fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/compat.h>
  34 #include <linux/bit_spinlock.h>
  35 #include <linux/xattr.h>
  36 #include <linux/posix_acl.h>
  37 #include <linux/falloc.h>
  38 #include <linux/slab.h>
  39 #include <linux/ratelimit.h>
  40 #include <linux/mount.h>
  41 #include <linux/btrfs.h>
  42 #include <linux/blkdev.h>
  43 #include <linux/posix_acl_xattr.h>
  44 #include <linux/uio.h>
  45 #include <linux/magic.h>
  46 #include "ctree.h"
  47 #include "disk-io.h"
  48 #include "transaction.h"
  49 #include "btrfs_inode.h"
  50 #include "print-tree.h"
  51 #include "ordered-data.h"
  52 #include "xattr.h"
  53 #include "tree-log.h"
  54 #include "volumes.h"
  55 #include "compression.h"
  56 #include "locking.h"
  57 #include "free-space-cache.h"
  58 #include "inode-map.h"
  59 #include "backref.h"
  60 #include "hash.h"
  61 #include "props.h"
  62 #include "qgroup.h"
  63 #include "dedupe.h"
  64
  65 struct btrfs_iget_args {
  66         struct btrfs_key *location;
  67         struct btrfs_root *root;
  68 };
  69
  70 struct btrfs_dio_data {
  71         u64 reserve;
  72         u64 unsubmitted_oe_range_start;
  73         u64 unsubmitted_oe_range_end;
  74         int overwrite;
  75 };
  76
  77 static const struct inode_operations btrfs_dir_inode_operations;
  78 static const struct inode_operations btrfs_symlink_inode_operations;
  79 static const struct inode_operations btrfs_dir_ro_inode_operations;
  80 static const struct inode_operations btrfs_special_inode_operations;
  81 static const struct inode_operations btrfs_file_inode_operations;
  82 static const struct address_space_operations btrfs_aops;
  83 static const struct address_space_operations btrfs_symlink_aops;
  84 static const struct file_operations btrfs_dir_file_operations;
  85 static const struct extent_io_ops btrfs_extent_io_ops;
  86
  87 static struct kmem_cache *btrfs_inode_cachep;
  88 struct kmem_cache *btrfs_trans_handle_cachep;
  89 struct kmem_cache *btrfs_path_cachep;
  90 struct kmem_cache *btrfs_free_space_cachep;
  91
  92 #define S_SHIFT 12
  93 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  94         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  95         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  96         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  97         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  98         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  99         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 100         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 101 };
 102
 103 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 104 static int btrfs_truncate(struct inode *inode);
 105 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 106 static noinline int cow_file_range(struct inode *inode,
 107                                    struct page *locked_page,
 108                                    u64 start, u64 end, u64 delalloc_end,
 109                                    int *page_started, unsigned long *nr_written,
 110                                    int unlock, struct btrfs_dedupe_hash *hash);
 111 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
 112                                        u64 orig_start, u64 block_start,
 113                                        u64 block_len, u64 orig_block_len,
 114                                        u64 ram_bytes, int compress_type,
 115                                        int type);
 116
 117 static void __endio_write_update_ordered(struct inode *inode,
 118                                          const u64 offset, const u64 bytes,
 119                                          const bool uptodate);
 120
 121 /*
 122  * Cleanup all submitted ordered extents in specified range to handle errors
 123  * from the fill_dellaloc() callback.
 124  *
 125  * NOTE: caller must ensure that when an error happens, it can not call
 126  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 127  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 128  * to be released, which we want to happen only when finishing the ordered
 129  * extent (btrfs_finish_ordered_io()). Also note that the caller of the
 130  * fill_delalloc() callback already does proper cleanup for the first page of
 131  * the range, that is, it invokes the callback writepage_end_io_hook() for the
 132  * range of the first page.
 133  */
 134 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 135                                                  const u64 offset,
 136                                                  const u64 bytes)
 137 {
 138         unsigned long index = offset >> PAGE_SHIFT;
 139         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 140         struct page *page;
 141
 142         while (index <= end_index) {
 143                 page = find_get_page(inode->i_mapping, index);
 144                 index++;
 145                 if (!page)
 146                         continue;
 147                 ClearPagePrivate2(page);
 148                 put_page(page);
 149         }
 150         return __endio_write_update_ordered(inode, offset + PAGE_SIZE,
 151                                             bytes - PAGE_SIZE, false);
 152 }
 153
 154 static int btrfs_dirty_inode(struct inode *inode);
 155
 156 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 157 void btrfs_test_inode_set_ops(struct inode *inode)
 158 {
 159         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 160 }
 161 #endif
 162
 163 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 164                                      struct inode *inode,  struct inode *dir,
 165                                      const struct qstr *qstr)
 166 {
 167         int err;
 168
 169         err = btrfs_init_acl(trans, inode, dir);
 170         if (!err)
 171                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 172         return err;
 173 }
 174
 175 /*
 176  * this does all the hard work for inserting an inline extent into
 177  * the btree.  The caller should have done a btrfs_drop_extents so that
 178  * no overlapping inline items exist in the btree
 179  */
 180 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 181                                 struct btrfs_path *path, int extent_inserted,
 182                                 struct btrfs_root *root, struct inode *inode,
 183                                 u64 start, size_t size, size_t compressed_size,
 184                                 int compress_type,
 185                                 struct page **compressed_pages)
 186 {
 187         struct extent_buffer *leaf;
 188         struct page *page = NULL;
 189         char *kaddr;
 190         unsigned long ptr;
 191         struct btrfs_file_extent_item *ei;
 192         int ret;
 193         size_t cur_size = size;
 194         unsigned long offset;
 195
 196         if (compressed_size && compressed_pages)
 197                 cur_size = compressed_size;
 198
 199         inode_add_bytes(inode, size);
 200
 201         if (!extent_inserted) {
 202                 struct btrfs_key key;
 203                 size_t datasize;
 204
 205                 key.objectid = btrfs_ino(BTRFS_I(inode));
 206                 key.offset = start;
 207                 key.type = BTRFS_EXTENT_DATA_KEY;
 208
 209                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 210                 path->leave_spinning = 1;
 211                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 212                                               datasize);
 213                 if (ret)
 214                         goto fail;
 215         }
 216         leaf = path->nodes[0];
 217         ei = btrfs_item_ptr(leaf, path->slots[0],
 218                             struct btrfs_file_extent_item);
 219         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 220         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 221         btrfs_set_file_extent_encryption(leaf, ei, 0);
 222         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 223         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 224         ptr = btrfs_file_extent_inline_start(ei);
 225
 226         if (compress_type != BTRFS_COMPRESS_NONE) {
 227                 struct page *cpage;
 228                 int i = 0;
 229                 while (compressed_size > 0) {
 230                         cpage = compressed_pages[i];
 231                         cur_size = min_t(unsigned long, compressed_size,
 232                                        PAGE_SIZE);
 233
 234                         kaddr = kmap_atomic(cpage);
 235                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 236                         kunmap_atomic(kaddr);
 237
 238                         i++;
 239                         ptr += cur_size;
 240                         compressed_size -= cur_size;
 241                 }
 242                 btrfs_set_file_extent_compression(leaf, ei,
 243                                                   compress_type);
 244         } else {
 245                 page = find_get_page(inode->i_mapping,
 246                                      start >> PAGE_SHIFT);
 247                 btrfs_set_file_extent_compression(leaf, ei, 0);
 248                 kaddr = kmap_atomic(page);
 249                 offset = start & (PAGE_SIZE - 1);
 250                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 251                 kunmap_atomic(kaddr);
 252                 put_page(page);
 253         }
 254         btrfs_mark_buffer_dirty(leaf);
 255         btrfs_release_path(path);
 256
 257         /*
 258          * we're an inline extent, so nobody can
 259          * extend the file past i_size without locking
 260          * a page we already have locked.
 261          *
 262          * We must do any isize and inode updates
 263          * before we unlock the pages.  Otherwise we
 264          * could end up racing with unlink.
 265          */
 266         BTRFS_I(inode)->disk_i_size = inode->i_size;
 267         ret = btrfs_update_inode(trans, root, inode);
 268
 269 fail:
 270         return ret;
 271 }
 272
 273
 274 /*
 275  * conditionally insert an inline extent into the file.  This
 276  * does the checks required to make sure the data is small enough
 277  * to fit as an inline extent.
 278  */
 279 static noinline int cow_file_range_inline(struct btrfs_root *root,
 280                                           struct inode *inode, u64 start,
 281                                           u64 end, size_t compressed_size,
 282                                           int compress_type,
 283                                           struct page **compressed_pages)
 284 {
 285         struct btrfs_fs_info *fs_info = root->fs_info;
 286         struct btrfs_trans_handle *trans;
 287         u64 isize = i_size_read(inode);
 288         u64 actual_end = min(end + 1, isize);
 289         u64 inline_len = actual_end - start;
 290         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 291         u64 data_len = inline_len;
 292         int ret;
 293         struct btrfs_path *path;
 294         int extent_inserted = 0;
 295         u32 extent_item_size;
 296
 297         if (compressed_size)
 298                 data_len = compressed_size;
 299
 300         if (start > 0 ||
 301             actual_end > fs_info->sectorsize ||
 302             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 303             (!compressed_size &&
 304             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 305             end + 1 < isize ||
 306             data_len > fs_info->max_inline) {
 307                 return 1;
 308         }
 309
 310         path = btrfs_alloc_path();
 311         if (!path)
 312                 return -ENOMEM;
 313
 314         trans = btrfs_join_transaction(root);
 315         if (IS_ERR(trans)) {
 316                 btrfs_free_path(path);
 317                 return PTR_ERR(trans);
 318         }
 319         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 320
 321         if (compressed_size && compressed_pages)
 322                 extent_item_size = btrfs_file_extent_calc_inline_size(
 323                    compressed_size);
 324         else
 325                 extent_item_size = btrfs_file_extent_calc_inline_size(
 326                     inline_len);
 327
 328         ret = __btrfs_drop_extents(trans, root, inode, path,
 329                                    start, aligned_end, NULL,
 330                                    1, 1, extent_item_size, &extent_inserted);
 331         if (ret) {
 332                 btrfs_abort_transaction(trans, ret);
 333                 goto out;
 334         }
 335
 336         if (isize > actual_end)
 337                 inline_len = min_t(u64, isize, actual_end);
 338         ret = insert_inline_extent(trans, path, extent_inserted,
 339                                    root, inode, start,
 340                                    inline_len, compressed_size,
 341                                    compress_type, compressed_pages);
 342         if (ret && ret != -ENOSPC) {
 343                 btrfs_abort_transaction(trans, ret);
 344                 goto out;
 345         } else if (ret == -ENOSPC) {
 346                 ret = 1;
 347                 goto out;
 348         }
 349
 350         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 351         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 352 out:
 353         /*
 354          * Don't forget to free the reserved space, as for inlined extent
 355          * it won't count as data extent, free them directly here.
 356          * And at reserve time, it's always aligned to page size, so
 357          * just free one page here.
 358          */
 359         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 360         btrfs_free_path(path);
 361         btrfs_end_transaction(trans);
 362         return ret;
 363 }
 364
 365 struct async_extent {
 366         u64 start;
 367         u64 ram_size;
 368         u64 compressed_size;
 369         struct page **pages;
 370         unsigned long nr_pages;
 371         int compress_type;
 372         struct list_head list;
 373 };
 374
 375 struct async_cow {
 376         struct inode *inode;
 377         struct btrfs_root *root;
 378         struct page *locked_page;
 379         u64 start;
 380         u64 end;
 381         unsigned int write_flags;
 382         struct list_head extents;
 383         struct btrfs_work work;
 384 };
 385
 386 static noinline int add_async_extent(struct async_cow *cow,
 387                                      u64 start, u64 ram_size,
 388                                      u64 compressed_size,
 389                                      struct page **pages,
 390                                      unsigned long nr_pages,
 391                                      int compress_type)
 392 {
 393         struct async_extent *async_extent;
 394
 395         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 396         BUG_ON(!async_extent); /* -ENOMEM */
 397         async_extent->start = start;
 398         async_extent->ram_size = ram_size;
 399         async_extent->compressed_size = compressed_size;
 400         async_extent->pages = pages;
 401         async_extent->nr_pages = nr_pages;
 402         async_extent->compress_type = compress_type;
 403         list_add_tail(&async_extent->list, &cow->extents);
 404         return 0;
 405 }
 406
 407 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 408 {
 409         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 410
 411         /* force compress */
 412         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 413                 return 1;
 414         /* defrag ioctl */
 415         if (BTRFS_I(inode)->defrag_compress)
 416                 return 1;
 417         /* bad compression ratios */
 418         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 419                 return 0;
 420         if (btrfs_test_opt(fs_info, COMPRESS) ||
 421             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 422             BTRFS_I(inode)->prop_compress)
 423                 return btrfs_compress_heuristic(inode, start, end);
 424         return 0;
 425 }
 426
 427 static inline void inode_should_defrag(struct btrfs_inode *inode,
 428                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 429 {
 430         /* If this is a small write inside eof, kick off a defrag */
 431         if (num_bytes < small_write &&
 432             (start > 0 || end + 1 < inode->disk_i_size))
 433                 btrfs_add_inode_defrag(NULL, inode);
 434 }
 435
 436 /*
 437  * we create compressed extents in two phases.  The first
 438  * phase compresses a range of pages that have already been
 439  * locked (both pages and state bits are locked).
 440  *
 441  * This is done inside an ordered work queue, and the compression
 442  * is spread across many cpus.  The actual IO submission is step
 443  * two, and the ordered work queue takes care of making sure that
 444  * happens in the same order things were put onto the queue by
 445  * writepages and friends.
 446  *
 447  * If this code finds it can't get good compression, it puts an
 448  * entry onto the work queue to write the uncompressed bytes.  This
 449  * makes sure that both compressed inodes and uncompressed inodes
 450  * are written in the same order that the flusher thread sent them
 451  * down.
 452  */
 453 static noinline void compress_file_range(struct inode *inode,
 454                                         struct page *locked_page,
 455                                         u64 start, u64 end,
 456                                         struct async_cow *async_cow,
 457                                         int *num_added)
 458 {
 459         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 460         struct btrfs_root *root = BTRFS_I(inode)->root;
 461         u64 blocksize = fs_info->sectorsize;
 462         u64 actual_end;
 463         u64 isize = i_size_read(inode);
 464         int ret = 0;
 465         struct page **pages = NULL;
 466         unsigned long nr_pages;
 467         unsigned long total_compressed = 0;
 468         unsigned long total_in = 0;
 469         int i;
 470         int will_compress;
 471         int compress_type = fs_info->compress_type;
 472         int redirty = 0;
 473
 474         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 475                         SZ_16K);
 476
 477         actual_end = min_t(u64, isize, end + 1);
 478 again:
 479         will_compress = 0;
 480         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 481         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 482         nr_pages = min_t(unsigned long, nr_pages,
 483                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 484
 485         /*
 486          * we don't want to send crud past the end of i_size through
 487          * compression, that's just a waste of CPU time.  So, if the
 488          * end of the file is before the start of our current
 489          * requested range of bytes, we bail out to the uncompressed
 490          * cleanup code that can deal with all of this.
 491          *
 492          * It isn't really the fastest way to fix things, but this is a
 493          * very uncommon corner.
 494          */
 495         if (actual_end <= start)
 496                 goto cleanup_and_bail_uncompressed;
 497
 498         total_compressed = actual_end - start;
 499
 500         /*
 501          * skip compression for a small file range(<=blocksize) that
 502          * isn't an inline extent, since it doesn't save disk space at all.
 503          */
 504         if (total_compressed <= blocksize &&
 505            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 506                 goto cleanup_and_bail_uncompressed;
 507
 508         total_compressed = min_t(unsigned long, total_compressed,
 509                         BTRFS_MAX_UNCOMPRESSED);
 510         total_in = 0;
 511         ret = 0;
 512
 513         /*
 514          * we do compression for mount -o compress and when the
 515          * inode has not been flagged as nocompress.  This flag can
 516          * change at any time if we discover bad compression ratios.
 517          */
 518         if (inode_need_compress(inode, start, end)) {
 519                 WARN_ON(pages);
 520                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 521                 if (!pages) {
 522                         /* just bail out to the uncompressed code */
 523                         goto cont;
 524                 }
 525
 526                 if (BTRFS_I(inode)->defrag_compress)
 527                         compress_type = BTRFS_I(inode)->defrag_compress;
 528                 else if (BTRFS_I(inode)->prop_compress)
 529                         compress_type = BTRFS_I(inode)->prop_compress;
 530
 531                 /*
 532                  * we need to call clear_page_dirty_for_io on each
 533                  * page in the range.  Otherwise applications with the file
 534                  * mmap'd can wander in and change the page contents while
 535                  * we are compressing them.
 536                  *
 537                  * If the compression fails for any reason, we set the pages
 538                  * dirty again later on.
 539                  *
 540                  * Note that the remaining part is redirtied, the start pointer
 541                  * has moved, the end is the original one.
 542                  */
 543                 if (!redirty) {
 544                         extent_range_clear_dirty_for_io(inode, start, end);
 545                         redirty = 1;
 546                 }
 547
 548                 /* Compression level is applied here and only here */
 549                 ret = btrfs_compress_pages(
 550                         compress_type | (fs_info->compress_level << 4),
 551                                            inode->i_mapping, start,
 552                                            pages,
 553                                            &nr_pages,
 554                                            &total_in,
 555                                            &total_compressed);
 556
 557                 if (!ret) {
 558                         unsigned long offset = total_compressed &
 559                                 (PAGE_SIZE - 1);
 560                         struct page *page = pages[nr_pages - 1];
 561                         char *kaddr;
 562
 563                         /* zero the tail end of the last page, we might be
 564                          * sending it down to disk
 565                          */
 566                         if (offset) {
 567                                 kaddr = kmap_atomic(page);
 568                                 memset(kaddr + offset, 0,
 569                                        PAGE_SIZE - offset);
 570                                 kunmap_atomic(kaddr);
 571                         }
 572                         will_compress = 1;
 573                 }
 574         }
 575 cont:
 576         if (start == 0) {
 577                 /* lets try to make an inline extent */
 578                 if (ret || total_in < actual_end) {
 579                         /* we didn't compress the entire range, try
 580                          * to make an uncompressed inline extent.
 581                          */
 582                         ret = cow_file_range_inline(root, inode, start, end,
 583                                             0, BTRFS_COMPRESS_NONE, NULL);
 584                 } else {
 585                         /* try making a compressed inline extent */
 586                         ret = cow_file_range_inline(root, inode, start, end,
 587                                                     total_compressed,
 588                                                     compress_type, pages);
 589                 }
 590                 if (ret <= 0) {
 591                         unsigned long clear_flags = EXTENT_DELALLOC |
 592                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 593                                 EXTENT_DO_ACCOUNTING;
 594                         unsigned long page_error_op;
 595
 596                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 597
 598                         /*
 599                          * inline extent creation worked or returned error,
 600                          * we don't need to create any more async work items.
 601                          * Unlock and free up our temp pages.
 602                          *
 603                          * We use DO_ACCOUNTING here because we need the
 604                          * delalloc_release_metadata to be done _after_ we drop
 605                          * our outstanding extent for clearing delalloc for this
 606                          * range.
 607                          */
 608                         extent_clear_unlock_delalloc(inode, start, end, end,
 609                                                      NULL, clear_flags,
 610                                                      PAGE_UNLOCK |
 611                                                      PAGE_CLEAR_DIRTY |
 612                                                      PAGE_SET_WRITEBACK |
 613                                                      page_error_op |
 614                                                      PAGE_END_WRITEBACK);
 615                         goto free_pages_out;
 616                 }
 617         }
 618
 619         if (will_compress) {
 620                 /*
 621                  * we aren't doing an inline extent round the compressed size
 622                  * up to a block size boundary so the allocator does sane
 623                  * things
 624                  */
 625                 total_compressed = ALIGN(total_compressed, blocksize);
 626
 627                 /*
 628                  * one last check to make sure the compression is really a
 629                  * win, compare the page count read with the blocks on disk,
 630                  * compression must free at least one sector size
 631                  */
 632                 total_in = ALIGN(total_in, PAGE_SIZE);
 633                 if (total_compressed + blocksize <= total_in) {
 634                         *num_added += 1;
 635
 636                         /*
 637                          * The async work queues will take care of doing actual
 638                          * allocation on disk for these compressed pages, and
 639                          * will submit them to the elevator.
