                         hole_em->ram_bytes = hole_size;
5019                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
5020                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
5021                         hole_em->generation = fs_info->generation;
5022
5023                         while (1) {
5024                                 write_lock(&em_tree->lock);
5025                                 err = add_extent_mapping(em_tree, hole_em, 1);
5026                                 write_unlock(&em_tree->lock);
5027                                 if (err != -EEXIST)
5028                                         break;
5029                                 btrfs_drop_extent_cache(BTRFS_I(inode),
5030                                                         cur_offset,
5031                                                         cur_offset +
5032                                                         hole_size - 1, 0);
5033                         }
5034                         free_extent_map(hole_em);
5035                 }
5036 next:
5037                 free_extent_map(em);
5038                 em = NULL;
5039                 cur_offset = last_byte;
5040                 if (cur_offset >= block_end)
5041                         break;
5042         }
5043         free_extent_map(em);
5044         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state);
5045         return err;
5046 }
5047
5048 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5049 {
5050         struct btrfs_root *root = BTRFS_I(inode)->root;
5051         struct btrfs_trans_handle *trans;
5052         loff_t oldsize = i_size_read(inode);
5053         loff_t newsize = attr->ia_size;
5054         int mask = attr->ia_valid;
5055         int ret;
5056
5057         /*
5058          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5059          * special case where we need to update the times despite not having
5060          * these flags set.  For all other operations the VFS set these flags
5061          * explicitly if it wants a timestamp update.
5062          */
5063         if (newsize != oldsize) {
5064                 inode_inc_iversion(inode);
5065                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5066                         inode->i_ctime = inode->i_mtime =
5067                                 current_time(inode);
5068         }
5069
5070         if (newsize > oldsize) {
5071                 /*
5072                  * Don't do an expanding truncate while snapshotting is ongoing.
5073                  * This is to ensure the snapshot captures a fully consistent
5074                  * state of this file - if the snapshot captures this expanding
5075                  * truncation, it must capture all writes that happened before
5076                  * this truncation.
5077                  */
5078                 btrfs_wait_for_snapshot_creation(root);
5079                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5080                 if (ret) {
5081                         btrfs_end_write_no_snapshotting(root);
5082                         return ret;
5083                 }
5084
5085                 trans = btrfs_start_transaction(root, 1);
5086                 if (IS_ERR(trans)) {
5087                         btrfs_end_write_no_snapshotting(root);
5088                         return PTR_ERR(trans);
5089                 }
5090
5091                 i_size_write(inode, newsize);
5092                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5093                 pagecache_isize_extended(inode, oldsize, newsize);
5094                 ret = btrfs_update_inode(trans, root, inode);
5095                 btrfs_end_write_no_snapshotting(root);
5096                 btrfs_end_transaction(trans);
5097         } else {
5098
5099                 /*
5100                  * We're truncating a file that used to have good data down to
5101                  * zero. Make sure it gets into the ordered flush list so that
5102                  * any new writes get down to disk quickly.
5103                  */
5104                 if (newsize == 0)
5105                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5106                                 &BTRFS_I(inode)->runtime_flags);
5107
5108                 truncate_setsize(inode, newsize);
5109
5110                 /* Disable nonlocked read DIO to avoid the endless truncate */
5111                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5112                 inode_dio_wait(inode);
5113                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5114
5115                 ret = btrfs_truncate(inode, newsize == oldsize);
5116                 if (ret && inode->i_nlink) {
5117                         int err;
5118
5119                         /*
5120                          * Truncate failed, so fix up the in-memory size. We
5121                          * adjusted disk_i_size down as we removed extents, so
5122                          * wait for disk_i_size to be stable and then update the
5123                          * in-memory size to match.
5124                          */
5125                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5126                         if (err)
5127                                 return err;
5128                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5129                 }
5130         }
5131
5132         return ret;
5133 }
5134
5135 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5136 {
5137         struct inode *inode = d_inode(dentry);
5138         struct btrfs_root *root = BTRFS_I(inode)->root;
5139         int err;
5140
5141         if (btrfs_root_readonly(root))
5142                 return -EROFS;
5143
5144         err = setattr_prepare(dentry, attr);
5145         if (err)
5146                 return err;
5147
5148         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5149                 err = btrfs_setsize(inode, attr);
5150                 if (err)
5151                         return err;
5152         }
5153
5154         if (attr->ia_valid) {
5155                 setattr_copy(inode, attr);
5156                 inode_inc_iversion(inode);
5157                 err = btrfs_dirty_inode(inode);
5158
5159                 if (!err && attr->ia_valid & ATTR_MODE)
5160                         err = posix_acl_chmod(inode, inode->i_mode);
5161         }
5162
5163         return err;
5164 }
5165
5166 /*
5167  * While truncating the inode pages during eviction, we get the VFS calling
5168  * btrfs_invalidatepage() against each page of the inode. This is slow because
5169  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5170  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5171  * extent_state structures over and over, wasting lots of time.
5172  *
5173  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5174  * those expensive operations on a per page basis and do only the ordered io
5175  * finishing, while we release here the extent_map and extent_state structures,
5176  * without the excessive merging and splitting.
5177  */
5178 static void evict_inode_truncate_pages(struct inode *inode)
5179 {
5180         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5181         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5182         struct rb_node *node;
5183
5184         ASSERT(inode->i_state & I_FREEING);
5185         truncate_inode_pages_final(&inode->i_data);
5186
5187         write_lock(&map_tree->lock);
5188         while (!RB_EMPTY_ROOT(&map_tree->map.rb_root)) {
5189                 struct extent_map *em;
5190
5191                 node = rb_first_cached(&map_tree->map);
5192                 em = rb_entry(node, struct extent_map, rb_node);
5193                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5194                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5195                 remove_extent_mapping(map_tree, em);
5196                 free_extent_map(em);
5197                 if (need_resched()) {
5198                         write_unlock(&map_tree->lock);
5199                         cond_resched();
5200                         write_lock(&map_tree->lock);
5201                 }
5202         }
5203         write_unlock(&map_tree->lock);
5204
5205         /*
5206          * Keep looping until we have no more ranges in the io tree.
5207          * We can have ongoing bios started by readpages (called from readahead)
5208          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5209          * still in progress (unlocked the pages in the bio but did not yet
5210          * unlocked the ranges in the io tree). Therefore this means some
5211          * ranges can still be locked and eviction started because before
5212          * submitting those bios, which are executed by a separate task (work
5213          * queue kthread), inode references (inode->i_count) were not taken
5214          * (which would be dropped in the end io callback of each bio).
5215          * Therefore here we effectively end up waiting for those bios and
5216          * anyone else holding locked ranges without having bumped the inode's
5217          * reference count - if we don't do it, when they access the inode's
5218          * io_tree to unlock a range it may be too late, leading to an
5219          * use-after-free issue.
5220          */
5221         spin_lock(&io_tree->lock);
5222         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5223                 struct extent_state *state;
5224                 struct extent_state *cached_state = NULL;
5225                 u64 start;
5226                 u64 end;
5227                 unsigned state_flags;
5228
5229                 node = rb_first(&io_tree->state);
5230                 state = rb_entry(node, struct extent_state, rb_node);
5231                 start = state->start;
5232                 end = state->end;
5233                 state_flags = state->state;
5234                 spin_unlock(&io_tree->lock);
5235
5236                 lock_extent_bits(io_tree, start, end, &cached_state);
5237
5238                 /*
5239                  * If still has DELALLOC flag, the extent didn't reach disk,
5240                  * and its reserved space won't be freed by delayed_ref.
5241                  * So we need to free its reserved space here.
5242                  * (Refer to comment in btrfs_invalidatepage, case 2)
5243                  *
5244                  * Note, end is the bytenr of last byte, so we need + 1 here.
5245                  */
5246                 if (state_flags & EXTENT_DELALLOC)
5247                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5248
5249                 clear_extent_bit(io_tree, start, end,
5250                                  EXTENT_LOCKED | EXTENT_DIRTY |
5251                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5252                                  EXTENT_DEFRAG, 1, 1, &cached_state);
5253
5254                 cond_resched();
5255                 spin_lock(&io_tree->lock);
5256         }
5257         spin_unlock(&io_tree->lock);
5258 }
5259
5260 static struct btrfs_trans_handle *evict_refill_and_join(struct btrfs_root *root,
5261                                                         struct btrfs_block_rsv *rsv)
5262 {
5263         struct btrfs_fs_info *fs_info = root->fs_info;
5264         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5265         int failures = 0;
5266
5267         for (;;) {
5268                 struct btrfs_trans_handle *trans;
5269                 int ret;
5270
5271                 ret = btrfs_block_rsv_refill(root, rsv, rsv->size,
5272                                              BTRFS_RESERVE_FLUSH_LIMIT);
5273
5274                 if (ret && ++failures > 2) {
5275                         btrfs_warn(fs_info,
5276                                    "could not allocate space for a delete; will truncate on mount");
5277                         return ERR_PTR(-ENOSPC);
5278                 }
5279
5280                 trans = btrfs_join_transaction(root);
5281                 if (IS_ERR(trans) || !ret)
5282                         return trans;
5283
5284                 /*
5285                  * Try to steal from the global reserve if there is space for
5286                  * it.
5287                  */
5288                 if (!btrfs_check_space_for_delayed_refs(fs_info) &&
5289                     !btrfs_block_rsv_migrate(global_rsv, rsv, rsv->size, 0))
5290                         return trans;
5291
5292                 /* If not, commit and try again. */
5293                 ret = btrfs_commit_transaction(trans);
5294                 if (ret)
5295                         return ERR_PTR(ret);
5296         }
5297 }
5298
5299 void btrfs_evict_inode(struct inode *inode)
5300 {
5301         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5302         struct btrfs_trans_handle *trans;
5303         struct btrfs_root *root = BTRFS_I(inode)->root;
5304         struct btrfs_block_rsv *rsv;
5305         int ret;
5306
5307         trace_btrfs_inode_evict(inode);
5308
5309         if (!root) {
5310                 clear_inode(inode);
5311                 return;
5312         }
5313
5314         evict_inode_truncate_pages(inode);
5315
5316         if (inode->i_nlink &&
5317             ((btrfs_root_refs(&root->root_item) != 0 &&
5318               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5319              btrfs_is_free_space_inode(BTRFS_I(inode))))
5320                 goto no_delete;
5321
5322         if (is_bad_inode(inode))
5323                 goto no_delete;
5324
5325         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5326
5327         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
5328                 goto no_delete;
5329
5330         if (inode->i_nlink > 0) {
5331                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5332                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5333                 goto no_delete;
5334         }
5335
5336         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5337         if (ret)
5338                 goto no_delete;
5339
5340         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5341         if (!rsv)
5342                 goto no_delete;
5343         rsv->size = btrfs_calc_trunc_metadata_size(fs_info, 1);
5344         rsv->failfast = 1;
5345
5346         btrfs_i_size_write(BTRFS_I(inode), 0);
5347
5348         while (1) {
5349                 trans = evict_refill_and_join(root, rsv);
5350                 if (IS_ERR(trans))
5351                         goto free_rsv;
5352
5353                 trans->block_rsv = rsv;
5354
5355                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5356                 trans->block_rsv = &fs_info->trans_block_rsv;
5357                 btrfs_end_transaction(trans);
5358                 btrfs_btree_balance_dirty(fs_info);
5359                 if (ret && ret != -ENOSPC && ret != -EAGAIN)
5360                         goto free_rsv;
5361                 else if (!ret)
5362                         break;
5363         }
5364
5365         /*
5366          * Errors here aren't a big deal, it just means we leave orphan items in
5367          * the tree. They will be cleaned up on the next mount. If the inode
5368          * number gets reused, cleanup deletes the orphan item without doing
5369          * anything, and unlink reuses the existing orphan item.
5370          *
5371          * If it turns out that we are dropping too many of these, we might want
5372          * to add a mechanism for retrying these after a commit.
5373          */
5374         trans = evict_refill_and_join(root, rsv);
5375         if (!IS_ERR(trans)) {
5376                 trans->block_rsv = rsv;
5377                 btrfs_orphan_del(trans, BTRFS_I(inode));
5378                 trans->block_rsv = &fs_info->trans_block_rsv;
5379                 btrfs_end_transaction(trans);
5380         }
5381
5382         if (!(root == fs_info->tree_root ||
5383               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5384                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5385
5386 free_rsv:
5387         btrfs_free_block_rsv(fs_info, rsv);
5388 no_delete:
5389         /*
5390          * If we didn't successfully delete, the orphan item will still be in
5391          * the tree and we'll retry on the next mount. Again, we might also want
5392          * to retry these periodically in the future.
5393          */
5394         btrfs_remove_delayed_node(BTRFS_I(inode));
5395         clear_inode(inode);
5396 }
5397
5398 /*
5399  * this returns the key found in the dir entry in the location pointer.
5400  * If no dir entries were found, returns -ENOENT.
5401  * If found a corrupted location in dir entry, returns -EUCLEAN.
5402  */
5403 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5404                                struct btrfs_key *location)
5405 {
5406         const char *name = dentry->d_name.name;
5407         int namelen = dentry->d_name.len;
5408         struct btrfs_dir_item *di;
5409         struct btrfs_path *path;
5410         struct btrfs_root *root = BTRFS_I(dir)->root;
5411         int ret = 0;
5412
5413         path = btrfs_alloc_path();
5414         if (!path)
5415                 return -ENOMEM;
5416
5417         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5418                         name, namelen, 0);
5419         if (IS_ERR_OR_NULL(di)) {
5420                 ret = di ? PTR_ERR(di) : -ENOENT;
5421                 goto out;
5422         }
5423
5424         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5425         if (location->type != BTRFS_INODE_ITEM_KEY &&
5426             location->type != BTRFS_ROOT_ITEM_KEY) {
5427                 ret = -EUCLEAN;
5428                 btrfs_warn(root->fs_info,
5429 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5430                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5431                            location->objectid, location->type, location->offset);
5432         }
5433 out:
5434         btrfs_free_path(path);
5435         return ret;
5436 }
5437
5438 /*
5439  * when we hit a tree root in a directory, the btrfs part of the inode
5440  * needs to be changed to reflect the root directory of the tree root.  This
5441  * is kind of like crossing a mount point.
5442  */
5443 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5444                                     struct inode *dir,
5445                                     struct dentry *dentry,
5446                                     struct btrfs_key *location,
5447                                     struct btrfs_root **sub_root)
5448 {
5449         struct btrfs_path *path;
5450         struct btrfs_root *new_root;
5451         struct btrfs_root_ref *ref;
5452         struct extent_buffer *leaf;
5453         struct btrfs_key key;
5454         int ret;
5455         int err = 0;
5456
5457         path = btrfs_alloc_path();
5458         if (!path) {
5459                 err = -ENOMEM;
5460                 goto out;
5461         }
5462
5463         err = -ENOENT;
5464         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5465         key.type = BTRFS_ROOT_REF_KEY;
5466         key.offset = location->objectid;
5467
5468         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5469         if (ret) {
5470                 if (ret < 0)
5471                         err = ret;
5472                 goto out;
5473         }
5474
5475         leaf = path->nodes[0];
5476         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5477         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5478             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5479                 goto out;
5480
5481         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5482                                    (unsigned long)(ref + 1),
5483                                    dentry->d_name.len);
5484         if (ret)
5485                 goto out;
5486
5487         btrfs_release_path(path);
5488
5489         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5490         if (IS_ERR(new_root)) {
5491                 err = PTR_ERR(new_root);
5492                 goto out;
5493         }
5494
5495         *sub_root = new_root;
5496         location->objectid = btrfs_root_dirid(&new_root->root_item);
5497         location->type = BTRFS_INODE_ITEM_KEY;
5498         location->offset = 0;
5499         err = 0;
5500 out:
5501         btrfs_free_path(path);
5502         return err;
5503 }
5504
5505 static void inode_tree_add(struct inode *inode)
5506 {
5507         struct btrfs_root *root = BTRFS_I(inode)->root;
5508         struct btrfs_inode *entry;
5509         struct rb_node **p;
5510         struct rb_node *parent;
5511         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5512         u64 ino = btrfs_ino(BTRFS_I(inode));
5513
5514         if (inode_unhashed(inode))
5515                 return;
5516         parent = NULL;
5517         spin_lock(&root->inode_lock);
5518         p = &root->inode_tree.rb_node;
5519         while (*p) {
5520                 parent = *p;
5521                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5522
5523                 if (ino < btrfs_ino(entry))
5524                         p = &parent->rb_left;
5525                 else if (ino > btrfs_ino(entry))
5526                         p = &parent->rb_right;
5527                 else {
5528                         WARN_ON(!(entry->vfs_inode.i_state &
5529                                   (I_WILL_FREE | I_FREEING)));
5530                         rb_replace_node(parent, new, &root->inode_tree);
5531                         RB_CLEAR_NODE(parent);
5532                         spin_unlock(&root->inode_lock);
5533                         return;
5534                 }
5535         }
5536         rb_link_node(new, parent, p);
5537         rb_insert_color(new, &root->inode_tree);
5538         spin_unlock(&root->inode_lock);
5539 }
5540
5541 static void inode_tree_del(struct inode *inode)
5542 {
5543         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5544         struct btrfs_root *root = BTRFS_I(inode)->root;
5545         int empty = 0;
5546
5547         spin_lock(&root->inode_lock);
5548         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5549                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5550                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5551                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5552         }
5553         spin_unlock(&root->inode_lock);
5554
5555         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5556                 synchronize_srcu(&fs_info->subvol_srcu);
5557                 spin_lock(&root->inode_lock);
5558                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5559                 spin_unlock(&root->inode_lock);
5560                 if (empty)
5561                         btrfs_add_dead_root(root);
5562         }
5563 }
5564
5565
5566 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5567 {
5568         struct btrfs_iget_args *args = p;
5569         inode->i_ino = args->location->objectid;
5570         memcpy(&BTRFS_I(inode)->location, args->location,
5571                sizeof(*args->location));
5572         BTRFS_I(inode)->root = args->root;
5573         return 0;
5574 }
5575
5576 static int btrfs_find_actor(struct inode *inode, void *opaque)
5577 {
5578         struct btrfs_iget_args *args = opaque;
5579         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5580                 args->root == BTRFS_I(inode)->root;
5581 }
5582
5583 static struct inode *btrfs_iget_locked(struct super_block *s,
5584                                        struct btrfs_key *location,
5585                                        struct btrfs_root *root)
5586 {
5587         struct inode *inode;
5588         struct btrfs_iget_args args;
5589         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5590
5591         args.location = location;
5592         args.root = root;
5593
5594         inode = iget5_locked(s, hashval, btrfs_find_actor,
5595                              btrfs_init_locked_inode,
5596                              (void *)&args);
5597         return inode;
5598 }
5599
5600 /* Get an inode object given its location and corresponding root.
5601  * Returns in *is_new if the inode was read from disk
5602  */
5603 struct inode *btrfs_iget_path(struct super_block *s, struct btrfs_key *location,
5604                               struct btrfs_root *root, int *new,
5605                               struct btrfs_path *path)
5606 {
5607         struct inode *inode;
5608
5609         inode = btrfs_iget_locked(s, location, root);
5610         if (!inode)
5611                 return ERR_PTR(-ENOMEM);
5612
5613         if (inode->i_state & I_NEW) {
5614                 int ret;
5615
5616                 ret = btrfs_read_locked_inode(inode, path);
5617                 if (!ret) {
5618                         inode_tree_add(inode);
5619                         unlock_new_inode(inode);
5620                         if (new)
5621                                 *new = 1;
5622                 } else {
5623                         iget_failed(inode);
5624                         /*
5625                          * ret > 0 can come from btrfs_search_slot called by
5626                          * btrfs_read_locked_inode, this means the inode item
5627                          * was not found.
5628                          */
5629                         if (ret > 0)
5630                                 ret = -ENOENT;
5631                         inode = ERR_PTR(ret);
5632                 }
5633         }
5634
5635         return inode;
5636 }
5637
5638 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5639                          struct btrfs_root *root, int *new)
5640 {
5641         return btrfs_iget_path(s, location, root, new, NULL);
5642 }
5643
5644 static struct inode *new_simple_dir(struct super_block *s,
5645                                     struct btrfs_key *key,
5646                                     struct btrfs_root *root)
5647 {
5648         struct inode *inode = new_inode(s);
5649
5650         if (!inode)
5651                 return ERR_PTR(-ENOMEM);
5652
5653         BTRFS_I(inode)->root = root;
5654         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5655         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5656
5657         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5658         inode->i_op = &btrfs_dir_ro_inode_operations;
5659         inode->i_opflags &= ~IOP_XATTR;
5660         inode->i_fop = &simple_dir_operations;
5661         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5662         inode->i_mtime = current_time(inode);
5663         inode->i_atime = inode->i_mtime;
5664         inode->i_ctime = inode->i_mtime;
5665         BTRFS_I(inode)->i_otime = inode->i_mtime;
5666
5667         return inode;
5668 }
5669
5670 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5671 {
5672         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5673         struct inode *inode;
5674         struct btrfs_root *root = BTRFS_I(dir)->root;
5675         struct btrfs_root *sub_root = root;
5676         struct btrfs_key location;
5677         int index;
5678         int ret = 0;
5679
5680         if (dentry->d_name.len > BTRFS_NAME_LEN)
5681                 return ERR_PTR(-ENAMETOOLONG);
5682
5683         ret = btrfs_inode_by_name(dir, dentry, &location);
5684         if (ret < 0)
5685                 return ERR_PTR(ret);
5686
5687         if (location.type == BTRFS_INODE_ITEM_KEY) {
5688                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5689                 return inode;
5690         }
5691
5692         index = srcu_read_lock(&fs_info->subvol_srcu);
5693         ret = fixup_tree_root_location(fs_info, dir, dentry,
5694                                        &location, &sub_root);
5695         if (ret < 0) {
5696                 if (ret != -ENOENT)
5697                         inode = ERR_PTR(ret);
5698                 else
5699                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5700         } else {
5701                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5702         }
5703         srcu_read_unlock(&fs_info->subvol_srcu, index);
5704
5705         if (!IS_ERR(inode) && root != sub_root) {
5706                 down_read(&fs_info->cleanup_work_sem);
5707                 if (!sb_rdonly(inode->i_sb))
5708                         ret = btrfs_orphan_cleanup(sub_root);
5709                 up_read(&fs_info->cleanup_work_sem);
5710                 if (ret) {
5711                         iput(inode);
5712                         inode = ERR_PTR(ret);
5713                 }
5714         }
5715
5716         return inode;
5717 }
5718
5719 static int btrfs_dentry_delete(const struct dentry *dentry)
5720 {
5721         struct btrfs_root *root;
5722         struct inode *inode = d_inode(dentry);
5723
5724         if (!inode && !IS_ROOT(dentry))
5725                 inode = d_inode(dentry->d_parent);
5726
5727         if (inode) {
5728                 root = BTRFS_I(inode)->root;
5729                 if (btrfs_root_refs(&root->root_item) == 0)
5730                         return 1;
5731
5732                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5733                         return 1;
5734         }
5735         return 0;
5736 }
5737
5738 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5739                                    unsigned int flags)
5740 {
5741         struct inode *inode = btrfs_lookup_dentry(dir, dentry);
5742
5743         if (inode == ERR_PTR(-ENOENT))
5744                 inode = NULL;
5745         return d_splice_alias(inode, dentry);
5746 }
5747
5748 unsigned char btrfs_filetype_table[] = {
5749         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5750 };
5751
5752 /*
5753  * All this infrastructure exists because dir_emit can fault, and we are holding
5754  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5755  * our information into that, and then dir_emit from the buffer.  This is
5756  * similar to what NFS does, only we don't keep the buffer around in pagecache
5757  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5758  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5759  * tree lock.
