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