 640                          */
 641                         add_async_extent(async_cow, start, total_in,
 642                                         total_compressed, pages, nr_pages,
 643                                         compress_type);
 644
 645                         if (start + total_in < end) {
 646                                 start += total_in;
 647                                 pages = NULL;
 648                                 cond_resched();
 649                                 goto again;
 650                         }
 651                         return;
 652                 }
 653         }
 654         if (pages) {
 655                 /*
 656                  * the compression code ran but failed to make things smaller,
 657                  * free any pages it allocated and our page pointer array
 658                  */
 659                 for (i = 0; i < nr_pages; i++) {
 660                         WARN_ON(pages[i]->mapping);
 661                         put_page(pages[i]);
 662                 }
 663                 kfree(pages);
 664                 pages = NULL;
 665                 total_compressed = 0;
 666                 nr_pages = 0;
 667
 668                 /* flag the file so we don't compress in the future */
 669                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 670                     !(BTRFS_I(inode)->prop_compress)) {
 671                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 672                 }
 673         }
 674 cleanup_and_bail_uncompressed:
 675         /*
 676          * No compression, but we still need to write the pages in the file
 677          * we've been given so far.  redirty the locked page if it corresponds
 678          * to our extent and set things up for the async work queue to run
 679          * cow_file_range to do the normal delalloc dance.
 680          */
 681         if (page_offset(locked_page) >= start &&
 682             page_offset(locked_page) <= end)
 683                 __set_page_dirty_nobuffers(locked_page);
 684                 /* unlocked later on in the async handlers */
 685
 686         if (redirty)
 687                 extent_range_redirty_for_io(inode, start, end);
 688         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 689                          BTRFS_COMPRESS_NONE);
 690         *num_added += 1;
 691
 692         return;
 693
 694 free_pages_out:
 695         for (i = 0; i < nr_pages; i++) {
 696                 WARN_ON(pages[i]->mapping);
 697                 put_page(pages[i]);
 698         }
 699         kfree(pages);
 700 }
 701
 702 static void free_async_extent_pages(struct async_extent *async_extent)
 703 {
 704         int i;
 705
 706         if (!async_extent->pages)
 707                 return;
 708
 709         for (i = 0; i < async_extent->nr_pages; i++) {
 710                 WARN_ON(async_extent->pages[i]->mapping);
 711                 put_page(async_extent->pages[i]);
 712         }
 713         kfree(async_extent->pages);
 714         async_extent->nr_pages = 0;
 715         async_extent->pages = NULL;
 716 }
 717
 718 /*
 719  * phase two of compressed writeback.  This is the ordered portion
 720  * of the code, which only gets called in the order the work was
 721  * queued.  We walk all the async extents created by compress_file_range
 722  * and send them down to the disk.
 723  */
 724 static noinline void submit_compressed_extents(struct inode *inode,
 725                                               struct async_cow *async_cow)
 726 {
 727         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 728         struct async_extent *async_extent;
 729         u64 alloc_hint = 0;
 730         struct btrfs_key ins;
 731         struct extent_map *em;
 732         struct btrfs_root *root = BTRFS_I(inode)->root;
 733         struct extent_io_tree *io_tree;
 734         int ret = 0;
 735
 736 again:
 737         while (!list_empty(&async_cow->extents)) {
 738                 async_extent = list_entry(async_cow->extents.next,
 739                                           struct async_extent, list);
 740                 list_del(&async_extent->list);
 741
 742                 io_tree = &BTRFS_I(inode)->io_tree;
 743
 744 retry:
 745                 /* did the compression code fall back to uncompressed IO? */
 746                 if (!async_extent->pages) {
 747                         int page_started = 0;
 748                         unsigned long nr_written = 0;
 749
 750                         lock_extent(io_tree, async_extent->start,
 751                                          async_extent->start +
 752                                          async_extent->ram_size - 1);
 753
 754                         /* allocate blocks */
 755                         ret = cow_file_range(inode, async_cow->locked_page,
 756                                              async_extent->start,
 757                                              async_extent->start +
 758                                              async_extent->ram_size - 1,
 759                                              async_extent->start +
 760                                              async_extent->ram_size - 1,
 761                                              &page_started, &nr_written, 0,
 762                                              NULL);
 763
 764                         /* JDM XXX */
 765
 766                         /*
 767                          * if page_started, cow_file_range inserted an
 768                          * inline extent and took care of all the unlocking
 769                          * and IO for us.  Otherwise, we need to submit
 770                          * all those pages down to the drive.
 771                          */
 772                         if (!page_started && !ret)
 773                                 extent_write_locked_range(inode,
 774                                                   async_extent->start,
 775                                                   async_extent->start +
 776                                                   async_extent->ram_size - 1,
 777                                                   WB_SYNC_ALL);
 778                         else if (ret)
 779                                 unlock_page(async_cow->locked_page);
 780                         kfree(async_extent);
 781                         cond_resched();
 782                         continue;
 783                 }
 784
 785                 lock_extent(io_tree, async_extent->start,
 786                             async_extent->start + async_extent->ram_size - 1);
 787
 788                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 789                                            async_extent->compressed_size,
 790                                            async_extent->compressed_size,
 791                                            0, alloc_hint, &ins, 1, 1);
 792                 if (ret) {
 793                         free_async_extent_pages(async_extent);
 794
 795                         if (ret == -ENOSPC) {
 796                                 unlock_extent(io_tree, async_extent->start,
 797                                               async_extent->start +
 798                                               async_extent->ram_size - 1);
 799
 800                                 /*
 801                                  * we need to redirty the pages if we decide to
 802                                  * fallback to uncompressed IO, otherwise we
 803                                  * will not submit these pages down to lower
 804                                  * layers.
 805                                  */
 806                                 extent_range_redirty_for_io(inode,
 807                                                 async_extent->start,
 808                                                 async_extent->start +
 809                                                 async_extent->ram_size - 1);
 810
 811                                 goto retry;
 812                         }
 813                         goto out_free;
 814                 }
 815                 /*
 816                  * here we're doing allocation and writeback of the
 817                  * compressed pages
 818                  */
 819                 em = create_io_em(inode, async_extent->start,
 820                                   async_extent->ram_size, /* len */
 821                                   async_extent->start, /* orig_start */
 822                                   ins.objectid, /* block_start */
 823                                   ins.offset, /* block_len */
 824                                   ins.offset, /* orig_block_len */
 825                                   async_extent->ram_size, /* ram_bytes */
 826                                   async_extent->compress_type,
 827                                   BTRFS_ORDERED_COMPRESSED);
 828                 if (IS_ERR(em))
 829                         /* ret value is not necessary due to void function */
 830                         goto out_free_reserve;
 831                 free_extent_map(em);
 832
 833                 ret = btrfs_add_ordered_extent_compress(inode,
 834                                                 async_extent->start,
 835                                                 ins.objectid,
 836                                                 async_extent->ram_size,
 837                                                 ins.offset,
 838                                                 BTRFS_ORDERED_COMPRESSED,
 839                                                 async_extent->compress_type);
 840                 if (ret) {
 841                         btrfs_drop_extent_cache(BTRFS_I(inode),
 842                                                 async_extent->start,
 843                                                 async_extent->start +
 844                                                 async_extent->ram_size - 1, 0);
 845                         goto out_free_reserve;
 846                 }
 847                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 848
 849                 /*
 850                  * clear dirty, set writeback and unlock the pages.
 851                  */
 852                 extent_clear_unlock_delalloc(inode, async_extent->start,
 853                                 async_extent->start +
 854                                 async_extent->ram_size - 1,
 855                                 async_extent->start +
 856                                 async_extent->ram_size - 1,
 857                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 858                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 859                                 PAGE_SET_WRITEBACK);
 860                 if (btrfs_submit_compressed_write(inode,
 861                                     async_extent->start,
 862                                     async_extent->ram_size,
 863                                     ins.objectid,
 864                                     ins.offset, async_extent->pages,
 865                                     async_extent->nr_pages,
 866                                     async_cow->write_flags)) {
 867                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 868                         struct page *p = async_extent->pages[0];
 869                         const u64 start = async_extent->start;
 870                         const u64 end = start + async_extent->ram_size - 1;
 871
 872                         p->mapping = inode->i_mapping;
 873                         tree->ops->writepage_end_io_hook(p, start, end,
 874                                                          NULL, 0);
 875                         p->mapping = NULL;
 876                         extent_clear_unlock_delalloc(inode, start, end, end,
 877                                                      NULL, 0,
 878                                                      PAGE_END_WRITEBACK |
 879                                                      PAGE_SET_ERROR);
 880                         free_async_extent_pages(async_extent);
 881                 }
 882                 alloc_hint = ins.objectid + ins.offset;
 883                 kfree(async_extent);
 884                 cond_resched();
 885         }
 886         return;
 887 out_free_reserve:
 888         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 889         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 890 out_free:
 891         extent_clear_unlock_delalloc(inode, async_extent->start,
 892                                      async_extent->start +
 893                                      async_extent->ram_size - 1,
 894                                      async_extent->start +
 895                                      async_extent->ram_size - 1,
 896                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 897                                      EXTENT_DELALLOC_NEW |
 898                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 899                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 900                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 901                                      PAGE_SET_ERROR);
 902         free_async_extent_pages(async_extent);
 903         kfree(async_extent);
 904         goto again;
 905 }
 906
 907 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 908                                       u64 num_bytes)
 909 {
 910         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 911         struct extent_map *em;
 912         u64 alloc_hint = 0;
 913
 914         read_lock(&em_tree->lock);
 915         em = search_extent_mapping(em_tree, start, num_bytes);
 916         if (em) {
 917                 /*
 918                  * if block start isn't an actual block number then find the
 919                  * first block in this inode and use that as a hint.  If that
 920                  * block is also bogus then just don't worry about it.
 921                  */
 922                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 923                         free_extent_map(em);
 924                         em = search_extent_mapping(em_tree, 0, 0);
 925                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 926                                 alloc_hint = em->block_start;
 927                         if (em)
 928                                 free_extent_map(em);
 929                 } else {
 930                         alloc_hint = em->block_start;
 931                         free_extent_map(em);
 932                 }
 933         }
 934         read_unlock(&em_tree->lock);
 935
 936         return alloc_hint;
 937 }
 938
 939 /*
 940  * when extent_io.c finds a delayed allocation range in the file,
 941  * the call backs end up in this code.  The basic idea is to
 942  * allocate extents on disk for the range, and create ordered data structs
 943  * in ram to track those extents.
 944  *
 945  * locked_page is the page that writepage had locked already.  We use
 946  * it to make sure we don't do extra locks or unlocks.
 947  *
 948  * *page_started is set to one if we unlock locked_page and do everything
 949  * required to start IO on it.  It may be clean and already done with
 950  * IO when we return.
 951  */
 952 static noinline int cow_file_range(struct inode *inode,
 953                                    struct page *locked_page,
 954                                    u64 start, u64 end, u64 delalloc_end,
 955                                    int *page_started, unsigned long *nr_written,
 956                                    int unlock, struct btrfs_dedupe_hash *hash)
 957 {
 958         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 959         struct btrfs_root *root = BTRFS_I(inode)->root;
 960         u64 alloc_hint = 0;
 961         u64 num_bytes;
 962         unsigned long ram_size;
 963         u64 disk_num_bytes;
 964         u64 cur_alloc_size = 0;
 965         u64 blocksize = fs_info->sectorsize;
 966         struct btrfs_key ins;
 967         struct extent_map *em;
 968         unsigned clear_bits;
 969         unsigned long page_ops;
 970         bool extent_reserved = false;
 971         int ret = 0;
 972
 973         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 974                 WARN_ON_ONCE(1);
 975                 ret = -EINVAL;
 976                 goto out_unlock;
 977         }
 978
 979         num_bytes = ALIGN(end - start + 1, blocksize);
 980         num_bytes = max(blocksize,  num_bytes);
 981         disk_num_bytes = num_bytes;
 982
 983         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 984
 985         if (start == 0) {
 986                 /* lets try to make an inline extent */
 987                 ret = cow_file_range_inline(root, inode, start, end, 0,
 988                                         BTRFS_COMPRESS_NONE, NULL);
 989                 if (ret == 0) {
 990                         /*
 991                          * We use DO_ACCOUNTING here because we need the
 992                          * delalloc_release_metadata to be run _after_ we drop
 993                          * our outstanding extent for clearing delalloc for this
 994                          * range.
 995                          */
 996                         extent_clear_unlock_delalloc(inode, start, end,
 997                                      delalloc_end, NULL,
 998                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 999                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
1000                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1001                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1002                                      PAGE_END_WRITEBACK);
1003                         *nr_written = *nr_written +
1004                              (end - start + PAGE_SIZE) / PAGE_SIZE;
1005                         *page_started = 1;
1006                         goto out;
1007                 } else if (ret < 0) {
1008                         goto out_unlock;
1009                 }
1010         }
1011
1012         BUG_ON(disk_num_bytes >
1013                btrfs_super_total_bytes(fs_info->super_copy));
1014
1015         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1016         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1017                         start + num_bytes - 1, 0);
1018
1019         while (disk_num_bytes > 0) {
1020                 cur_alloc_size = disk_num_bytes;
1021                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1022                                            fs_info->sectorsize, 0, alloc_hint,
1023                                            &ins, 1, 1);
1024                 if (ret < 0)
1025                         goto out_unlock;
1026                 cur_alloc_size = ins.offset;
1027                 extent_reserved = true;
1028
1029                 ram_size = ins.offset;
1030                 em = create_io_em(inode, start, ins.offset, /* len */
1031                                   start, /* orig_start */
1032                                   ins.objectid, /* block_start */
1033                                   ins.offset, /* block_len */
1034                                   ins.offset, /* orig_block_len */
1035                                   ram_size, /* ram_bytes */
1036                                   BTRFS_COMPRESS_NONE, /* compress_type */
1037                                   BTRFS_ORDERED_REGULAR /* type */);
1038                 if (IS_ERR(em))
1039                         goto out_reserve;
1040                 free_extent_map(em);
1041
1042                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1043                                                ram_size, cur_alloc_size, 0);
1044                 if (ret)
1045                         goto out_drop_extent_cache;
1046
1047                 if (root->root_key.objectid ==
1048                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1049                         ret = btrfs_reloc_clone_csums(inode, start,
1050                                                       cur_alloc_size);
1051                         /*
1052                          * Only drop cache here, and process as normal.
1053                          *
1054                          * We must not allow extent_clear_unlock_delalloc()
1055                          * at out_unlock label to free meta of this ordered
1056                          * extent, as its meta should be freed by
1057                          * btrfs_finish_ordered_io().
1058                          *
1059                          * So we must continue until @start is increased to
1060                          * skip current ordered extent.
1061                          */
1062                         if (ret)
1063                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1064                                                 start + ram_size - 1, 0);
1065                 }
1066
1067                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1068
1069                 /* we're not doing compressed IO, don't unlock the first
1070                  * page (which the caller expects to stay locked), don't
1071                  * clear any dirty bits and don't set any writeback bits
1072                  *
1073                  * Do set the Private2 bit so we know this page was properly
1074                  * setup for writepage
1075                  */
1076                 page_ops = unlock ? PAGE_UNLOCK : 0;
1077                 page_ops |= PAGE_SET_PRIVATE2;
1078
1079                 extent_clear_unlock_delalloc(inode, start,
1080                                              start + ram_size - 1,
1081                                              delalloc_end, locked_page,
1082                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1083                                              page_ops);
1084                 if (disk_num_bytes < cur_alloc_size)
1085                         disk_num_bytes = 0;
1086                 else
1087                         disk_num_bytes -= cur_alloc_size;
1088                 num_bytes -= cur_alloc_size;
1089                 alloc_hint = ins.objectid + ins.offset;
1090                 start += cur_alloc_size;
1091                 extent_reserved = false;
1092
1093                 /*
1094                  * btrfs_reloc_clone_csums() error, since start is increased
1095                  * extent_clear_unlock_delalloc() at out_unlock label won't
1096                  * free metadata of current ordered extent, we're OK to exit.
1097                  */
1098                 if (ret)
1099                         goto out_unlock;
1100         }
1101 out:
1102         return ret;
1103
1104 out_drop_extent_cache:
1105         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1106 out_reserve:
1107         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1108         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1109 out_unlock:
1110         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1111                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1112         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1113                 PAGE_END_WRITEBACK;
1114         /*
1115          * If we reserved an extent for our delalloc range (or a subrange) and
1116          * failed to create the respective ordered extent, then it means that
1117          * when we reserved the extent we decremented the extent's size from
1118          * the data space_info's bytes_may_use counter and incremented the
1119          * space_info's bytes_reserved counter by the same amount. We must make
1120          * sure extent_clear_unlock_delalloc() does not try to decrement again
1121          * the data space_info's bytes_may_use counter, therefore we do not pass
1122          * it the flag EXTENT_CLEAR_DATA_RESV.
1123          */
1124         if (extent_reserved) {
1125                 extent_clear_unlock_delalloc(inode, start,
1126                                              start + cur_alloc_size,
1127                                              start + cur_alloc_size,
1128                                              locked_page,
1129                                              clear_bits,
1130                                              page_ops);
1131                 start += cur_alloc_size;
1132                 if (start >= end)
1133                         goto out;
1134         }
1135         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1136                                      locked_page,
1137                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1138                                      page_ops);
1139         goto out;
1140 }
1141
1142 /*
1143  * work queue call back to started compression on a file and pages
1144  */
1145 static noinline void async_cow_start(struct btrfs_work *work)
1146 {
1147         struct async_cow *async_cow;
1148         int num_added = 0;
1149         async_cow = container_of(work, struct async_cow, work);
1150
1151         compress_file_range(async_cow->inode, async_cow->locked_page,
1152                             async_cow->start, async_cow->end, async_cow,
1153                             &num_added);
1154         if (num_added == 0) {
1155                 btrfs_add_delayed_iput(async_cow->inode);
1156                 async_cow->inode = NULL;
1157         }
1158 }
1159
1160 /*
1161  * work queue call back to submit previously compressed pages
1162  */
1163 static noinline void async_cow_submit(struct btrfs_work *work)
1164 {
1165         struct btrfs_fs_info *fs_info;
1166         struct async_cow *async_cow;
1167         struct btrfs_root *root;
1168         unsigned long nr_pages;
1169
1170         async_cow = container_of(work, struct async_cow, work);
1171
1172         root = async_cow->root;
1173         fs_info = root->fs_info;
1174         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1175                 PAGE_SHIFT;
1176
1177         /*
1178          * atomic_sub_return implies a barrier for waitqueue_active
1179          */
1180         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1181             5 * SZ_1M &&
1182             waitqueue_active(&fs_info->async_submit_wait))
1183                 wake_up(&fs_info->async_submit_wait);
1184
1185         if (async_cow->inode)
1186                 submit_compressed_extents(async_cow->inode, async_cow);
1187 }
1188
1189 static noinline void async_cow_free(struct btrfs_work *work)
1190 {
1191         struct async_cow *async_cow;
1192         async_cow = container_of(work, struct async_cow, work);
1193         if (async_cow->inode)
1194                 btrfs_add_delayed_iput(async_cow->inode);
1195         kfree(async_cow);
1196 }
1197
1198 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1199                                 u64 start, u64 end, int *page_started,
1200                                 unsigned long *nr_written,
1201                                 unsigned int write_flags)
1202 {
1203         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1204         struct async_cow *async_cow;
1205         struct btrfs_root *root = BTRFS_I(inode)->root;
1206         unsigned long nr_pages;
1207         u64 cur_end;
1208
1209         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1210                          1, 0, NULL);
1211         while (start < end) {
1212                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1213                 BUG_ON(!async_cow); /* -ENOMEM */
1214                 async_cow->inode = igrab(inode);
1215                 async_cow->root = root;
1216                 async_cow->locked_page = locked_page;
1217                 async_cow->start = start;
1218                 async_cow->write_flags = write_flags;
1219
1220                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1221                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1222                         cur_end = end;
1223                 else
1224                         cur_end = min(end, start + SZ_512K - 1);
1225
1226                 async_cow->end = cur_end;
1227                 INIT_LIST_HEAD(&async_cow->extents);
1228
1229                 btrfs_init_work(&async_cow->work,
1230                                 btrfs_delalloc_helper,
1231                                 async_cow_start, async_cow_submit,
1232                                 async_cow_free);
1233
1234                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1235                         PAGE_SHIFT;
1236                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1237
1238                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1239
1240                 *nr_written += nr_pages;
1241                 start = cur_end + 1;
1242         }
1243         *page_started = 1;
1244         return 0;
1245 }
1246
1247 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1248                                         u64 bytenr, u64 num_bytes)
1249 {
1250         int ret;
1251         struct btrfs_ordered_sum *sums;
1252         LIST_HEAD(list);
1253
1254         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1255                                        bytenr + num_bytes - 1, &list, 0);
1256         if (ret == 0 && list_empty(&list))
1257                 return 0;
1258
1259         while (!list_empty(&list)) {
1260                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1261                 list_del(&sums->list);
1262                 kfree(sums);
1263         }
1264         return 1;
1265 }
1266
1267 /*
1268  * when nowcow writeback call back.  This checks for snapshots or COW copies
1269  * of the extents that exist in the file, and COWs the file as required.