5760  */
5761 static int btrfs_opendir(struct inode *inode, struct file *file)
5762 {
5763         struct btrfs_file_private *private;
5764
5765         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5766         if (!private)
5767                 return -ENOMEM;
5768         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5769         if (!private->filldir_buf) {
5770                 kfree(private);
5771                 return -ENOMEM;
5772         }
5773         file->private_data = private;
5774         return 0;
5775 }
5776
5777 struct dir_entry {
5778         u64 ino;
5779         u64 offset;
5780         unsigned type;
5781         int name_len;
5782 };
5783
5784 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5785 {
5786         while (entries--) {
5787                 struct dir_entry *entry = addr;
5788                 char *name = (char *)(entry + 1);
5789
5790                 ctx->pos = get_unaligned(&entry->offset);
5791                 if (!dir_emit(ctx, name, get_unaligned(&entry->name_len),
5792                                          get_unaligned(&entry->ino),
5793                                          get_unaligned(&entry->type)))
5794                         return 1;
5795                 addr += sizeof(struct dir_entry) +
5796                         get_unaligned(&entry->name_len);
5797                 ctx->pos++;
5798         }
5799         return 0;
5800 }
5801
5802 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5803 {
5804         struct inode *inode = file_inode(file);
5805         struct btrfs_root *root = BTRFS_I(inode)->root;
5806         struct btrfs_file_private *private = file->private_data;
5807         struct btrfs_dir_item *di;
5808         struct btrfs_key key;
5809         struct btrfs_key found_key;
5810         struct btrfs_path *path;
5811         void *addr;
5812         struct list_head ins_list;
5813         struct list_head del_list;
5814         int ret;
5815         struct extent_buffer *leaf;
5816         int slot;
5817         char *name_ptr;
5818         int name_len;
5819         int entries = 0;
5820         int total_len = 0;
5821         bool put = false;
5822         struct btrfs_key location;
5823
5824         if (!dir_emit_dots(file, ctx))
5825                 return 0;
5826
5827         path = btrfs_alloc_path();
5828         if (!path)
5829                 return -ENOMEM;
5830
5831         addr = private->filldir_buf;
5832         path->reada = READA_FORWARD;
5833
5834         INIT_LIST_HEAD(&ins_list);
5835         INIT_LIST_HEAD(&del_list);
5836         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5837
5838 again:
5839         key.type = BTRFS_DIR_INDEX_KEY;
5840         key.offset = ctx->pos;
5841         key.objectid = btrfs_ino(BTRFS_I(inode));
5842
5843         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5844         if (ret < 0)
5845                 goto err;
5846
5847         while (1) {
5848                 struct dir_entry *entry;
5849
5850                 leaf = path->nodes[0];
5851                 slot = path->slots[0];
5852                 if (slot >= btrfs_header_nritems(leaf)) {
5853                         ret = btrfs_next_leaf(root, path);
5854                         if (ret < 0)
5855                                 goto err;
5856                         else if (ret > 0)
5857                                 break;
5858                         continue;
5859                 }
5860
5861                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5862
5863                 if (found_key.objectid != key.objectid)
5864                         break;
5865                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
5866                         break;
5867                 if (found_key.offset < ctx->pos)
5868                         goto next;
5869                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
5870                         goto next;
5871                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5872                 name_len = btrfs_dir_name_len(leaf, di);
5873                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
5874                     PAGE_SIZE) {
5875                         btrfs_release_path(path);
5876                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5877                         if (ret)
5878                                 goto nopos;
5879                         addr = private->filldir_buf;
5880                         entries = 0;
5881                         total_len = 0;
5882                         goto again;
5883                 }
5884
5885                 entry = addr;
5886                 put_unaligned(name_len, &entry->name_len);
5887                 name_ptr = (char *)(entry + 1);
5888                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
5889                                    name_len);
5890                 put_unaligned(btrfs_filetype_table[btrfs_dir_type(leaf, di)],
5891                                 &entry->type);
5892                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
5893                 put_unaligned(location.objectid, &entry->ino);
5894                 put_unaligned(found_key.offset, &entry->offset);
5895                 entries++;
5896                 addr += sizeof(struct dir_entry) + name_len;
5897                 total_len += sizeof(struct dir_entry) + name_len;
5898 next:
5899                 path->slots[0]++;
5900         }
5901         btrfs_release_path(path);
5902
5903         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5904         if (ret)
5905                 goto nopos;
5906
5907         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5908         if (ret)
5909                 goto nopos;
5910
5911         /*
5912          * Stop new entries from being returned after we return the last
5913          * entry.
5914          *
5915          * New directory entries are assigned a strictly increasing
5916          * offset.  This means that new entries created during readdir
5917          * are *guaranteed* to be seen in the future by that readdir.
5918          * This has broken buggy programs which operate on names as
5919          * they're returned by readdir.  Until we re-use freed offsets
5920          * we have this hack to stop new entries from being returned
5921          * under the assumption that they'll never reach this huge
5922          * offset.
5923          *
5924          * This is being careful not to overflow 32bit loff_t unless the
5925          * last entry requires it because doing so has broken 32bit apps
5926          * in the past.
5927          */
5928         if (ctx->pos >= INT_MAX)
5929                 ctx->pos = LLONG_MAX;
5930         else
5931                 ctx->pos = INT_MAX;
5932 nopos:
5933         ret = 0;
5934 err:
5935         if (put)
5936                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
5937         btrfs_free_path(path);
5938         return ret;
5939 }
5940
5941 /*
5942  * This is somewhat expensive, updating the tree every time the
5943  * inode changes.  But, it is most likely to find the inode in cache.
5944  * FIXME, needs more benchmarking...there are no reasons other than performance
5945  * to keep or drop this code.
5946  */
5947 static int btrfs_dirty_inode(struct inode *inode)
5948 {
5949         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5950         struct btrfs_root *root = BTRFS_I(inode)->root;
5951         struct btrfs_trans_handle *trans;
5952         int ret;
5953
5954         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5955                 return 0;
5956
5957         trans = btrfs_join_transaction(root);
5958         if (IS_ERR(trans))
5959                 return PTR_ERR(trans);
5960
5961         ret = btrfs_update_inode(trans, root, inode);
5962         if (ret && ret == -ENOSPC) {
5963                 /* whoops, lets try again with the full transaction */
5964                 btrfs_end_transaction(trans);
5965                 trans = btrfs_start_transaction(root, 1);
5966                 if (IS_ERR(trans))
5967                         return PTR_ERR(trans);
5968
5969                 ret = btrfs_update_inode(trans, root, inode);
5970         }
5971         btrfs_end_transaction(trans);
5972         if (BTRFS_I(inode)->delayed_node)
5973                 btrfs_balance_delayed_items(fs_info);
5974
5975         return ret;
5976 }
5977
5978 /*
5979  * This is a copy of file_update_time.  We need this so we can return error on
5980  * ENOSPC for updating the inode in the case of file write and mmap writes.
5981  */
5982 static int btrfs_update_time(struct inode *inode, struct timespec64 *now,
5983                              int flags)
5984 {
5985         struct btrfs_root *root = BTRFS_I(inode)->root;
5986         bool dirty = flags & ~S_VERSION;
5987
5988         if (btrfs_root_readonly(root))
5989                 return -EROFS;
5990
5991         if (flags & S_VERSION)
5992                 dirty |= inode_maybe_inc_iversion(inode, dirty);
5993         if (flags & S_CTIME)
5994                 inode->i_ctime = *now;
5995         if (flags & S_MTIME)
5996                 inode->i_mtime = *now;
5997         if (flags & S_ATIME)
5998                 inode->i_atime = *now;
5999         return dirty ? btrfs_dirty_inode(inode) : 0;
6000 }
6001
6002 /*
6003  * find the highest existing sequence number in a directory
6004  * and then set the in-memory index_cnt variable to reflect
6005  * free sequence numbers
6006  */
6007 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6008 {
6009         struct btrfs_root *root = inode->root;
6010         struct btrfs_key key, found_key;
6011         struct btrfs_path *path;
6012         struct extent_buffer *leaf;
6013         int ret;
6014
6015         key.objectid = btrfs_ino(inode);
6016         key.type = BTRFS_DIR_INDEX_KEY;
6017         key.offset = (u64)-1;
6018
6019         path = btrfs_alloc_path();
6020         if (!path)
6021                 return -ENOMEM;
6022
6023         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6024         if (ret < 0)
6025                 goto out;
6026         /* FIXME: we should be able to handle this */
6027         if (ret == 0)
6028                 goto out;
6029         ret = 0;
6030
6031         /*
6032          * MAGIC NUMBER EXPLANATION:
6033          * since we search a directory based on f_pos we have to start at 2
6034          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6035          * else has to start at 2
6036          */
6037         if (path->slots[0] == 0) {
6038                 inode->index_cnt = 2;
6039                 goto out;
6040         }
6041
6042         path->slots[0]--;
6043
6044         leaf = path->nodes[0];
6045         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6046
6047         if (found_key.objectid != btrfs_ino(inode) ||
6048             found_key.type != BTRFS_DIR_INDEX_KEY) {
6049                 inode->index_cnt = 2;
6050                 goto out;
6051         }
6052
6053         inode->index_cnt = found_key.offset + 1;
6054 out:
6055         btrfs_free_path(path);
6056         return ret;
6057 }
6058
6059 /*
6060  * helper to find a free sequence number in a given directory.  This current
6061  * code is very simple, later versions will do smarter things in the btree
6062  */
6063 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6064 {
6065         int ret = 0;
6066
6067         if (dir->index_cnt == (u64)-1) {
6068                 ret = btrfs_inode_delayed_dir_index_count(dir);
6069                 if (ret) {
6070                         ret = btrfs_set_inode_index_count(dir);
6071                         if (ret)
6072                                 return ret;
6073                 }
6074         }
6075
6076         *index = dir->index_cnt;
6077         dir->index_cnt++;
6078
6079         return ret;
6080 }
6081
6082 static int btrfs_insert_inode_locked(struct inode *inode)
6083 {
6084         struct btrfs_iget_args args;
6085         args.location = &BTRFS_I(inode)->location;
6086         args.root = BTRFS_I(inode)->root;
6087
6088         return insert_inode_locked4(inode,
6089                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6090                    btrfs_find_actor, &args);
6091 }
6092
6093 /*
6094  * Inherit flags from the parent inode.
6095  *
6096  * Currently only the compression flags and the cow flags are inherited.
6097  */
6098 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6099 {
6100         unsigned int flags;
6101
6102         if (!dir)
6103                 return;
6104
6105         flags = BTRFS_I(dir)->flags;
6106
6107         if (flags & BTRFS_INODE_NOCOMPRESS) {
6108                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6109                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6110         } else if (flags & BTRFS_INODE_COMPRESS) {
6111                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6112                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6113         }
6114
6115         if (flags & BTRFS_INODE_NODATACOW) {
6116                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6117                 if (S_ISREG(inode->i_mode))
6118                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6119         }
6120
6121         btrfs_sync_inode_flags_to_i_flags(inode);
6122 }
6123
6124 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6125                                      struct btrfs_root *root,
6126                                      struct inode *dir,
6127                                      const char *name, int name_len,
6128                                      u64 ref_objectid, u64 objectid,
6129                                      umode_t mode, u64 *index)
6130 {
6131         struct btrfs_fs_info *fs_info = root->fs_info;
6132         struct inode *inode;
6133         struct btrfs_inode_item *inode_item;
6134         struct btrfs_key *location;
6135         struct btrfs_path *path;
6136         struct btrfs_inode_ref *ref;
6137         struct btrfs_key key[2];
6138         u32 sizes[2];
6139         int nitems = name ? 2 : 1;
6140         unsigned long ptr;
6141         int ret;
6142
6143         path = btrfs_alloc_path();
6144         if (!path)
6145                 return ERR_PTR(-ENOMEM);
6146
6147         inode = new_inode(fs_info->sb);
6148         if (!inode) {
6149                 btrfs_free_path(path);
6150                 return ERR_PTR(-ENOMEM);
6151         }
6152
6153         /*
6154          * O_TMPFILE, set link count to 0, so that after this point,
6155          * we fill in an inode item with the correct link count.
6156          */
6157         if (!name)
6158                 set_nlink(inode, 0);
6159
6160         /*
6161          * we have to initialize this early, so we can reclaim the inode
6162          * number if we fail afterwards in this function.
6163          */
6164         inode->i_ino = objectid;
6165
6166         if (dir && name) {
6167                 trace_btrfs_inode_request(dir);
6168
6169                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6170                 if (ret) {
6171                         btrfs_free_path(path);
6172                         iput(inode);
6173                         return ERR_PTR(ret);
6174                 }
6175         } else if (dir) {
6176                 *index = 0;
6177         }
6178         /*
6179          * index_cnt is ignored for everything but a dir,
6180          * btrfs_set_inode_index_count has an explanation for the magic
6181          * number
6182          */
6183         BTRFS_I(inode)->index_cnt = 2;
6184         BTRFS_I(inode)->dir_index = *index;
6185         BTRFS_I(inode)->root = root;
6186         BTRFS_I(inode)->generation = trans->transid;
6187         inode->i_generation = BTRFS_I(inode)->generation;
6188
6189         /*
6190          * We could have gotten an inode number from somebody who was fsynced
6191          * and then removed in this same transaction, so let's just set full
6192          * sync since it will be a full sync anyway and this will blow away the
6193          * old info in the log.
6194          */
6195         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6196
6197         key[0].objectid = objectid;
6198         key[0].type = BTRFS_INODE_ITEM_KEY;
6199         key[0].offset = 0;
6200
6201         sizes[0] = sizeof(struct btrfs_inode_item);
6202
6203         if (name) {
6204                 /*
6205                  * Start new inodes with an inode_ref. This is slightly more
6206                  * efficient for small numbers of hard links since they will
6207                  * be packed into one item. Extended refs will kick in if we
6208                  * add more hard links than can fit in the ref item.
6209                  */
6210                 key[1].objectid = objectid;
6211                 key[1].type = BTRFS_INODE_REF_KEY;
6212                 key[1].offset = ref_objectid;
6213
6214                 sizes[1] = name_len + sizeof(*ref);
6215         }
6216
6217         location = &BTRFS_I(inode)->location;
6218         location->objectid = objectid;
6219         location->offset = 0;
6220         location->type = BTRFS_INODE_ITEM_KEY;
6221
6222         ret = btrfs_insert_inode_locked(inode);
6223         if (ret < 0) {
6224                 iput(inode);
6225                 goto fail;
6226         }
6227
6228         path->leave_spinning = 1;
6229         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6230         if (ret != 0)
6231                 goto fail_unlock;
6232
6233         inode_init_owner(inode, dir, mode);
6234         inode_set_bytes(inode, 0);
6235
6236         inode->i_mtime = current_time(inode);
6237         inode->i_atime = inode->i_mtime;
6238         inode->i_ctime = inode->i_mtime;
6239         BTRFS_I(inode)->i_otime = inode->i_mtime;
6240
6241         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6242                                   struct btrfs_inode_item);
6243         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6244                              sizeof(*inode_item));
6245         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6246
6247         if (name) {
6248                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6249                                      struct btrfs_inode_ref);
6250                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6251                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6252                 ptr = (unsigned long)(ref + 1);
6253                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6254         }
6255
6256         btrfs_mark_buffer_dirty(path->nodes[0]);
6257         btrfs_free_path(path);
6258
6259         btrfs_inherit_iflags(inode, dir);
6260
6261         if (S_ISREG(mode)) {
6262                 if (btrfs_test_opt(fs_info, NODATASUM))
6263                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6264                 if (btrfs_test_opt(fs_info, NODATACOW))
6265                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6266                                 BTRFS_INODE_NODATASUM;
6267         }
6268
6269         inode_tree_add(inode);
6270
6271         trace_btrfs_inode_new(inode);
6272         btrfs_set_inode_last_trans(trans, inode);
6273
6274         btrfs_update_root_times(trans, root);
6275
6276         ret = btrfs_inode_inherit_props(trans, inode, dir);
6277         if (ret)
6278                 btrfs_err(fs_info,
6279                           "error inheriting props for ino %llu (root %llu): %d",
6280                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6281
6282         return inode;
6283
6284 fail_unlock:
6285         discard_new_inode(inode);
6286 fail:
6287         if (dir && name)
6288                 BTRFS_I(dir)->index_cnt--;
6289         btrfs_free_path(path);
6290         return ERR_PTR(ret);
6291 }
6292
6293 static inline u8 btrfs_inode_type(struct inode *inode)
6294 {
6295         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6296 }
6297
6298 /*
6299  * utility function to add 'inode' into 'parent_inode' with
6300  * a give name and a given sequence number.
6301  * if 'add_backref' is true, also insert a backref from the
6302  * inode to the parent directory.
6303  */
6304 int btrfs_add_link(struct btrfs_trans_handle *trans,
6305                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6306                    const char *name, int name_len, int add_backref, u64 index)
6307 {
6308         int ret = 0;
6309         struct btrfs_key key;
6310         struct btrfs_root *root = parent_inode->root;
6311         u64 ino = btrfs_ino(inode);
6312         u64 parent_ino = btrfs_ino(parent_inode);
6313
6314         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6315                 memcpy(&key, &inode->root->root_key, sizeof(key));
6316         } else {
6317                 key.objectid = ino;
6318                 key.type = BTRFS_INODE_ITEM_KEY;
6319                 key.offset = 0;
6320         }
6321
6322         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6323                 ret = btrfs_add_root_ref(trans, key.objectid,
6324                                          root->root_key.objectid, parent_ino,
6325                                          index, name, name_len);
6326         } else if (add_backref) {
6327                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6328                                              parent_ino, index);
6329         }
6330
6331         /* Nothing to clean up yet */
6332         if (ret)
6333                 return ret;
6334
6335         ret = btrfs_insert_dir_item(trans, name, name_len, parent_inode, &key,
6336                                     btrfs_inode_type(&inode->vfs_inode), index);
6337         if (ret == -EEXIST || ret == -EOVERFLOW)
6338                 goto fail_dir_item;
6339         else if (ret) {
6340                 btrfs_abort_transaction(trans, ret);
6341                 return ret;
6342         }
6343
6344         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6345                            name_len * 2);
6346         inode_inc_iversion(&parent_inode->vfs_inode);
6347         parent_inode->vfs_inode.i_mtime = parent_inode->vfs_inode.i_ctime =
6348                 current_time(&parent_inode->vfs_inode);
6349         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6350         if (ret)
6351                 btrfs_abort_transaction(trans, ret);
6352         return ret;
6353
6354 fail_dir_item:
6355         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6356                 u64 local_index;
6357                 int err;
6358                 err = btrfs_del_root_ref(trans, key.objectid,
6359                                          root->root_key.objectid, parent_ino,
6360                                          &local_index, name, name_len);
6361                 if (err)
6362                         btrfs_abort_transaction(trans, err);
6363         } else if (add_backref) {
6364                 u64 local_index;
6365                 int err;
6366
6367                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6368                                           ino, parent_ino, &local_index);
6369                 if (err)
6370                         btrfs_abort_transaction(trans, err);
6371         }
6372
6373         /* Return the original error code */
6374         return ret;
6375 }
6376
6377 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6378                             struct btrfs_inode *dir, struct dentry *dentry,
6379                             struct btrfs_inode *inode, int backref, u64 index)
6380 {
6381         int err = btrfs_add_link(trans, dir, inode,
6382                                  dentry->d_name.name, dentry->d_name.len,
6383                                  backref, index);
6384         if (err > 0)
6385                 err = -EEXIST;
6386         return err;
6387 }
6388
6389 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6390                         umode_t mode, dev_t rdev)
6391 {
6392         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6393         struct btrfs_trans_handle *trans;
6394         struct btrfs_root *root = BTRFS_I(dir)->root;
6395         struct inode *inode = NULL;
6396         int err;
6397         u64 objectid;
6398         u64 index = 0;
6399
6400         /*
6401          * 2 for inode item and ref
6402          * 2 for dir items
6403          * 1 for xattr if selinux is on
6404          */
6405         trans = btrfs_start_transaction(root, 5);
6406         if (IS_ERR(trans))
6407                 return PTR_ERR(trans);
6408
6409         err = btrfs_find_free_ino(root, &objectid);
6410         if (err)
6411                 goto out_unlock;
6412
6413         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6414                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6415                         mode, &index);
6416         if (IS_ERR(inode)) {
6417                 err = PTR_ERR(inode);
6418                 inode = NULL;
6419                 goto out_unlock;
6420         }
6421
6422         /*
6423         * If the active LSM wants to access the inode during
6424         * d_instantiate it needs these. Smack checks to see
6425         * if the filesystem supports xattrs by looking at the
6426         * ops vector.
6427         */
6428         inode->i_op = &btrfs_special_inode_operations;
6429         init_special_inode(inode, inode->i_mode, rdev);
6430
6431         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6432         if (err)
6433                 goto out_unlock;
6434
6435         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6436                         0, index);
6437         if (err)
6438                 goto out_unlock;
6439
6440         btrfs_update_inode(trans, root, inode);
6441         d_instantiate_new(dentry, inode);
6442
6443 out_unlock:
6444         btrfs_end_transaction(trans);
6445         btrfs_btree_balance_dirty(fs_info);
6446         if (err && inode) {
6447                 inode_dec_link_count(inode);
6448                 discard_new_inode(inode);
6449         }
6450         return err;
6451 }
6452
6453 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6454                         umode_t mode, bool excl)
6455 {
6456         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6457         struct btrfs_trans_handle *trans;
6458         struct btrfs_root *root = BTRFS_I(dir)->root;
6459         struct inode *inode = NULL;
6460         int err;
6461         u64 objectid;
6462         u64 index = 0;
6463
6464         /*
6465          * 2 for inode item and ref
6466          * 2 for dir items
6467          * 1 for xattr if selinux is on
6468          */
6469         trans = btrfs_start_transaction(root, 5);
6470         if (IS_ERR(trans))
6471                 return PTR_ERR(trans);
6472
6473         err = btrfs_find_free_ino(root, &objectid);
6474         if (err)
6475                 goto out_unlock;
6476
6477         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6478                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6479                         mode, &index);
6480         if (IS_ERR(inode)) {
6481                 err = PTR_ERR(inode);
6482                 inode = NULL;
6483                 goto out_unlock;
6484         }
6485         /*
6486         * If the active LSM wants to access the inode during
6487         * d_instantiate it needs these. Smack checks to see
6488         * if the filesystem supports xattrs by looking at the
6489         * ops vector.
6490         */
6491         inode->i_fop = &btrfs_file_operations;
6492         inode->i_op = &btrfs_file_inode_operations;
6493         inode->i_mapping->a_ops = &btrfs_aops;
6494
6495         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6496         if (err)
6497                 goto out_unlock;
6498
6499         err = btrfs_update_inode(trans, root, inode);
6500         if (err)
6501                 goto out_unlock;
6502
6503         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6504                         0, index);
6505         if (err)
6506                 goto out_unlock;
6507
6508         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6509         d_instantiate_new(dentry, inode);
6510
6511 out_unlock:
6512         btrfs_end_transaction(trans);
6513         if (err && inode) {
6514                 inode_dec_link_count(inode);
6515                 discard_new_inode(inode);
6516         }
6517         btrfs_btree_balance_dirty(fs_info);
6518         return err;
6519 }
6520
6521 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6522                       struct dentry *dentry)
6523 {
6524         struct btrfs_trans_handle *trans = NULL;
6525         struct btrfs_root *root = BTRFS_I(dir)->root;
6526         struct inode *inode = d_inode(old_dentry);
6527         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6528         u64 index;
6529         int err;
6530         int drop_inode = 0;
6531
6532         /* do not allow sys_link's with other subvols of the same device */
6533         if (root->root_key.objectid != BTRFS_I(inode)->root->root_key.objectid)
6534                 return -EXDEV;
6535
6536         if (inode->i_nlink >= BTRFS_LINK_MAX)
6537                 return -EMLINK;
6538
6539         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6540         if (err)
6541                 goto fail;
6542
6543         /*
6544          * 2 items for inode and inode ref
6545          * 2 items for dir items
6546          * 1 item for parent inode
6547          * 1 item for orphan item deletion if O_TMPFILE
6548          */
6549         trans = btrfs_start_transaction(root, inode->i_nlink ? 5 : 6);
6550         if (IS_ERR(trans)) {
6551                 err = PTR_ERR(trans);
6552                 trans = NULL;
6553                 goto fail;
6554         }
6555
6556         /* There are several dir indexes for this inode, clear the cache. */
6557         BTRFS_I(inode)->dir_index = 0ULL;
6558         inc_nlink(inode);
6559         inode_inc_iversion(inode);
6560         inode->i_ctime = current_time(inode);
6561         ihold(inode);
6562         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6563
6564         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6565                         1, index);
6566
6567         if (err) {
6568                 drop_inode = 1;
6569         } else {
6570                 struct dentry *parent = dentry->d_parent;
6571                 int ret;
6572
6573                 err = btrfs_update_inode(trans, root, inode);
6574                 if (err)
6575                         goto fail;
6576                 if (inode->i_nlink == 1) {
6577                         /*
6578                          * If new hard link count is 1, it's a file created
6579                          * with open(2) O_TMPFILE flag.
6580                          */
6581                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6582                         if (err)
6583                                 goto fail;
6584                 }
6585                 BTRFS_I(inode)->last_link_trans = trans->transid;
6586                 d_instantiate(dentry, inode);
6587                 ret = btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent,
6588                                          true, NULL);
6589                 if (ret == BTRFS_NEED_TRANS_COMMIT) {
6590                         err = btrfs_commit_transaction(trans);
6591                         trans = NULL;
6592                 }
6593         }
6594
6595 fail:
6596         if (trans)
6597                 btrfs_end_transaction(trans);
6598         if (drop_inode) {
6599                 inode_dec_link_count(inode);
6600                 iput(inode);
6601         }
6602         btrfs_btree_balance_dirty(fs_info);
6603         return err;
6604 }
6605
6606 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6607 {
6608         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6609         struct inode *inode = NULL;
6610         struct btrfs_trans_handle *trans;
6611         struct btrfs_root *root = BTRFS_I(dir)->root;
6612         int err = 0;
6613         u64 objectid = 0;
6614         u64 index = 0;
6615
6616         /*
6617          * 2 items for inode and ref
6618          * 2 items for dir items
6619          * 1 for xattr if selinux is on
6620          */
6621         trans = btrfs_start_transaction(root, 5);
6622         if (IS_ERR(trans))
6623                 return PTR_ERR(trans);
6624
6625         err = btrfs_find_free_ino(root, &objectid);
6626         if (err)
6627                 goto out_fail;
6628
6629         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6630                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6631                         S_IFDIR | mode, &index);
6632         if (IS_ERR(inode)) {
6633                 err = PTR_ERR(inode);
6634                 inode = NULL;
6635                 goto out_fail;
6636         }
6637
6638         /* these must be set before we unlock the inode */
6639         inode->i_op = &btrfs_dir_inode_operations;
6640         inode->i_fop = &btrfs_dir_file_operations;
6641
6642         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6643         if (err)
6644                 goto out_fail;
6645
6646         btrfs_i_size_write(BTRFS_I(inode), 0);
6647         err = btrfs_update_inode(trans, root, inode);
6648         if (err)
6649                 goto out_fail;
6650
6651         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6652                         dentry->d_name.name,
6653                         dentry->d_name.len, 0, index);
6654         if (err)
6655                 goto out_fail;
6656
6657         d_instantiate_new(dentry, inode);
6658
6659 out_fail:
6660         btrfs_end_transaction(trans);
6661         if (err && inode) {
6662                 inode_dec_link_count(inode);
6663                 discard_new_inode(inode);
6664         }
6665         btrfs_btree_balance_dirty(fs_info);
6666         return err;
6667 }
6668
6669 static noinline int uncompress_inline(struct btrfs_path *path,
6670                                       struct page *page,
6671                                       size_t pg_offset, u64 extent_offset,
6672                                       struct btrfs_file_extent_item *item)
6673 {
6674         int ret;
6675         struct extent_buffer *leaf = path->nodes[0];
6676         char *tmp;
6677         size_t max_size;
6678         unsigned long inline_size;
6679         unsigned long ptr;
6680         int compress_type;
6681
6682         WARN_ON(pg_offset != 0);
6683         compress_type = btrfs_file_extent_compression(leaf, item);
6684         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6685         inline_size = btrfs_file_extent_inline_item_len(leaf,
6686                                         btrfs_item_nr(path->slots[0]));
6687         tmp = kmalloc(inline_size, GFP_NOFS);
6688         if (!tmp)
6689                 return -ENOMEM;
6690         ptr = btrfs_file_extent_inline_start(item);
6691
6692         read_extent_buffer(leaf, tmp, ptr, inline_size);
6693
6694         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6695         ret = btrfs_decompress(compress_type, tmp, page,
6696                                extent_offset, inline_size, max_size);
6697
6698         /*
6699          * decompression code contains a memset to fill in any space between the end
6700          * of the uncompressed data and the end of max_size in case the decompressed
6701          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6702          * the end of an inline extent and the beginning of the next block, so we
6703          * cover that region here.
6704          */
6705
6706         if (max_size + pg_offset < PAGE_SIZE) {
6707                 char *map = kmap(page);
6708                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6709                 kunmap(page);
6710         }
6711         kfree(tmp);
6712         return ret;
6713 }
6714
6715 /*
6716  * a bit scary, this does extent mapping from logical file offset to the disk.
6717  * the ugly parts come from merging extents from the disk with the in-ram
6718  * representation.  This gets more complex because of the data=ordered code,
6719  * where the in-ram extents might be locked pending data=ordered completion.
6720  *
6721  * This also copies inline extents directly into the page.
6722  */
6723 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6724                                     struct page *page,
6725                                     size_t pg_offset, u64 start, u64 len,
6726                                     int create)
6727 {
6728         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6729         int ret;
6730         int err = 0;
6731         u64 extent_start = 0;
6732         u64 extent_end = 0;
6733         u64 objectid = btrfs_ino(inode);
6734         u32 found_type;
6735         struct btrfs_path *path = NULL;
6736         struct btrfs_root *root = inode->root;
6737         struct btrfs_file_extent_item *item;
6738         struct extent_buffer *leaf;
6739         struct btrfs_key found_key;
6740         struct extent_map *em = NULL;
6741         struct extent_map_tree *em_tree = &inode->extent_tree;
6742         struct extent_io_tree *io_tree = &inode->io_tree;
6743         const bool new_inline = !page || create;
6744
6745         read_lock(&em_tree->lock);
6746         em = lookup_extent_mapping(em_tree, start, len);
6747         if (em)
6748                 em->bdev = fs_info->fs_devices->latest_bdev;
6749         read_unlock(&em_tree->lock);
6750
6751         if (em) {
6752                 if (em->start > start || em->start + em->len <= start)
6753                         free_extent_map(em);
6754                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6755                         free_extent_map(em);
6756                 else
6757                         goto out;
6758         }
6759         em = alloc_extent_map();
6760         if (!em) {
6761                 err = -ENOMEM;
6762                 goto out;
6763         }
6764         em->bdev = fs_info->fs_devices->latest_bdev;
6765         em->start = EXTENT_MAP_HOLE;
6766         em->orig_start = EXTENT_MAP_HOLE;
6767         em->len = (u64)-1;
6768         em->block_len = (u64)-1;
6769
6770         path = btrfs_alloc_path();
6771         if (!path) {
6772                 err = -ENOMEM;
6773                 goto out;
6774         }
6775
6776         /* Chances are we'll be called again, so go ahead and do readahead */
6777         path->reada = READA_FORWARD;
6778
6779         /*
6780          * Unless we're going to uncompress the inline extent, no sleep would
6781          * happen.
6782          */
6783         path->leave_spinning = 1;
6784
6785         ret = btrfs_lookup_file_extent(NULL, root, path, objectid, start, 0);
6786         if (ret < 0) {
6787                 err = ret;
6788                 goto out;
6789         }
6790
6791         if (ret != 0) {
6792                 if (path->slots[0] == 0)
6793                         goto not_found;
6794                 path->slots[0]--;
6795         }
6796
6797         leaf = path->nodes[0];
6798         item = btrfs_item_ptr(leaf, path->slots[0],
6799                               struct btrfs_file_extent_item);
6800         /* are we inside the extent that was found? */
6801         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6802         found_type = found_key.type;
6803         if (found_key.objectid != objectid ||
6804             found_type != BTRFS_EXTENT_DATA_KEY) {
6805                 /*
6806                  * If we backup past the first extent we want to move forward
6807                  * and see if there is an extent in front of us, otherwise we'll
6808                  * say there is a hole for our whole search range which can
6809                  * cause problems.
6810                  */
6811                 extent_end = start;
6812                 goto next;
6813         }
6814
6815         found_type = btrfs_file_extent_type(leaf, item);
6816         extent_start = found_key.offset;
6817         if (found_type == BTRFS_FILE_EXTENT_REG ||
6818             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6819                 extent_end = extent_start +
6820                        btrfs_file_extent_num_bytes(leaf, item);
6821
6822                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
6823                                                        extent_start);
6824         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6825                 size_t size;
6826
6827                 size = btrfs_file_extent_ram_bytes(leaf, item);
6828                 extent_end = ALIGN(extent_start + size,
6829                                    fs_info->sectorsize);
6830
6831                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
6832                                                       path->slots[0],
6833                                                       extent_start);
6834         }
6835 next:
6836         if (start >= extent_end) {
6837                 path->slots[0]++;
6838                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6839                         ret = btrfs_next_leaf(root, path);
6840                         if (ret < 0) {
6841                                 err = ret;
6842                                 goto out;
6843                         }
6844                         if (ret > 0)
6845                                 goto not_found;
6846                         leaf = path->nodes[0];
6847                 }
6848                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6849                 if (found_key.objectid != objectid ||
6850                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6851                         goto not_found;
6852                 if (start + len <= found_key.offset)
6853                         goto not_found;
6854                 if (start > found_key.offset)
6855                         goto next;
6856                 em->start = start;
6857                 em->orig_start = start;
6858                 em->len = found_key.offset - start;
6859                 goto not_found_em;
6860         }
6861
6862         btrfs_extent_item_to_extent_map(inode, path, item,
6863                         new_inline, em);
6864
6865         if (found_type == BTRFS_FILE_EXTENT_REG ||
6866             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6867                 goto insert;
6868         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6869                 unsigned long ptr;
6870                 char *map;
6871                 size_t size;
6872                 size_t extent_offset;
6873                 size_t copy_size;
6874
6875                 if (new_inline)
6876                         goto out;
6877
6878                 size = btrfs_file_extent_ram_bytes(leaf, item);
6879                 extent_offset = page_offset(page) + pg_offset - extent_start;
6880                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6881                                   size - extent_offset);
6882                 em->start = extent_start + extent_offset;
6883                 em->len = ALIGN(copy_size, fs_info->sectorsize);
6884                 em->orig_block_len = em->len;
6885                 em->orig_start = em->start;
6886                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6887
6888                 btrfs_set_path_blocking(path);
6889                 if (!PageUptodate(page)) {
6890                         if (btrfs_file_extent_compression(leaf, item) !=
6891                             BTRFS_COMPRESS_NONE) {
6892                                 ret = uncompress_inline(path, page, pg_offset,
6893                                                         extent_offset, item);
6894                                 if (ret) {
6895                                         err = ret;
6896                                         goto out;
6897                                 }
6898                         } else {
6899                                 map = kmap(page);
6900                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6901                                                    copy_size);
6902                                 if (pg_offset + copy_size < PAGE_SIZE) {
6903                                         memset(map + pg_offset + copy_size, 0,
6904                                                PAGE_SIZE - pg_offset -
6905                                                copy_size);
6906                                 }
6907                                 kunmap(page);
6908                         }
6909                         flush_dcache_page(page);
6910                 }
6911                 set_extent_uptodate(io_tree, em->start,
6912                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6913                 goto insert;
6914         }
6915 not_found:
6916         em->start = start;
6917         em->orig_start = start;
6918         em->len = len;
6919 not_found_em:
6920         em->block_start = EXTENT_MAP_HOLE;
6921 insert:
6922         btrfs_release_path(path);
6923         if (em->start > start || extent_map_end(em) <= start) {
6924                 btrfs_err(fs_info,
6925                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
6926                           em->start, em->len, start, len);
6927                 err = -EIO;
6928                 goto out;
6929         }
6930
6931         err = 0;
6932         write_lock(&em_tree->lock);
6933         err = btrfs_add_extent_mapping(fs_info, em_tree, &em, start, len);
6934         write_unlock(&em_tree->lock);
6935 out:
6936         btrfs_free_path(path);
6937
6938         trace_btrfs_get_extent(root, inode, em);
6939
6940         if (err) {
6941                 free_extent_map(em);
6942                 return ERR_PTR(err);
6943         }
6944         BUG_ON(!em); /* Error is always set */
6945         return em;
6946 }
6947
6948 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
6949                 struct page *page,
6950                 size_t pg_offset, u64 start, u64 len,
6951                 int create)
6952 {
6953         struct extent_map *em;
6954         struct extent_map *hole_em = NULL;
6955         u64 range_start = start;
6956         u64 end;
6957         u64 found;
6958         u64 found_end;
6959         int err = 0;
6960
6961         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
6962         if (IS_ERR(em))
6963                 return em;
6964         /*
6965          * If our em maps to:
6966          * - a hole or
6967          * - a pre-alloc extent,
6968          * there might actually be delalloc bytes behind it.
6969          */
6970         if (em->block_start != EXTENT_MAP_HOLE &&
6971             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
6972                 return em;
6973         else
6974                 hole_em = em;
6975
6976         /* check to see if we've wrapped (len == -1 or similar) */
6977         end = start + len;
6978         if (end < start)
6979                 end = (u64)-1;
6980         else
6981                 end -= 1;
6982
6983         em = NULL;
6984
6985         /* ok, we didn't find anything, lets look for delalloc */
6986         found = count_range_bits(&inode->io_tree, &range_start,
6987                                  end, len, EXTENT_DELALLOC, 1);
6988         found_end = range_start + found;
6989         if (found_end < range_start)
6990                 found_end = (u64)-1;
6991
6992         /*
6993          * we didn't find anything useful, return
6994          * the original results from get_extent()
6995          */
6996         if (range_start > end || found_end <= start) {
6997                 em = hole_em;
6998                 hole_em = NULL;
6999                 goto out;
7000         }
7001
7002         /* adjust the range_start to make sure it doesn't
7003          * go backwards from the start they passed in
7004          */
7005         range_start = max(start, range_start);
7006         found = found_end - range_start;
7007
7008         if (found > 0) {
7009                 u64 hole_start = start;
7010                 u64 hole_len = len;
7011
7012                 em = alloc_extent_map();
7013                 if (!em) {
7014                         err = -ENOMEM;
7015                         goto out;
7016                 }
7017                 /*
7018                  * when btrfs_get_extent can't find anything it
7019                  * returns one huge hole
7020                  *
7021                  * make sure what it found really fits our range, and
7022                  * adjust to make sure it is based on the start from
7023                  * the caller
7024                  */
7025                 if (hole_em) {
7026                         u64 calc_end = extent_map_end(hole_em);
7027
7028                         if (calc_end <= start || (hole_em->start > end)) {
7029                                 free_extent_map(hole_em);
7030                                 hole_em = NULL;
7031                         } else {
7032                                 hole_start = max(hole_em->start, start);
7033                                 hole_len = calc_end - hole_start;
7034                         }
7035                 }
7036                 em->bdev = NULL;
7037                 if (hole_em && range_start > hole_start) {
7038                         /* our hole starts before our delalloc, so we
7039                          * have to return just the parts of the hole
7040                          * that go until  the delalloc starts
7041                          */
7042                         em->len = min(hole_len,
7043                                       range_start - hole_start);
7044                         em->start = hole_start;
7045                         em->orig_start = hole_start;
7046                         /*
7047                          * don't adjust block start at all,
7048                          * it is fixed at EXTENT_MAP_HOLE
7049                          */
7050                         em->block_start = hole_em->block_start;
7051                         em->block_len = hole_len;
7052                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7053                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7054                 } else {
7055                         em->start = range_start;
7056                         em->len = found;
7057                         em->orig_start = range_start;
7058                         em->block_start = EXTENT_MAP_DELALLOC;
7059                         em->block_len = found;
7060                 }
7061         } else {
7062                 return hole_em;
7063         }
7064 out:
7065
7066         free_extent_map(hole_em);
7067         if (err) {
7068                 free_extent_map(em);
7069                 return ERR_PTR(err);
7070         }
7071         return em;
7072 }
7073
7074 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7075                                                   const u64 start,
7076                                                   const u64 len,
7077                                                   const u64 orig_start,
7078                                                   const u64 block_start,
7079                                                   const u64 block_len,
7080                                                   const u64 orig_block_len,
7081                                                   const u64 ram_bytes,
7082                                                   const int type)
7083 {
7084         struct extent_map *em = NULL;
7085         int ret;
7086
7087         if (type != BTRFS_ORDERED_NOCOW) {
7088                 em = create_io_em(inode, start, len, orig_start,
7089                                   block_start, block_len, orig_block_len,
7090                                   ram_bytes,
7091                                   BTRFS_COMPRESS_NONE, /* compress_type */
7092                                   type);
7093                 if (IS_ERR(em))
7094                         goto out;
7095         }
7096         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7097                                            len, block_len, type);
7098         if (ret) {
7099                 if (em) {
7100                         free_extent_map(em);
7101                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7102                                                 start + len - 1, 0);
7103                 }
7104                 em = ERR_PTR(ret);
7105         }
7106  out:
7107
7108         return em;
7109 }
7110
7111 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7112                                                   u64 start, u64 len)
7113 {
7114         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7115         struct btrfs_root *root = BTRFS_I(inode)->root;
7116         struct extent_map *em;
7117         struct btrfs_key ins;
7118         u64 alloc_hint;
7119         int ret;
7120
7121         alloc_hint = get_extent_allocation_hint(inode, start, len);
7122         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7123                                    0, alloc_hint, &ins, 1, 1);
7124         if (ret)
7125                 return ERR_PTR(ret);
7126
7127         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7128                                      ins.objectid, ins.offset, ins.offset,
7129                                      ins.offset, BTRFS_ORDERED_REGULAR);
7130         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7131         if (IS_ERR(em))
7132                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7133                                            ins.offset, 1);
7134
7135         return em;
7136 }
7137
7138 /*
7139  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7140  * block must be cow'd
7141  */
7142 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7143                               u64 *orig_start, u64 *orig_block_len,
7144                               u64 *ram_bytes)
7145 {
7146         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7147         struct btrfs_path *path;
7148         int ret;
7149         struct extent_buffer *leaf;
7150         struct btrfs_root *root = BTRFS_I(inode)->root;
7151         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7152         struct btrfs_file_extent_item *fi;
7153         struct btrfs_key key;
7154         u64 disk_bytenr;
7155         u64 backref_offset;
7156         u64 extent_end;
7157         u64 num_bytes;
7158         int slot;
7159         int found_type;
7160         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7161
7162         path = btrfs_alloc_path();
7163         if (!path)
7164                 return -ENOMEM;
7165
7166         ret = btrfs_lookup_file_extent(NULL, root, path,
7167                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7168         if (ret < 0)
7169                 goto out;
7170
7171         slot = path->slots[0];
7172         if (ret == 1) {
7173                 if (slot == 0) {
7174                         /* can't find the item, must cow */
7175                         ret = 0;
7176                         goto out;
7177                 }
7178                 slot--;
7179         }
7180         ret = 0;
7181         leaf = path->nodes[0];
7182         btrfs_item_key_to_cpu(leaf, &key, slot);
7183         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7184             key.type != BTRFS_EXTENT_DATA_KEY) {
7185                 /* not our file or wrong item type, must cow */
7186                 goto out;
7187         }
7188
7189         if (key.offset > offset) {
7190                 /* Wrong offset, must cow */
7191                 goto out;
7192         }
7193
7194         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7195         found_type = btrfs_file_extent_type(leaf, fi);
7196         if (found_type != BTRFS_FILE_EXTENT_REG &&
7197             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7198                 /* not a regular extent, must cow */
7199                 goto out;
7200         }
7201
7202         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7203                 goto out;
7204
7205         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7206         if (extent_end <= offset)
7207                 goto out;
7208
7209         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7210         if (disk_bytenr == 0)
7211                 goto out;
7212
7213         if (btrfs_file_extent_compression(leaf, fi) ||
7214             btrfs_file_extent_encryption(leaf, fi) ||
7215             btrfs_file_extent_other_encoding(leaf, fi))
7216                 goto out;
7217
7218         /*
7219          * Do the same check as in btrfs_cross_ref_exist but without the
7220          * unnecessary search.
7221          */
7222         if (btrfs_file_extent_generation(leaf, fi) <=
7223             btrfs_root_last_snapshot(&root->root_item))
7224                 goto out;
7225
7226         backref_offset = btrfs_file_extent_offset(leaf, fi);
7227
7228         if (orig_start) {
7229                 *orig_start = key.offset - backref_offset;
7230                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7231                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7232         }
7233
7234         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7235                 goto out;
7236
7237         num_bytes = min(offset + *len, extent_end) - offset;
7238         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7239                 u64 range_end;
7240
7241                 range_end = round_up(offset + num_bytes,
7242                                      root->fs_info->sectorsize) - 1;
7243                 ret = test_range_bit(io_tree, offset, range_end,
7244                                      EXTENT_DELALLOC, 0, NULL);
7245                 if (ret) {
7246                         ret = -EAGAIN;
7247                         goto out;
7248                 }
7249         }
7250
7251         btrfs_release_path(path);
7252
7253         /*
7254          * look for other files referencing this extent, if we
7255          * find any we must cow
7256          */
7257
7258         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7259                                     key.offset - backref_offset, disk_bytenr);
7260         if (ret) {
7261                 ret = 0;
7262                 goto out;
7263         }
7264
7265         /*
7266          * adjust disk_bytenr and num_bytes to cover just the bytes
7267          * in this extent we are about to write.  If there
7268          * are any csums in that range we have to cow in order
7269          * to keep the csums correct
7270          */
7271         disk_bytenr += backref_offset;
7272         disk_bytenr += offset - key.offset;
7273         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7274                 goto out;
7275         /*
7276          * all of the above have passed, it is safe to overwrite this extent
7277          * without cow
7278          */
7279         *len = num_bytes;
7280         ret = 1;
7281 out:
7282         btrfs_free_path(path);
7283         return ret;
7284 }
7285
7286 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7287                               struct extent_state **cached_state, int writing)
7288 {
7289         struct btrfs_ordered_extent *ordered;
7290         int ret = 0;
7291
7292         while (1) {
7293                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7294                                  cached_state);
7295                 /*
7296                  * We're concerned with the entire range that we're going to be
7297                  * doing DIO to, so we need to make sure there's no ordered
7298                  * extents in this range.