1270  *
1271  * If no cow copies or snapshots exist, we write directly to the existing
1272  * blocks on disk
1273  */
1274 static noinline int run_delalloc_nocow(struct inode *inode,
1275                                        struct page *locked_page,
1276                               u64 start, u64 end, int *page_started, int force,
1277                               unsigned long *nr_written)
1278 {
1279         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1280         struct btrfs_root *root = BTRFS_I(inode)->root;
1281         struct extent_buffer *leaf;
1282         struct btrfs_path *path;
1283         struct btrfs_file_extent_item *fi;
1284         struct btrfs_key found_key;
1285         struct extent_map *em;
1286         u64 cow_start;
1287         u64 cur_offset;
1288         u64 extent_end;
1289         u64 extent_offset;
1290         u64 disk_bytenr;
1291         u64 num_bytes;
1292         u64 disk_num_bytes;
1293         u64 ram_bytes;
1294         int extent_type;
1295         int ret, err;
1296         int type;
1297         int nocow;
1298         int check_prev = 1;
1299         bool nolock;
1300         u64 ino = btrfs_ino(BTRFS_I(inode));
1301
1302         path = btrfs_alloc_path();
1303         if (!path) {
1304                 extent_clear_unlock_delalloc(inode, start, end, end,
1305                                              locked_page,
1306                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1307                                              EXTENT_DO_ACCOUNTING |
1308                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1309                                              PAGE_CLEAR_DIRTY |
1310                                              PAGE_SET_WRITEBACK |
1311                                              PAGE_END_WRITEBACK);
1312                 return -ENOMEM;
1313         }
1314
1315         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1316
1317         cow_start = (u64)-1;
1318         cur_offset = start;
1319         while (1) {
1320                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1321                                                cur_offset, 0);
1322                 if (ret < 0)
1323                         goto error;
1324                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1325                         leaf = path->nodes[0];
1326                         btrfs_item_key_to_cpu(leaf, &found_key,
1327                                               path->slots[0] - 1);
1328                         if (found_key.objectid == ino &&
1329                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1330                                 path->slots[0]--;
1331                 }
1332                 check_prev = 0;
1333 next_slot:
1334                 leaf = path->nodes[0];
1335                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1336                         ret = btrfs_next_leaf(root, path);
1337                         if (ret < 0) {
1338                                 if (cow_start != (u64)-1)
1339                                         cur_offset = cow_start;
1340                                 goto error;
1341                         }
1342                         if (ret > 0)
1343                                 break;
1344                         leaf = path->nodes[0];
1345                 }
1346
1347                 nocow = 0;
1348                 disk_bytenr = 0;
1349                 num_bytes = 0;
1350                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1351
1352                 if (found_key.objectid > ino)
1353                         break;
1354                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1355                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1356                         path->slots[0]++;
1357                         goto next_slot;
1358                 }
1359                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1360                     found_key.offset > end)
1361                         break;
1362
1363                 if (found_key.offset > cur_offset) {
1364                         extent_end = found_key.offset;
1365                         extent_type = 0;
1366                         goto out_check;
1367                 }
1368
1369                 fi = btrfs_item_ptr(leaf, path->slots[0],
1370                                     struct btrfs_file_extent_item);
1371                 extent_type = btrfs_file_extent_type(leaf, fi);
1372
1373                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1374                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1375                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1376                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1377                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1378                         extent_end = found_key.offset +
1379                                 btrfs_file_extent_num_bytes(leaf, fi);
1380                         disk_num_bytes =
1381                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1382                         if (extent_end <= start) {
1383                                 path->slots[0]++;
1384                                 goto next_slot;
1385                         }
1386                         if (disk_bytenr == 0)
1387                                 goto out_check;
1388                         if (btrfs_file_extent_compression(leaf, fi) ||
1389                             btrfs_file_extent_encryption(leaf, fi) ||
1390                             btrfs_file_extent_other_encoding(leaf, fi))
1391                                 goto out_check;
1392                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1393                                 goto out_check;
1394                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1395                                 goto out_check;
1396                         if (btrfs_cross_ref_exist(root, ino,
1397                                                   found_key.offset -
1398                                                   extent_offset, disk_bytenr))
1399                                 goto out_check;
1400                         disk_bytenr += extent_offset;
1401                         disk_bytenr += cur_offset - found_key.offset;
1402                         num_bytes = min(end + 1, extent_end) - cur_offset;
1403                         /*
1404                          * if there are pending snapshots for this root,
1405                          * we fall into common COW way.
1406                          */
1407                         if (!nolock) {
1408                                 err = btrfs_start_write_no_snapshotting(root);
1409                                 if (!err)
1410                                         goto out_check;
1411                         }
1412                         /*
1413                          * force cow if csum exists in the range.
1414                          * this ensure that csum for a given extent are
1415                          * either valid or do not exist.
1416                          */
1417                         if (csum_exist_in_range(fs_info, disk_bytenr,
1418                                                 num_bytes)) {
1419                                 if (!nolock)
1420                                         btrfs_end_write_no_snapshotting(root);
1421                                 goto out_check;
1422                         }
1423                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr)) {
1424                                 if (!nolock)
1425                                         btrfs_end_write_no_snapshotting(root);
1426                                 goto out_check;
1427                         }
1428                         nocow = 1;
1429                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1430                         extent_end = found_key.offset +
1431                                 btrfs_file_extent_inline_len(leaf,
1432                                                      path->slots[0], fi);
1433                         extent_end = ALIGN(extent_end,
1434                                            fs_info->sectorsize);
1435                 } else {
1436                         BUG_ON(1);
1437                 }
1438 out_check:
1439                 if (extent_end <= start) {
1440                         path->slots[0]++;
1441                         if (!nolock && nocow)
1442                                 btrfs_end_write_no_snapshotting(root);
1443                         if (nocow)
1444                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1445                         goto next_slot;
1446                 }
1447                 if (!nocow) {
1448                         if (cow_start == (u64)-1)
1449                                 cow_start = cur_offset;
1450                         cur_offset = extent_end;
1451                         if (cur_offset > end)
1452                                 break;
1453                         path->slots[0]++;
1454                         goto next_slot;
1455                 }
1456
1457                 btrfs_release_path(path);
1458                 if (cow_start != (u64)-1) {
1459                         ret = cow_file_range(inode, locked_page,
1460                                              cow_start, found_key.offset - 1,
1461                                              end, page_started, nr_written, 1,
1462                                              NULL);
1463                         if (ret) {
1464                                 if (!nolock && nocow)
1465                                         btrfs_end_write_no_snapshotting(root);
1466                                 if (nocow)
1467                                         btrfs_dec_nocow_writers(fs_info,
1468                                                                 disk_bytenr);
1469                                 goto error;
1470                         }
1471                         cow_start = (u64)-1;
1472                 }
1473
1474                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1475                         u64 orig_start = found_key.offset - extent_offset;
1476
1477                         em = create_io_em(inode, cur_offset, num_bytes,
1478                                           orig_start,
1479                                           disk_bytenr, /* block_start */
1480                                           num_bytes, /* block_len */
1481                                           disk_num_bytes, /* orig_block_len */
1482                                           ram_bytes, BTRFS_COMPRESS_NONE,
1483                                           BTRFS_ORDERED_PREALLOC);
1484                         if (IS_ERR(em)) {
1485                                 if (!nolock && nocow)
1486                                         btrfs_end_write_no_snapshotting(root);
1487                                 if (nocow)
1488                                         btrfs_dec_nocow_writers(fs_info,
1489                                                                 disk_bytenr);
1490                                 ret = PTR_ERR(em);
1491                                 goto error;
1492                         }
1493                         free_extent_map(em);
1494                 }
1495
1496                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1497                         type = BTRFS_ORDERED_PREALLOC;
1498                 } else {
1499                         type = BTRFS_ORDERED_NOCOW;
1500                 }
1501
1502                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1503                                                num_bytes, num_bytes, type);
1504                 if (nocow)
1505                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1506                 BUG_ON(ret); /* -ENOMEM */
1507
1508                 if (root->root_key.objectid ==
1509                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1510                         /*
1511                          * Error handled later, as we must prevent
1512                          * extent_clear_unlock_delalloc() in error handler
1513                          * from freeing metadata of created ordered extent.
1514                          */
1515                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1516                                                       num_bytes);
1517
1518                 extent_clear_unlock_delalloc(inode, cur_offset,
1519                                              cur_offset + num_bytes - 1, end,
1520                                              locked_page, EXTENT_LOCKED |
1521                                              EXTENT_DELALLOC |
1522                                              EXTENT_CLEAR_DATA_RESV,
1523                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1524
1525                 if (!nolock && nocow)
1526                         btrfs_end_write_no_snapshotting(root);
1527                 cur_offset = extent_end;
1528
1529                 /*
1530                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1531                  * handler, as metadata for created ordered extent will only
1532                  * be freed by btrfs_finish_ordered_io().
1533                  */
1534                 if (ret)
1535                         goto error;
1536                 if (cur_offset > end)
1537                         break;
1538         }
1539         btrfs_release_path(path);
1540
1541         if (cur_offset <= end && cow_start == (u64)-1) {
1542                 cow_start = cur_offset;
1543                 cur_offset = end;
1544         }
1545
1546         if (cow_start != (u64)-1) {
1547                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1548                                      page_started, nr_written, 1, NULL);
1549                 if (ret)
1550                         goto error;
1551         }
1552
1553 error:
1554         if (ret && cur_offset < end)
1555                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1556                                              locked_page, EXTENT_LOCKED |
1557                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1558                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1559                                              PAGE_CLEAR_DIRTY |
1560                                              PAGE_SET_WRITEBACK |
1561                                              PAGE_END_WRITEBACK);
1562         btrfs_free_path(path);
1563         return ret;
1564 }
1565
1566 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1567 {
1568
1569         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1570             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1571                 return 0;
1572
1573         /*
1574          * @defrag_bytes is a hint value, no spinlock held here,
1575          * if is not zero, it means the file is defragging.
1576          * Force cow if given extent needs to be defragged.
1577          */
1578         if (BTRFS_I(inode)->defrag_bytes &&
1579             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1580                            EXTENT_DEFRAG, 0, NULL))
1581                 return 1;
1582
1583         return 0;
1584 }
1585
1586 /*
1587  * extent_io.c call back to do delayed allocation processing
1588  */
1589 static int run_delalloc_range(void *private_data, struct page *locked_page,
1590                               u64 start, u64 end, int *page_started,
1591                               unsigned long *nr_written,
1592                               struct writeback_control *wbc)
1593 {
1594         struct inode *inode = private_data;
1595         int ret;
1596         int force_cow = need_force_cow(inode, start, end);
1597         unsigned int write_flags = wbc_to_write_flags(wbc);
1598
1599         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1600                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1601                                          page_started, 1, nr_written);
1602         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1603                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1604                                          page_started, 0, nr_written);
1605         } else if (!inode_need_compress(inode, start, end)) {
1606                 ret = cow_file_range(inode, locked_page, start, end, end,
1607                                       page_started, nr_written, 1, NULL);
1608         } else {
1609                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1610                         &BTRFS_I(inode)->runtime_flags);
1611                 ret = cow_file_range_async(inode, locked_page, start, end,
1612                                            page_started, nr_written,
1613                                            write_flags);
1614         }
1615         if (ret)
1616                 btrfs_cleanup_ordered_extents(inode, start, end - start + 1);
1617         return ret;
1618 }
1619
1620 static void btrfs_split_extent_hook(void *private_data,
1621                                     struct extent_state *orig, u64 split)
1622 {
1623         struct inode *inode = private_data;
1624         u64 size;
1625
1626         /* not delalloc, ignore it */
1627         if (!(orig->state & EXTENT_DELALLOC))
1628                 return;
1629
1630         size = orig->end - orig->start + 1;
1631         if (size > BTRFS_MAX_EXTENT_SIZE) {
1632                 u32 num_extents;
1633                 u64 new_size;
1634
1635                 /*
1636                  * See the explanation in btrfs_merge_extent_hook, the same
1637                  * applies here, just in reverse.
1638                  */
1639                 new_size = orig->end - split + 1;
1640                 num_extents = count_max_extents(new_size);
1641                 new_size = split - orig->start;
1642                 num_extents += count_max_extents(new_size);
1643                 if (count_max_extents(size) >= num_extents)
1644                         return;
1645         }
1646
1647         spin_lock(&BTRFS_I(inode)->lock);
1648         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1649         spin_unlock(&BTRFS_I(inode)->lock);
1650 }
1651
1652 /*
1653  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1654  * extents so we can keep track of new extents that are just merged onto old
1655  * extents, such as when we are doing sequential writes, so we can properly
1656  * account for the metadata space we'll need.
1657  */
1658 static void btrfs_merge_extent_hook(void *private_data,
1659                                     struct extent_state *new,
1660                                     struct extent_state *other)
1661 {
1662         struct inode *inode = private_data;
1663         u64 new_size, old_size;
1664         u32 num_extents;
1665
1666         /* not delalloc, ignore it */
1667         if (!(other->state & EXTENT_DELALLOC))
1668                 return;
1669
1670         if (new->start > other->start)
1671                 new_size = new->end - other->start + 1;
1672         else
1673                 new_size = other->end - new->start + 1;
1674
1675         /* we're not bigger than the max, unreserve the space and go */
1676         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1677                 spin_lock(&BTRFS_I(inode)->lock);
1678                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1679                 spin_unlock(&BTRFS_I(inode)->lock);
1680                 return;
1681         }
1682
1683         /*
1684          * We have to add up either side to figure out how many extents were
1685          * accounted for before we merged into one big extent.  If the number of
1686          * extents we accounted for is <= the amount we need for the new range
1687          * then we can return, otherwise drop.  Think of it like this
1688          *
1689          * [ 4k][MAX_SIZE]
1690          *
1691          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1692          * need 2 outstanding extents, on one side we have 1 and the other side
1693          * we have 1 so they are == and we can return.  But in this case
1694          *
1695          * [MAX_SIZE+4k][MAX_SIZE+4k]
1696          *
1697          * Each range on their own accounts for 2 extents, but merged together
1698          * they are only 3 extents worth of accounting, so we need to drop in
1699          * this case.
1700          */
1701         old_size = other->end - other->start + 1;
1702         num_extents = count_max_extents(old_size);
1703         old_size = new->end - new->start + 1;
1704         num_extents += count_max_extents(old_size);
1705         if (count_max_extents(new_size) >= num_extents)
1706                 return;
1707
1708         spin_lock(&BTRFS_I(inode)->lock);
1709         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1710         spin_unlock(&BTRFS_I(inode)->lock);
1711 }
1712
1713 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1714                                       struct inode *inode)
1715 {
1716         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1717
1718         spin_lock(&root->delalloc_lock);
1719         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1720                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1721                               &root->delalloc_inodes);
1722                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1723                         &BTRFS_I(inode)->runtime_flags);
1724                 root->nr_delalloc_inodes++;
1725                 if (root->nr_delalloc_inodes == 1) {
1726                         spin_lock(&fs_info->delalloc_root_lock);
1727                         BUG_ON(!list_empty(&root->delalloc_root));
1728                         list_add_tail(&root->delalloc_root,
1729                                       &fs_info->delalloc_roots);
1730                         spin_unlock(&fs_info->delalloc_root_lock);
1731                 }
1732         }
1733         spin_unlock(&root->delalloc_lock);
1734 }
1735
1736 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1737                                      struct btrfs_inode *inode)
1738 {
1739         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1740
1741         spin_lock(&root->delalloc_lock);
1742         if (!list_empty(&inode->delalloc_inodes)) {
1743                 list_del_init(&inode->delalloc_inodes);
1744                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1745                           &inode->runtime_flags);
1746                 root->nr_delalloc_inodes--;
1747                 if (!root->nr_delalloc_inodes) {
1748                         spin_lock(&fs_info->delalloc_root_lock);
1749                         BUG_ON(list_empty(&root->delalloc_root));
1750                         list_del_init(&root->delalloc_root);
1751                         spin_unlock(&fs_info->delalloc_root_lock);
1752                 }
1753         }
1754         spin_unlock(&root->delalloc_lock);
1755 }
1756
1757 /*
1758  * extent_io.c set_bit_hook, used to track delayed allocation
1759  * bytes in this file, and to maintain the list of inodes that
1760  * have pending delalloc work to be done.
1761  */
1762 static void btrfs_set_bit_hook(void *private_data,
1763                                struct extent_state *state, unsigned *bits)
1764 {
1765         struct inode *inode = private_data;
1766
1767         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1768
1769         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1770                 WARN_ON(1);
1771         /*
1772          * set_bit and clear bit hooks normally require _irqsave/restore
1773          * but in this case, we are only testing for the DELALLOC
1774          * bit, which is only set or cleared with irqs on
1775          */
1776         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1777                 struct btrfs_root *root = BTRFS_I(inode)->root;
1778                 u64 len = state->end + 1 - state->start;
1779                 u32 num_extents = count_max_extents(len);
1780                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1781
1782                 spin_lock(&BTRFS_I(inode)->lock);
1783                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1784                 spin_unlock(&BTRFS_I(inode)->lock);
1785
1786                 /* For sanity tests */
1787                 if (btrfs_is_testing(fs_info))
1788                         return;
1789
1790                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1791                                          fs_info->delalloc_batch);
1792                 spin_lock(&BTRFS_I(inode)->lock);
1793                 BTRFS_I(inode)->delalloc_bytes += len;
1794                 if (*bits & EXTENT_DEFRAG)
1795                         BTRFS_I(inode)->defrag_bytes += len;
1796                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1797                                          &BTRFS_I(inode)->runtime_flags))
1798                         btrfs_add_delalloc_inodes(root, inode);
1799                 spin_unlock(&BTRFS_I(inode)->lock);
1800         }
1801
1802         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1803             (*bits & EXTENT_DELALLOC_NEW)) {
1804                 spin_lock(&BTRFS_I(inode)->lock);
1805                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1806                         state->start;
1807                 spin_unlock(&BTRFS_I(inode)->lock);
1808         }
1809 }
1810
1811 /*
1812  * extent_io.c clear_bit_hook, see set_bit_hook for why
1813  */
1814 static void btrfs_clear_bit_hook(void *private_data,
1815                                  struct extent_state *state,
1816                                  unsigned *bits)
1817 {
1818         struct btrfs_inode *inode = BTRFS_I((struct inode *)private_data);
1819         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1820         u64 len = state->end + 1 - state->start;
1821         u32 num_extents = count_max_extents(len);
1822
1823         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1824                 spin_lock(&inode->lock);
1825                 inode->defrag_bytes -= len;
1826                 spin_unlock(&inode->lock);
1827         }
1828
1829         /*
1830          * set_bit and clear bit hooks normally require _irqsave/restore
1831          * but in this case, we are only testing for the DELALLOC
1832          * bit, which is only set or cleared with irqs on
1833          */
1834         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1835                 struct btrfs_root *root = inode->root;
1836                 bool do_list = !btrfs_is_free_space_inode(inode);
1837
1838                 spin_lock(&inode->lock);
1839                 btrfs_mod_outstanding_extents(inode, -num_extents);
1840                 spin_unlock(&inode->lock);
1841
1842                 /*
1843                  * We don't reserve metadata space for space cache inodes so we
1844                  * don't need to call dellalloc_release_metadata if there is an
1845                  * error.