7299                  */
7300                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7301                                                      lockend - lockstart + 1);
7302
7303                 /*
7304                  * We need to make sure there are no buffered pages in this
7305                  * range either, we could have raced between the invalidate in
7306                  * generic_file_direct_write and locking the extent.  The
7307                  * invalidate needs to happen so that reads after a write do not
7308                  * get stale data.
7309                  */
7310                 if (!ordered &&
7311                     (!writing || !filemap_range_has_page(inode->i_mapping,
7312                                                          lockstart, lockend)))
7313                         break;
7314
7315                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7316                                      cached_state);
7317
7318                 if (ordered) {
7319                         /*
7320                          * If we are doing a DIO read and the ordered extent we
7321                          * found is for a buffered write, we can not wait for it
7322                          * to complete and retry, because if we do so we can
7323                          * deadlock with concurrent buffered writes on page
7324                          * locks. This happens only if our DIO read covers more
7325                          * than one extent map, if at this point has already
7326                          * created an ordered extent for a previous extent map
7327                          * and locked its range in the inode's io tree, and a
7328                          * concurrent write against that previous extent map's
7329                          * range and this range started (we unlock the ranges
7330                          * in the io tree only when the bios complete and
7331                          * buffered writes always lock pages before attempting
7332                          * to lock range in the io tree).
7333                          */
7334                         if (writing ||
7335                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7336                                 btrfs_start_ordered_extent(inode, ordered, 1);
7337                         else
7338                                 ret = -ENOTBLK;
7339                         btrfs_put_ordered_extent(ordered);
7340                 } else {
7341                         /*
7342                          * We could trigger writeback for this range (and wait
7343                          * for it to complete) and then invalidate the pages for
7344                          * this range (through invalidate_inode_pages2_range()),
7345                          * but that can lead us to a deadlock with a concurrent
7346                          * call to readpages() (a buffered read or a defrag call
7347                          * triggered a readahead) on a page lock due to an
7348                          * ordered dio extent we created before but did not have
7349                          * yet a corresponding bio submitted (whence it can not
7350                          * complete), which makes readpages() wait for that
7351                          * ordered extent to complete while holding a lock on
7352                          * that page.
7353                          */
7354                         ret = -ENOTBLK;
7355                 }
7356
7357                 if (ret)
7358                         break;
7359
7360                 cond_resched();
7361         }
7362
7363         return ret;
7364 }
7365
7366 /* The callers of this must take lock_extent() */
7367 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7368                                        u64 orig_start, u64 block_start,
7369                                        u64 block_len, u64 orig_block_len,
7370                                        u64 ram_bytes, int compress_type,
7371                                        int type)
7372 {
7373         struct extent_map_tree *em_tree;
7374         struct extent_map *em;
7375         struct btrfs_root *root = BTRFS_I(inode)->root;
7376         int ret;
7377
7378         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7379                type == BTRFS_ORDERED_COMPRESSED ||
7380                type == BTRFS_ORDERED_NOCOW ||
7381                type == BTRFS_ORDERED_REGULAR);
7382
7383         em_tree = &BTRFS_I(inode)->extent_tree;
7384         em = alloc_extent_map();
7385         if (!em)
7386                 return ERR_PTR(-ENOMEM);
7387
7388         em->start = start;
7389         em->orig_start = orig_start;
7390         em->len = len;
7391         em->block_len = block_len;
7392         em->block_start = block_start;
7393         em->bdev = root->fs_info->fs_devices->latest_bdev;
7394         em->orig_block_len = orig_block_len;
7395         em->ram_bytes = ram_bytes;
7396         em->generation = -1;
7397         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7398         if (type == BTRFS_ORDERED_PREALLOC) {
7399                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7400         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7401                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7402                 em->compress_type = compress_type;
7403         }
7404
7405         do {
7406                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7407                                 em->start + em->len - 1, 0);
7408                 write_lock(&em_tree->lock);
7409                 ret = add_extent_mapping(em_tree, em, 1);
7410                 write_unlock(&em_tree->lock);
7411                 /*
7412                  * The caller has taken lock_extent(), who could race with us
7413                  * to add em?
7414                  */
7415         } while (ret == -EEXIST);
7416
7417         if (ret) {
7418                 free_extent_map(em);
7419                 return ERR_PTR(ret);
7420         }
7421
7422         /* em got 2 refs now, callers needs to do free_extent_map once. */
7423         return em;
7424 }
7425
7426
7427 static int btrfs_get_blocks_direct_read(struct extent_map *em,
7428                                         struct buffer_head *bh_result,
7429                                         struct inode *inode,
7430                                         u64 start, u64 len)
7431 {
7432         if (em->block_start == EXTENT_MAP_HOLE ||
7433                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7434                 return -ENOENT;
7435
7436         len = min(len, em->len - (start - em->start));
7437
7438         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7439                 inode->i_blkbits;
7440         bh_result->b_size = len;
7441         bh_result->b_bdev = em->bdev;
7442         set_buffer_mapped(bh_result);
7443
7444         return 0;
7445 }
7446
7447 static int btrfs_get_blocks_direct_write(struct extent_map **map,
7448                                          struct buffer_head *bh_result,
7449                                          struct inode *inode,
7450                                          struct btrfs_dio_data *dio_data,
7451                                          u64 start, u64 len)
7452 {
7453         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7454         struct extent_map *em = *map;
7455         int ret = 0;
7456
7457         /*
7458          * We don't allocate a new extent in the following cases
7459          *
7460          * 1) The inode is marked as NODATACOW. In this case we'll just use the
7461          * existing extent.
7462          * 2) The extent is marked as PREALLOC. We're good to go here and can
7463          * just use the extent.
7464          *
7465          */
7466         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7467             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7468              em->block_start != EXTENT_MAP_HOLE)) {
7469                 int type;
7470                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7471
7472                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7473                         type = BTRFS_ORDERED_PREALLOC;
7474                 else
7475                         type = BTRFS_ORDERED_NOCOW;
7476                 len = min(len, em->len - (start - em->start));
7477                 block_start = em->block_start + (start - em->start);
7478
7479                 if (can_nocow_extent(inode, start, &len, &orig_start,
7480                                      &orig_block_len, &ram_bytes) == 1 &&
7481                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7482                         struct extent_map *em2;
7483
7484                         em2 = btrfs_create_dio_extent(inode, start, len,
7485                                                       orig_start, block_start,
7486                                                       len, orig_block_len,
7487                                                       ram_bytes, type);
7488                         btrfs_dec_nocow_writers(fs_info, block_start);
7489                         if (type == BTRFS_ORDERED_PREALLOC) {
7490                                 free_extent_map(em);
7491                                 *map = em = em2;
7492                         }
7493
7494                         if (em2 && IS_ERR(em2)) {
7495                                 ret = PTR_ERR(em2);
7496                                 goto out;
7497                         }
7498                         /*
7499                          * For inode marked NODATACOW or extent marked PREALLOC,
7500                          * use the existing or preallocated extent, so does not
7501                          * need to adjust btrfs_space_info's bytes_may_use.
7502                          */
7503                         btrfs_free_reserved_data_space_noquota(inode, start,
7504                                                                len);
7505                         goto skip_cow;
7506                 }
7507         }
7508
7509         /* this will cow the extent */
7510         len = bh_result->b_size;
7511         free_extent_map(em);
7512         *map = em = btrfs_new_extent_direct(inode, start, len);
7513         if (IS_ERR(em)) {
7514                 ret = PTR_ERR(em);
7515                 goto out;
7516         }
7517
7518         len = min(len, em->len - (start - em->start));
7519
7520 skip_cow:
7521         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7522                 inode->i_blkbits;
7523         bh_result->b_size = len;
7524         bh_result->b_bdev = em->bdev;
7525         set_buffer_mapped(bh_result);
7526
7527         if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7528                 set_buffer_new(bh_result);
7529
7530         /*
7531          * Need to update the i_size under the extent lock so buffered
7532          * readers will get the updated i_size when we unlock.
7533          */
7534         if (!dio_data->overwrite && start + len > i_size_read(inode))
7535                 i_size_write(inode, start + len);
7536
7537         WARN_ON(dio_data->reserve < len);
7538         dio_data->reserve -= len;
7539         dio_data->unsubmitted_oe_range_end = start + len;
7540         current->journal_info = dio_data;
7541 out:
7542         return ret;
7543 }
7544
7545 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7546                                    struct buffer_head *bh_result, int create)
7547 {
7548         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7549         struct extent_map *em;
7550         struct extent_state *cached_state = NULL;
7551         struct btrfs_dio_data *dio_data = NULL;
7552         u64 start = iblock << inode->i_blkbits;
7553         u64 lockstart, lockend;
7554         u64 len = bh_result->b_size;
7555         int unlock_bits = EXTENT_LOCKED;
7556         int ret = 0;
7557
7558         if (create)
7559                 unlock_bits |= EXTENT_DIRTY;
7560         else
7561                 len = min_t(u64, len, fs_info->sectorsize);
7562
7563         lockstart = start;
7564         lockend = start + len - 1;
7565
7566         if (current->journal_info) {
7567                 /*
7568                  * Need to pull our outstanding extents and set journal_info to NULL so
7569                  * that anything that needs to check if there's a transaction doesn't get
7570                  * confused.
7571                  */
7572                 dio_data = current->journal_info;
7573                 current->journal_info = NULL;
7574         }
7575
7576         /*
7577          * If this errors out it's because we couldn't invalidate pagecache for
7578          * this range and we need to fallback to buffered.
7579          */
7580         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7581                                create)) {
7582                 ret = -ENOTBLK;
7583                 goto err;
7584         }
7585
7586         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7587         if (IS_ERR(em)) {
7588                 ret = PTR_ERR(em);
7589                 goto unlock_err;
7590         }
7591
7592         /*
7593          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7594          * io.  INLINE is special, and we could probably kludge it in here, but
7595          * it's still buffered so for safety lets just fall back to the generic
7596          * buffered path.
7597          *
7598          * For COMPRESSED we _have_ to read the entire extent in so we can
7599          * decompress it, so there will be buffering required no matter what we
7600          * do, so go ahead and fallback to buffered.
7601          *
7602          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7603          * to buffered IO.  Don't blame me, this is the price we pay for using
7604          * the generic code.
7605          */
7606         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7607             em->block_start == EXTENT_MAP_INLINE) {
7608                 free_extent_map(em);
7609                 ret = -ENOTBLK;
7610                 goto unlock_err;
7611         }
7612
7613         if (create) {
7614                 ret = btrfs_get_blocks_direct_write(&em, bh_result, inode,
7615                                                     dio_data, start, len);
7616                 if (ret < 0)
7617                         goto unlock_err;
7618
7619                 /* clear and unlock the entire range */
7620                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7621                                  unlock_bits, 1, 0, &cached_state);
7622         } else {
7623                 ret = btrfs_get_blocks_direct_read(em, bh_result, inode,
7624                                                    start, len);
7625                 /* Can be negative only if we read from a hole */
7626                 if (ret < 0) {
7627                         ret = 0;
7628                         free_extent_map(em);
7629                         goto unlock_err;
7630                 }
7631                 /*
7632                  * We need to unlock only the end area that we aren't using.
7633                  * The rest is going to be unlocked by the endio routine.
7634                  */
7635                 lockstart = start + bh_result->b_size;
7636                 if (lockstart < lockend) {
7637                         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7638                                          lockend, unlock_bits, 1, 0,
7639                                          &cached_state);
7640                 } else {
7641                         free_extent_state(cached_state);
7642                 }
7643         }
7644
7645         free_extent_map(em);
7646
7647         return 0;
7648
7649 unlock_err:
7650         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7651                          unlock_bits, 1, 0, &cached_state);
7652 err:
7653         if (dio_data)
7654                 current->journal_info = dio_data;
7655         return ret;
7656 }
7657
7658 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7659                                                  struct bio *bio,
7660                                                  int mirror_num)
7661 {
7662         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7663         blk_status_t ret;
7664
7665         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7666
7667         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7668         if (ret)
7669                 return ret;
7670
7671         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
7672
7673         return ret;
7674 }
7675
7676 static int btrfs_check_dio_repairable(struct inode *inode,
7677                                       struct bio *failed_bio,
7678                                       struct io_failure_record *failrec,
7679                                       int failed_mirror)
7680 {
7681         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7682         int num_copies;
7683
7684         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7685         if (num_copies == 1) {
7686                 /*
7687                  * we only have a single copy of the data, so don't bother with
7688                  * all the retry and error correction code that follows. no
7689                  * matter what the error is, it is very likely to persist.
7690                  */
7691                 btrfs_debug(fs_info,
7692                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7693                         num_copies, failrec->this_mirror, failed_mirror);
7694                 return 0;
7695         }
7696
7697         failrec->failed_mirror = failed_mirror;
7698         failrec->this_mirror++;
7699         if (failrec->this_mirror == failed_mirror)
7700                 failrec->this_mirror++;
7701
7702         if (failrec->this_mirror > num_copies) {
7703                 btrfs_debug(fs_info,
7704                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7705                         num_copies, failrec->this_mirror, failed_mirror);
7706                 return 0;
7707         }
7708
7709         return 1;
7710 }
7711
7712 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
7713                                    struct page *page, unsigned int pgoff,
7714                                    u64 start, u64 end, int failed_mirror,
7715                                    bio_end_io_t *repair_endio, void *repair_arg)
7716 {
7717         struct io_failure_record *failrec;
7718         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7719         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
7720         struct bio *bio;
7721         int isector;
7722         unsigned int read_mode = 0;
7723         int segs;
7724         int ret;
7725         blk_status_t status;
7726         struct bio_vec bvec;
7727
7728         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7729
7730         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7731         if (ret)
7732                 return errno_to_blk_status(ret);
7733
7734         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7735                                          failed_mirror);
7736         if (!ret) {
7737                 free_io_failure(failure_tree, io_tree, failrec);
7738                 return BLK_STS_IOERR;
7739         }
7740
7741         segs = bio_segments(failed_bio);
7742         bio_get_first_bvec(failed_bio, &bvec);
7743         if (segs > 1 ||
7744             (bvec.bv_len > btrfs_inode_sectorsize(inode)))
7745                 read_mode |= REQ_FAILFAST_DEV;
7746
7747         isector = start - btrfs_io_bio(failed_bio)->logical;
7748         isector >>= inode->i_sb->s_blocksize_bits;
7749         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7750                                 pgoff, isector, repair_endio, repair_arg);
7751         bio->bi_opf = REQ_OP_READ | read_mode;
7752
7753         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7754                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
7755                     read_mode, failrec->this_mirror, failrec->in_validation);
7756
7757         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7758         if (status) {
7759                 free_io_failure(failure_tree, io_tree, failrec);
7760                 bio_put(bio);
7761         }
7762
7763         return status;
7764 }
7765
7766 struct btrfs_retry_complete {
7767         struct completion done;
7768         struct inode *inode;
7769         u64 start;
7770         int uptodate;
7771 };
7772
7773 static void btrfs_retry_endio_nocsum(struct bio *bio)
7774 {
7775         struct btrfs_retry_complete *done = bio->bi_private;
7776         struct inode *inode = done->inode;
7777         struct bio_vec *bvec;
7778         struct extent_io_tree *io_tree, *failure_tree;
7779         int i;
7780
7781         if (bio->bi_status)
7782                 goto end;
7783
7784         ASSERT(bio->bi_vcnt == 1);
7785         io_tree = &BTRFS_I(inode)->io_tree;
7786         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7787         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(inode));
7788
7789         done->uptodate = 1;
7790         ASSERT(!bio_flagged(bio, BIO_CLONED));
7791         bio_for_each_segment_all(bvec, bio, i)
7792                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
7793                                  io_tree, done->start, bvec->bv_page,
7794                                  btrfs_ino(BTRFS_I(inode)), 0);
7795 end:
7796         complete(&done->done);
7797         bio_put(bio);
7798 }
7799
7800 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
7801                                                 struct btrfs_io_bio *io_bio)
7802 {
7803         struct btrfs_fs_info *fs_info;
7804         struct bio_vec bvec;
7805         struct bvec_iter iter;
7806         struct btrfs_retry_complete done;
7807         u64 start;
7808         unsigned int pgoff;
7809         u32 sectorsize;
7810         int nr_sectors;
7811         blk_status_t ret;
7812         blk_status_t err = BLK_STS_OK;
7813
7814         fs_info = BTRFS_I(inode)->root->fs_info;
7815         sectorsize = fs_info->sectorsize;
7816
7817         start = io_bio->logical;
7818         done.inode = inode;
7819         io_bio->bio.bi_iter = io_bio->iter;
7820
7821         bio_for_each_segment(bvec, &io_bio->bio, iter) {
7822                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
7823                 pgoff = bvec.bv_offset;
7824
7825 next_block_or_try_again:
7826                 done.uptodate = 0;
7827                 done.start = start;
7828                 init_completion(&done.done);
7829
7830                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
7831                                 pgoff, start, start + sectorsize - 1,
7832                                 io_bio->mirror_num,
7833                                 btrfs_retry_endio_nocsum, &done);
7834                 if (ret) {
7835                         err = ret;
7836                         goto next;
7837                 }
7838
7839                 wait_for_completion_io(&done.done);
7840
7841                 if (!done.uptodate) {
7842                         /* We might have another mirror, so try again */
7843                         goto next_block_or_try_again;
7844                 }
7845
7846 next:
7847                 start += sectorsize;
7848
7849                 nr_sectors--;
7850                 if (nr_sectors) {
7851                         pgoff += sectorsize;
7852                         ASSERT(pgoff < PAGE_SIZE);
7853                         goto next_block_or_try_again;
7854                 }
7855         }
7856
7857         return err;
7858 }
7859
7860 static void btrfs_retry_endio(struct bio *bio)
7861 {
7862         struct btrfs_retry_complete *done = bio->bi_private;
7863         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7864         struct extent_io_tree *io_tree, *failure_tree;
7865         struct inode *inode = done->inode;
7866         struct bio_vec *bvec;
7867         int uptodate;
7868         int ret;
7869         int i;
7870
7871         if (bio->bi_status)
7872                 goto end;
7873
7874         uptodate = 1;
7875
7876         ASSERT(bio->bi_vcnt == 1);
7877         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(done->inode));
7878
7879         io_tree = &BTRFS_I(inode)->io_tree;
7880         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7881
7882         ASSERT(!bio_flagged(bio, BIO_CLONED));
7883         bio_for_each_segment_all(bvec, bio, i) {
7884                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
7885                                              bvec->bv_offset, done->start,
7886                                              bvec->bv_len);
7887                 if (!ret)
7888                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
7889                                          failure_tree, io_tree, done->start,
7890                                          bvec->bv_page,
7891                                          btrfs_ino(BTRFS_I(inode)),
7892                                          bvec->bv_offset);
7893                 else
7894                         uptodate = 0;
7895         }
7896
7897         done->uptodate = uptodate;
7898 end:
7899         complete(&done->done);
7900         bio_put(bio);
7901 }
7902
7903 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
7904                 struct btrfs_io_bio *io_bio, blk_status_t err)
7905 {
7906         struct btrfs_fs_info *fs_info;
7907         struct bio_vec bvec;
7908         struct bvec_iter iter;
7909         struct btrfs_retry_complete done;
7910         u64 start;
7911         u64 offset = 0;
7912         u32 sectorsize;
7913         int nr_sectors;
7914         unsigned int pgoff;
7915         int csum_pos;
7916         bool uptodate = (err == 0);
7917         int ret;
7918         blk_status_t status;
7919
7920         fs_info = BTRFS_I(inode)->root->fs_info;
7921         sectorsize = fs_info->sectorsize;
7922
7923         err = BLK_STS_OK;
7924         start = io_bio->logical;
7925         done.inode = inode;
7926         io_bio->bio.bi_iter = io_bio->iter;
7927
7928         bio_for_each_segment(bvec, &io_bio->bio, iter) {
7929                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
7930
7931                 pgoff = bvec.bv_offset;
7932 next_block:
7933                 if (uptodate) {
7934                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
7935                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
7936                                         bvec.bv_page, pgoff, start, sectorsize);
7937                         if (likely(!ret))
7938                                 goto next;
7939                 }
7940 try_again:
7941                 done.uptodate = 0;
7942                 done.start = start;
7943                 init_completion(&done.done);
7944
7945                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
7946                                         pgoff, start, start + sectorsize - 1,
7947                                         io_bio->mirror_num, btrfs_retry_endio,
7948                                         &done);
7949                 if (status) {
7950                         err = status;
7951                         goto next;
7952                 }
7953
7954                 wait_for_completion_io(&done.done);
7955
7956                 if (!done.uptodate) {
7957                         /* We might have another mirror, so try again */
7958                         goto try_again;
7959                 }
7960 next:
7961                 offset += sectorsize;
7962                 start += sectorsize;
7963
7964                 ASSERT(nr_sectors);
7965
7966                 nr_sectors--;
7967                 if (nr_sectors) {
7968                         pgoff += sectorsize;
7969                         ASSERT(pgoff < PAGE_SIZE);
7970                         goto next_block;
7971                 }
7972         }
7973
7974         return err;
7975 }
7976
7977 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
7978                 struct btrfs_io_bio *io_bio, blk_status_t err)
7979 {
7980         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
7981
7982         if (skip_csum) {
7983                 if (unlikely(err))
7984                         return __btrfs_correct_data_nocsum(inode, io_bio);
7985                 else
7986                         return BLK_STS_OK;
7987         } else {
7988                 return __btrfs_subio_endio_read(inode, io_bio, err);
7989         }
7990 }
7991
7992 static void btrfs_endio_direct_read(struct bio *bio)
7993 {
7994         struct btrfs_dio_private *dip = bio->bi_private;
7995         struct inode *inode = dip->inode;
7996         struct bio *dio_bio;
7997         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7998         blk_status_t err = bio->bi_status;
7999
8000         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8001                 err = btrfs_subio_endio_read(inode, io_bio, err);
8002
8003         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8004                       dip->logical_offset + dip->bytes - 1);
8005         dio_bio = dip->dio_bio;
8006
8007         kfree(dip);
8008
8009         dio_bio->bi_status = err;
8010         dio_end_io(dio_bio);
8011         btrfs_io_bio_free_csum(io_bio);
8012         bio_put(bio);
8013 }
8014
8015 static void __endio_write_update_ordered(struct inode *inode,
8016                                          const u64 offset, const u64 bytes,
8017                                          const bool uptodate)
8018 {
8019         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8020         struct btrfs_ordered_extent *ordered = NULL;
8021         struct btrfs_workqueue *wq;
8022         btrfs_work_func_t func;
8023         u64 ordered_offset = offset;
8024         u64 ordered_bytes = bytes;
8025         u64 last_offset;
8026
8027         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
8028                 wq = fs_info->endio_freespace_worker;
8029                 func = btrfs_freespace_write_helper;
8030         } else {
8031                 wq = fs_info->endio_write_workers;
8032                 func = btrfs_endio_write_helper;
8033         }
8034
8035         while (ordered_offset < offset + bytes) {
8036                 last_offset = ordered_offset;
8037                 if (btrfs_dec_test_first_ordered_pending(inode, &ordered,
8038                                                            &ordered_offset,
8039                                                            ordered_bytes,
8040                                                            uptodate)) {
8041                         btrfs_init_work(&ordered->work, func,
8042                                         finish_ordered_fn,
8043                                         NULL, NULL);
8044                         btrfs_queue_work(wq, &ordered->work);
8045                 }
8046                 /*
8047                  * If btrfs_dec_test_ordered_pending does not find any ordered
8048                  * extent in the range, we can exit.