1846                  */
1847                 if (*bits & EXTENT_CLEAR_META_RESV &&
1848                     root != fs_info->tree_root)
1849                         btrfs_delalloc_release_metadata(inode, len);
1850
1851                 /* For sanity tests. */
1852                 if (btrfs_is_testing(fs_info))
1853                         return;
1854
1855                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1856                     do_list && !(state->state & EXTENT_NORESERVE) &&
1857                     (*bits & EXTENT_CLEAR_DATA_RESV))
1858                         btrfs_free_reserved_data_space_noquota(
1859                                         &inode->vfs_inode,
1860                                         state->start, len);
1861
1862                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1863                                          fs_info->delalloc_batch);
1864                 spin_lock(&inode->lock);
1865                 inode->delalloc_bytes -= len;
1866                 if (do_list && inode->delalloc_bytes == 0 &&
1867                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1868                                         &inode->runtime_flags))
1869                         btrfs_del_delalloc_inode(root, inode);
1870                 spin_unlock(&inode->lock);
1871         }
1872
1873         if ((state->state & EXTENT_DELALLOC_NEW) &&
1874             (*bits & EXTENT_DELALLOC_NEW)) {
1875                 spin_lock(&inode->lock);
1876                 ASSERT(inode->new_delalloc_bytes >= len);
1877                 inode->new_delalloc_bytes -= len;
1878                 spin_unlock(&inode->lock);
1879         }
1880 }
1881
1882 /*
1883  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1884  * we don't create bios that span stripes or chunks
1885  *
1886  * return 1 if page cannot be merged to bio
1887  * return 0 if page can be merged to bio
1888  * return error otherwise
1889  */
1890 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1891                          size_t size, struct bio *bio,
1892                          unsigned long bio_flags)
1893 {
1894         struct inode *inode = page->mapping->host;
1895         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1896         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1897         u64 length = 0;
1898         u64 map_length;
1899         int ret;
1900
1901         if (bio_flags & EXTENT_BIO_COMPRESSED)
1902                 return 0;
1903
1904         length = bio->bi_iter.bi_size;
1905         map_length = length;
1906         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1907                               NULL, 0);
1908         if (ret < 0)
1909                 return ret;
1910         if (map_length < length + size)
1911                 return 1;
1912         return 0;
1913 }
1914
1915 /*
1916  * in order to insert checksums into the metadata in large chunks,
1917  * we wait until bio submission time.   All the pages in the bio are
1918  * checksummed and sums are attached onto the ordered extent record.
1919  *
1920  * At IO completion time the cums attached on the ordered extent record
1921  * are inserted into the btree
1922  */
1923 static blk_status_t __btrfs_submit_bio_start(void *private_data, struct bio *bio,
1924                                     int mirror_num, unsigned long bio_flags,
1925                                     u64 bio_offset)
1926 {
1927         struct inode *inode = private_data;
1928         blk_status_t ret = 0;
1929
1930         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1931         BUG_ON(ret); /* -ENOMEM */
1932         return 0;
1933 }
1934
1935 /*
1936  * in order to insert checksums into the metadata in large chunks,
1937  * we wait until bio submission time.   All the pages in the bio are
1938  * checksummed and sums are attached onto the ordered extent record.
1939  *
1940  * At IO completion time the cums attached on the ordered extent record
1941  * are inserted into the btree
1942  */
1943 static blk_status_t __btrfs_submit_bio_done(void *private_data, struct bio *bio,
1944                           int mirror_num, unsigned long bio_flags,
1945                           u64 bio_offset)
1946 {
1947         struct inode *inode = private_data;
1948         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1949         blk_status_t ret;
1950
1951         ret = btrfs_map_bio(fs_info, bio, mirror_num, 1);
1952         if (ret) {
1953                 bio->bi_status = ret;
1954                 bio_endio(bio);
1955         }
1956         return ret;
1957 }
1958
1959 /*
1960  * extent_io.c submission hook. This does the right thing for csum calculation
1961  * on write, or reading the csums from the tree before a read.
1962  *
1963  * Rules about async/sync submit,
1964  * a) read:                             sync submit
1965  *
1966  * b) write without checksum:           sync submit
1967  *
1968  * c) write with checksum:
1969  *    c-1) if bio is issued by fsync:   sync submit
1970  *         (sync_writers != 0)
1971  *
1972  *    c-2) if root is reloc root:       sync submit
1973  *         (only in case of buffered IO)
1974  *
1975  *    c-3) otherwise:                   async submit
1976  */
1977 static blk_status_t btrfs_submit_bio_hook(void *private_data, struct bio *bio,
1978                                  int mirror_num, unsigned long bio_flags,
1979                                  u64 bio_offset)
1980 {
1981         struct inode *inode = private_data;
1982         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1983         struct btrfs_root *root = BTRFS_I(inode)->root;
1984         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1985         blk_status_t ret = 0;
1986         int skip_sum;
1987         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1988
1989         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1990
1991         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
1992                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1993
1994         if (bio_op(bio) != REQ_OP_WRITE) {
1995                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
1996                 if (ret)
1997                         goto out;
1998
1999                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
2000                         ret = btrfs_submit_compressed_read(inode, bio,
2001                                                            mirror_num,
2002                                                            bio_flags);
2003                         goto out;
2004                 } else if (!skip_sum) {
2005                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
2006                         if (ret)
2007                                 goto out;
2008                 }
2009                 goto mapit;
2010         } else if (async && !skip_sum) {
2011                 /* csum items have already been cloned */
2012                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2013                         goto mapit;
2014                 /* we're doing a write, do the async checksumming */
2015                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2016                                           bio_offset, inode,
2017                                           __btrfs_submit_bio_start,
2018                                           __btrfs_submit_bio_done);
2019                 goto out;
2020         } else if (!skip_sum) {
2021                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2022                 if (ret)
2023                         goto out;
2024         }
2025
2026 mapit:
2027         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
2028
2029 out:
2030         if (ret) {
2031                 bio->bi_status = ret;
2032                 bio_endio(bio);
2033         }
2034         return ret;
2035 }
2036
2037 /*
2038  * given a list of ordered sums record them in the inode.  This happens
2039  * at IO completion time based on sums calculated at bio submission time.
2040  */
2041 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2042                              struct inode *inode, struct list_head *list)
2043 {
2044         struct btrfs_ordered_sum *sum;
2045
2046         list_for_each_entry(sum, list, list) {
2047                 trans->adding_csums = true;
2048                 btrfs_csum_file_blocks(trans,
2049                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2050                 trans->adding_csums = false;
2051         }
2052         return 0;
2053 }
2054
2055 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2056                               unsigned int extra_bits,
2057                               struct extent_state **cached_state, int dedupe)
2058 {
2059         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
2060         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2061                                    extra_bits, cached_state);
2062 }
2063
2064 /* see btrfs_writepage_start_hook for details on why this is required */
2065 struct btrfs_writepage_fixup {
2066         struct page *page;
2067         struct btrfs_work work;
2068 };
2069
2070 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2071 {
2072         struct btrfs_writepage_fixup *fixup;
2073         struct btrfs_ordered_extent *ordered;
2074         struct extent_state *cached_state = NULL;
2075         struct extent_changeset *data_reserved = NULL;
2076         struct page *page;
2077         struct inode *inode;
2078         u64 page_start;
2079         u64 page_end;
2080         int ret;
2081
2082         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2083         page = fixup->page;
2084 again:
2085         lock_page(page);
2086         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2087                 ClearPageChecked(page);
2088                 goto out_page;
2089         }
2090
2091         inode = page->mapping->host;
2092         page_start = page_offset(page);
2093         page_end = page_offset(page) + PAGE_SIZE - 1;
2094
2095         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2096                          &cached_state);
2097
2098         /* already ordered? We're done */
2099         if (PagePrivate2(page))
2100                 goto out;
2101
2102         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2103                                         PAGE_SIZE);
2104         if (ordered) {
2105                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2106                                      page_end, &cached_state);
2107                 unlock_page(page);
2108                 btrfs_start_ordered_extent(inode, ordered, 1);
2109                 btrfs_put_ordered_extent(ordered);
2110                 goto again;
2111         }
2112
2113         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2114                                            PAGE_SIZE);
2115         if (ret) {
2116                 mapping_set_error(page->mapping, ret);
2117                 end_extent_writepage(page, ret, page_start, page_end);
2118                 ClearPageChecked(page);
2119                 goto out;
2120          }
2121
2122         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2123                                         &cached_state, 0);
2124         if (ret) {
2125                 mapping_set_error(page->mapping, ret);
2126                 end_extent_writepage(page, ret, page_start, page_end);
2127                 ClearPageChecked(page);
2128                 goto out;
2129         }
2130
2131         ClearPageChecked(page);
2132         set_page_dirty(page);
2133         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2134 out:
2135         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2136                              &cached_state);
2137 out_page:
2138         unlock_page(page);
2139         put_page(page);
2140         kfree(fixup);
2141         extent_changeset_free(data_reserved);
2142 }
2143
2144 /*
2145  * There are a few paths in the higher layers of the kernel that directly
2146  * set the page dirty bit without asking the filesystem if it is a
2147  * good idea.  This causes problems because we want to make sure COW
2148  * properly happens and the data=ordered rules are followed.
2149  *
2150  * In our case any range that doesn't have the ORDERED bit set
2151  * hasn't been properly setup for IO.  We kick off an async process
2152  * to fix it up.  The async helper will wait for ordered extents, set
2153  * the delalloc bit and make it safe to write the page.
2154  */
2155 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2156 {
2157         struct inode *inode = page->mapping->host;
2158         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2159         struct btrfs_writepage_fixup *fixup;
2160
2161         /* this page is properly in the ordered list */
2162         if (TestClearPagePrivate2(page))
2163                 return 0;
2164
2165         if (PageChecked(page))
2166                 return -EAGAIN;
2167
2168         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2169         if (!fixup)
2170                 return -EAGAIN;
2171
2172         SetPageChecked(page);
2173         get_page(page);
2174         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2175                         btrfs_writepage_fixup_worker, NULL, NULL);
2176         fixup->page = page;
2177         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2178         return -EBUSY;
2179 }
2180
2181 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2182                                        struct inode *inode, u64 file_pos,
2183                                        u64 disk_bytenr, u64 disk_num_bytes,
2184                                        u64 num_bytes, u64 ram_bytes,
2185                                        u8 compression, u8 encryption,
2186                                        u16 other_encoding, int extent_type)
2187 {
2188         struct btrfs_root *root = BTRFS_I(inode)->root;
2189         struct btrfs_file_extent_item *fi;
2190         struct btrfs_path *path;
2191         struct extent_buffer *leaf;
2192         struct btrfs_key ins;
2193         u64 qg_released;
2194         int extent_inserted = 0;
2195         int ret;
2196
2197         path = btrfs_alloc_path();
2198         if (!path)
2199                 return -ENOMEM;
2200
2201         /*
2202          * we may be replacing one extent in the tree with another.
2203          * The new extent is pinned in the extent map, and we don't want
2204          * to drop it from the cache until it is completely in the btree.
2205          *
2206          * So, tell btrfs_drop_extents to leave this extent in the cache.
2207          * the caller is expected to unpin it and allow it to be merged
2208          * with the others.
2209          */
2210         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2211                                    file_pos + num_bytes, NULL, 0,
2212                                    1, sizeof(*fi), &extent_inserted);
2213         if (ret)
2214                 goto out;
2215
2216         if (!extent_inserted) {
2217                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2218                 ins.offset = file_pos;
2219                 ins.type = BTRFS_EXTENT_DATA_KEY;
2220
2221                 path->leave_spinning = 1;
2222                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2223                                               sizeof(*fi));
2224                 if (ret)
2225                         goto out;
2226         }
2227         leaf = path->nodes[0];
2228         fi = btrfs_item_ptr(leaf, path->slots[0],
2229                             struct btrfs_file_extent_item);
2230         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2231         btrfs_set_file_extent_type(leaf, fi, extent_type);
2232         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2233         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2234         btrfs_set_file_extent_offset(leaf, fi, 0);
2235         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2236         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2237         btrfs_set_file_extent_compression(leaf, fi, compression);
2238         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2239         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2240
2241         btrfs_mark_buffer_dirty(leaf);
2242         btrfs_release_path(path);
2243
2244         inode_add_bytes(inode, num_bytes);
2245
2246         ins.objectid = disk_bytenr;
2247         ins.offset = disk_num_bytes;
2248         ins.type = BTRFS_EXTENT_ITEM_KEY;
2249
2250         /*
2251          * Release the reserved range from inode dirty range map, as it is
2252          * already moved into delayed_ref_head
2253          */
2254         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2255         if (ret < 0)
2256                 goto out;
2257         qg_released = ret;
2258         ret = btrfs_alloc_reserved_file_extent(trans, root,
2259                                                btrfs_ino(BTRFS_I(inode)),
2260                                                file_pos, qg_released, &ins);
2261 out:
2262         btrfs_free_path(path);
2263
2264         return ret;
2265 }
2266
2267 /* snapshot-aware defrag */
2268 struct sa_defrag_extent_backref {
2269         struct rb_node node;
2270         struct old_sa_defrag_extent *old;
2271         u64 root_id;
2272         u64 inum;
2273         u64 file_pos;
2274         u64 extent_offset;
2275         u64 num_bytes;
2276         u64 generation;
2277 };
2278
2279 struct old_sa_defrag_extent {
2280         struct list_head list;
2281         struct new_sa_defrag_extent *new;
2282
2283         u64 extent_offset;
2284         u64 bytenr;
2285         u64 offset;
2286         u64 len;
2287         int count;
2288 };
2289
2290 struct new_sa_defrag_extent {
2291         struct rb_root root;
2292         struct list_head head;
2293         struct btrfs_path *path;
2294         struct inode *inode;
2295         u64 file_pos;
2296         u64 len;
2297         u64 bytenr;
2298         u64 disk_len;
2299         u8 compress_type;
2300 };
2301
2302 static int backref_comp(struct sa_defrag_extent_backref *b1,
2303                         struct sa_defrag_extent_backref *b2)
2304 {
2305         if (b1->root_id < b2->root_id)
2306                 return -1;
2307         else if (b1->root_id > b2->root_id)
2308                 return 1;
2309
2310         if (b1->inum < b2->inum)
2311                 return -1;
2312         else if (b1->inum > b2->inum)
2313                 return 1;
2314
2315         if (b1->file_pos < b2->file_pos)
2316                 return -1;
2317         else if (b1->file_pos > b2->file_pos)
2318                 return 1;
2319
2320         /*
2321          * [------------------------------] ===> (a range of space)
2322          *     |<--->|   |<---->| =============> (fs/file tree A)
2323          * |<---------------------------->| ===> (fs/file tree B)
2324          *
2325          * A range of space can refer to two file extents in one tree while
2326          * refer to only one file extent in another tree.
2327          *
2328          * So we may process a disk offset more than one time(two extents in A)
2329          * and locate at the same extent(one extent in B), then insert two same
2330          * backrefs(both refer to the extent in B).
2331          */
2332         return 0;
2333 }
2334
2335 static void backref_insert(struct rb_root *root,
2336                            struct sa_defrag_extent_backref *backref)
2337 {
2338         struct rb_node **p = &root->rb_node;
2339         struct rb_node *parent = NULL;
2340         struct sa_defrag_extent_backref *entry;
2341         int ret;
2342
2343         while (*p) {
2344                 parent = *p;
2345                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2346
2347                 ret = backref_comp(backref, entry);
2348                 if (ret < 0)
2349                         p = &(*p)->rb_left;
2350                 else
2351                         p = &(*p)->rb_right;
2352         }
2353
2354         rb_link_node(&backref->node, parent, p);
2355         rb_insert_color(&backref->node, root);
2356 }
2357
2358 /*
2359  * Note the backref might has changed, and in this case we just return 0.
2360  */
2361 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2362                                        void *ctx)
2363 {
2364         struct btrfs_file_extent_item *extent;
2365         struct old_sa_defrag_extent *old = ctx;
2366         struct new_sa_defrag_extent *new = old->new;
2367         struct btrfs_path *path = new->path;
2368         struct btrfs_key key;
2369         struct btrfs_root *root;
2370         struct sa_defrag_extent_backref *backref;
2371         struct extent_buffer *leaf;
2372         struct inode *inode = new->inode;
2373         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2374         int slot;
2375         int ret;
2376         u64 extent_offset;
2377         u64 num_bytes;
2378
2379         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2380             inum == btrfs_ino(BTRFS_I(inode)))
2381                 return 0;
2382
2383         key.objectid = root_id;
2384         key.type = BTRFS_ROOT_ITEM_KEY;
2385         key.offset = (u64)-1;
2386
2387         root = btrfs_read_fs_root_no_name(fs_info, &key);
2388         if (IS_ERR(root)) {
2389                 if (PTR_ERR(root) == -ENOENT)
2390                         return 0;
2391                 WARN_ON(1);
2392                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2393                          inum, offset, root_id);
2394                 return PTR_ERR(root);
2395         }
2396
2397         key.objectid = inum;
2398         key.type = BTRFS_EXTENT_DATA_KEY;
2399         if (offset > (u64)-1 << 32)
2400                 key.offset = 0;
2401         else
2402                 key.offset = offset;
2403
2404         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2405         if (WARN_ON(ret < 0))
2406                 return ret;
2407         ret = 0;
2408
2409         while (1) {
2410                 cond_resched();
2411
2412                 leaf = path->nodes[0];
2413                 slot = path->slots[0];
2414
2415                 if (slot >= btrfs_header_nritems(leaf)) {
2416                         ret = btrfs_next_leaf(root, path);
2417                         if (ret < 0) {
2418                                 goto out;
2419                         } else if (ret > 0) {
2420                                 ret = 0;
2421                                 goto out;
2422                         }
2423                         continue;
2424                 }
2425
2426                 path->slots[0]++;
2427
2428                 btrfs_item_key_to_cpu(leaf, &key, slot);
2429
2430                 if (key.objectid > inum)
2431                         goto out;
2432
2433                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2434                         continue;
2435
2436                 extent = btrfs_item_ptr(leaf, slot,
2437                                         struct btrfs_file_extent_item);
2438
2439                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2440                         continue;
2441
2442                 /*
2443                  * 'offset' refers to the exact key.offset,
2444                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2445                  * (key.offset - extent_offset).
2446                  */
2447                 if (key.offset != offset)
2448                         continue;
2449
2450                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2451                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2452
2453                 if (extent_offset >= old->extent_offset + old->offset +
2454                     old->len || extent_offset + num_bytes <=
2455                     old->extent_offset + old->offset)
2456                         continue;
2457                 break;
2458         }
2459
2460         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2461         if (!backref) {
2462                 ret = -ENOENT;
2463                 goto out;
2464         }
2465
2466         backref->root_id = root_id;
2467         backref->inum = inum;
2468         backref->file_pos = offset;
2469         backref->num_bytes = num_bytes;
2470         backref->extent_offset = extent_offset;
2471         backref->generation = btrfs_file_extent_generation(leaf, extent);
2472         backref->old = old;
2473         backref_insert(&new->root, backref);
2474         old->count++;
2475 out:
2476         btrfs_release_path(path);
2477         WARN_ON(ret);
2478         return ret;
2479 }
2480
2481 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2482                                    struct new_sa_defrag_extent *new)
2483 {
2484         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2485         struct old_sa_defrag_extent *old, *tmp;
2486         int ret;
2487
2488         new->path = path;
2489
2490         list_for_each_entry_safe(old, tmp, &new->head, list) {
2491                 ret = iterate_inodes_from_logical(old->bytenr +
2492                                                   old->extent_offset, fs_info,
2493                                                   path, record_one_backref,
2494                                                   old, false);
2495                 if (ret < 0 && ret != -ENOENT)
2496                         return false;
2497
2498                 /* no backref to be processed for this extent */
2499                 if (!old->count) {
2500                         list_del(&old->list);
2501                         kfree(old);
2502                 }
2503         }
2504
2505         if (list_empty(&new->head))
2506                 return false;
2507
2508         return true;
2509 }
2510
2511 static int relink_is_mergable(struct extent_buffer *leaf,
2512                               struct btrfs_file_extent_item *fi,
2513                               struct new_sa_defrag_extent *new)
2514 {
2515         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2516                 return 0;
2517
2518         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2519                 return 0;
2520
2521         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2522                 return 0;
2523
2524         if (btrfs_file_extent_encryption(leaf, fi) ||
2525             btrfs_file_extent_other_encoding(leaf, fi))
2526                 return 0;
2527
2528         return 1;
2529 }
2530
2531 /*
2532  * Note the backref might has changed, and in this case we just return 0.
2533  */
2534 static noinline int relink_extent_backref(struct btrfs_path *path,
2535                                  struct sa_defrag_extent_backref *prev,
2536                                  struct sa_defrag_extent_backref *backref)
2537 {
2538         struct btrfs_file_extent_item *extent;
2539         struct btrfs_file_extent_item *item;
2540         struct btrfs_ordered_extent *ordered;
2541         struct btrfs_trans_handle *trans;
2542         struct btrfs_root *root;
2543         struct btrfs_key key;
2544         struct extent_buffer *leaf;
2545         struct old_sa_defrag_extent *old = backref->old;
2546         struct new_sa_defrag_extent *new = old->new;
2547         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2548         struct inode *inode;
2549         struct extent_state *cached = NULL;
2550         int ret = 0;
2551         u64 start;
2552         u64 len;
2553         u64 lock_start;
2554         u64 lock_end;
2555         bool merge = false;
2556         int index;
2557
2558         if (prev && prev->root_id == backref->root_id &&
2559             prev->inum == backref->inum &&
2560             prev->file_pos + prev->num_bytes == backref->file_pos)
2561                 merge = true;
2562
2563         /* step 1: get root */
2564         key.objectid = backref->root_id;
2565         key.type = BTRFS_ROOT_ITEM_KEY;
2566         key.offset = (u64)-1;
2567
2568         index = srcu_read_lock(&fs_info->subvol_srcu);
2569
2570         root = btrfs_read_fs_root_no_name(fs_info, &key);
2571         if (IS_ERR(root)) {
2572                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2573                 if (PTR_ERR(root) == -ENOENT)
2574                         return 0;
2575                 return PTR_ERR(root);
2576         }
2577
2578         if (btrfs_root_readonly(root)) {
2579                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2580                 return 0;
2581         }
2582
2583         /* step 2: get inode */
2584         key.objectid = backref->inum;
2585         key.type = BTRFS_INODE_ITEM_KEY;
2586         key.offset = 0;
2587
2588         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2589         if (IS_ERR(inode)) {
2590                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2591                 return 0;
2592         }
2593
2594         srcu_read_unlock(&fs_info->subvol_srcu, index);
2595
2596         /* step 3: relink backref */
2597         lock_start = backref->file_pos;
2598         lock_end = backref->file_pos + backref->num_bytes - 1;
2599         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2600                          &cached);
2601
2602         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2603         if (ordered) {
2604                 btrfs_put_ordered_extent(ordered);
2605                 goto out_unlock;
2606         }
2607
2608         trans = btrfs_join_transaction(root);
2609         if (IS_ERR(trans)) {
2610                 ret = PTR_ERR(trans);
2611                 goto out_unlock;
2612         }
2613
2614         key.objectid = backref->inum;
2615         key.type = BTRFS_EXTENT_DATA_KEY;
2616         key.offset = backref->file_pos;
2617
2618         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2619         if (ret < 0) {
2620                 goto out_free_path;
2621         } else if (ret > 0) {
2622                 ret = 0;
2623                 goto out_free_path;
2624         }
2625
2626         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2627                                 struct btrfs_file_extent_item);
2628
2629         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2630             backref->generation)
2631                 goto out_free_path;
2632
2633         btrfs_release_path(path);
2634
2635         start = backref->file_pos;
2636         if (backref->extent_offset < old->extent_offset + old->offset)
2637                 start += old->extent_offset + old->offset -
2638                          backref->extent_offset;
2639
2640         len = min(backref->extent_offset + backref->num_bytes,
2641                   old->extent_offset + old->offset + old->len);
2642         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2643
2644         ret = btrfs_drop_extents(trans, root, inode, start,
2645                                  start + len, 1);
2646         if (ret)
2647                 goto out_free_path;
2648 again:
2649         key.objectid = btrfs_ino(BTRFS_I(inode));
2650         key.type = BTRFS_EXTENT_DATA_KEY;
2651         key.offset = start;
2652
2653         path->leave_spinning = 1;
2654         if (merge) {
2655                 struct btrfs_file_extent_item *fi;
2656                 u64 extent_len;
2657                 struct btrfs_key found_key;
2658
2659                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2660                 if (ret < 0)
2661                         goto out_free_path;
2662
2663                 path->slots[0]--;
2664                 leaf = path->nodes[0];
2665                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2666
2667                 fi = btrfs_item_ptr(leaf, path->slots[0],
2668                                     struct btrfs_file_extent_item);
2669                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2670
2671                 if (extent_len + found_key.offset == start &&
2672                     relink_is_mergable(leaf, fi, new)) {
2673                         btrfs_set_file_extent_num_bytes(leaf, fi,
2674                                                         extent_len + len);
2675                         btrfs_mark_buffer_dirty(leaf);
2676                         inode_add_bytes(inode, len);
2677
2678                         ret = 1;
2679                         goto out_free_path;
2680                 } else {
2681                         merge = false;
2682                         btrfs_release_path(path);
2683                         goto again;
2684                 }
2685         }
2686
2687         ret = btrfs_insert_empty_item(trans, root, path, &key,
2688                                         sizeof(*extent));
2689         if (ret) {
2690                 btrfs_abort_transaction(trans, ret);
2691                 goto out_free_path;
2692         }
2693
2694         leaf = path->nodes[0];
2695         item = btrfs_item_ptr(leaf, path->slots[0],
2696                                 struct btrfs_file_extent_item);
2697         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2698         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2699         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2700         btrfs_set_file_extent_num_bytes(leaf, item, len);
2701         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2702         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2703         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2704         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2705         btrfs_set_file_extent_encryption(leaf, item, 0);
2706         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2707
2708         btrfs_mark_buffer_dirty(leaf);
2709         inode_add_bytes(inode, len);
2710         btrfs_release_path(path);
2711
2712         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2713                         new->disk_len, 0,
2714                         backref->root_id, backref->inum,
2715                         new->file_pos); /* start - extent_offset */
2716         if (ret) {
2717                 btrfs_abort_transaction(trans, ret);
2718                 goto out_free_path;
2719         }
2720
2721         ret = 1;
2722 out_free_path:
2723         btrfs_release_path(path);
2724         path->leave_spinning = 0;
2725         btrfs_end_transaction(trans);
2726 out_unlock:
2727         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2728                              &cached);
2729         iput(inode);
2730         return ret;
2731 }
2732
2733 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2734 {
2735         struct old_sa_defrag_extent *old, *tmp;
2736
2737         if (!new)
2738                 return;
2739
2740         list_for_each_entry_safe(old, tmp, &new->head, list) {
2741                 kfree(old);
2742         }
2743         kfree(new);
2744 }
2745
2746 static void relink_file_extents(struct new_sa_defrag_extent *new)
2747 {
2748         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2749         struct btrfs_path *path;
2750         struct sa_defrag_extent_backref *backref;
2751         struct sa_defrag_extent_backref *prev = NULL;
2752         struct inode *inode;
2753         struct btrfs_root *root;
2754         struct rb_node *node;
2755         int ret;
2756
2757         inode = new->inode;
2758         root = BTRFS_I(inode)->root;
2759
2760         path = btrfs_alloc_path();
2761         if (!path)
2762                 return;
2763
2764         if (!record_extent_backrefs(path, new)) {
2765                 btrfs_free_path(path);
2766                 goto out;
2767         }
2768         btrfs_release_path(path);
2769
2770         while (1) {
2771                 node = rb_first(&new->root);
2772                 if (!node)
2773                         break;
2774                 rb_erase(node, &new->root);
2775
2776                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2777
2778                 ret = relink_extent_backref(path, prev, backref);
2779                 WARN_ON(ret < 0);
2780
2781                 kfree(prev);
2782
2783                 if (ret == 1)
2784                         prev = backref;
2785                 else
2786                         prev = NULL;
2787                 cond_resched();
2788         }
2789         kfree(prev);
2790
2791         btrfs_free_path(path);
2792 out:
2793         free_sa_defrag_extent(new);
2794
2795         atomic_dec(&fs_info->defrag_running);
2796         wake_up(&fs_info->transaction_wait);
2797 }
2798
2799 static struct new_sa_defrag_extent *
2800 record_old_file_extents(struct inode *inode,
2801                         struct btrfs_ordered_extent *ordered)
2802 {
2803         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2804         struct btrfs_root *root = BTRFS_I(inode)->root;
2805         struct btrfs_path *path;
2806         struct btrfs_key key;
2807         struct old_sa_defrag_extent *old;
2808         struct new_sa_defrag_extent *new;
2809         int ret;
2810
2811         new = kmalloc(sizeof(*new), GFP_NOFS);
2812         if (!new)
2813                 return NULL;
2814
2815         new->inode = inode;
2816         new->file_pos = ordered->file_offset;
2817         new->len = ordered->len;
2818         new->bytenr = ordered->start;
2819         new->disk_len = ordered->disk_len;
2820         new->compress_type = ordered->compress_type;
2821         new->root = RB_ROOT;
2822         INIT_LIST_HEAD(&new->head);
2823
2824         path = btrfs_alloc_path();
2825         if (!path)
2826                 goto out_kfree;
2827
2828         key.objectid = btrfs_ino(BTRFS_I(inode));
2829         key.type = BTRFS_EXTENT_DATA_KEY;
2830         key.offset = new->file_pos;
2831
2832         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2833         if (ret < 0)
2834                 goto out_free_path;
2835         if (ret > 0 && path->slots[0] > 0)
2836                 path->slots[0]--;
2837
2838         /* find out all the old extents for the file range */
2839         while (1) {
2840                 struct btrfs_file_extent_item *extent;
2841                 struct extent_buffer *l;
2842                 int slot;
2843                 u64 num_bytes;
2844                 u64 offset;
2845                 u64 end;
2846                 u64 disk_bytenr;
2847                 u64 extent_offset;
2848
2849                 l = path->nodes[0];
2850                 slot = path->slots[0];
2851
2852                 if (slot >= btrfs_header_nritems(l)) {
2853                         ret = btrfs_next_leaf(root, path);
2854                         if (ret < 0)
2855                                 goto out_free_path;
2856                         else if (ret > 0)
2857                                 break;
2858                         continue;
2859                 }
2860
2861                 btrfs_item_key_to_cpu(l, &key, slot);
2862
2863                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2864                         break;
2865                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2866                         break;
2867                 if (key.offset >= new->file_pos + new->len)
2868                         break;
2869
2870                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2871
2872                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2873                 if (key.offset + num_bytes < new->file_pos)
2874                         goto next;
2875
2876                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2877                 if (!disk_bytenr)
2878                         goto next;
2879
2880                 extent_offset = btrfs_file_extent_offset(l, extent);
2881
2882                 old = kmalloc(sizeof(*old), GFP_NOFS);
2883                 if (!old)
2884                         goto out_free_path;
2885
2886                 offset = max(new->file_pos, key.offset);
2887                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2888
2889                 old->bytenr = disk_bytenr;
2890                 old->extent_offset = extent_offset;
2891                 old->offset = offset - key.offset;
2892                 old->len = end - offset;
2893                 old->new = new;
2894                 old->count = 0;
2895                 list_add_tail(&old->list, &new->head);
2896 next:
2897                 path->slots[0]++;
2898                 cond_resched();
2899         }
2900
2901         btrfs_free_path(path);
2902         atomic_inc(&fs_info->defrag_running);
2903
2904         return new;
2905
2906 out_free_path:
2907         btrfs_free_path(path);
2908 out_kfree:
2909         free_sa_defrag_extent(new);
2910         return NULL;
2911 }
2912
2913 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2914                                          u64 start, u64 len)
2915 {
2916         struct btrfs_block_group_cache *cache;
2917
2918         cache = btrfs_lookup_block_group(fs_info, start);
2919         ASSERT(cache);
2920
2921         spin_lock(&cache->lock);
2922         cache->delalloc_bytes -= len;
2923         spin_unlock(&cache->lock);
2924
2925         btrfs_put_block_group(cache);
2926 }
2927
2928 /* as ordered data IO finishes, this gets called so we can finish
2929  * an ordered extent if the range of bytes in the file it covers are
2930  * fully written.
2931  */
2932 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2933 {
2934         struct inode *inode = ordered_extent->inode;
2935         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2936         struct btrfs_root *root = BTRFS_I(inode)->root;
2937         struct btrfs_trans_handle *trans = NULL;
2938         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2939         struct extent_state *cached_state = NULL;
2940         struct new_sa_defrag_extent *new = NULL;
2941         int compress_type = 0;
2942         int ret = 0;
2943         u64 logical_len = ordered_extent->len;
2944         bool nolock;
2945         bool truncated = false;
2946         bool range_locked = false;
2947         bool clear_new_delalloc_bytes = false;
2948
2949         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2950             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2951             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2952                 clear_new_delalloc_bytes = true;
2953
2954         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2955
2956         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2957                 ret = -EIO;
2958                 goto out;
2959         }
2960
2961         btrfs_free_io_failure_record(BTRFS_I(inode),
2962                         ordered_extent->file_offset,
2963                         ordered_extent->file_offset +
2964                         ordered_extent->len - 1);
2965
2966         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2967                 truncated = true;
2968                 logical_len = ordered_extent->truncated_len;
2969                 /* Truncated the entire extent, don't bother adding */
2970                 if (!logical_len)
2971                         goto out;
2972         }
2973
2974         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2975                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2976
2977                 /*
2978                  * For mwrite(mmap + memset to write) case, we still reserve
2979                  * space for NOCOW range.
2980                  * As NOCOW won't cause a new delayed ref, just free the space
2981                  */
2982                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
2983                                        ordered_extent->len);
2984                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2985                 if (nolock)
2986                         trans = btrfs_join_transaction_nolock(root);
2987                 else
2988                         trans = btrfs_join_transaction(root);
2989                 if (IS_ERR(trans)) {
2990                         ret = PTR_ERR(trans);
2991                         trans = NULL;
2992                         goto out;
2993                 }
2994                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
2995                 ret = btrfs_update_inode_fallback(trans, root, inode);
2996                 if (ret) /* -ENOMEM or corruption */
2997                         btrfs_abort_transaction(trans, ret);
2998                 goto out;
2999         }
3000
3001         range_locked = true;
3002         lock_extent_bits(io_tree, ordered_extent->file_offset,
3003                          ordered_extent->file_offset + ordered_extent->len - 1,
3004                          &cached_state);
3005
3006         ret = test_range_bit(io_tree, ordered_extent->file_offset,
3007                         ordered_extent->file_offset + ordered_extent->len - 1,
3008                         EXTENT_DEFRAG, 0, cached_state);
3009         if (ret) {
3010                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
3011                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
3012                         /* the inode is shared */
3013                         new = record_old_file_extents(inode, ordered_extent);
3014
3015                 clear_extent_bit(io_tree, ordered_extent->file_offset,
3016                         ordered_extent->file_offset + ordered_extent->len - 1,
3017                         EXTENT_DEFRAG, 0, 0, &cached_state);
3018         }
3019
3020         if (nolock)
3021                 trans = btrfs_join_transaction_nolock(root);
3022         else
3023                 trans = btrfs_join_transaction(root);
3024         if (IS_ERR(trans)) {
3025                 ret = PTR_ERR(trans);
3026                 trans = NULL;
3027                 goto out;
3028         }
3029
3030         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3031
3032         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3033                 compress_type = ordered_extent->compress_type;
3034         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3035                 BUG_ON(compress_type);
3036                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3037                                        ordered_extent->len);
3038                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3039                                                 ordered_extent->file_offset,
3040                                                 ordered_extent->file_offset +
3041                                                 logical_len);
3042         } else {
3043                 BUG_ON(root == fs_info->tree_root);
3044                 ret = insert_reserved_file_extent(trans, inode,
3045                                                 ordered_extent->file_offset,
3046                                                 ordered_extent->start,
3047                                                 ordered_extent->disk_len,
3048                                                 logical_len, logical_len,
3049                                                 compress_type, 0, 0,
3050                                                 BTRFS_FILE_EXTENT_REG);
3051                 if (!ret)
3052                         btrfs_release_delalloc_bytes(fs_info,
3053                                                      ordered_extent->start,
3054                                                      ordered_extent->disk_len);
3055         }
3056         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3057                            ordered_extent->file_offset, ordered_extent->len,
3058                            trans->transid);
3059         if (ret < 0) {
3060                 btrfs_abort_transaction(trans, ret);
3061                 goto out;
3062         }
3063
3064         add_pending_csums(trans, inode, &ordered_extent->list);
3065
3066         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3067         ret = btrfs_update_inode_fallback(trans, root, inode);
3068         if (ret) { /* -ENOMEM or corruption */
3069                 btrfs_abort_transaction(trans, ret);
3070                 goto out;
3071         }
3072         ret = 0;
3073 out:
3074         if (range_locked || clear_new_delalloc_bytes) {
3075                 unsigned int clear_bits = 0;
3076
3077                 if (range_locked)
3078                         clear_bits |= EXTENT_LOCKED;
3079                 if (clear_new_delalloc_bytes)
3080                         clear_bits |= EXTENT_DELALLOC_NEW;
3081                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3082                                  ordered_extent->file_offset,
3083                                  ordered_extent->file_offset +
3084                                  ordered_extent->len - 1,
3085                                  clear_bits,
3086                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3087                                  0, &cached_state);
3088         }
3089
3090         if (trans)
3091                 btrfs_end_transaction(trans);
3092
3093         if (ret || truncated) {
3094                 u64 start, end;
3095
3096                 if (truncated)
3097                         start = ordered_extent->file_offset + logical_len;
3098                 else
3099                         start = ordered_extent->file_offset;
3100                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3101                 clear_extent_uptodate(io_tree, start, end, NULL);
3102
3103                 /* Drop the cache for the part of the extent we didn't write. */
3104                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3105
3106                 /*
3107                  * If the ordered extent had an IOERR or something else went
3108                  * wrong we need to return the space for this ordered extent
3109                  * back to the allocator.  We only free the extent in the
3110                  * truncated case if we didn't write out the extent at all.
3111                  */
3112                 if ((ret || !logical_len) &&
3113                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3114                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3115                         btrfs_free_reserved_extent(fs_info,
3116                                                    ordered_extent->start,
3117                                                    ordered_extent->disk_len, 1);
3118         }
3119
3120
3121         /*
3122          * This needs to be done to make sure anybody waiting knows we are done
3123          * updating everything for this ordered extent.