8049                  */
8050                 if (ordered_offset == last_offset)
8051                         return;
8052                 /*
8053                  * Our bio might span multiple ordered extents. In this case
8054                  * we keep going until we have accounted the whole dio.
8055                  */
8056                 if (ordered_offset < offset + bytes) {
8057                         ordered_bytes = offset + bytes - ordered_offset;
8058                         ordered = NULL;
8059                 }
8060         }
8061 }
8062
8063 static void btrfs_endio_direct_write(struct bio *bio)
8064 {
8065         struct btrfs_dio_private *dip = bio->bi_private;
8066         struct bio *dio_bio = dip->dio_bio;
8067
8068         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8069                                      dip->bytes, !bio->bi_status);
8070
8071         kfree(dip);
8072
8073         dio_bio->bi_status = bio->bi_status;
8074         dio_end_io(dio_bio);
8075         bio_put(bio);
8076 }
8077
8078 static blk_status_t btrfs_submit_bio_start_direct_io(void *private_data,
8079                                     struct bio *bio, u64 offset)
8080 {
8081         struct inode *inode = private_data;
8082         blk_status_t ret;
8083         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8084         BUG_ON(ret); /* -ENOMEM */
8085         return 0;
8086 }
8087
8088 static void btrfs_end_dio_bio(struct bio *bio)
8089 {
8090         struct btrfs_dio_private *dip = bio->bi_private;
8091         blk_status_t err = bio->bi_status;
8092
8093         if (err)
8094                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8095                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8096                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8097                            bio->bi_opf,
8098                            (unsigned long long)bio->bi_iter.bi_sector,
8099                            bio->bi_iter.bi_size, err);
8100
8101         if (dip->subio_endio)
8102                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8103
8104         if (err) {
8105                 /*
8106                  * We want to perceive the errors flag being set before
8107                  * decrementing the reference count. We don't need a barrier
8108                  * since atomic operations with a return value are fully
8109                  * ordered as per atomic_t.txt
8110                  */
8111                 dip->errors = 1;
8112         }
8113
8114         /* if there are more bios still pending for this dio, just exit */
8115         if (!atomic_dec_and_test(&dip->pending_bios))
8116                 goto out;
8117
8118         if (dip->errors) {
8119                 bio_io_error(dip->orig_bio);
8120         } else {
8121                 dip->dio_bio->bi_status = BLK_STS_OK;
8122                 bio_endio(dip->orig_bio);
8123         }
8124 out:
8125         bio_put(bio);
8126 }
8127
8128 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8129                                                  struct btrfs_dio_private *dip,
8130                                                  struct bio *bio,
8131                                                  u64 file_offset)
8132 {
8133         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8134         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8135         blk_status_t ret;
8136
8137         /*
8138          * We load all the csum data we need when we submit
8139          * the first bio to reduce the csum tree search and
8140          * contention.
8141          */
8142         if (dip->logical_offset == file_offset) {
8143                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8144                                                 file_offset);
8145                 if (ret)
8146                         return ret;
8147         }
8148
8149         if (bio == dip->orig_bio)
8150                 return 0;
8151
8152         file_offset -= dip->logical_offset;
8153         file_offset >>= inode->i_sb->s_blocksize_bits;
8154         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8155
8156         return 0;
8157 }
8158
8159 static inline blk_status_t btrfs_submit_dio_bio(struct bio *bio,
8160                 struct inode *inode, u64 file_offset, int async_submit)
8161 {
8162         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8163         struct btrfs_dio_private *dip = bio->bi_private;
8164         bool write = bio_op(bio) == REQ_OP_WRITE;
8165         blk_status_t ret;
8166
8167         /* Check btrfs_submit_bio_hook() for rules about async submit. */
8168         if (async_submit)
8169                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8170
8171         if (!write) {
8172                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8173                 if (ret)
8174                         goto err;
8175         }
8176
8177         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8178                 goto map;
8179
8180         if (write && async_submit) {
8181                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8182                                           file_offset, inode,
8183                                           btrfs_submit_bio_start_direct_io);
8184                 goto err;
8185         } else if (write) {
8186                 /*
8187                  * If we aren't doing async submit, calculate the csum of the
8188                  * bio now.
8189                  */
8190                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8191                 if (ret)
8192                         goto err;
8193         } else {
8194                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8195                                                      file_offset);
8196                 if (ret)
8197                         goto err;
8198         }
8199 map:
8200         ret = btrfs_map_bio(fs_info, bio, 0, 0);
8201 err:
8202         return ret;
8203 }
8204
8205 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8206 {
8207         struct inode *inode = dip->inode;
8208         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8209         struct bio *bio;
8210         struct bio *orig_bio = dip->orig_bio;
8211         u64 start_sector = orig_bio->bi_iter.bi_sector;
8212         u64 file_offset = dip->logical_offset;
8213         u64 map_length;
8214         int async_submit = 0;
8215         u64 submit_len;
8216         int clone_offset = 0;
8217         int clone_len;
8218         int ret;
8219         blk_status_t status;
8220
8221         map_length = orig_bio->bi_iter.bi_size;
8222         submit_len = map_length;
8223         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio), start_sector << 9,
8224                               &map_length, NULL, 0);
8225         if (ret)
8226                 return -EIO;
8227
8228         if (map_length >= submit_len) {
8229                 bio = orig_bio;
8230                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8231                 goto submit;
8232         }
8233
8234         /* async crcs make it difficult to collect full stripe writes. */
8235         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8236                 async_submit = 0;
8237         else
8238                 async_submit = 1;
8239
8240         /* bio split */
8241         ASSERT(map_length <= INT_MAX);
8242         atomic_inc(&dip->pending_bios);
8243         do {
8244                 clone_len = min_t(int, submit_len, map_length);
8245
8246                 /*
8247                  * This will never fail as it's passing GPF_NOFS and
8248                  * the allocation is backed by btrfs_bioset.
8249                  */
8250                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8251                                               clone_len);
8252                 bio->bi_private = dip;
8253                 bio->bi_end_io = btrfs_end_dio_bio;
8254                 btrfs_io_bio(bio)->logical = file_offset;
8255
8256                 ASSERT(submit_len >= clone_len);
8257                 submit_len -= clone_len;
8258                 if (submit_len == 0)
8259                         break;
8260
8261                 /*
8262                  * Increase the count before we submit the bio so we know
8263                  * the end IO handler won't happen before we increase the
8264                  * count. Otherwise, the dip might get freed before we're
8265                  * done setting it up.
8266                  */
8267                 atomic_inc(&dip->pending_bios);
8268
8269                 status = btrfs_submit_dio_bio(bio, inode, file_offset,
8270                                                 async_submit);
8271                 if (status) {
8272                         bio_put(bio);
8273                         atomic_dec(&dip->pending_bios);
8274                         goto out_err;
8275                 }
8276
8277                 clone_offset += clone_len;
8278                 start_sector += clone_len >> 9;
8279                 file_offset += clone_len;
8280
8281                 map_length = submit_len;
8282                 ret = btrfs_map_block(fs_info, btrfs_op(orig_bio),
8283                                       start_sector << 9, &map_length, NULL, 0);
8284                 if (ret)
8285                         goto out_err;
8286         } while (submit_len > 0);
8287
8288 submit:
8289         status = btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8290         if (!status)
8291                 return 0;
8292
8293         bio_put(bio);
8294 out_err:
8295         dip->errors = 1;
8296         /*
8297          * Before atomic variable goto zero, we must  make sure dip->errors is
8298          * perceived to be set. This ordering is ensured by the fact that an
8299          * atomic operations with a return value are fully ordered as per
8300          * atomic_t.txt
8301          */
8302         if (atomic_dec_and_test(&dip->pending_bios))
8303                 bio_io_error(dip->orig_bio);
8304
8305         /* bio_end_io() will handle error, so we needn't return it */
8306         return 0;
8307 }
8308
8309 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8310                                 loff_t file_offset)
8311 {
8312         struct btrfs_dio_private *dip = NULL;
8313         struct bio *bio = NULL;
8314         struct btrfs_io_bio *io_bio;
8315         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8316         int ret = 0;
8317
8318         bio = btrfs_bio_clone(dio_bio);
8319
8320         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8321         if (!dip) {
8322                 ret = -ENOMEM;
8323                 goto free_ordered;
8324         }
8325
8326         dip->private = dio_bio->bi_private;
8327         dip->inode = inode;
8328         dip->logical_offset = file_offset;
8329         dip->bytes = dio_bio->bi_iter.bi_size;
8330         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8331         bio->bi_private = dip;
8332         dip->orig_bio = bio;
8333         dip->dio_bio = dio_bio;
8334         atomic_set(&dip->pending_bios, 0);
8335         io_bio = btrfs_io_bio(bio);
8336         io_bio->logical = file_offset;
8337
8338         if (write) {
8339                 bio->bi_end_io = btrfs_endio_direct_write;
8340         } else {
8341                 bio->bi_end_io = btrfs_endio_direct_read;
8342                 dip->subio_endio = btrfs_subio_endio_read;
8343         }
8344
8345         /*
8346          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8347          * even if we fail to submit a bio, because in such case we do the
8348          * corresponding error handling below and it must not be done a second
8349          * time by btrfs_direct_IO().
8350          */
8351         if (write) {
8352                 struct btrfs_dio_data *dio_data = current->journal_info;
8353
8354                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8355                         dip->bytes;
8356                 dio_data->unsubmitted_oe_range_start =
8357                         dio_data->unsubmitted_oe_range_end;
8358         }
8359
8360         ret = btrfs_submit_direct_hook(dip);
8361         if (!ret)
8362                 return;
8363
8364         btrfs_io_bio_free_csum(io_bio);
8365
8366 free_ordered:
8367         /*
8368          * If we arrived here it means either we failed to submit the dip
8369          * or we either failed to clone the dio_bio or failed to allocate the
8370          * dip. If we cloned the dio_bio and allocated the dip, we can just
8371          * call bio_endio against our io_bio so that we get proper resource
8372          * cleanup if we fail to submit the dip, otherwise, we must do the
8373          * same as btrfs_endio_direct_[write|read] because we can't call these
8374          * callbacks - they require an allocated dip and a clone of dio_bio.
8375          */
8376         if (bio && dip) {
8377                 bio_io_error(bio);
8378                 /*
8379                  * The end io callbacks free our dip, do the final put on bio
8380                  * and all the cleanup and final put for dio_bio (through
8381                  * dio_end_io()).
8382                  */
8383                 dip = NULL;
8384                 bio = NULL;
8385         } else {
8386                 if (write)
8387                         __endio_write_update_ordered(inode,
8388                                                 file_offset,
8389                                                 dio_bio->bi_iter.bi_size,
8390                                                 false);
8391                 else
8392                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8393                               file_offset + dio_bio->bi_iter.bi_size - 1);
8394
8395                 dio_bio->bi_status = BLK_STS_IOERR;
8396                 /*
8397                  * Releases and cleans up our dio_bio, no need to bio_put()
8398                  * nor bio_endio()/bio_io_error() against dio_bio.
8399                  */
8400                 dio_end_io(dio_bio);
8401         }
8402         if (bio)
8403                 bio_put(bio);
8404         kfree(dip);
8405 }
8406
8407 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8408                                const struct iov_iter *iter, loff_t offset)
8409 {
8410         int seg;
8411         int i;
8412         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8413         ssize_t retval = -EINVAL;
8414
8415         if (offset & blocksize_mask)
8416                 goto out;
8417
8418         if (iov_iter_alignment(iter) & blocksize_mask)
8419                 goto out;
8420
8421         /* If this is a write we don't need to check anymore */
8422         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8423                 return 0;
8424         /*
8425          * Check to make sure we don't have duplicate iov_base's in this
8426          * iovec, if so return EINVAL, otherwise we'll get csum errors
8427          * when reading back.
8428          */
8429         for (seg = 0; seg < iter->nr_segs; seg++) {
8430                 for (i = seg + 1; i < iter->nr_segs; i++) {
8431                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8432                                 goto out;
8433                 }
8434         }
8435         retval = 0;
8436 out:
8437         return retval;
8438 }
8439
8440 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8441 {
8442         struct file *file = iocb->ki_filp;
8443         struct inode *inode = file->f_mapping->host;
8444         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8445         struct btrfs_dio_data dio_data = { 0 };
8446         struct extent_changeset *data_reserved = NULL;
8447         loff_t offset = iocb->ki_pos;
8448         size_t count = 0;
8449         int flags = 0;
8450         bool wakeup = true;
8451         bool relock = false;
8452         ssize_t ret;
8453
8454         if (check_direct_IO(fs_info, iter, offset))
8455                 return 0;
8456
8457         inode_dio_begin(inode);
8458
8459         /*
8460          * The generic stuff only does filemap_write_and_wait_range, which
8461          * isn't enough if we've written compressed pages to this area, so
8462          * we need to flush the dirty pages again to make absolutely sure
8463          * that any outstanding dirty pages are on disk.
8464          */
8465         count = iov_iter_count(iter);
8466         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8467                      &BTRFS_I(inode)->runtime_flags))
8468                 filemap_fdatawrite_range(inode->i_mapping, offset,
8469                                          offset + count - 1);
8470
8471         if (iov_iter_rw(iter) == WRITE) {
8472                 /*
8473                  * If the write DIO is beyond the EOF, we need update
8474                  * the isize, but it is protected by i_mutex. So we can
8475                  * not unlock the i_mutex at this case.
8476                  */
8477                 if (offset + count <= inode->i_size) {
8478                         dio_data.overwrite = 1;
8479                         inode_unlock(inode);
8480                         relock = true;
8481                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8482                         ret = -EAGAIN;
8483                         goto out;
8484                 }
8485                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8486                                                    offset, count);
8487                 if (ret)
8488                         goto out;
8489
8490                 /*
8491                  * We need to know how many extents we reserved so that we can
8492                  * do the accounting properly if we go over the number we
8493                  * originally calculated.  Abuse current->journal_info for this.
8494                  */
8495                 dio_data.reserve = round_up(count,
8496                                             fs_info->sectorsize);
8497                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8498                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8499                 current->journal_info = &dio_data;
8500                 down_read(&BTRFS_I(inode)->dio_sem);
8501         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8502                                      &BTRFS_I(inode)->runtime_flags)) {
8503                 inode_dio_end(inode);
8504                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8505                 wakeup = false;
8506         }
8507
8508         ret = __blockdev_direct_IO(iocb, inode,
8509                                    fs_info->fs_devices->latest_bdev,
8510                                    iter, btrfs_get_blocks_direct, NULL,
8511                                    btrfs_submit_direct, flags);
8512         if (iov_iter_rw(iter) == WRITE) {
8513                 up_read(&BTRFS_I(inode)->dio_sem);
8514                 current->journal_info = NULL;
8515                 if (ret < 0 && ret != -EIOCBQUEUED) {
8516                         if (dio_data.reserve)
8517                                 btrfs_delalloc_release_space(inode, data_reserved,
8518                                         offset, dio_data.reserve, true);
8519                         /*
8520                          * On error we might have left some ordered extents
8521                          * without submitting corresponding bios for them, so
8522                          * cleanup them up to avoid other tasks getting them
8523                          * and waiting for them to complete forever.
8524                          */
8525                         if (dio_data.unsubmitted_oe_range_start <
8526                             dio_data.unsubmitted_oe_range_end)
8527                                 __endio_write_update_ordered(inode,
8528                                         dio_data.unsubmitted_oe_range_start,
8529                                         dio_data.unsubmitted_oe_range_end -
8530                                         dio_data.unsubmitted_oe_range_start,
8531                                         false);
8532                 } else if (ret >= 0 && (size_t)ret < count)
8533                         btrfs_delalloc_release_space(inode, data_reserved,
8534                                         offset, count - (size_t)ret, true);
8535                 btrfs_delalloc_release_extents(BTRFS_I(inode), count, false);
8536         }
8537 out:
8538         if (wakeup)
8539                 inode_dio_end(inode);
8540         if (relock)
8541                 inode_lock(inode);
8542
8543         extent_changeset_free(data_reserved);
8544         return ret;
8545 }
8546
8547 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8548
8549 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8550                 __u64 start, __u64 len)
8551 {
8552         int     ret;
8553
8554         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8555         if (ret)
8556                 return ret;
8557
8558         return extent_fiemap(inode, fieinfo, start, len);
8559 }
8560
8561 int btrfs_readpage(struct file *file, struct page *page)
8562 {
8563         struct extent_io_tree *tree;
8564         tree = &BTRFS_I(page->mapping->host)->io_tree;
8565         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8566 }
8567
8568 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8569 {
8570         struct inode *inode = page->mapping->host;
8571         int ret;
8572
8573         if (current->flags & PF_MEMALLOC) {
8574                 redirty_page_for_writepage(wbc, page);
8575                 unlock_page(page);
8576                 return 0;
8577         }
8578
8579         /*
8580          * If we are under memory pressure we will call this directly from the
8581          * VM, we need to make sure we have the inode referenced for the ordered
8582          * extent.  If not just return like we didn't do anything.
8583          */
8584         if (!igrab(inode)) {
8585                 redirty_page_for_writepage(wbc, page);
8586                 return AOP_WRITEPAGE_ACTIVATE;
8587         }
8588         ret = extent_write_full_page(page, wbc);
8589         btrfs_add_delayed_iput(inode);
8590         return ret;
8591 }
8592
8593 static int btrfs_writepages(struct address_space *mapping,
8594                             struct writeback_control *wbc)
8595 {
8596         return extent_writepages(mapping, wbc);
8597 }
8598
8599 static int
8600 btrfs_readpages(struct file *file, struct address_space *mapping,
8601                 struct list_head *pages, unsigned nr_pages)
8602 {
8603         return extent_readpages(mapping, pages, nr_pages);
8604 }
8605
8606 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8607 {
8608         int ret = try_release_extent_mapping(page, gfp_flags);
8609         if (ret == 1) {
8610                 ClearPagePrivate(page);
8611                 set_page_private(page, 0);
8612                 put_page(page);
8613         }
8614         return ret;
8615 }
8616
8617 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8618 {
8619         if (PageWriteback(page) || PageDirty(page))
8620                 return 0;
8621         return __btrfs_releasepage(page, gfp_flags);
8622 }
8623
8624 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8625                                  unsigned int length)
8626 {
8627         struct inode *inode = page->mapping->host;
8628         struct extent_io_tree *tree;
8629         struct btrfs_ordered_extent *ordered;
8630         struct extent_state *cached_state = NULL;
8631         u64 page_start = page_offset(page);
8632         u64 page_end = page_start + PAGE_SIZE - 1;
8633         u64 start;
8634         u64 end;
8635         int inode_evicting = inode->i_state & I_FREEING;
8636
8637         /*
8638          * we have the page locked, so new writeback can't start,
8639          * and the dirty bit won't be cleared while we are here.
8640          *
8641          * Wait for IO on this page so that we can safely clear
8642          * the PagePrivate2 bit and do ordered accounting
8643          */
8644         wait_on_page_writeback(page);
8645
8646         tree = &BTRFS_I(inode)->io_tree;
8647         if (offset) {
8648                 btrfs_releasepage(page, GFP_NOFS);
8649                 return;
8650         }
8651
8652         if (!inode_evicting)
8653                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8654 again:
8655         start = page_start;
8656         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8657                                         page_end - start + 1);
8658         if (ordered) {
8659                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8660                 /*
8661                  * IO on this page will never be started, so we need
8662                  * to account for any ordered extents now
8663                  */
8664                 if (!inode_evicting)
8665                         clear_extent_bit(tree, start, end,
8666                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8667                                          EXTENT_DELALLOC_NEW |
8668                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8669                                          EXTENT_DEFRAG, 1, 0, &cached_state);
8670                 /*
8671                  * whoever cleared the private bit is responsible
8672                  * for the finish_ordered_io
8673                  */
8674                 if (TestClearPagePrivate2(page)) {
8675                         struct btrfs_ordered_inode_tree *tree;
8676                         u64 new_len;
8677
8678                         tree = &BTRFS_I(inode)->ordered_tree;
8679
8680                         spin_lock_irq(&tree->lock);
8681                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8682                         new_len = start - ordered->file_offset;
8683                         if (new_len < ordered->truncated_len)
8684                                 ordered->truncated_len = new_len;
8685                         spin_unlock_irq(&tree->lock);
8686
8687                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8688                                                            start,
8689                                                            end - start + 1, 1))
8690                                 btrfs_finish_ordered_io(ordered);
8691                 }
8692                 btrfs_put_ordered_extent(ordered);
8693                 if (!inode_evicting) {
8694                         cached_state = NULL;
8695                         lock_extent_bits(tree, start, end,
8696                                          &cached_state);
8697                 }
8698
8699                 start = end + 1;
8700                 if (start < page_end)
8701                         goto again;
8702         }
8703
8704         /*
8705          * Qgroup reserved space handler
8706          * Page here will be either
8707          * 1) Already written to disk
8708          *    In this case, its reserved space is released from data rsv map
8709          *    and will be freed by delayed_ref handler finally.
8710          *    So even we call qgroup_free_data(), it won't decrease reserved
8711          *    space.
8712          * 2) Not written to disk
8713          *    This means the reserved space should be freed here. However,
8714          *    if a truncate invalidates the page (by clearing PageDirty)
8715          *    and the page is accounted for while allocating extent
8716          *    in btrfs_check_data_free_space() we let delayed_ref to
8717          *    free the entire extent.
8718          */
8719         if (PageDirty(page))
8720                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
8721         if (!inode_evicting) {
8722                 clear_extent_bit(tree, page_start, page_end,
8723                                  EXTENT_LOCKED | EXTENT_DIRTY |
8724                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8725                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
8726                                  &cached_state);
8727
8728                 __btrfs_releasepage(page, GFP_NOFS);
8729         }
8730
8731         ClearPageChecked(page);
8732         if (PagePrivate(page)) {
8733                 ClearPagePrivate(page);
8734                 set_page_private(page, 0);
8735                 put_page(page);
8736         }
8737 }
8738
8739 /*
8740  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8741  * called from a page fault handler when a page is first dirtied. Hence we must
8742  * be careful to check for EOF conditions here. We set the page up correctly
8743  * for a written page which means we get ENOSPC checking when writing into
8744  * holes and correct delalloc and unwritten extent mapping on filesystems that
8745  * support these features.
8746  *
8747  * We are not allowed to take the i_mutex here so we have to play games to
8748  * protect against truncate races as the page could now be beyond EOF.  Because
8749  * truncate_setsize() writes the inode size before removing pages, once we have
8750  * the page lock we can determine safely if the page is beyond EOF. If it is not
8751  * beyond EOF, then the page is guaranteed safe against truncation until we
8752  * unlock the page.
8753  */
8754 vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
8755 {
8756         struct page *page = vmf->page;
8757         struct inode *inode = file_inode(vmf->vma->vm_file);
8758         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8759         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8760         struct btrfs_ordered_extent *ordered;
8761         struct extent_state *cached_state = NULL;
8762         struct extent_changeset *data_reserved = NULL;
8763         char *kaddr;
8764         unsigned long zero_start;
8765         loff_t size;
8766         vm_fault_t ret;
8767         int ret2;
8768         int reserved = 0;
8769         u64 reserved_space;
8770         u64 page_start;
8771         u64 page_end;
8772         u64 end;
8773
8774         reserved_space = PAGE_SIZE;
8775
8776         sb_start_pagefault(inode->i_sb);
8777         page_start = page_offset(page);
8778         page_end = page_start + PAGE_SIZE - 1;
8779         end = page_end;
8780
8781         /*
8782          * Reserving delalloc space after obtaining the page lock can lead to
8783          * deadlock. For example, if a dirty page is locked by this function
8784          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8785          * dirty page write out, then the btrfs_writepage() function could
8786          * end up waiting indefinitely to get a lock on the page currently
8787          * being processed by btrfs_page_mkwrite() function.