3124          */
3125         btrfs_remove_ordered_extent(inode, ordered_extent);
3126
3127         /* for snapshot-aware defrag */
3128         if (new) {
3129                 if (ret) {
3130                         free_sa_defrag_extent(new);
3131                         atomic_dec(&fs_info->defrag_running);
3132                 } else {
3133                         relink_file_extents(new);
3134                 }
3135         }
3136
3137         /* once for us */
3138         btrfs_put_ordered_extent(ordered_extent);
3139         /* once for the tree */
3140         btrfs_put_ordered_extent(ordered_extent);
3141
3142         return ret;
3143 }
3144
3145 static void finish_ordered_fn(struct btrfs_work *work)
3146 {
3147         struct btrfs_ordered_extent *ordered_extent;
3148         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3149         btrfs_finish_ordered_io(ordered_extent);
3150 }
3151
3152 static void btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3153                                 struct extent_state *state, int uptodate)
3154 {
3155         struct inode *inode = page->mapping->host;
3156         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3157         struct btrfs_ordered_extent *ordered_extent = NULL;
3158         struct btrfs_workqueue *wq;
3159         btrfs_work_func_t func;
3160
3161         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3162
3163         ClearPagePrivate2(page);
3164         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3165                                             end - start + 1, uptodate))
3166                 return;
3167
3168         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
3169                 wq = fs_info->endio_freespace_worker;
3170                 func = btrfs_freespace_write_helper;
3171         } else {
3172                 wq = fs_info->endio_write_workers;
3173                 func = btrfs_endio_write_helper;
3174         }
3175
3176         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3177                         NULL);
3178         btrfs_queue_work(wq, &ordered_extent->work);
3179 }
3180
3181 static int __readpage_endio_check(struct inode *inode,
3182                                   struct btrfs_io_bio *io_bio,
3183                                   int icsum, struct page *page,
3184                                   int pgoff, u64 start, size_t len)
3185 {
3186         char *kaddr;
3187         u32 csum_expected;
3188         u32 csum = ~(u32)0;
3189
3190         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3191
3192         kaddr = kmap_atomic(page);
3193         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3194         btrfs_csum_final(csum, (u8 *)&csum);
3195         if (csum != csum_expected)
3196                 goto zeroit;
3197
3198         kunmap_atomic(kaddr);
3199         return 0;
3200 zeroit:
3201         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3202                                     io_bio->mirror_num);
3203         memset(kaddr + pgoff, 1, len);
3204         flush_dcache_page(page);
3205         kunmap_atomic(kaddr);
3206         return -EIO;
3207 }
3208
3209 /*
3210  * when reads are done, we need to check csums to verify the data is correct
3211  * if there's a match, we allow the bio to finish.  If not, the code in
3212  * extent_io.c will try to find good copies for us.
3213  */
3214 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3215                                       u64 phy_offset, struct page *page,
3216                                       u64 start, u64 end, int mirror)
3217 {
3218         size_t offset = start - page_offset(page);
3219         struct inode *inode = page->mapping->host;
3220         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3221         struct btrfs_root *root = BTRFS_I(inode)->root;
3222
3223         if (PageChecked(page)) {
3224                 ClearPageChecked(page);
3225                 return 0;
3226         }
3227
3228         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3229                 return 0;
3230
3231         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3232             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3233                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3234                 return 0;
3235         }
3236
3237         phy_offset >>= inode->i_sb->s_blocksize_bits;
3238         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3239                                       start, (size_t)(end - start + 1));
3240 }
3241
3242 void btrfs_add_delayed_iput(struct inode *inode)
3243 {
3244         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3245         struct btrfs_inode *binode = BTRFS_I(inode);
3246
3247         if (atomic_add_unless(&inode->i_count, -1, 1))
3248                 return;
3249
3250         spin_lock(&fs_info->delayed_iput_lock);
3251         if (binode->delayed_iput_count == 0) {
3252                 ASSERT(list_empty(&binode->delayed_iput));
3253                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3254         } else {
3255                 binode->delayed_iput_count++;
3256         }
3257         spin_unlock(&fs_info->delayed_iput_lock);
3258 }
3259
3260 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3261 {
3262
3263         spin_lock(&fs_info->delayed_iput_lock);
3264         while (!list_empty(&fs_info->delayed_iputs)) {
3265                 struct btrfs_inode *inode;
3266
3267                 inode = list_first_entry(&fs_info->delayed_iputs,
3268                                 struct btrfs_inode, delayed_iput);
3269                 if (inode->delayed_iput_count) {
3270                         inode->delayed_iput_count--;
3271                         list_move_tail(&inode->delayed_iput,
3272                                         &fs_info->delayed_iputs);
3273                 } else {
3274                         list_del_init(&inode->delayed_iput);
3275                 }
3276                 spin_unlock(&fs_info->delayed_iput_lock);
3277                 iput(&inode->vfs_inode);
3278                 spin_lock(&fs_info->delayed_iput_lock);
3279         }
3280         spin_unlock(&fs_info->delayed_iput_lock);
3281 }
3282
3283 /*
3284  * This is called in transaction commit time. If there are no orphan
3285  * files in the subvolume, it removes orphan item and frees block_rsv
3286  * structure.
3287  */
3288 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3289                               struct btrfs_root *root)
3290 {
3291         struct btrfs_fs_info *fs_info = root->fs_info;
3292         struct btrfs_block_rsv *block_rsv;
3293         int ret;
3294
3295         if (atomic_read(&root->orphan_inodes) ||
3296             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3297                 return;
3298
3299         spin_lock(&root->orphan_lock);
3300         if (atomic_read(&root->orphan_inodes)) {
3301                 spin_unlock(&root->orphan_lock);
3302                 return;
3303         }
3304
3305         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3306                 spin_unlock(&root->orphan_lock);
3307                 return;
3308         }
3309
3310         block_rsv = root->orphan_block_rsv;
3311         root->orphan_block_rsv = NULL;
3312         spin_unlock(&root->orphan_lock);
3313
3314         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3315             btrfs_root_refs(&root->root_item) > 0) {
3316                 ret = btrfs_del_orphan_item(trans, fs_info->tree_root,
3317                                             root->root_key.objectid);
3318                 if (ret)
3319                         btrfs_abort_transaction(trans, ret);
3320                 else
3321                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3322                                   &root->state);
3323         }
3324
3325         if (block_rsv) {
3326                 WARN_ON(block_rsv->size > 0);
3327                 btrfs_free_block_rsv(fs_info, block_rsv);
3328         }
3329 }
3330
3331 /*
3332  * This creates an orphan entry for the given inode in case something goes
3333  * wrong in the middle of an unlink/truncate.
3334  *
3335  * NOTE: caller of this function should reserve 5 units of metadata for
3336  *       this function.
3337  */
3338 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3339                 struct btrfs_inode *inode)
3340 {
3341         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
3342         struct btrfs_root *root = inode->root;
3343         struct btrfs_block_rsv *block_rsv = NULL;
3344         int reserve = 0;
3345         int insert = 0;
3346         int ret;
3347
3348         if (!root->orphan_block_rsv) {
3349                 block_rsv = btrfs_alloc_block_rsv(fs_info,
3350                                                   BTRFS_BLOCK_RSV_TEMP);
3351                 if (!block_rsv)
3352                         return -ENOMEM;
3353         }
3354
3355         spin_lock(&root->orphan_lock);
3356         if (!root->orphan_block_rsv) {
3357                 root->orphan_block_rsv = block_rsv;
3358         } else if (block_rsv) {
3359                 btrfs_free_block_rsv(fs_info, block_rsv);
3360                 block_rsv = NULL;
3361         }
3362
3363         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3364                               &inode->runtime_flags)) {
3365 #if 0
3366                 /*
3367                  * For proper ENOSPC handling, we should do orphan
3368                  * cleanup when mounting. But this introduces backward
3369                  * compatibility issue.
3370                  */
3371                 if (!xchg(&root->orphan_item_inserted, 1))
3372                         insert = 2;
3373                 else
3374                         insert = 1;
3375 #endif
3376                 insert = 1;
3377                 atomic_inc(&root->orphan_inodes);
3378         }
3379
3380         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3381                               &inode->runtime_flags))
3382                 reserve = 1;
3383         spin_unlock(&root->orphan_lock);
3384
3385         /* grab metadata reservation from transaction handle */
3386         if (reserve) {
3387                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3388                 ASSERT(!ret);
3389                 if (ret) {
3390                         atomic_dec(&root->orphan_inodes);
3391                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3392                                   &inode->runtime_flags);
3393                         if (insert)
3394                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3395                                           &inode->runtime_flags);
3396                         return ret;
3397                 }
3398         }
3399
3400         /* insert an orphan item to track this unlinked/truncated file */
3401         if (insert >= 1) {
3402                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3403                 if (ret) {
3404                         atomic_dec(&root->orphan_inodes);
3405                         if (reserve) {
3406                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3407                                           &inode->runtime_flags);
3408                                 btrfs_orphan_release_metadata(inode);
3409                         }
3410                         if (ret != -EEXIST) {
3411                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3412                                           &inode->runtime_flags);
3413                                 btrfs_abort_transaction(trans, ret);
3414                                 return ret;
3415                         }
3416                 }
3417                 ret = 0;
3418         }
3419
3420         /* insert an orphan item to track subvolume contains orphan files */
3421         if (insert >= 2) {
3422                 ret = btrfs_insert_orphan_item(trans, fs_info->tree_root,
3423                                                root->root_key.objectid);
3424                 if (ret && ret != -EEXIST) {
3425                         btrfs_abort_transaction(trans, ret);
3426                         return ret;
3427                 }
3428         }
3429         return 0;
3430 }
3431
3432 /*
3433  * We have done the truncate/delete so we can go ahead and remove the orphan
3434  * item for this particular inode.
3435  */
3436 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3437                             struct btrfs_inode *inode)
3438 {
3439         struct btrfs_root *root = inode->root;
3440         int delete_item = 0;
3441         int release_rsv = 0;
3442         int ret = 0;
3443
3444         spin_lock(&root->orphan_lock);
3445         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3446                                &inode->runtime_flags))
3447                 delete_item = 1;
3448
3449         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3450                                &inode->runtime_flags))
3451                 release_rsv = 1;
3452         spin_unlock(&root->orphan_lock);
3453
3454         if (delete_item) {
3455                 atomic_dec(&root->orphan_inodes);
3456                 if (trans)
3457                         ret = btrfs_del_orphan_item(trans, root,
3458                                                     btrfs_ino(inode));
3459         }
3460
3461         if (release_rsv)
3462                 btrfs_orphan_release_metadata(inode);
3463
3464         return ret;
3465 }
3466
3467 /*
3468  * this cleans up any orphans that may be left on the list from the last use
3469  * of this root.
3470  */
3471 int btrfs_orphan_cleanup(struct btrfs_root *root)
3472 {
3473         struct btrfs_fs_info *fs_info = root->fs_info;
3474         struct btrfs_path *path;
3475         struct extent_buffer *leaf;
3476         struct btrfs_key key, found_key;
3477         struct btrfs_trans_handle *trans;
3478         struct inode *inode;
3479         u64 last_objectid = 0;
3480         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3481
3482         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3483                 return 0;
3484
3485         path = btrfs_alloc_path();
3486         if (!path) {
3487                 ret = -ENOMEM;
3488                 goto out;
3489         }
3490         path->reada = READA_BACK;
3491
3492         key.objectid = BTRFS_ORPHAN_OBJECTID;
3493         key.type = BTRFS_ORPHAN_ITEM_KEY;
3494         key.offset = (u64)-1;
3495
3496         while (1) {
3497                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3498                 if (ret < 0)
3499                         goto out;
3500
3501                 /*
3502                  * if ret == 0 means we found what we were searching for, which
3503                  * is weird, but possible, so only screw with path if we didn't
3504                  * find the key and see if we have stuff that matches
3505                  */
3506                 if (ret > 0) {
3507                         ret = 0;
3508                         if (path->slots[0] == 0)
3509                                 break;
3510                         path->slots[0]--;
3511                 }
3512
3513                 /* pull out the item */
3514                 leaf = path->nodes[0];
3515                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3516
3517                 /* make sure the item matches what we want */
3518                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3519                         break;
3520                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3521                         break;
3522
3523                 /* release the path since we're done with it */
3524                 btrfs_release_path(path);
3525
3526                 /*
3527                  * this is where we are basically btrfs_lookup, without the
3528                  * crossing root thing.  we store the inode number in the
3529                  * offset of the orphan item.
3530                  */
3531
3532                 if (found_key.offset == last_objectid) {
3533                         btrfs_err(fs_info,
3534                                   "Error removing orphan entry, stopping orphan cleanup");
3535                         ret = -EINVAL;
3536                         goto out;
3537                 }
3538
3539                 last_objectid = found_key.offset;
3540
3541                 found_key.objectid = found_key.offset;
3542                 found_key.type = BTRFS_INODE_ITEM_KEY;
3543                 found_key.offset = 0;
3544                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3545                 ret = PTR_ERR_OR_ZERO(inode);
3546                 if (ret && ret != -ENOENT)
3547                         goto out;
3548
3549                 if (ret == -ENOENT && root == fs_info->tree_root) {
3550                         struct btrfs_root *dead_root;
3551                         struct btrfs_fs_info *fs_info = root->fs_info;
3552                         int is_dead_root = 0;
3553
3554                         /*
3555                          * this is an orphan in the tree root. Currently these
3556                          * could come from 2 sources:
3557                          *  a) a snapshot deletion in progress
3558                          *  b) a free space cache inode
3559                          * We need to distinguish those two, as the snapshot
3560                          * orphan must not get deleted.
3561                          * find_dead_roots already ran before us, so if this
3562                          * is a snapshot deletion, we should find the root
3563                          * in the dead_roots list
3564                          */
3565                         spin_lock(&fs_info->trans_lock);
3566                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3567                                             root_list) {
3568                                 if (dead_root->root_key.objectid ==
3569                                     found_key.objectid) {
3570                                         is_dead_root = 1;
3571                                         break;
3572                                 }
3573                         }
3574                         spin_unlock(&fs_info->trans_lock);
3575                         if (is_dead_root) {
3576                                 /* prevent this orphan from being found again */
3577                                 key.offset = found_key.objectid - 1;
3578                                 continue;
3579                         }
3580                 }
3581                 /*
3582                  * Inode is already gone but the orphan item is still there,
3583                  * kill the orphan item.
3584                  */
3585                 if (ret == -ENOENT) {
3586                         trans = btrfs_start_transaction(root, 1);
3587                         if (IS_ERR(trans)) {
3588                                 ret = PTR_ERR(trans);
3589                                 goto out;
3590                         }
3591                         btrfs_debug(fs_info, "auto deleting %Lu",
3592                                     found_key.objectid);
3593                         ret = btrfs_del_orphan_item(trans, root,
3594                                                     found_key.objectid);
3595                         btrfs_end_transaction(trans);
3596                         if (ret)
3597                                 goto out;
3598                         continue;
3599                 }
3600
3601                 /*
3602                  * add this inode to the orphan list so btrfs_orphan_del does
3603                  * the proper thing when we hit it
3604                  */
3605                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3606                         &BTRFS_I(inode)->runtime_flags);
3607                 atomic_inc(&root->orphan_inodes);
3608
3609                 /* if we have links, this was a truncate, lets do that */
3610                 if (inode->i_nlink) {
3611                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3612                                 iput(inode);
3613                                 continue;
3614                         }
3615                         nr_truncate++;
3616
3617                         /* 1 for the orphan item deletion. */
3618                         trans = btrfs_start_transaction(root, 1);
3619                         if (IS_ERR(trans)) {
3620                                 iput(inode);
3621                                 ret = PTR_ERR(trans);
3622                                 goto out;
3623                         }
3624                         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
3625                         btrfs_end_transaction(trans);
3626                         if (ret) {
3627                                 iput(inode);
3628                                 goto out;
3629                         }
3630
3631                         ret = btrfs_truncate(inode);
3632                         if (ret)
3633                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
3634                 } else {
3635                         nr_unlink++;
3636                 }
3637
3638                 /* this will do delete_inode and everything for us */
3639                 iput(inode);
3640                 if (ret)
3641                         goto out;
3642         }
3643         /* release the path since we're done with it */
3644         btrfs_release_path(path);
3645
3646         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3647
3648         if (root->orphan_block_rsv)
3649                 btrfs_block_rsv_release(fs_info, root->orphan_block_rsv,
3650                                         (u64)-1);
3651
3652         if (root->orphan_block_rsv ||
3653             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3654                 trans = btrfs_join_transaction(root);
3655                 if (!IS_ERR(trans))
3656                         btrfs_end_transaction(trans);
3657         }
3658
3659         if (nr_unlink)
3660                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3661         if (nr_truncate)
3662                 btrfs_debug(fs_info, "truncated %d orphans", nr_truncate);
3663
3664 out:
3665         if (ret)
3666                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3667         btrfs_free_path(path);
3668         return ret;
3669 }
3670
3671 /*
3672  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3673  * don't find any xattrs, we know there can't be any acls.
3674  *
3675  * slot is the slot the inode is in, objectid is the objectid of the inode
3676  */
3677 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3678                                           int slot, u64 objectid,
3679                                           int *first_xattr_slot)
3680 {
3681         u32 nritems = btrfs_header_nritems(leaf);
3682         struct btrfs_key found_key;
3683         static u64 xattr_access = 0;
3684         static u64 xattr_default = 0;
3685         int scanned = 0;
3686
3687         if (!xattr_access) {
3688                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3689                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3690                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3691                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3692         }
3693
3694         slot++;
3695         *first_xattr_slot = -1;
3696         while (slot < nritems) {
3697                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3698
3699                 /* we found a different objectid, there must not be acls */
3700                 if (found_key.objectid != objectid)
3701                         return 0;
3702
3703                 /* we found an xattr, assume we've got an acl */
3704                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3705                         if (*first_xattr_slot == -1)
3706                                 *first_xattr_slot = slot;
3707                         if (found_key.offset == xattr_access ||
3708                             found_key.offset == xattr_default)
3709                                 return 1;
3710                 }
3711
3712                 /*
3713                  * we found a key greater than an xattr key, there can't
3714                  * be any acls later on
3715                  */
3716                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3717                         return 0;
3718
3719                 slot++;
3720                 scanned++;
3721
3722                 /*
3723                  * it goes inode, inode backrefs, xattrs, extents,
3724                  * so if there are a ton of hard links to an inode there can
3725                  * be a lot of backrefs.  Don't waste time searching too hard,
3726                  * this is just an optimization
3727                  */
3728                 if (scanned >= 8)
3729                         break;
3730         }
3731         /* we hit the end of the leaf before we found an xattr or
3732          * something larger than an xattr.  We have to assume the inode
3733          * has acls
3734          */
3735         if (*first_xattr_slot == -1)
3736                 *first_xattr_slot = slot;
3737         return 1;
3738 }
3739
3740 /*
3741  * read an inode from the btree into the in-memory inode
3742  */
3743 static int btrfs_read_locked_inode(struct inode *inode)
3744 {
3745         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3746         struct btrfs_path *path;
3747         struct extent_buffer *leaf;
3748         struct btrfs_inode_item *inode_item;
3749         struct btrfs_root *root = BTRFS_I(inode)->root;
3750         struct btrfs_key location;
3751         unsigned long ptr;
3752         int maybe_acls;
3753         u32 rdev;
3754         int ret;
3755         bool filled = false;
3756         int first_xattr_slot;
3757
3758         ret = btrfs_fill_inode(inode, &rdev);
3759         if (!ret)
3760                 filled = true;
3761
3762         path = btrfs_alloc_path();
3763         if (!path) {
3764                 ret = -ENOMEM;
3765                 goto make_bad;
3766         }
3767
3768         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3769
3770         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3771         if (ret) {
3772                 if (ret > 0)
3773                         ret = -ENOENT;
3774                 goto make_bad;
3775         }
3776
3777         leaf = path->nodes[0];
3778
3779         if (filled)
3780                 goto cache_index;
3781
3782         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3783                                     struct btrfs_inode_item);
3784         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3785         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3786         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3787         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3788         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3789
3790         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3791         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3792
3793         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3794         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3795
3796         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3797         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3798
3799         BTRFS_I(inode)->i_otime.tv_sec =
3800                 btrfs_timespec_sec(leaf, &inode_item->otime);
3801         BTRFS_I(inode)->i_otime.tv_nsec =
3802                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3803
3804         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3805         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3806         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3807
3808         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3809         inode->i_generation = BTRFS_I(inode)->generation;
3810         inode->i_rdev = 0;
3811         rdev = btrfs_inode_rdev(leaf, inode_item);
3812
3813         BTRFS_I(inode)->index_cnt = (u64)-1;
3814         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3815
3816 cache_index:
3817         /*
3818          * If we were modified in the current generation and evicted from memory
3819          * and then re-read we need to do a full sync since we don't have any
3820          * idea about which extents were modified before we were evicted from
3821          * cache.
3822          *
3823          * This is required for both inode re-read from disk and delayed inode
3824          * in delayed_nodes_tree.
3825          */
3826         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3827                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3828                         &BTRFS_I(inode)->runtime_flags);
3829
3830         /*
3831          * We don't persist the id of the transaction where an unlink operation
3832          * against the inode was last made. So here we assume the inode might
3833          * have been evicted, and therefore the exact value of last_unlink_trans
3834          * lost, and set it to last_trans to avoid metadata inconsistencies
3835          * between the inode and its parent if the inode is fsync'ed and the log
3836          * replayed. For example, in the scenario:
3837          *
3838          * touch mydir/foo
3839          * ln mydir/foo mydir/bar
3840          * sync
3841          * unlink mydir/bar
3842          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3843          * xfs_io -c fsync mydir/foo
3844          * <power failure>
3845          * mount fs, triggers fsync log replay
3846          *
3847          * We must make sure that when we fsync our inode foo we also log its
3848          * parent inode, otherwise after log replay the parent still has the
3849          * dentry with the "bar" name but our inode foo has a link count of 1
3850          * and doesn't have an inode ref with the name "bar" anymore.
3851          *
3852          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3853          * but it guarantees correctness at the expense of occasional full
3854          * transaction commits on fsync if our inode is a directory, or if our
3855          * inode is not a directory, logging its parent unnecessarily.
3856          */
3857         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3858
3859         path->slots[0]++;
3860         if (inode->i_nlink != 1 ||
3861             path->slots[0] >= btrfs_header_nritems(leaf))
3862                 goto cache_acl;
3863
3864         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3865         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3866                 goto cache_acl;
3867
3868         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3869         if (location.type == BTRFS_INODE_REF_KEY) {
3870                 struct btrfs_inode_ref *ref;
3871
3872                 ref = (struct btrfs_inode_ref *)ptr;
3873                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3874         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3875                 struct btrfs_inode_extref *extref;
3876
3877                 extref = (struct btrfs_inode_extref *)ptr;
3878                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3879                                                                      extref);
3880         }
3881 cache_acl:
3882         /*
3883          * try to precache a NULL acl entry for files that don't have
3884          * any xattrs or acls
3885          */
3886         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3887                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3888         if (first_xattr_slot != -1) {
3889                 path->slots[0] = first_xattr_slot;
3890                 ret = btrfs_load_inode_props(inode, path);
3891                 if (ret)
3892                         btrfs_err(fs_info,
3893                                   "error loading props for ino %llu (root %llu): %d",
3894                                   btrfs_ino(BTRFS_I(inode)),
3895                                   root->root_key.objectid, ret);
3896         }
3897         btrfs_free_path(path);
3898
3899         if (!maybe_acls)
3900                 cache_no_acl(inode);
3901
3902         switch (inode->i_mode & S_IFMT) {
3903         case S_IFREG:
3904                 inode->i_mapping->a_ops = &btrfs_aops;
3905                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3906                 inode->i_fop = &btrfs_file_operations;
3907                 inode->i_op = &btrfs_file_inode_operations;
3908                 break;
3909         case S_IFDIR:
3910                 inode->i_fop = &btrfs_dir_file_operations;
3911                 inode->i_op = &btrfs_dir_inode_operations;
3912                 break;
3913         case S_IFLNK:
3914                 inode->i_op = &btrfs_symlink_inode_operations;
3915                 inode_nohighmem(inode);
3916                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3917                 break;
3918         default:
3919                 inode->i_op = &btrfs_special_inode_operations;
3920                 init_special_inode(inode, inode->i_mode, rdev);
3921                 break;
3922         }
3923
3924         btrfs_update_iflags(inode);
3925         return 0;
3926
3927 make_bad:
3928         btrfs_free_path(path);
3929         make_bad_inode(inode);
3930         return ret;
3931 }
3932
3933 /*
3934  * given a leaf and an inode, copy the inode fields into the leaf
3935  */
3936 static void fill_inode_item(struct btrfs_trans_handle *trans,
3937                             struct extent_buffer *leaf,
3938                             struct btrfs_inode_item *item,
3939                             struct inode *inode)
3940 {
3941         struct btrfs_map_token token;
3942
3943         btrfs_init_map_token(&token);
3944
3945         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3946         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3947         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3948                                    &token);
3949         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3950         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3951
3952         btrfs_set_token_timespec_sec(leaf, &item->atime,
3953                                      inode->i_atime.tv_sec, &token);
3954         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3955                                       inode->i_atime.tv_nsec, &token);
3956
3957         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3958                                      inode->i_mtime.tv_sec, &token);
3959         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3960                                       inode->i_mtime.tv_nsec, &token);
3961
3962         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3963                                      inode->i_ctime.tv_sec, &token);
3964         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3965                                       inode->i_ctime.tv_nsec, &token);
3966
3967         btrfs_set_token_timespec_sec(leaf, &item->otime,
3968                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3969         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3970                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3971
3972         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3973                                      &token);
3974         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3975                                          &token);
3976         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3977         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3978         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3979         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3980         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3981 }
3982
3983 /*
3984  * copy everything in the in-memory inode into the btree.
3985  */
3986 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3987                                 struct btrfs_root *root, struct inode *inode)
3988 {
3989         struct btrfs_inode_item *inode_item;
3990         struct btrfs_path *path;
3991         struct extent_buffer *leaf;
3992         int ret;
3993
3994         path = btrfs_alloc_path();
3995         if (!path)
3996                 return -ENOMEM;
3997
3998         path->leave_spinning = 1;
3999         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
4000                                  1);
4001         if (ret) {
4002                 if (ret > 0)
4003                         ret = -ENOENT;
4004                 goto failed;
4005         }
4006
4007         leaf = path->nodes[0];
4008         inode_item = btrfs_item_ptr(leaf, path->slots[0],
4009                                     struct btrfs_inode_item);
4010
4011         fill_inode_item(trans, leaf, inode_item, inode);
4012         btrfs_mark_buffer_dirty(leaf);
4013         btrfs_set_inode_last_trans(trans, inode);
4014         ret = 0;
4015 failed:
4016         btrfs_free_path(path);
4017         return ret;
4018 }
4019
4020 /*
4021  * copy everything in the in-memory inode into the btree.
4022  */
4023 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
4024                                 struct btrfs_root *root, struct inode *inode)
4025 {
4026         struct btrfs_fs_info *fs_info = root->fs_info;
4027         int ret;
4028
4029         /*
4030          * If the inode is a free space inode, we can deadlock during commit
4031          * if we put it into the delayed code.
4032          *
4033          * The data relocation inode should also be directly updated
4034          * without delay
4035          */
4036         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
4037             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
4038             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
4039                 btrfs_update_root_times(trans, root);
4040
4041                 ret = btrfs_delayed_update_inode(trans, root, inode);
4042                 if (!ret)
4043                         btrfs_set_inode_last_trans(trans, inode);
4044                 return ret;
4045         }
4046
4047         return btrfs_update_inode_item(trans, root, inode);
4048 }
4049
4050 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
4051                                          struct btrfs_root *root,
4052                                          struct inode *inode)
4053 {
4054         int ret;
4055
4056         ret = btrfs_update_inode(trans, root, inode);
4057         if (ret == -ENOSPC)
4058                 return btrfs_update_inode_item(trans, root, inode);
4059         return ret;
4060 }
4061
4062 /*
4063  * unlink helper that gets used here in inode.c and in the tree logging
4064  * recovery code.  It remove a link in a directory with a given name, and
4065  * also drops the back refs in the inode to the directory
4066  */
4067 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4068                                 struct btrfs_root *root,
4069                                 struct btrfs_inode *dir,
4070                                 struct btrfs_inode *inode,
4071                                 const char *name, int name_len)
4072 {
4073         struct btrfs_fs_info *fs_info = root->fs_info;
4074         struct btrfs_path *path;
4075         int ret = 0;
4076         struct extent_buffer *leaf;
4077         struct btrfs_dir_item *di;
4078         struct btrfs_key key;
4079         u64 index;
4080         u64 ino = btrfs_ino(inode);
4081         u64 dir_ino = btrfs_ino(dir);
4082
4083         path = btrfs_alloc_path();
4084         if (!path) {
4085                 ret = -ENOMEM;
4086                 goto out;
4087         }
4088
4089         path->leave_spinning = 1;
4090         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4091                                     name, name_len, -1);
4092         if (IS_ERR(di)) {
4093                 ret = PTR_ERR(di);
4094                 goto err;
4095         }
4096         if (!di) {
4097                 ret = -ENOENT;
4098                 goto err;
4099         }
4100         leaf = path->nodes[0];
4101         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4102         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4103         if (ret)
4104                 goto err;
4105         btrfs_release_path(path);
4106
4107         /*
4108          * If we don't have dir index, we have to get it by looking up
4109          * the inode ref, since we get the inode ref, remove it directly,
4110          * it is unnecessary to do delayed deletion.
4111          *
4112          * But if we have dir index, needn't search inode ref to get it.
4113          * Since the inode ref is close to the inode item, it is better
4114          * that we delay to delete it, and just do this deletion when
4115          * we update the inode item.
4116          */
4117         if (inode->dir_index) {
4118                 ret = btrfs_delayed_delete_inode_ref(inode);
4119                 if (!ret) {
4120                         index = inode->dir_index;
4121                         goto skip_backref;
4122                 }
4123         }
4124
4125         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4126                                   dir_ino, &index);
4127         if (ret) {
4128                 btrfs_info(fs_info,
4129                         "failed to delete reference to %.*s, inode %llu parent %llu",
4130                         name_len, name, ino, dir_ino);
4131                 btrfs_abort_transaction(trans, ret);
4132                 goto err;
4133         }
4134 skip_backref:
4135         ret = btrfs_delete_delayed_dir_index(trans, fs_info, dir, index);
4136         if (ret) {
4137                 btrfs_abort_transaction(trans, ret);
4138                 goto err;
4139         }
4140
4141         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
4142                         dir_ino);
4143         if (ret != 0 && ret != -ENOENT) {
4144                 btrfs_abort_transaction(trans, ret);
4145                 goto err;
4146         }
4147
4148         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
4149                         index);
4150         if (ret == -ENOENT)
4151                 ret = 0;
4152         else if (ret)
4153                 btrfs_abort_transaction(trans, ret);
4154 err:
4155         btrfs_free_path(path);
4156         if (ret)
4157                 goto out;
4158
4159         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
4160         inode_inc_iversion(&inode->vfs_inode);
4161         inode_inc_iversion(&dir->vfs_inode);
4162         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4163                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4164         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4165 out:
4166         return ret;
4167 }
4168
4169 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4170                        struct btrfs_root *root,
4171                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4172                        const char *name, int name_len)
4173 {
4174         int ret;
4175         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4176         if (!ret) {
4177                 drop_nlink(&inode->vfs_inode);
4178                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4179         }
4180         return ret;
4181 }
4182
4183 /*
4184  * helper to start transaction for unlink and rmdir.
4185  *
4186  * unlink and rmdir are special in btrfs, they do not always free space, so
4187  * if we cannot make our reservations the normal way try and see if there is
4188  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4189  * allow the unlink to occur.
4190  */
4191 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4192 {
4193         struct btrfs_root *root = BTRFS_I(dir)->root;
4194
4195         /*
4196          * 1 for the possible orphan item
4197          * 1 for the dir item
4198          * 1 for the dir index
4199          * 1 for the inode ref
4200          * 1 for the inode
4201          */
4202         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4203 }
4204
4205 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4206 {
4207         struct btrfs_root *root = BTRFS_I(dir)->root;
4208         struct btrfs_trans_handle *trans;
4209         struct inode *inode = d_inode(dentry);
4210         int ret;
4211
4212         trans = __unlink_start_trans(dir);
4213         if (IS_ERR(trans))
4214                 return PTR_ERR(trans);
4215
4216         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4217                         0);
4218
4219         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4220                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4221                         dentry->d_name.len);
4222         if (ret)
4223                 goto out;
4224
4225         if (inode->i_nlink == 0) {
4226                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4227                 if (ret)
4228                         goto out;
4229         }
4230
4231 out:
4232         btrfs_end_transaction(trans);
4233         btrfs_btree_balance_dirty(root->fs_info);
4234         return ret;
4235 }
4236
4237 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4238                         struct btrfs_root *root,
4239                         struct inode *dir, u64 objectid,
4240                         const char *name, int name_len)
4241 {
4242         struct btrfs_fs_info *fs_info = root->fs_info;
4243         struct btrfs_path *path;
4244         struct extent_buffer *leaf;
4245         struct btrfs_dir_item *di;
4246         struct btrfs_key key;
4247         u64 index;
4248         int ret;
4249         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4250
4251         path = btrfs_alloc_path();
4252         if (!path)
4253                 return -ENOMEM;
4254
4255         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4256                                    name, name_len, -1);
4257         if (IS_ERR_OR_NULL(di)) {
4258                 if (!di)
4259                         ret = -ENOENT;
4260                 else
4261                         ret = PTR_ERR(di);
4262                 goto out;
4263         }
4264
4265         leaf = path->nodes[0];
4266         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4267         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4268         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4269         if (ret) {
4270                 btrfs_abort_transaction(trans, ret);
4271                 goto out;
4272         }
4273         btrfs_release_path(path);
4274
4275         ret = btrfs_del_root_ref(trans, fs_info, objectid,
4276                                  root->root_key.objectid, dir_ino,
4277                                  &index, name, name_len);
4278         if (ret < 0) {
4279                 if (ret != -ENOENT) {
4280                         btrfs_abort_transaction(trans, ret);
4281                         goto out;
4282                 }
4283                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4284                                                  name, name_len);
4285                 if (IS_ERR_OR_NULL(di)) {
4286                         if (!di)
4287                                 ret = -ENOENT;
4288                         else
4289                                 ret = PTR_ERR(di);
4290                         btrfs_abort_transaction(trans, ret);
4291                         goto out;
4292                 }
4293
4294                 leaf = path->nodes[0];
4295                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4296                 btrfs_release_path(path);
4297                 index = key.offset;
4298         }
4299         btrfs_release_path(path);
4300
4301         ret = btrfs_delete_delayed_dir_index(trans, fs_info, BTRFS_I(dir), index);
4302         if (ret) {
4303                 btrfs_abort_transaction(trans, ret);
4304                 goto out;
4305         }
4306
4307         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4308         inode_inc_iversion(dir);
4309         dir->i_mtime = dir->i_ctime = current_time(dir);
4310         ret = btrfs_update_inode_fallback(trans, root, dir);
4311         if (ret)
4312                 btrfs_abort_transaction(trans, ret);
4313 out:
4314         btrfs_free_path(path);
4315         return ret;
4316 }
4317
4318 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4319 {
4320         struct inode *inode = d_inode(dentry);
4321         int err = 0;
4322         struct btrfs_root *root = BTRFS_I(dir)->root;
4323         struct btrfs_trans_handle *trans;
4324         u64 last_unlink_trans;
4325
4326         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4327                 return -ENOTEMPTY;
4328         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4329                 return -EPERM;
4330
4331         trans = __unlink_start_trans(dir);
4332         if (IS_ERR(trans))
4333                 return PTR_ERR(trans);
4334
4335         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4336                 err = btrfs_unlink_subvol(trans, root, dir,
4337                                           BTRFS_I(inode)->location.objectid,
4338                                           dentry->d_name.name,
4339                                           dentry->d_name.len);
4340                 goto out;
4341         }
4342
4343         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4344         if (err)
4345                 goto out;
4346
4347         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4348
4349         /* now the directory is empty */
4350         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4351                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4352                         dentry->d_name.len);
4353         if (!err) {
4354                 btrfs_i_size_write(BTRFS_I(inode), 0);
4355                 /*
4356                  * Propagate the last_unlink_trans value of the deleted dir to
4357                  * its parent directory. This is to prevent an unrecoverable
4358                  * log tree in the case we do something like this:
4359                  * 1) create dir foo
4360                  * 2) create snapshot under dir foo
4361                  * 3) delete the snapshot
4362                  * 4) rmdir foo
4363                  * 5) mkdir foo
4364                  * 6) fsync foo or some file inside foo
4365                  */
4366                 if (last_unlink_trans >= trans->transid)
4367                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4368         }
4369 out:
4370         btrfs_end_transaction(trans);
4371         btrfs_btree_balance_dirty(root->fs_info);
4372
4373         return err;
4374 }
4375
4376 static int truncate_space_check(struct btrfs_trans_handle *trans,
4377                                 struct btrfs_root *root,
4378                                 u64 bytes_deleted)
4379 {
4380         struct btrfs_fs_info *fs_info = root->fs_info;
4381         int ret;
4382
4383         /*
4384          * This is only used to apply pressure to the enospc system, we don't
4385          * intend to use this reservation at all.
4386          */
4387         bytes_deleted = btrfs_csum_bytes_to_leaves(fs_info, bytes_deleted);
4388         bytes_deleted *= fs_info->nodesize;
4389         ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
4390                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4391         if (!ret) {
4392                 trace_btrfs_space_reservation(fs_info, "transaction",
4393                                               trans->transid,
4394                                               bytes_deleted, 1);
4395                 trans->bytes_reserved += bytes_deleted;
4396         }
4397         return ret;
4398
4399 }
4400
4401 /*
4402  * Return this if we need to call truncate_block for the last bit of the
4403  * truncate.
4404  */
4405 #define NEED_TRUNCATE_BLOCK 1
4406
4407 /*
4408  * this can truncate away extent items, csum items and directory items.
4409  * It starts at a high offset and removes keys until it can't find
4410  * any higher than new_size
4411  *
4412  * csum items that cross the new i_size are truncated to the new size
4413  * as well.
4414  *
4415  * min_type is the minimum key type to truncate down to.  If set to 0, this
4416  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4417  */
4418 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4419                                struct btrfs_root *root,
4420                                struct inode *inode,
4421                                u64 new_size, u32 min_type)
4422 {
4423         struct btrfs_fs_info *fs_info = root->fs_info;
4424         struct btrfs_path *path;
4425         struct extent_buffer *leaf;
4426         struct btrfs_file_extent_item *fi;
4427         struct btrfs_key key;
4428         struct btrfs_key found_key;
4429         u64 extent_start = 0;
4430         u64 extent_num_bytes = 0;
4431         u64 extent_offset = 0;
4432         u64 item_end = 0;
4433         u64 last_size = new_size;
4434         u32 found_type = (u8)-1;
4435         int found_extent;
4436         int del_item;
4437         int pending_del_nr = 0;
4438         int pending_del_slot = 0;
4439         int extent_type = -1;
4440         int ret;
4441         int err = 0;
4442         u64 ino = btrfs_ino(BTRFS_I(inode));
4443         u64 bytes_deleted = 0;
4444         bool be_nice = false;
4445         bool should_throttle = false;
4446         bool should_end = false;
4447
4448         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4449
4450         /*
4451          * for non-free space inodes and ref cows, we want to back off from
4452          * time to time
4453          */
4454         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4455             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4456                 be_nice = true;
4457
4458         path = btrfs_alloc_path();
4459         if (!path)
4460                 return -ENOMEM;
4461         path->reada = READA_BACK;
4462
4463         /*
4464          * We want to drop from the next block forward in case this new size is
4465          * not block aligned since we will be keeping the last block of the
4466          * extent just the way it is.