8788          */
8789         ret2 = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
8790                                            reserved_space);
8791         if (!ret2) {
8792                 ret2 = file_update_time(vmf->vma->vm_file);
8793                 reserved = 1;
8794         }
8795         if (ret2) {
8796                 ret = vmf_error(ret2);
8797                 if (reserved)
8798                         goto out;
8799                 goto out_noreserve;
8800         }
8801
8802         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8803 again:
8804         lock_page(page);
8805         size = i_size_read(inode);
8806
8807         if ((page->mapping != inode->i_mapping) ||
8808             (page_start >= size)) {
8809                 /* page got truncated out from underneath us */
8810                 goto out_unlock;
8811         }
8812         wait_on_page_writeback(page);
8813
8814         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8815         set_page_extent_mapped(page);
8816
8817         /*
8818          * we can't set the delalloc bits if there are pending ordered
8819          * extents.  Drop our locks and wait for them to finish
8820          */
8821         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
8822                         PAGE_SIZE);
8823         if (ordered) {
8824                 unlock_extent_cached(io_tree, page_start, page_end,
8825                                      &cached_state);
8826                 unlock_page(page);
8827                 btrfs_start_ordered_extent(inode, ordered, 1);
8828                 btrfs_put_ordered_extent(ordered);
8829                 goto again;
8830         }
8831
8832         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
8833                 reserved_space = round_up(size - page_start,
8834                                           fs_info->sectorsize);
8835                 if (reserved_space < PAGE_SIZE) {
8836                         end = page_start + reserved_space - 1;
8837                         btrfs_delalloc_release_space(inode, data_reserved,
8838                                         page_start, PAGE_SIZE - reserved_space,
8839                                         true);
8840                 }
8841         }
8842
8843         /*
8844          * page_mkwrite gets called when the page is firstly dirtied after it's
8845          * faulted in, but write(2) could also dirty a page and set delalloc
8846          * bits, thus in this case for space account reason, we still need to
8847          * clear any delalloc bits within this page range since we have to
8848          * reserve data&meta space before lock_page() (see above comments).
8849          */
8850         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
8851                           EXTENT_DIRTY | EXTENT_DELALLOC |
8852                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
8853                           0, 0, &cached_state);
8854
8855         ret2 = btrfs_set_extent_delalloc(inode, page_start, end, 0,
8856                                         &cached_state, 0);
8857         if (ret2) {
8858                 unlock_extent_cached(io_tree, page_start, page_end,
8859                                      &cached_state);
8860                 ret = VM_FAULT_SIGBUS;
8861                 goto out_unlock;
8862         }
8863         ret2 = 0;
8864
8865         /* page is wholly or partially inside EOF */
8866         if (page_start + PAGE_SIZE > size)
8867                 zero_start = offset_in_page(size);
8868         else
8869                 zero_start = PAGE_SIZE;
8870
8871         if (zero_start != PAGE_SIZE) {
8872                 kaddr = kmap(page);
8873                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
8874                 flush_dcache_page(page);
8875                 kunmap(page);
8876         }
8877         ClearPageChecked(page);
8878         set_page_dirty(page);
8879         SetPageUptodate(page);
8880
8881         BTRFS_I(inode)->last_trans = fs_info->generation;
8882         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
8883         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
8884
8885         unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
8886
8887         if (!ret2) {
8888                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, true);
8889                 sb_end_pagefault(inode->i_sb);
8890                 extent_changeset_free(data_reserved);
8891                 return VM_FAULT_LOCKED;
8892         }
8893
8894 out_unlock:
8895         unlock_page(page);
8896 out:
8897         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, (ret != 0));
8898         btrfs_delalloc_release_space(inode, data_reserved, page_start,
8899                                      reserved_space, (ret != 0));
8900 out_noreserve:
8901         sb_end_pagefault(inode->i_sb);
8902         extent_changeset_free(data_reserved);
8903         return ret;
8904 }
8905
8906 static int btrfs_truncate(struct inode *inode, bool skip_writeback)
8907 {
8908         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8909         struct btrfs_root *root = BTRFS_I(inode)->root;
8910         struct btrfs_block_rsv *rsv;
8911         int ret;
8912         struct btrfs_trans_handle *trans;
8913         u64 mask = fs_info->sectorsize - 1;
8914         u64 min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
8915
8916         if (!skip_writeback) {
8917                 ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
8918                                                (u64)-1);
8919                 if (ret)
8920                         return ret;
8921         }
8922
8923         /*
8924          * Yes ladies and gentlemen, this is indeed ugly.  We have a couple of
8925          * things going on here:
8926          *
8927          * 1) We need to reserve space to update our inode.
8928          *
8929          * 2) We need to have something to cache all the space that is going to
8930          * be free'd up by the truncate operation, but also have some slack
8931          * space reserved in case it uses space during the truncate (thank you
8932          * very much snapshotting).
8933          *
8934          * And we need these to be separate.  The fact is we can use a lot of
8935          * space doing the truncate, and we have no earthly idea how much space
8936          * we will use, so we need the truncate reservation to be separate so it
8937          * doesn't end up using space reserved for updating the inode.  We also
8938          * need to be able to stop the transaction and start a new one, which
8939          * means we need to be able to update the inode several times, and we
8940          * have no idea of knowing how many times that will be, so we can't just
8941          * reserve 1 item for the entirety of the operation, so that has to be
8942          * done separately as well.
8943          *
8944          * So that leaves us with
8945          *
8946          * 1) rsv - for the truncate reservation, which we will steal from the
8947          * transaction reservation.
8948          * 2) fs_info->trans_block_rsv - this will have 1 items worth left for
8949          * updating the inode.
8950          */
8951         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
8952         if (!rsv)
8953                 return -ENOMEM;
8954         rsv->size = min_size;
8955         rsv->failfast = 1;
8956
8957         /*
8958          * 1 for the truncate slack space
8959          * 1 for updating the inode.
8960          */
8961         trans = btrfs_start_transaction(root, 2);
8962         if (IS_ERR(trans)) {
8963                 ret = PTR_ERR(trans);
8964                 goto out;
8965         }
8966
8967         /* Migrate the slack space for the truncate to our reserve */
8968         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
8969                                       min_size, false);
8970         BUG_ON(ret);
8971
8972         /*
8973          * So if we truncate and then write and fsync we normally would just
8974          * write the extents that changed, which is a problem if we need to
8975          * first truncate that entire inode.  So set this flag so we write out
8976          * all of the extents in the inode to the sync log so we're completely
8977          * safe.
8978          */
8979         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
8980         trans->block_rsv = rsv;
8981
8982         while (1) {
8983                 ret = btrfs_truncate_inode_items(trans, root, inode,
8984                                                  inode->i_size,
8985                                                  BTRFS_EXTENT_DATA_KEY);
8986                 trans->block_rsv = &fs_info->trans_block_rsv;
8987                 if (ret != -ENOSPC && ret != -EAGAIN)
8988                         break;
8989
8990                 ret = btrfs_update_inode(trans, root, inode);
8991                 if (ret)
8992                         break;
8993
8994                 btrfs_end_transaction(trans);
8995                 btrfs_btree_balance_dirty(fs_info);
8996
8997                 trans = btrfs_start_transaction(root, 2);
8998                 if (IS_ERR(trans)) {
8999                         ret = PTR_ERR(trans);
9000                         trans = NULL;
9001                         break;
9002                 }
9003
9004                 btrfs_block_rsv_release(fs_info, rsv, -1);
9005                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9006                                               rsv, min_size, false);
9007                 BUG_ON(ret);    /* shouldn't happen */
9008                 trans->block_rsv = rsv;
9009         }
9010
9011         /*
9012          * We can't call btrfs_truncate_block inside a trans handle as we could
9013          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
9014          * we've truncated everything except the last little bit, and can do
9015          * btrfs_truncate_block and then update the disk_i_size.
9016          */
9017         if (ret == NEED_TRUNCATE_BLOCK) {
9018                 btrfs_end_transaction(trans);
9019                 btrfs_btree_balance_dirty(fs_info);
9020
9021                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
9022                 if (ret)
9023                         goto out;
9024                 trans = btrfs_start_transaction(root, 1);
9025                 if (IS_ERR(trans)) {
9026                         ret = PTR_ERR(trans);
9027                         goto out;
9028                 }
9029                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
9030         }
9031
9032         if (trans) {
9033                 int ret2;
9034
9035                 trans->block_rsv = &fs_info->trans_block_rsv;
9036                 ret2 = btrfs_update_inode(trans, root, inode);
9037                 if (ret2 && !ret)
9038                         ret = ret2;
9039
9040                 ret2 = btrfs_end_transaction(trans);
9041                 if (ret2 && !ret)
9042                         ret = ret2;
9043                 btrfs_btree_balance_dirty(fs_info);
9044         }
9045 out:
9046         btrfs_free_block_rsv(fs_info, rsv);
9047
9048         return ret;
9049 }
9050
9051 /*
9052  * create a new subvolume directory/inode (helper for the ioctl).
9053  */
9054 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9055                              struct btrfs_root *new_root,
9056                              struct btrfs_root *parent_root,
9057                              u64 new_dirid)
9058 {
9059         struct inode *inode;
9060         int err;
9061         u64 index = 0;
9062
9063         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9064                                 new_dirid, new_dirid,
9065                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9066                                 &index);
9067         if (IS_ERR(inode))
9068                 return PTR_ERR(inode);
9069         inode->i_op = &btrfs_dir_inode_operations;
9070         inode->i_fop = &btrfs_dir_file_operations;
9071
9072         set_nlink(inode, 1);
9073         btrfs_i_size_write(BTRFS_I(inode), 0);
9074         unlock_new_inode(inode);
9075
9076         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9077         if (err)
9078                 btrfs_err(new_root->fs_info,
9079                           "error inheriting subvolume %llu properties: %d",
9080                           new_root->root_key.objectid, err);
9081
9082         err = btrfs_update_inode(trans, new_root, inode);
9083
9084         iput(inode);
9085         return err;
9086 }
9087
9088 struct inode *btrfs_alloc_inode(struct super_block *sb)
9089 {
9090         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
9091         struct btrfs_inode *ei;
9092         struct inode *inode;
9093
9094         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_KERNEL);
9095         if (!ei)
9096                 return NULL;
9097
9098         ei->root = NULL;
9099         ei->generation = 0;
9100         ei->last_trans = 0;
9101         ei->last_sub_trans = 0;
9102         ei->logged_trans = 0;
9103         ei->delalloc_bytes = 0;
9104         ei->new_delalloc_bytes = 0;
9105         ei->defrag_bytes = 0;
9106         ei->disk_i_size = 0;
9107         ei->flags = 0;
9108         ei->csum_bytes = 0;
9109         ei->index_cnt = (u64)-1;
9110         ei->dir_index = 0;
9111         ei->last_unlink_trans = 0;
9112         ei->last_link_trans = 0;
9113         ei->last_log_commit = 0;
9114
9115         spin_lock_init(&ei->lock);
9116         ei->outstanding_extents = 0;
9117         if (sb->s_magic != BTRFS_TEST_MAGIC)
9118                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
9119                                               BTRFS_BLOCK_RSV_DELALLOC);
9120         ei->runtime_flags = 0;
9121         ei->prop_compress = BTRFS_COMPRESS_NONE;
9122         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9123
9124         ei->delayed_node = NULL;
9125
9126         ei->i_otime.tv_sec = 0;
9127         ei->i_otime.tv_nsec = 0;
9128
9129         inode = &ei->vfs_inode;
9130         extent_map_tree_init(&ei->extent_tree);
9131         extent_io_tree_init(&ei->io_tree, inode);
9132         extent_io_tree_init(&ei->io_failure_tree, inode);
9133         ei->io_tree.track_uptodate = 1;
9134         ei->io_failure_tree.track_uptodate = 1;
9135         atomic_set(&ei->sync_writers, 0);
9136         mutex_init(&ei->log_mutex);
9137         mutex_init(&ei->delalloc_mutex);
9138         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9139         INIT_LIST_HEAD(&ei->delalloc_inodes);
9140         INIT_LIST_HEAD(&ei->delayed_iput);
9141         RB_CLEAR_NODE(&ei->rb_node);
9142         init_rwsem(&ei->dio_sem);
9143
9144         return inode;
9145 }
9146
9147 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9148 void btrfs_test_destroy_inode(struct inode *inode)
9149 {
9150         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9151         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9152 }
9153 #endif
9154
9155 static void btrfs_i_callback(struct rcu_head *head)
9156 {
9157         struct inode *inode = container_of(head, struct inode, i_rcu);
9158         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9159 }
9160
9161 void btrfs_destroy_inode(struct inode *inode)
9162 {
9163         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9164         struct btrfs_ordered_extent *ordered;
9165         struct btrfs_root *root = BTRFS_I(inode)->root;
9166
9167         WARN_ON(!hlist_empty(&inode->i_dentry));
9168         WARN_ON(inode->i_data.nrpages);
9169         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
9170         WARN_ON(BTRFS_I(inode)->block_rsv.size);
9171         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9172         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9173         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9174         WARN_ON(BTRFS_I(inode)->csum_bytes);
9175         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9176
9177         /*
9178          * This can happen where we create an inode, but somebody else also
9179          * created the same inode and we need to destroy the one we already
9180          * created.
9181          */
9182         if (!root)
9183                 goto free;
9184
9185         while (1) {
9186                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9187                 if (!ordered)
9188                         break;
9189                 else {
9190                         btrfs_err(fs_info,
9191                                   "found ordered extent %llu %llu on inode cleanup",
9192                                   ordered->file_offset, ordered->len);
9193                         btrfs_remove_ordered_extent(inode, ordered);
9194                         btrfs_put_ordered_extent(ordered);
9195                         btrfs_put_ordered_extent(ordered);
9196                 }
9197         }
9198         btrfs_qgroup_check_reserved_leak(inode);
9199         inode_tree_del(inode);
9200         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9201 free:
9202         call_rcu(&inode->i_rcu, btrfs_i_callback);
9203 }
9204
9205 int btrfs_drop_inode(struct inode *inode)
9206 {
9207         struct btrfs_root *root = BTRFS_I(inode)->root;
9208
9209         if (root == NULL)
9210                 return 1;
9211
9212         /* the snap/subvol tree is on deleting */
9213         if (btrfs_root_refs(&root->root_item) == 0)
9214                 return 1;
9215         else
9216                 return generic_drop_inode(inode);
9217 }
9218
9219 static void init_once(void *foo)
9220 {
9221         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9222
9223         inode_init_once(&ei->vfs_inode);
9224 }
9225
9226 void __cold btrfs_destroy_cachep(void)
9227 {
9228         /*
9229          * Make sure all delayed rcu free inodes are flushed before we
9230          * destroy cache.
9231          */
9232         rcu_barrier();
9233         kmem_cache_destroy(btrfs_inode_cachep);
9234         kmem_cache_destroy(btrfs_trans_handle_cachep);
9235         kmem_cache_destroy(btrfs_path_cachep);
9236         kmem_cache_destroy(btrfs_free_space_cachep);
9237 }
9238
9239 int __init btrfs_init_cachep(void)
9240 {
9241         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9242                         sizeof(struct btrfs_inode), 0,
9243                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9244                         init_once);
9245         if (!btrfs_inode_cachep)
9246                 goto fail;
9247
9248         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9249                         sizeof(struct btrfs_trans_handle), 0,
9250                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9251         if (!btrfs_trans_handle_cachep)
9252                 goto fail;
9253
9254         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9255                         sizeof(struct btrfs_path), 0,
9256                         SLAB_MEM_SPREAD, NULL);
9257         if (!btrfs_path_cachep)
9258                 goto fail;
9259
9260         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9261                         sizeof(struct btrfs_free_space), 0,
9262                         SLAB_MEM_SPREAD, NULL);
9263         if (!btrfs_free_space_cachep)
9264                 goto fail;
9265
9266         return 0;
9267 fail:
9268         btrfs_destroy_cachep();
9269         return -ENOMEM;
9270 }
9271
9272 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9273                          u32 request_mask, unsigned int flags)
9274 {
9275         u64 delalloc_bytes;
9276         struct inode *inode = d_inode(path->dentry);
9277         u32 blocksize = inode->i_sb->s_blocksize;
9278         u32 bi_flags = BTRFS_I(inode)->flags;
9279
9280         stat->result_mask |= STATX_BTIME;
9281         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9282         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9283         if (bi_flags & BTRFS_INODE_APPEND)
9284                 stat->attributes |= STATX_ATTR_APPEND;
9285         if (bi_flags & BTRFS_INODE_COMPRESS)
9286                 stat->attributes |= STATX_ATTR_COMPRESSED;
9287         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9288                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9289         if (bi_flags & BTRFS_INODE_NODUMP)
9290                 stat->attributes |= STATX_ATTR_NODUMP;
9291
9292         stat->attributes_mask |= (STATX_ATTR_APPEND |
9293                                   STATX_ATTR_COMPRESSED |
9294                                   STATX_ATTR_IMMUTABLE |
9295                                   STATX_ATTR_NODUMP);
9296
9297         generic_fillattr(inode, stat);
9298         stat->dev = BTRFS_I(inode)->root->anon_dev;
9299
9300         spin_lock(&BTRFS_I(inode)->lock);
9301         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9302         spin_unlock(&BTRFS_I(inode)->lock);
9303         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9304                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9305         return 0;
9306 }
9307
9308 static int btrfs_rename_exchange(struct inode *old_dir,
9309                               struct dentry *old_dentry,
9310                               struct inode *new_dir,
9311                               struct dentry *new_dentry)
9312 {
9313         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9314         struct btrfs_trans_handle *trans;
9315         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9316         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9317         struct inode *new_inode = new_dentry->d_inode;
9318         struct inode *old_inode = old_dentry->d_inode;
9319         struct timespec64 ctime = current_time(old_inode);
9320         struct dentry *parent;
9321         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9322         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9323         u64 old_idx = 0;
9324         u64 new_idx = 0;
9325         u64 root_objectid;
9326         int ret;
9327         bool root_log_pinned = false;
9328         bool dest_log_pinned = false;
9329         struct btrfs_log_ctx ctx_root;
9330         struct btrfs_log_ctx ctx_dest;
9331         bool sync_log_root = false;
9332         bool sync_log_dest = false;
9333         bool commit_transaction = false;
9334
9335         /* we only allow rename subvolume link between subvolumes */
9336         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9337                 return -EXDEV;
9338
9339         btrfs_init_log_ctx(&ctx_root, old_inode);
9340         btrfs_init_log_ctx(&ctx_dest, new_inode);
9341
9342         /* close the race window with snapshot create/destroy ioctl */
9343         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9344                 down_read(&fs_info->subvol_sem);
9345         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9346                 down_read(&fs_info->subvol_sem);
9347
9348         /*
9349          * We want to reserve the absolute worst case amount of items.  So if
9350          * both inodes are subvols and we need to unlink them then that would
9351          * require 4 item modifications, but if they are both normal inodes it
9352          * would require 5 item modifications, so we'll assume their normal
9353          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9354          * should cover the worst case number of items we'll modify.
9355          */
9356         trans = btrfs_start_transaction(root, 12);
9357         if (IS_ERR(trans)) {
9358                 ret = PTR_ERR(trans);
9359                 goto out_notrans;
9360         }
9361
9362         /*
9363          * We need to find a free sequence number both in the source and
9364          * in the destination directory for the exchange.
9365          */
9366         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9367         if (ret)
9368                 goto out_fail;
9369         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9370         if (ret)
9371                 goto out_fail;
9372
9373         BTRFS_I(old_inode)->dir_index = 0ULL;
9374         BTRFS_I(new_inode)->dir_index = 0ULL;
9375
9376         /* Reference for the source. */
9377         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9378                 /* force full log commit if subvolume involved. */
9379                 btrfs_set_log_full_commit(fs_info, trans);
9380         } else {
9381                 btrfs_pin_log_trans(root);
9382                 root_log_pinned = true;
9383                 ret = btrfs_insert_inode_ref(trans, dest,
9384                                              new_dentry->d_name.name,
9385                                              new_dentry->d_name.len,
9386                                              old_ino,
9387                                              btrfs_ino(BTRFS_I(new_dir)),
9388                                              old_idx);
9389                 if (ret)
9390                         goto out_fail;
9391         }
9392
9393         /* And now for the dest. */
9394         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9395                 /* force full log commit if subvolume involved. */
9396                 btrfs_set_log_full_commit(fs_info, trans);
9397         } else {
9398                 btrfs_pin_log_trans(dest);
9399                 dest_log_pinned = true;
9400                 ret = btrfs_insert_inode_ref(trans, root,
9401                                              old_dentry->d_name.name,
9402                                              old_dentry->d_name.len,
9403                                              new_ino,
9404                                              btrfs_ino(BTRFS_I(old_dir)),
9405                                              new_idx);
9406                 if (ret)
9407                         goto out_fail;
9408         }
9409
9410         /* Update inode version and ctime/mtime. */
9411         inode_inc_iversion(old_dir);
9412         inode_inc_iversion(new_dir);
9413         inode_inc_iversion(old_inode);
9414         inode_inc_iversion(new_inode);
9415         old_dir->i_ctime = old_dir->i_mtime = ctime;
9416         new_dir->i_ctime = new_dir->i_mtime = ctime;
9417         old_inode->i_ctime = ctime;
9418         new_inode->i_ctime = ctime;
9419
9420         if (old_dentry->d_parent != new_dentry->d_parent) {
9421                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9422                                 BTRFS_I(old_inode), 1);
9423                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9424                                 BTRFS_I(new_inode), 1);
9425         }
9426
9427         /* src is a subvolume */
9428         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9429                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9430                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9431                                           old_dentry->d_name.name,
9432                                           old_dentry->d_name.len);
9433         } else { /* src is an inode */
9434                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9435                                            BTRFS_I(old_dentry->d_inode),
9436                                            old_dentry->d_name.name,
9437                                            old_dentry->d_name.len);
9438                 if (!ret)
9439                         ret = btrfs_update_inode(trans, root, old_inode);
9440         }
9441         if (ret) {
9442                 btrfs_abort_transaction(trans, ret);
9443                 goto out_fail;
9444         }
9445
9446         /* dest is a subvolume */
9447         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9448                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9449                 ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
9450                                           new_dentry->d_name.name,
9451                                           new_dentry->d_name.len);
9452         } else { /* dest is an inode */
9453                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9454                                            BTRFS_I(new_dentry->d_inode),
9455                                            new_dentry->d_name.name,
9456                                            new_dentry->d_name.len);
9457                 if (!ret)
9458                         ret = btrfs_update_inode(trans, dest, new_inode);
9459         }
9460         if (ret) {
9461                 btrfs_abort_transaction(trans, ret);
9462                 goto out_fail;
9463         }
9464
9465         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9466                              new_dentry->d_name.name,
9467                              new_dentry->d_name.len, 0, old_idx);
9468         if (ret) {
9469                 btrfs_abort_transaction(trans, ret);
9470                 goto out_fail;
9471         }
9472
9473         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9474                              old_dentry->d_name.name,
9475                              old_dentry->d_name.len, 0, new_idx);
9476         if (ret) {
9477                 btrfs_abort_transaction(trans, ret);
9478                 goto out_fail;
9479         }
9480
9481         if (old_inode->i_nlink == 1)
9482                 BTRFS_I(old_inode)->dir_index = old_idx;
9483         if (new_inode->i_nlink == 1)
9484                 BTRFS_I(new_inode)->dir_index = new_idx;
9485
9486         if (root_log_pinned) {
9487                 parent = new_dentry->d_parent;
9488                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
9489                                          BTRFS_I(old_dir), parent,
9490                                          false, &ctx_root);
9491                 if (ret == BTRFS_NEED_LOG_SYNC)
9492                         sync_log_root = true;
9493                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
9494                         commit_transaction = true;
9495                 ret = 0;
9496                 btrfs_end_log_trans(root);
9497                 root_log_pinned = false;
9498         }
9499         if (dest_log_pinned) {
9500                 if (!commit_transaction) {
9501                         parent = old_dentry->d_parent;
9502                         ret = btrfs_log_new_name(trans, BTRFS_I(new_inode),
9503                                                  BTRFS_I(new_dir), parent,
9504                                                  false, &ctx_dest);
9505                         if (ret == BTRFS_NEED_LOG_SYNC)
9506                                 sync_log_dest = true;
9507                         else if (ret == BTRFS_NEED_TRANS_COMMIT)
9508                                 commit_transaction = true;
9509                         ret = 0;
9510                 }
9511                 btrfs_end_log_trans(dest);
9512                 dest_log_pinned = false;
9513         }
9514 out_fail:
9515         /*
9516          * If we have pinned a log and an error happened, we unpin tasks
9517          * trying to sync the log and force them to fallback to a transaction
9518          * commit if the log currently contains any of the inodes involved in
9519          * this rename operation (to ensure we do not persist a log with an
9520          * inconsistent state for any of these inodes or leading to any
9521          * inconsistencies when replayed). If the transaction was aborted, the
9522          * abortion reason is propagated to userspace when attempting to commit
9523          * the transaction. If the log does not contain any of these inodes, we
9524          * allow the tasks to sync it.
9525          */
9526         if (ret && (root_log_pinned || dest_log_pinned)) {
9527                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9528                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9529                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9530                     (new_inode &&
9531                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9532                         btrfs_set_log_full_commit(fs_info, trans);
9533
9534                 if (root_log_pinned) {
9535                         btrfs_end_log_trans(root);
9536                         root_log_pinned = false;
9537                 }
9538                 if (dest_log_pinned) {
9539                         btrfs_end_log_trans(dest);
9540                         dest_log_pinned = false;
9541                 }
9542         }
9543         if (!ret && sync_log_root && !commit_transaction) {
9544                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root,
9545                                      &ctx_root);
9546                 if (ret)
9547                         commit_transaction = true;
9548         }
9549         if (!ret && sync_log_dest && !commit_transaction) {
9550                 ret = btrfs_sync_log(trans, BTRFS_I(new_inode)->root,
9551                                      &ctx_dest);
9552                 if (ret)
9553                         commit_transaction = true;
9554         }
9555         if (commit_transaction) {
9556                 ret = btrfs_commit_transaction(trans);
9557         } else {
9558                 int ret2;
9559
9560                 ret2 = btrfs_end_transaction(trans);
9561                 ret = ret ? ret : ret2;
9562         }
9563 out_notrans:
9564         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9565                 up_read(&fs_info->subvol_sem);
9566         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9567                 up_read(&fs_info->subvol_sem);
9568
9569         return ret;
9570 }
9571
9572 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9573                                      struct btrfs_root *root,
9574                                      struct inode *dir,
9575                                      struct dentry *dentry)
9576 {
9577         int ret;
9578         struct inode *inode;
9579         u64 objectid;
9580         u64 index;
9581
9582         ret = btrfs_find_free_ino(root, &objectid);
9583         if (ret)
9584                 return ret;
9585
9586         inode = btrfs_new_inode(trans, root, dir,
9587                                 dentry->d_name.name,
9588                                 dentry->d_name.len,
9589                                 btrfs_ino(BTRFS_I(dir)),
9590                                 objectid,
9591                                 S_IFCHR | WHITEOUT_MODE,
9592                                 &index);
9593
9594         if (IS_ERR(inode)) {
9595                 ret = PTR_ERR(inode);
9596                 return ret;
9597         }
9598
9599         inode->i_op = &btrfs_special_inode_operations;
9600         init_special_inode(inode, inode->i_mode,
9601                 WHITEOUT_DEV);
9602
9603         ret = btrfs_init_inode_security(trans, inode, dir,
9604                                 &dentry->d_name);
9605         if (ret)
9606                 goto out;
9607
9608         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9609                                 BTRFS_I(inode), 0, index);
9610         if (ret)
9611                 goto out;
9612
9613         ret = btrfs_update_inode(trans, root, inode);
9614 out:
9615         unlock_new_inode(inode);
9616         if (ret)
9617                 inode_dec_link_count(inode);
9618         iput(inode);
9619
9620         return ret;
9621 }
9622
9623 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9624                            struct inode *new_dir, struct dentry *new_dentry,
9625                            unsigned int flags)
9626 {
9627         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9628         struct btrfs_trans_handle *trans;
9629         unsigned int trans_num_items;
9630         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9631         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9632         struct inode *new_inode = d_inode(new_dentry);
9633         struct inode *old_inode = d_inode(old_dentry);
9634         u64 index = 0;
9635         u64 root_objectid;
9636         int ret;
9637         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9638         bool log_pinned = false;
9639         struct btrfs_log_ctx ctx;
9640         bool sync_log = false;
9641         bool commit_transaction = false;
9642
9643         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9644                 return -EPERM;
9645
9646         /* we only allow rename subvolume link between subvolumes */
9647         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9648                 return -EXDEV;
9649
9650         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9651             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9652                 return -ENOTEMPTY;
9653
9654         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9655             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9656                 return -ENOTEMPTY;
9657
9658
9659         /* check for collisions, even if the  name isn't there */
9660         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9661                              new_dentry->d_name.name,
9662                              new_dentry->d_name.len);
9663
9664         if (ret) {
9665                 if (ret == -EEXIST) {
9666                         /* we shouldn't get
9667                          * eexist without a new_inode */
9668                         if (WARN_ON(!new_inode)) {
9669                                 return ret;
9670                         }
9671                 } else {
9672                         /* maybe -EOVERFLOW */
9673                         return ret;
9674                 }
9675         }
9676         ret = 0;
9677
9678         /*
9679          * we're using rename to replace one file with another.  Start IO on it
9680          * now so  we don't add too much work to the end of the transaction
9681          */
9682         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9683                 filemap_flush(old_inode->i_mapping);
9684
9685         /* close the racy window with snapshot create/destroy ioctl */
9686         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9687                 down_read(&fs_info->subvol_sem);
9688         /*
9689          * We want to reserve the absolute worst case amount of items.  So if
9690          * both inodes are subvols and we need to unlink them then that would
9691          * require 4 item modifications, but if they are both normal inodes it
9692          * would require 5 item modifications, so we'll assume they are normal
9693          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9694          * should cover the worst case number of items we'll modify.
9695          * If our rename has the whiteout flag, we need more 5 units for the
9696          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9697          * when selinux is enabled).
9698          */
9699         trans_num_items = 11;
9700         if (flags & RENAME_WHITEOUT)
9701                 trans_num_items += 5;
9702         trans = btrfs_start_transaction(root, trans_num_items);
9703         if (IS_ERR(trans)) {
9704                 ret = PTR_ERR(trans);
9705                 goto out_notrans;
9706         }
9707
9708         if (dest != root)
9709                 btrfs_record_root_in_trans(trans, dest);
9710
9711         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9712         if (ret)
9713                 goto out_fail;
9714
9715         BTRFS_I(old_inode)->dir_index = 0ULL;
9716         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9717                 /* force full log commit if subvolume involved. */
9718                 btrfs_set_log_full_commit(fs_info, trans);
9719         } else {
9720                 btrfs_pin_log_trans(root);
9721                 log_pinned = true;
9722                 ret = btrfs_insert_inode_ref(trans, dest,
9723                                              new_dentry->d_name.name,
9724                                              new_dentry->d_name.len,
9725                                              old_ino,
9726                                              btrfs_ino(BTRFS_I(new_dir)), index);
9727                 if (ret)
9728                         goto out_fail;
9729         }
9730
9731         inode_inc_iversion(old_dir);
9732         inode_inc_iversion(new_dir);
9733         inode_inc_iversion(old_inode);
9734         old_dir->i_ctime = old_dir->i_mtime =
9735         new_dir->i_ctime = new_dir->i_mtime =
9736         old_inode->i_ctime = current_time(old_dir);
9737
9738         if (old_dentry->d_parent != new_dentry->d_parent)
9739                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9740                                 BTRFS_I(old_inode), 1);
9741
9742         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9743                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9744                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9745                                         old_dentry->d_name.name,
9746                                         old_dentry->d_name.len);
9747         } else {
9748                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9749                                         BTRFS_I(d_inode(old_dentry)),
9750                                         old_dentry->d_name.name,
9751                                         old_dentry->d_name.len);
9752                 if (!ret)
9753                         ret = btrfs_update_inode(trans, root, old_inode);
9754         }
9755         if (ret) {
9756                 btrfs_abort_transaction(trans, ret);
9757                 goto out_fail;
9758         }
9759
9760         if (new_inode) {
9761                 inode_inc_iversion(new_inode);
9762                 new_inode->i_ctime = current_time(new_inode);
9763                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
9764                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9765                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9766                         ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
9767                                                 new_dentry->d_name.name,
9768                                                 new_dentry->d_name.len);
9769                         BUG_ON(new_inode->i_nlink == 0);
9770                 } else {
9771                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9772                                                  BTRFS_I(d_inode(new_dentry)),
9773                                                  new_dentry->d_name.name,
9774                                                  new_dentry->d_name.len);
9775                 }
9776                 if (!ret && new_inode->i_nlink == 0)
9777                         ret = btrfs_orphan_add(trans,
9778                                         BTRFS_I(d_inode(new_dentry)));
9779                 if (ret) {
9780                         btrfs_abort_transaction(trans, ret);
9781                         goto out_fail;
9782                 }
9783         }
9784
9785         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9786                              new_dentry->d_name.name,
9787                              new_dentry->d_name.len, 0, index);
9788         if (ret) {
9789                 btrfs_abort_transaction(trans, ret);
9790                 goto out_fail;
9791         }
9792
9793         if (old_inode->i_nlink == 1)
9794                 BTRFS_I(old_inode)->dir_index = index;
9795
9796         if (log_pinned) {
9797                 struct dentry *parent = new_dentry->d_parent;
9798
9799                 btrfs_init_log_ctx(&ctx, old_inode);
9800                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
9801                                          BTRFS_I(old_dir), parent,
9802                                          false, &ctx);
9803                 if (ret == BTRFS_NEED_LOG_SYNC)
9804                         sync_log = true;
9805                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
9806                         commit_transaction = true;
9807                 ret = 0;
9808                 btrfs_end_log_trans(root);
9809                 log_pinned = false;
9810         }
9811
9812         if (flags & RENAME_WHITEOUT) {
9813                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9814                                                 old_dentry);
9815
9816                 if (ret) {
9817                         btrfs_abort_transaction(trans, ret);
9818                         goto out_fail;
9819                 }
9820         }
9821 out_fail:
9822         /*
9823          * If we have pinned the log and an error happened, we unpin tasks
9824          * trying to sync the log and force them to fallback to a transaction
9825          * commit if the log currently contains any of the inodes involved in
9826          * this rename operation (to ensure we do not persist a log with an
9827          * inconsistent state for any of these inodes or leading to any
9828          * inconsistencies when replayed). If the transaction was aborted, the
9829          * abortion reason is propagated to userspace when attempting to commit
9830          * the transaction. If the log does not contain any of these inodes, we
9831          * allow the tasks to sync it.
9832          */
9833         if (ret && log_pinned) {
9834                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9835                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9836                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9837                     (new_inode &&
9838                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9839                         btrfs_set_log_full_commit(fs_info, trans);
9840
9841                 btrfs_end_log_trans(root);
9842                 log_pinned = false;
9843         }
9844         if (!ret && sync_log) {
9845                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root, &ctx);
9846                 if (ret)
9847                         commit_transaction = true;
9848         }
9849         if (commit_transaction) {
9850                 ret = btrfs_commit_transaction(trans);
9851         } else {
9852                 int ret2;
9853
9854                 ret2 = btrfs_end_transaction(trans);
9855                 ret = ret ? ret : ret2;
9856         }
9857 out_notrans:
9858         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9859                 up_read(&fs_info->subvol_sem);
9860
9861         return ret;
9862 }
9863
9864 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9865                          struct inode *new_dir, struct dentry *new_dentry,
9866                          unsigned int flags)
9867 {
9868         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9869                 return -EINVAL;
9870
9871         if (flags & RENAME_EXCHANGE)
9872                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9873                                           new_dentry);
9874
9875         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
9876 }
9877
9878 struct btrfs_delalloc_work {
9879         struct inode *inode;
9880         struct completion completion;
9881         struct list_head list;
9882         struct btrfs_work work;
9883 };
9884
9885 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9886 {
9887         struct btrfs_delalloc_work *delalloc_work;
9888         struct inode *inode;
9889
9890         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9891                                      work);
9892         inode = delalloc_work->inode;
9893         filemap_flush(inode->i_mapping);
9894         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9895                                 &BTRFS_I(inode)->runtime_flags))
9896                 filemap_flush(inode->i_mapping);
9897
9898         iput(inode);
9899         complete(&delalloc_work->completion);
9900 }
9901
9902 static struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode)
9903 {
9904         struct btrfs_delalloc_work *work;
9905
9906         work = kmalloc(sizeof(*work), GFP_NOFS);
9907         if (!work)
9908                 return NULL;
9909
9910         init_completion(&work->completion);
9911         INIT_LIST_HEAD(&work->list);
9912         work->inode = inode;
9913         WARN_ON_ONCE(!inode);
9914         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9915                         btrfs_run_delalloc_work, NULL, NULL);
9916
9917         return work;
9918 }
9919
9920 /*
9921  * some fairly slow code that needs optimization. This walks the list
9922  * of all the inodes with pending delalloc and forces them to disk.
9923  */
9924 static int start_delalloc_inodes(struct btrfs_root *root, int nr, bool snapshot)
9925 {
9926         struct btrfs_inode *binode;
9927         struct inode *inode;
9928         struct btrfs_delalloc_work *work, *next;
9929         struct list_head works;
9930         struct list_head splice;
9931         int ret = 0;
9932
9933         INIT_LIST_HEAD(&works);
9934         INIT_LIST_HEAD(&splice);
9935
9936         mutex_lock(&root->delalloc_mutex);
9937         spin_lock(&root->delalloc_lock);
9938         list_splice_init(&root->delalloc_inodes, &splice);
9939         while (!list_empty(&splice)) {
9940                 binode = list_entry(splice.next, struct btrfs_inode,
9941                                     delalloc_inodes);
9942
9943                 list_move_tail(&binode->delalloc_inodes,
9944                                &root->delalloc_inodes);
9945                 inode = igrab(&binode->vfs_inode);
9946                 if (!inode) {
9947                         cond_resched_lock(&root->delalloc_lock);
9948                         continue;
9949                 }
9950                 spin_unlock(&root->delalloc_lock);
9951
9952                 if (snapshot)
9953                         set_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
9954                                 &binode->runtime_flags);
9955                 work = btrfs_alloc_delalloc_work(inode);
9956                 if (!work) {
9957                         iput(inode);
9958                         ret = -ENOMEM;
9959                         goto out;
9960                 }
9961                 list_add_tail(&work->list, &works);
9962                 btrfs_queue_work(root->fs_info->flush_workers,
9963                                  &work->work);
9964                 ret++;
9965                 if (nr != -1 && ret >= nr)
9966                         goto out;
9967                 cond_resched();
9968                 spin_lock(&root->delalloc_lock);
9969         }
9970         spin_unlock(&root->delalloc_lock);
9971
9972 out:
9973         list_for_each_entry_safe(work, next, &works, list) {
9974                 list_del_init(&work->list);
9975                 wait_for_completion(&work->completion);
9976                 kfree(work);
9977         }
9978
9979         if (!list_empty(&splice)) {
9980                 spin_lock(&root->delalloc_lock);
9981                 list_splice_tail(&splice, &root->delalloc_inodes);
9982                 spin_unlock(&root->delalloc_lock);
9983         }
9984         mutex_unlock(&root->delalloc_mutex);
9985         return ret;
9986 }
9987
9988 int btrfs_start_delalloc_snapshot(struct btrfs_root *root)
9989 {
9990         struct btrfs_fs_info *fs_info = root->fs_info;
9991         int ret;
9992
9993         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
9994                 return -EROFS;
9995
9996         ret = start_delalloc_inodes(root, -1, true);
9997         if (ret > 0)
9998                 ret = 0;
9999         return ret;
10000 }
10001
10002 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int nr)
10003 {
10004         struct btrfs_root *root;
10005         struct list_head splice;
10006         int ret;
10007
10008         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10009                 return -EROFS;
10010
10011         INIT_LIST_HEAD(&splice);
10012
10013         mutex_lock(&fs_info->delalloc_root_mutex);
10014         spin_lock(&fs_info->delalloc_root_lock);
10015         list_splice_init(&fs_info->delalloc_roots, &splice);
10016         while (!list_empty(&splice) && nr) {
10017                 root = list_first_entry(&splice, struct btrfs_root,
10018                                         delalloc_root);
10019                 root = btrfs_grab_fs_root(root);
10020                 BUG_ON(!root);
10021                 list_move_tail(&root->delalloc_root,
10022                                &fs_info->delalloc_roots);
10023                 spin_unlock(&fs_info->delalloc_root_lock);
10024
10025                 ret = start_delalloc_inodes(root, nr, false);
10026                 btrfs_put_fs_root(root);
10027                 if (ret < 0)
10028                         goto out;
10029
10030                 if (nr != -1) {
10031                         nr -= ret;
10032                         WARN_ON(nr < 0);
10033                 }
10034                 spin_lock(&fs_info->delalloc_root_lock);
10035         }
10036         spin_unlock(&fs_info->delalloc_root_lock);
10037
10038         ret = 0;
10039 out:
10040         if (!list_empty(&splice)) {
10041                 spin_lock(&fs_info->delalloc_root_lock);
10042                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10043                 spin_unlock(&fs_info->delalloc_root_lock);
10044         }
10045         mutex_unlock(&fs_info->delalloc_root_mutex);
10046         return ret;
10047 }
10048
10049 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10050                          const char *symname)
10051 {
10052         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10053         struct btrfs_trans_handle *trans;
10054         struct btrfs_root *root = BTRFS_I(dir)->root;
10055         struct btrfs_path *path;
10056         struct btrfs_key key;
10057         struct inode *inode = NULL;
10058         int err;
10059         u64 objectid;
10060         u64 index = 0;
10061         int name_len;
10062         int datasize;
10063         unsigned long ptr;
10064         struct btrfs_file_extent_item *ei;
10065         struct extent_buffer *leaf;
10066
10067         name_len = strlen(symname);
10068         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10069                 return -ENAMETOOLONG;
10070
10071         /*
10072          * 2 items for inode item and ref
10073          * 2 items for dir items
10074          * 1 item for updating parent inode item
10075          * 1 item for the inline extent item
10076          * 1 item for xattr if selinux is on
10077          */
10078         trans = btrfs_start_transaction(root, 7);
10079         if (IS_ERR(trans))
10080                 return PTR_ERR(trans);
10081
10082         err = btrfs_find_free_ino(root, &objectid);
10083         if (err)
10084                 goto out_unlock;
10085
10086         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10087                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10088                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10089         if (IS_ERR(inode)) {
10090                 err = PTR_ERR(inode);
10091                 inode = NULL;
10092                 goto out_unlock;
10093         }
10094
10095         /*
10096         * If the active LSM wants to access the inode during
10097         * d_instantiate it needs these. Smack checks to see
10098         * if the filesystem supports xattrs by looking at the
10099         * ops vector.
10100         */
10101         inode->i_fop = &btrfs_file_operations;
10102         inode->i_op = &btrfs_file_inode_operations;
10103         inode->i_mapping->a_ops = &btrfs_aops;
10104         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10105
10106         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10107         if (err)
10108                 goto out_unlock;
10109
10110         path = btrfs_alloc_path();
10111         if (!path) {
10112                 err = -ENOMEM;
10113                 goto out_unlock;
10114         }
10115         key.objectid = btrfs_ino(BTRFS_I(inode));
10116         key.offset = 0;
10117         key.type = BTRFS_EXTENT_DATA_KEY;
10118         datasize = btrfs_file_extent_calc_inline_size(name_len);
10119         err = btrfs_insert_empty_item(trans, root, path, &key,
10120                                       datasize);
10121         if (err) {
10122                 btrfs_free_path(path);
10123                 goto out_unlock;
10124         }
10125         leaf = path->nodes[0];
10126         ei = btrfs_item_ptr(leaf, path->slots[0],
10127                             struct btrfs_file_extent_item);
10128         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10129         btrfs_set_file_extent_type(leaf, ei,
10130                                    BTRFS_FILE_EXTENT_INLINE);
10131         btrfs_set_file_extent_encryption(leaf, ei, 0);
10132         btrfs_set_file_extent_compression(leaf, ei, 0);
10133         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10134         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10135
10136         ptr = btrfs_file_extent_inline_start(ei);
10137         write_extent_buffer(leaf, symname, ptr, name_len);
10138         btrfs_mark_buffer_dirty(leaf);
10139         btrfs_free_path(path);
10140
10141         inode->i_op = &btrfs_symlink_inode_operations;
10142         inode_nohighmem(inode);
10143         inode->i_mapping->a_ops = &btrfs_aops;
10144         inode_set_bytes(inode, name_len);
10145         btrfs_i_size_write(BTRFS_I(inode), name_len);
10146         err = btrfs_update_inode(trans, root, inode);
10147         /*
10148          * Last step, add directory indexes for our symlink inode. This is the
10149          * last step to avoid extra cleanup of these indexes if an error happens
10150          * elsewhere above.
10151          */
10152         if (!err)
10153                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10154                                 BTRFS_I(inode), 0, index);
10155         if (err)
10156                 goto out_unlock;
10157
10158         d_instantiate_new(dentry, inode);
10159
10160 out_unlock:
10161         btrfs_end_transaction(trans);
10162         if (err && inode) {
10163                 inode_dec_link_count(inode);
10164                 discard_new_inode(inode);
10165         }
10166         btrfs_btree_balance_dirty(fs_info);
10167         return err;
10168 }
10169
10170 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10171                                        u64 start, u64 num_bytes, u64 min_size,
10172                                        loff_t actual_len, u64 *alloc_hint,
10173                                        struct btrfs_trans_handle *trans)
10174 {
10175         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10176         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10177         struct extent_map *em;
10178         struct btrfs_root *root = BTRFS_I(inode)->root;
10179         struct btrfs_key ins;
10180         u64 cur_offset = start;
10181         u64 i_size;
10182         u64 cur_bytes;
10183         u64 last_alloc = (u64)-1;
10184         int ret = 0;
10185         bool own_trans = true;
10186         u64 end = start + num_bytes - 1;
10187
10188         if (trans)
10189                 own_trans = false;
10190         while (num_bytes > 0) {
10191                 if (own_trans) {
10192                         trans = btrfs_start_transaction(root, 3);
10193                         if (IS_ERR(trans)) {
10194                                 ret = PTR_ERR(trans);
10195                                 break;
10196                         }
10197                 }
10198
10199                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10200                 cur_bytes = max(cur_bytes, min_size);
10201                 /*
10202                  * If we are severely fragmented we could end up with really
10203                  * small allocations, so if the allocator is returning small
10204                  * chunks lets make its job easier by only searching for those
10205                  * sized chunks.
10206                  */
10207                 cur_bytes = min(cur_bytes, last_alloc);
10208                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10209                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10210                 if (ret) {
10211                         if (own_trans)
10212                                 btrfs_end_transaction(trans);
10213                         break;
10214                 }
10215                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10216
10217                 last_alloc = ins.offset;
10218                 ret = insert_reserved_file_extent(trans, inode,
10219                                                   cur_offset, ins.objectid,
10220                                                   ins.offset, ins.offset,
10221                                                   ins.offset, 0, 0, 0,
10222                                                   BTRFS_FILE_EXTENT_PREALLOC);
10223                 if (ret) {
10224                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10225                                                    ins.offset, 0);
10226                         btrfs_abort_transaction(trans, ret);
10227                         if (own_trans)
10228                                 btrfs_end_transaction(trans);
10229                         break;
10230                 }
10231
10232                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10233                                         cur_offset + ins.offset -1, 0);
10234
10235                 em = alloc_extent_map();
10236                 if (!em) {
10237                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10238                                 &BTRFS_I(inode)->runtime_flags);
10239                         goto next;
10240                 }
10241
10242                 em->start = cur_offset;
10243                 em->orig_start = cur_offset;
10244                 em->len = ins.offset;
10245                 em->block_start = ins.objectid;
10246                 em->block_len = ins.offset;
10247                 em->orig_block_len = ins.offset;
10248                 em->ram_bytes = ins.offset;
10249                 em->bdev = fs_info->fs_devices->latest_bdev;
10250                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10251                 em->generation = trans->transid;
10252
10253                 while (1) {
10254                         write_lock(&em_tree->lock);
10255                         ret = add_extent_mapping(em_tree, em, 1);
10256                         write_unlock(&em_tree->lock);
10257                         if (ret != -EEXIST)
10258                                 break;
10259                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10260                                                 cur_offset + ins.offset - 1,
10261                                                 0);
10262                 }
10263                 free_extent_map(em);
10264 next:
10265                 num_bytes -= ins.offset;
10266                 cur_offset += ins.offset;
10267                 *alloc_hint = ins.objectid + ins.offset;
10268
10269                 inode_inc_iversion(inode);
10270                 inode->i_ctime = current_time(inode);
10271                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10272                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10273                     (actual_len > inode->i_size) &&
10274                     (cur_offset > inode->i_size)) {
10275                         if (cur_offset > actual_len)
10276                                 i_size = actual_len;
10277                         else
10278                                 i_size = cur_offset;
10279                         i_size_write(inode, i_size);
10280                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10281                 }
10282
10283                 ret = btrfs_update_inode(trans, root, inode);
10284
10285                 if (ret) {
10286                         btrfs_abort_transaction(trans, ret);
10287                         if (own_trans)
10288                                 btrfs_end_transaction(trans);
10289                         break;
10290                 }
10291
10292                 if (own_trans)
10293                         btrfs_end_transaction(trans);
10294         }
10295         if (cur_offset < end)
10296                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10297                         end - cur_offset + 1);
10298         return ret;
10299 }
10300
10301 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10302                               u64 start, u64 num_bytes, u64 min_size,
10303                               loff_t actual_len, u64 *alloc_hint)
10304 {
10305         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10306                                            min_size, actual_len, alloc_hint,
10307                                            NULL);
10308 }
10309
10310 int btrfs_prealloc_file_range_trans(struct inode *inode,
10311                                     struct btrfs_trans_handle *trans, int mode,
10312                                     u64 start, u64 num_bytes, u64 min_size,
10313                                     loff_t actual_len, u64 *alloc_hint)
10314 {
10315         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10316                                            min_size, actual_len, alloc_hint, trans);
10317 }
10318
10319 static int btrfs_set_page_dirty(struct page *page)
10320 {
10321         return __set_page_dirty_nobuffers(page);
10322 }
10323
10324 static int btrfs_permission(struct inode *inode, int mask)
10325 {
10326         struct btrfs_root *root = BTRFS_I(inode)->root;
10327         umode_t mode = inode->i_mode;
10328
10329         if (mask & MAY_WRITE &&
10330             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10331                 if (btrfs_root_readonly(root))
10332                         return -EROFS;
10333                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10334                         return -EACCES;
10335         }
10336         return generic_permission(inode, mask);
10337 }
10338
10339 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10340 {
10341         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10342         struct btrfs_trans_handle *trans;
10343         struct btrfs_root *root = BTRFS_I(dir)->root;
10344         struct inode *inode = NULL;
10345         u64 objectid;
10346         u64 index;
10347         int ret = 0;
10348
10349         /*
10350          * 5 units required for adding orphan entry
10351          */
10352         trans = btrfs_start_transaction(root, 5);
10353         if (IS_ERR(trans))
10354                 return PTR_ERR(trans);
10355
10356         ret = btrfs_find_free_ino(root, &objectid);
10357         if (ret)
10358                 goto out;
10359
10360         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10361                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10362         if (IS_ERR(inode)) {
10363                 ret = PTR_ERR(inode);
10364                 inode = NULL;
10365                 goto out;
10366         }
10367
10368         inode->i_fop = &btrfs_file_operations;
10369         inode->i_op = &btrfs_file_inode_operations;
10370
10371         inode->i_mapping->a_ops = &btrfs_aops;
10372         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10373
10374         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10375         if (ret)
10376                 goto out;
10377
10378         ret = btrfs_update_inode(trans, root, inode);
10379         if (ret)
10380                 goto out;
10381         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10382         if (ret)
10383                 goto out;
10384
10385         /*
10386          * We set number of links to 0 in btrfs_new_inode(), and here we set
10387          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10388          * through:
10389          *
10390          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10391          */
10392         set_nlink(inode, 1);
10393         d_tmpfile(dentry, inode);
10394         unlock_new_inode(inode);
10395         mark_inode_dirty(inode);
10396 out:
10397         btrfs_end_transaction(trans);
10398         if (ret && inode)
10399                 discard_new_inode(inode);
10400         btrfs_btree_balance_dirty(fs_info);
10401         return ret;
10402 }
10403
10404 void btrfs_set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
10405 {
10406         struct inode *inode = tree->private_data;
10407         unsigned long index = start >> PAGE_SHIFT;
10408         unsigned long end_index = end >> PAGE_SHIFT;
10409         struct page *page;
10410
10411         while (index <= end_index) {
10412                 page = find_get_page(inode->i_mapping, index);
10413                 ASSERT(page); /* Pages should be in the extent_io_tree */
10414                 set_page_writeback(page);
10415                 put_page(page);
10416                 index++;
10417         }
10418 }
10419
10420 #ifdef CONFIG_SWAP
10421 /*
10422  * Add an entry indicating a block group or device which is pinned by a
10423  * swapfile. Returns 0 on success, 1 if there is already an entry for it, or a
10424  * negative errno on failure.
10425  */
10426 static int btrfs_add_swapfile_pin(struct inode *inode, void *ptr,
10427                                   bool is_block_group)
10428 {
10429         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10430         struct btrfs_swapfile_pin *sp, *entry;
10431         struct rb_node **p;
10432         struct rb_node *parent = NULL;
10433
10434         sp = kmalloc(sizeof(*sp), GFP_NOFS);
10435         if (!sp)
10436                 return -ENOMEM;
10437         sp->ptr = ptr;
10438         sp->inode = inode;
10439         sp->is_block_group = is_block_group;
10440
10441         spin_lock(&fs_info->swapfile_pins_lock);
10442         p = &fs_info->swapfile_pins.rb_node;
10443         while (*p) {
10444                 parent = *p;
10445                 entry = rb_entry(parent, struct btrfs_swapfile_pin, node);
10446                 if (sp->ptr < entry->ptr ||
10447                     (sp->ptr == entry->ptr && sp->inode < entry->inode)) {
10448                         p = &(*p)->rb_left;
10449                 } else if (sp->ptr > entry->ptr ||
10450                            (sp->ptr == entry->ptr && sp->inode > entry->inode)) {
10451                         p = &(*p)->rb_right;
10452                 } else {
10453                         spin_unlock(&fs_info->swapfile_pins_lock);
10454                         kfree(sp);
10455                         return 1;
10456                 }
10457         }
10458         rb_link_node(&sp->node, parent, p);
10459         rb_insert_color(&sp->node, &fs_info->swapfile_pins);
10460         spin_unlock(&fs_info->swapfile_pins_lock);
10461         return 0;
10462 }
10463
10464 /* Free all of the entries pinned by this swapfile. */
10465 static void btrfs_free_swapfile_pins(struct inode *inode)
10466 {
10467         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10468         struct btrfs_swapfile_pin *sp;
10469         struct rb_node *node, *next;
10470
10471         spin_lock(&fs_info->swapfile_pins_lock);
10472         node = rb_first(&fs_info->swapfile_pins);
10473         while (node) {
10474                 next = rb_next(node);
10475                 sp = rb_entry(node, struct btrfs_swapfile_pin, node);
10476                 if (sp->inode == inode) {
10477                         rb_erase(&sp->node, &fs_info->swapfile_pins);
10478                         if (sp->is_block_group)
10479                                 btrfs_put_block_group(sp->ptr);
10480                         kfree(sp);
10481                 }
10482                 node = next;
10483         }
10484         spin_unlock(&fs_info->swapfile_pins_lock);
10485 }
10486
10487 struct btrfs_swap_info {
10488         u64 start;
10489         u64 block_start;
10490         u64 block_len;
10491         u64 lowest_ppage;
10492         u64 highest_ppage;
10493         unsigned long nr_pages;
10494         int nr_extents;
10495 };
10496
10497 static int btrfs_add_swap_extent(struct swap_info_struct *sis,
10498                                  struct btrfs_swap_info *bsi)
10499 {
10500         unsigned long nr_pages;
10501         u64 first_ppage, first_ppage_reported, next_ppage;
10502         int ret;
10503
10504         first_ppage = ALIGN(bsi->block_start, PAGE_SIZE) >> PAGE_SHIFT;
10505         next_ppage = ALIGN_DOWN(bsi->block_start + bsi->block_len,
10506                                 PAGE_SIZE) >> PAGE_SHIFT;
10507
10508         if (first_ppage >= next_ppage)
10509                 return 0;
10510         nr_pages = next_ppage - first_ppage;
10511
10512         first_ppage_reported = first_ppage;
10513         if (bsi->start == 0)
10514                 first_ppage_reported++;
10515         if (bsi->lowest_ppage > first_ppage_reported)
10516                 bsi->lowest_ppage = first_ppage_reported;
10517         if (bsi->highest_ppage < (next_ppage - 1))
10518                 bsi->highest_ppage = next_ppage - 1;
10519
10520         ret = add_swap_extent(sis, bsi->nr_pages, nr_pages, first_ppage);
10521         if (ret < 0)
10522                 return ret;
10523         bsi->nr_extents += ret;
10524         bsi->nr_pages += nr_pages;
10525         return 0;
10526 }
10527
10528 static void btrfs_swap_deactivate(struct file *file)
10529 {
10530         struct inode *inode = file_inode(file);
10531
10532         btrfs_free_swapfile_pins(inode);
10533         atomic_dec(&BTRFS_I(inode)->root->nr_swapfiles);
10534 }
10535
10536 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10537                                sector_t *span)
10538 {
10539         struct inode *inode = file_inode(file);
10540         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10541         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
10542         struct extent_state *cached_state = NULL;
10543         struct extent_map *em = NULL;
10544         struct btrfs_device *device = NULL;
10545         struct btrfs_swap_info bsi = {
10546                 .lowest_ppage = (sector_t)-1ULL,
10547         };
10548         int ret = 0;
10549         u64 isize;
10550         u64 start;
10551
10552         /*
10553          * If the swap file was just created, make sure delalloc is done. If the
10554          * file changes again after this, the user is doing something stupid and
10555          * we don't really care.
10556          */
10557         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
10558         if (ret)
10559                 return ret;
10560
10561         /*
10562          * The inode is locked, so these flags won't change after we check them.
10563          */
10564         if (BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS) {
10565                 btrfs_warn(fs_info, "swapfile must not be compressed");
10566                 return -EINVAL;
10567         }
10568         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW)) {
10569                 btrfs_warn(fs_info, "swapfile must not be copy-on-write");
10570                 return -EINVAL;
10571         }
10572         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
10573                 btrfs_warn(fs_info, "swapfile must not be checksummed");
10574                 return -EINVAL;
10575         }
10576
10577         /*
10578          * Balance or device remove/replace/resize can move stuff around from
10579          * under us. The EXCL_OP flag makes sure they aren't running/won't run
10580          * concurrently while we are mapping the swap extents, and
10581          * fs_info->swapfile_pins prevents them from running while the swap file
10582          * is active and moving the extents. Note that this also prevents a
10583          * concurrent device add which isn't actually necessary, but it's not
10584          * really worth the trouble to allow it.
10585          */
10586         if (test_and_set_bit(BTRFS_FS_EXCL_OP, &fs_info->flags)) {
10587                 btrfs_warn(fs_info,
10588            "cannot activate swapfile while exclusive operation is running");
10589                 return -EBUSY;
10590         }
10591         /*
10592          * Snapshots can create extents which require COW even if NODATACOW is
10593          * set. We use this counter to prevent snapshots. We must increment it
10594          * before walking the extents because we don't want a concurrent
10595          * snapshot to run after we've already checked the extents.
10596          */
10597         atomic_inc(&BTRFS_I(inode)->root->nr_swapfiles);
10598
10599         isize = ALIGN_DOWN(inode->i_size, fs_info->sectorsize);
10600
10601         lock_extent_bits(io_tree, 0, isize - 1, &cached_state);
10602         start = 0;
10603         while (start < isize) {
10604                 u64 logical_block_start, physical_block_start;
10605                 struct btrfs_block_group_cache *bg;
10606                 u64 len = isize - start;
10607
10608                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
10609                 if (IS_ERR(em)) {
10610                         ret = PTR_ERR(em);
10611                         goto out;
10612                 }
10613
10614                 if (em->block_start == EXTENT_MAP_HOLE) {
10615                         btrfs_warn(fs_info, "swapfile must not have holes");
10616                         ret = -EINVAL;
10617                         goto out;
10618                 }
10619                 if (em->block_start == EXTENT_MAP_INLINE) {
10620                         /*
10621                          * It's unlikely we'll ever actually find ourselves
10622                          * here, as a file small enough to fit inline won't be
10623                          * big enough to store more than the swap header, but in
10624                          * case something changes in the future, let's catch it
10625                          * here rather than later.
10626                          */
10627                         btrfs_warn(fs_info, "swapfile must not be inline");
10628                         ret = -EINVAL;
10629                         goto out;
10630                 }
10631                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
10632                         btrfs_warn(fs_info, "swapfile must not be compressed");
10633                         ret = -EINVAL;
10634                         goto out;
10635                 }
10636
10637                 logical_block_start = em->block_start + (start - em->start);
10638                 len = min(len, em->len - (start - em->start));
10639                 free_extent_map(em);
10640                 em = NULL;
10641
10642                 ret = can_nocow_extent(inode, start, &len, NULL, NULL, NULL);
10643                 if (ret < 0) {
10644                         goto out;
10645                 } else if (ret) {
10646                         ret = 0;
10647                 } else {
10648                         btrfs_warn(fs_info,
10649                                    "swapfile must not be copy-on-write");
10650                         ret = -EINVAL;
10651                         goto out;
10652                 }
10653
10654                 em = btrfs_get_chunk_map(fs_info, logical_block_start, len);
10655                 if (IS_ERR(em)) {
10656                         ret = PTR_ERR(em);
10657                         goto out;
10658                 }
10659
10660                 if (em->map_lookup->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
10661                         btrfs_warn(fs_info,
10662                                    "swapfile must have single data profile");
10663                         ret = -EINVAL;
10664                         goto out;
10665                 }
10666
10667                 if (device == NULL) {
10668                         device = em->map_lookup->stripes[0].dev;
10669                         ret = btrfs_add_swapfile_pin(inode, device, false);
10670                         if (ret == 1)
10671                                 ret = 0;
10672                         else if (ret)
10673                                 goto out;
10674                 } else if (device != em->map_lookup->stripes[0].dev) {
10675                         btrfs_warn(fs_info, "swapfile must be on one device");
10676                         ret = -EINVAL;
10677                         goto out;
10678                 }
10679
10680                 physical_block_start = (em->map_lookup->stripes[0].physical +
10681                                         (logical_block_start - em->start));
10682                 len = min(len, em->len - (logical_block_start - em->start));
10683                 free_extent_map(em);
10684                 em = NULL;
10685
10686                 bg = btrfs_lookup_block_group(fs_info, logical_block_start);
10687                 if (!bg) {
10688                         btrfs_warn(fs_info,
10689                            "could not find block group containing swapfile");
10690                         ret = -EINVAL;
10691                         goto out;
10692                 }
10693
10694                 ret = btrfs_add_swapfile_pin(inode, bg, true);
10695                 if (ret) {
10696                         btrfs_put_block_group(bg);
10697                         if (ret == 1)
10698                                 ret = 0;
10699                         else
10700                                 goto out;
10701                 }
10702
10703                 if (bsi.block_len &&
10704                     bsi.block_start + bsi.block_len == physical_block_start) {
10705                         bsi.block_len += len;
10706                 } else {
10707                         if (bsi.block_len) {
10708                                 ret = btrfs_add_swap_extent(sis, &bsi);
10709                                 if (ret)
10710                                         goto out;
10711                         }
10712                         bsi.start = start;
10713                         bsi.block_start = physical_block_start;
10714                         bsi.block_len = len;
10715                 }
10716
10717                 start += len;
10718         }
10719
10720         if (bsi.block_len)
10721                 ret = btrfs_add_swap_extent(sis, &bsi);
10722
10723 out:
10724         if (!IS_ERR_OR_NULL(em))
10725                 free_extent_map(em);
10726
10727         unlock_extent_cached(io_tree, 0, isize - 1, &cached_state);
10728
10729         if (ret)
10730                 btrfs_swap_deactivate(file);
10731
10732         clear_bit(BTRFS_FS_EXCL_OP, &fs_info->flags);
10733
10734         if (ret)
10735                 return ret;
10736
10737         if (device)
10738                 sis->bdev = device->bdev;
10739         *span = bsi.highest_ppage - bsi.lowest_ppage + 1;
10740         sis->max = bsi.nr_pages;
10741         sis->pages = bsi.nr_pages - 1;
10742         sis->highest_bit = bsi.nr_pages - 1;
10743         return bsi.nr_extents;
10744 }
10745 #else
10746 static void btrfs_swap_deactivate(struct file *file)
10747 {
10748 }
10749
10750 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10751                                sector_t *span)
10752 {
10753         return -EOPNOTSUPP;
10754 }
10755 #endif
10756
10757 static const struct inode_operations btrfs_dir_inode_operations = {
10758         .getattr        = btrfs_getattr,
10759         .lookup         = btrfs_lookup,
10760         .create         = btrfs_create,
10761         .unlink         = btrfs_unlink,
10762         .link           = btrfs_link,
10763         .mkdir          = btrfs_mkdir,
10764         .rmdir          = btrfs_rmdir,
10765         .rename         = btrfs_rename2,
10766         .symlink        = btrfs_symlink,
10767         .setattr        = btrfs_setattr,
10768         .mknod          = btrfs_mknod,
10769         .listxattr      = btrfs_listxattr,
10770         .permission     = btrfs_permission,
10771         .get_acl        = btrfs_get_acl,
10772         .set_acl        = btrfs_set_acl,
10773         .update_time    = btrfs_update_time,
10774         .tmpfile        = btrfs_tmpfile,
10775 };
10776 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10777         .lookup         = btrfs_lookup,
10778         .permission     = btrfs_permission,
10779         .update_time    = btrfs_update_time,
10780 };
10781
10782 static const struct file_operations btrfs_dir_file_operations = {
10783         .llseek         = generic_file_llseek,
10784         .read           = generic_read_dir,
10785         .iterate_shared = btrfs_real_readdir,
10786         .open           = btrfs_opendir,
10787         .unlocked_ioctl = btrfs_ioctl,
10788 #ifdef CONFIG_COMPAT
10789         .compat_ioctl   = btrfs_compat_ioctl,
10790 #endif
10791         .release        = btrfs_release_file,
10792         .fsync          = btrfs_sync_file,
10793 };
10794
10795 static const struct extent_io_ops btrfs_extent_io_ops = {
10796         /* mandatory callbacks */
10797         .submit_bio_hook = btrfs_submit_bio_hook,
10798         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10799 };
10800
10801 /*
10802  * btrfs doesn't support the bmap operation because swapfiles
10803  * use bmap to make a mapping of extents in the file.  They assume
10804  * these extents won't change over the life of the file and they
10805  * use the bmap result to do IO directly to the drive.
10806  *
10807  * the btrfs bmap call would return logical addresses that aren't
10808  * suitable for IO and they also will change frequently as COW
10809  * operations happen.  So, swapfile + btrfs == corruption.
10810  *
10811  * For now we're avoiding this by dropping bmap.
10812  */
10813 static const struct address_space_operations btrfs_aops = {
10814         .readpage       = btrfs_readpage,
10815         .writepage      = btrfs_writepage,
10816         .writepages     = btrfs_writepages,
10817         .readpages      = btrfs_readpages,
10818         .direct_IO      = btrfs_direct_IO,
10819         .invalidatepage = btrfs_invalidatepage,
10820         .releasepage    = btrfs_releasepage,
10821         .set_page_dirty = btrfs_set_page_dirty,
10822         .error_remove_page = generic_error_remove_page,
10823         .swap_activate  = btrfs_swap_activate,
10824         .swap_deactivate = btrfs_swap_deactivate,
10825 };
10826
10827 static const struct inode_operations btrfs_file_inode_operations = {
10828         .getattr        = btrfs_getattr,
10829         .setattr        = btrfs_setattr,
10830         .listxattr      = btrfs_listxattr,
10831         .permission     = btrfs_permission,
10832         .fiemap         = btrfs_fiemap,
10833         .get_acl        = btrfs_get_acl,
10834         .set_acl        = btrfs_set_acl,
10835         .update_time    = btrfs_update_time,
10836 };
10837 static const struct inode_operations btrfs_special_inode_operations = {
10838         .getattr        = btrfs_getattr,
10839         .setattr        = btrfs_setattr,
10840         .permission     = btrfs_permission,
10841         .listxattr      = btrfs_listxattr,
10842         .get_acl        = btrfs_get_acl,
10843         .set_acl        = btrfs_set_acl,
10844         .update_time    = btrfs_update_time,
10845 };
10846 static const struct inode_operations btrfs_symlink_inode_operations = {
10847         .get_link       = page_get_link,
10848         .getattr        = btrfs_getattr,
10849         .setattr        = btrfs_setattr,
10850         .permission     = btrfs_permission,
10851         .listxattr      = btrfs_listxattr,
10852         .update_time    = btrfs_update_time,
10853 };
10854
10855 const struct dentry_operations btrfs_dentry_operations = {
10856         .d_delete       = btrfs_dentry_delete,
10857 };