4467          */
4468         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4469             root == fs_info->tree_root)
4470                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4471                                         fs_info->sectorsize),
4472                                         (u64)-1, 0);
4473
4474         /*
4475          * This function is also used to drop the items in the log tree before
4476          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4477          * it is used to drop the loged items. So we shouldn't kill the delayed
4478          * items.
4479          */
4480         if (min_type == 0 && root == BTRFS_I(inode)->root)
4481                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4482
4483         key.objectid = ino;
4484         key.offset = (u64)-1;
4485         key.type = (u8)-1;
4486
4487 search_again:
4488         /*
4489          * with a 16K leaf size and 128MB extents, you can actually queue
4490          * up a huge file in a single leaf.  Most of the time that
4491          * bytes_deleted is > 0, it will be huge by the time we get here
4492          */
4493         if (be_nice && bytes_deleted > SZ_32M) {
4494                 if (btrfs_should_end_transaction(trans)) {
4495                         err = -EAGAIN;
4496                         goto error;
4497                 }
4498         }
4499
4500
4501         path->leave_spinning = 1;
4502         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4503         if (ret < 0) {
4504                 err = ret;
4505                 goto out;
4506         }
4507
4508         if (ret > 0) {
4509                 /* there are no items in the tree for us to truncate, we're
4510                  * done
4511                  */
4512                 if (path->slots[0] == 0)
4513                         goto out;
4514                 path->slots[0]--;
4515         }
4516
4517         while (1) {
4518                 fi = NULL;
4519                 leaf = path->nodes[0];
4520                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4521                 found_type = found_key.type;
4522
4523                 if (found_key.objectid != ino)
4524                         break;
4525
4526                 if (found_type < min_type)
4527                         break;
4528
4529                 item_end = found_key.offset;
4530                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4531                         fi = btrfs_item_ptr(leaf, path->slots[0],
4532                                             struct btrfs_file_extent_item);
4533                         extent_type = btrfs_file_extent_type(leaf, fi);
4534                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4535                                 item_end +=
4536                                     btrfs_file_extent_num_bytes(leaf, fi);
4537
4538                                 trace_btrfs_truncate_show_fi_regular(
4539                                         BTRFS_I(inode), leaf, fi,
4540                                         found_key.offset);
4541                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4542                                 item_end += btrfs_file_extent_inline_len(leaf,
4543                                                          path->slots[0], fi);
4544
4545                                 trace_btrfs_truncate_show_fi_inline(
4546                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4547                                         found_key.offset);
4548                         }
4549                         item_end--;
4550                 }
4551                 if (found_type > min_type) {
4552                         del_item = 1;
4553                 } else {
4554                         if (item_end < new_size)
4555                                 break;
4556                         if (found_key.offset >= new_size)
4557                                 del_item = 1;
4558                         else
4559                                 del_item = 0;
4560                 }
4561                 found_extent = 0;
4562                 /* FIXME, shrink the extent if the ref count is only 1 */
4563                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4564                         goto delete;
4565
4566                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4567                         u64 num_dec;
4568                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4569                         if (!del_item) {
4570                                 u64 orig_num_bytes =
4571                                         btrfs_file_extent_num_bytes(leaf, fi);
4572                                 extent_num_bytes = ALIGN(new_size -
4573                                                 found_key.offset,
4574                                                 fs_info->sectorsize);
4575                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4576                                                          extent_num_bytes);
4577                                 num_dec = (orig_num_bytes -
4578                                            extent_num_bytes);
4579                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4580                                              &root->state) &&
4581                                     extent_start != 0)
4582                                         inode_sub_bytes(inode, num_dec);
4583                                 btrfs_mark_buffer_dirty(leaf);
4584                         } else {
4585                                 extent_num_bytes =
4586                                         btrfs_file_extent_disk_num_bytes(leaf,
4587                                                                          fi);
4588                                 extent_offset = found_key.offset -
4589                                         btrfs_file_extent_offset(leaf, fi);
4590
4591                                 /* FIXME blocksize != 4096 */
4592                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4593                                 if (extent_start != 0) {
4594                                         found_extent = 1;
4595                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4596                                                      &root->state))
4597                                                 inode_sub_bytes(inode, num_dec);
4598                                 }
4599                         }
4600                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4601                         /*
4602                          * we can't truncate inline items that have had
4603                          * special encodings
4604                          */
4605                         if (!del_item &&
4606                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4607                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4608                             btrfs_file_extent_compression(leaf, fi) == 0) {
4609                                 u32 size = (u32)(new_size - found_key.offset);
4610
4611                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4612                                 size = btrfs_file_extent_calc_inline_size(size);
4613                                 btrfs_truncate_item(root->fs_info, path, size, 1);
4614                         } else if (!del_item) {
4615                                 /*
4616                                  * We have to bail so the last_size is set to
4617                                  * just before this extent.
4618                                  */
4619                                 err = NEED_TRUNCATE_BLOCK;
4620                                 break;
4621                         }
4622
4623                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4624                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4625                 }
4626 delete:
4627                 if (del_item)
4628                         last_size = found_key.offset;
4629                 else
4630                         last_size = new_size;
4631                 if (del_item) {
4632                         if (!pending_del_nr) {
4633                                 /* no pending yet, add ourselves */
4634                                 pending_del_slot = path->slots[0];
4635                                 pending_del_nr = 1;
4636                         } else if (pending_del_nr &&
4637                                    path->slots[0] + 1 == pending_del_slot) {
4638                                 /* hop on the pending chunk */
4639                                 pending_del_nr++;
4640                                 pending_del_slot = path->slots[0];
4641                         } else {
4642                                 BUG();
4643                         }
4644                 } else {
4645                         break;
4646                 }
4647                 should_throttle = false;
4648
4649                 if (found_extent &&
4650                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4651                      root == fs_info->tree_root)) {
4652                         btrfs_set_path_blocking(path);
4653                         bytes_deleted += extent_num_bytes;
4654                         ret = btrfs_free_extent(trans, root, extent_start,
4655                                                 extent_num_bytes, 0,
4656                                                 btrfs_header_owner(leaf),
4657                                                 ino, extent_offset);
4658                         BUG_ON(ret);
4659                         if (btrfs_should_throttle_delayed_refs(trans, fs_info))
4660                                 btrfs_async_run_delayed_refs(fs_info,
4661                                         trans->delayed_ref_updates * 2,
4662                                         trans->transid, 0);
4663                         if (be_nice) {
4664                                 if (truncate_space_check(trans, root,
4665                                                          extent_num_bytes)) {
4666                                         should_end = true;
4667                                 }
4668                                 if (btrfs_should_throttle_delayed_refs(trans,
4669                                                                        fs_info))
4670                                         should_throttle = true;
4671                         }
4672                 }
4673
4674                 if (found_type == BTRFS_INODE_ITEM_KEY)
4675                         break;
4676
4677                 if (path->slots[0] == 0 ||
4678                     path->slots[0] != pending_del_slot ||
4679                     should_throttle || should_end) {
4680                         if (pending_del_nr) {
4681                                 ret = btrfs_del_items(trans, root, path,
4682                                                 pending_del_slot,
4683                                                 pending_del_nr);
4684                                 if (ret) {
4685                                         btrfs_abort_transaction(trans, ret);
4686                                         goto error;
4687                                 }
4688                                 pending_del_nr = 0;
4689                         }
4690                         btrfs_release_path(path);
4691                         if (should_throttle) {
4692                                 unsigned long updates = trans->delayed_ref_updates;
4693                                 if (updates) {
4694                                         trans->delayed_ref_updates = 0;
4695                                         ret = btrfs_run_delayed_refs(trans,
4696                                                                    fs_info,
4697                                                                    updates * 2);
4698                                         if (ret && !err)
4699                                                 err = ret;
4700                                 }
4701                         }
4702                         /*
4703                          * if we failed to refill our space rsv, bail out
4704                          * and let the transaction restart
4705                          */
4706                         if (should_end) {
4707                                 err = -EAGAIN;
4708                                 goto error;
4709                         }
4710                         goto search_again;
4711                 } else {
4712                         path->slots[0]--;
4713                 }
4714         }
4715 out:
4716         if (pending_del_nr) {
4717                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4718                                       pending_del_nr);
4719                 if (ret)
4720                         btrfs_abort_transaction(trans, ret);
4721         }
4722 error:
4723         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4724                 ASSERT(last_size >= new_size);
4725                 if (!err && last_size > new_size)
4726                         last_size = new_size;
4727                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4728         }
4729
4730         btrfs_free_path(path);
4731
4732         if (be_nice && bytes_deleted > SZ_32M) {
4733                 unsigned long updates = trans->delayed_ref_updates;
4734                 if (updates) {
4735                         trans->delayed_ref_updates = 0;
4736                         ret = btrfs_run_delayed_refs(trans, fs_info,
4737                                                      updates * 2);
4738                         if (ret && !err)
4739                                 err = ret;
4740                 }
4741         }
4742         return err;
4743 }
4744
4745 /*
4746  * btrfs_truncate_block - read, zero a chunk and write a block
4747  * @inode - inode that we're zeroing
4748  * @from - the offset to start zeroing
4749  * @len - the length to zero, 0 to zero the entire range respective to the
4750  *      offset
4751  * @front - zero up to the offset instead of from the offset on
4752  *
4753  * This will find the block for the "from" offset and cow the block and zero the
4754  * part we want to zero.  This is used with truncate and hole punching.
4755  */
4756 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4757                         int front)
4758 {
4759         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4760         struct address_space *mapping = inode->i_mapping;
4761         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4762         struct btrfs_ordered_extent *ordered;
4763         struct extent_state *cached_state = NULL;
4764         struct extent_changeset *data_reserved = NULL;
4765         char *kaddr;
4766         u32 blocksize = fs_info->sectorsize;
4767         pgoff_t index = from >> PAGE_SHIFT;
4768         unsigned offset = from & (blocksize - 1);
4769         struct page *page;
4770         gfp_t mask = btrfs_alloc_write_mask(mapping);
4771         int ret = 0;
4772         u64 block_start;
4773         u64 block_end;
4774
4775         if (IS_ALIGNED(offset, blocksize) &&
4776             (!len || IS_ALIGNED(len, blocksize)))
4777                 goto out;
4778
4779         block_start = round_down(from, blocksize);
4780         block_end = block_start + blocksize - 1;
4781
4782         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4783                                            block_start, blocksize);
4784         if (ret)
4785                 goto out;
4786
4787 again:
4788         page = find_or_create_page(mapping, index, mask);
4789         if (!page) {
4790                 btrfs_delalloc_release_space(inode, data_reserved,
4791                                              block_start, blocksize);
4792                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4793                 ret = -ENOMEM;
4794                 goto out;
4795         }
4796
4797         if (!PageUptodate(page)) {
4798                 ret = btrfs_readpage(NULL, page);
4799                 lock_page(page);
4800                 if (page->mapping != mapping) {
4801                         unlock_page(page);
4802                         put_page(page);
4803                         goto again;
4804                 }
4805                 if (!PageUptodate(page)) {
4806                         ret = -EIO;
4807                         goto out_unlock;
4808                 }
4809         }
4810         wait_on_page_writeback(page);
4811
4812         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4813         set_page_extent_mapped(page);
4814
4815         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4816         if (ordered) {
4817                 unlock_extent_cached(io_tree, block_start, block_end,
4818                                      &cached_state);
4819                 unlock_page(page);
4820                 put_page(page);
4821                 btrfs_start_ordered_extent(inode, ordered, 1);
4822                 btrfs_put_ordered_extent(ordered);
4823                 goto again;
4824         }
4825
4826         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4827                           EXTENT_DIRTY | EXTENT_DELALLOC |
4828                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4829                           0, 0, &cached_state);
4830
4831         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
4832                                         &cached_state, 0);
4833         if (ret) {
4834                 unlock_extent_cached(io_tree, block_start, block_end,
4835                                      &cached_state);
4836                 goto out_unlock;
4837         }
4838
4839         if (offset != blocksize) {
4840                 if (!len)
4841                         len = blocksize - offset;
4842                 kaddr = kmap(page);
4843                 if (front)
4844                         memset(kaddr + (block_start - page_offset(page)),
4845                                 0, offset);
4846                 else
4847                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4848                                 0, len);
4849                 flush_dcache_page(page);
4850                 kunmap(page);
4851         }
4852         ClearPageChecked(page);
4853         set_page_dirty(page);
4854         unlock_extent_cached(io_tree, block_start, block_end, &cached_state);
4855
4856 out_unlock:
4857         if (ret)
4858                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
4859                                              blocksize);
4860         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4861         unlock_page(page);
4862         put_page(page);
4863 out:
4864         extent_changeset_free(data_reserved);
4865         return ret;
4866 }
4867
4868 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4869                              u64 offset, u64 len)
4870 {
4871         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4872         struct btrfs_trans_handle *trans;
4873         int ret;
4874
4875         /*
4876          * Still need to make sure the inode looks like it's been updated so
4877          * that any holes get logged if we fsync.
4878          */
4879         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4880                 BTRFS_I(inode)->last_trans = fs_info->generation;
4881                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4882                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4883                 return 0;
4884         }
4885
4886         /*
4887          * 1 - for the one we're dropping
4888          * 1 - for the one we're adding
4889          * 1 - for updating the inode.
4890          */
4891         trans = btrfs_start_transaction(root, 3);
4892         if (IS_ERR(trans))
4893                 return PTR_ERR(trans);
4894
4895         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4896         if (ret) {
4897                 btrfs_abort_transaction(trans, ret);
4898                 btrfs_end_transaction(trans);
4899                 return ret;
4900         }
4901
4902         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4903                         offset, 0, 0, len, 0, len, 0, 0, 0);
4904         if (ret)
4905                 btrfs_abort_transaction(trans, ret);
4906         else
4907                 btrfs_update_inode(trans, root, inode);
4908         btrfs_end_transaction(trans);
4909         return ret;
4910 }
4911
4912 /*
4913  * This function puts in dummy file extents for the area we're creating a hole
4914  * for.  So if we are truncating this file to a larger size we need to insert
4915  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4916  * the range between oldsize and size
4917  */
4918 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4919 {
4920         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4921         struct btrfs_root *root = BTRFS_I(inode)->root;
4922         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4923         struct extent_map *em = NULL;
4924         struct extent_state *cached_state = NULL;
4925         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4926         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
4927         u64 block_end = ALIGN(size, fs_info->sectorsize);
4928         u64 last_byte;
4929         u64 cur_offset;
4930         u64 hole_size;
4931         int err = 0;
4932
4933         /*
4934          * If our size started in the middle of a block we need to zero out the
4935          * rest of the block before we expand the i_size, otherwise we could
4936          * expose stale data.
4937          */
4938         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4939         if (err)
4940                 return err;
4941
4942         if (size <= hole_start)
4943                 return 0;
4944
4945         while (1) {
4946                 struct btrfs_ordered_extent *ordered;
4947
4948                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4949                                  &cached_state);
4950                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), hole_start,
4951                                                      block_end - hole_start);
4952                 if (!ordered)
4953                         break;
4954                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4955                                      &cached_state);
4956                 btrfs_start_ordered_extent(inode, ordered, 1);
4957                 btrfs_put_ordered_extent(ordered);
4958         }
4959
4960         cur_offset = hole_start;
4961         while (1) {
4962                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
4963                                 block_end - cur_offset, 0);
4964                 if (IS_ERR(em)) {
4965                         err = PTR_ERR(em);
4966                         em = NULL;
4967                         break;
4968                 }
4969                 last_byte = min(extent_map_end(em), block_end);
4970                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
4971                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4972                         struct extent_map *hole_em;
4973                         hole_size = last_byte - cur_offset;
4974
4975                         err = maybe_insert_hole(root, inode, cur_offset,
4976                                                 hole_size);
4977                         if (err)
4978                                 break;
4979                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
4980                                                 cur_offset + hole_size - 1, 0);
4981                         hole_em = alloc_extent_map();
4982                         if (!hole_em) {
4983                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4984                                         &BTRFS_I(inode)->runtime_flags);
4985                                 goto next;
4986                         }
4987                         hole_em->start = cur_offset;
4988                         hole_em->len = hole_size;
4989                         hole_em->orig_start = cur_offset;
4990
4991                         hole_em->block_start = EXTENT_MAP_HOLE;
4992                         hole_em->block_len = 0;
4993                         hole_em->orig_block_len = 0;
4994                         hole_em->ram_bytes = hole_size;
4995                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
4996                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4997                         hole_em->generation = fs_info->generation;
4998
4999                         while (1) {
5000                                 write_lock(&em_tree->lock);
5001                                 err = add_extent_mapping(em_tree, hole_em, 1);
5002                                 write_unlock(&em_tree->lock);
5003                                 if (err != -EEXIST)
5004                                         break;
5005                                 btrfs_drop_extent_cache(BTRFS_I(inode),
5006                                                         cur_offset,
5007                                                         cur_offset +
5008                                                         hole_size - 1, 0);
5009                         }
5010                         free_extent_map(hole_em);
5011                 }
5012 next:
5013                 free_extent_map(em);
5014                 em = NULL;
5015                 cur_offset = last_byte;
5016                 if (cur_offset >= block_end)
5017                         break;
5018         }
5019         free_extent_map(em);
5020         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state);
5021         return err;
5022 }
5023
5024 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5025 {
5026         struct btrfs_root *root = BTRFS_I(inode)->root;
5027         struct btrfs_trans_handle *trans;
5028         loff_t oldsize = i_size_read(inode);
5029         loff_t newsize = attr->ia_size;
5030         int mask = attr->ia_valid;
5031         int ret;
5032
5033         /*
5034          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5035          * special case where we need to update the times despite not having
5036          * these flags set.  For all other operations the VFS set these flags
5037          * explicitly if it wants a timestamp update.
5038          */
5039         if (newsize != oldsize) {
5040                 inode_inc_iversion(inode);
5041                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5042                         inode->i_ctime = inode->i_mtime =
5043                                 current_time(inode);
5044         }
5045
5046         if (newsize > oldsize) {
5047                 /*
5048                  * Don't do an expanding truncate while snapshotting is ongoing.
5049                  * This is to ensure the snapshot captures a fully consistent
5050                  * state of this file - if the snapshot captures this expanding
5051                  * truncation, it must capture all writes that happened before
5052                  * this truncation.
5053                  */
5054                 btrfs_wait_for_snapshot_creation(root);
5055                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5056                 if (ret) {
5057                         btrfs_end_write_no_snapshotting(root);
5058                         return ret;
5059                 }
5060
5061                 trans = btrfs_start_transaction(root, 1);
5062                 if (IS_ERR(trans)) {
5063                         btrfs_end_write_no_snapshotting(root);
5064                         return PTR_ERR(trans);
5065                 }
5066
5067                 i_size_write(inode, newsize);
5068                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5069                 pagecache_isize_extended(inode, oldsize, newsize);
5070                 ret = btrfs_update_inode(trans, root, inode);
5071                 btrfs_end_write_no_snapshotting(root);
5072                 btrfs_end_transaction(trans);
5073         } else {
5074
5075                 /*
5076                  * We're truncating a file that used to have good data down to
5077                  * zero. Make sure it gets into the ordered flush list so that
5078                  * any new writes get down to disk quickly.
5079                  */
5080                 if (newsize == 0)
5081                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5082                                 &BTRFS_I(inode)->runtime_flags);
5083
5084                 /*
5085                  * 1 for the orphan item we're going to add
5086                  * 1 for the orphan item deletion.
5087                  */
5088                 trans = btrfs_start_transaction(root, 2);
5089                 if (IS_ERR(trans))
5090                         return PTR_ERR(trans);
5091
5092                 /*
5093                  * We need to do this in case we fail at _any_ point during the
5094                  * actual truncate.  Once we do the truncate_setsize we could
5095                  * invalidate pages which forces any outstanding ordered io to
5096                  * be instantly completed which will give us extents that need
5097                  * to be truncated.  If we fail to get an orphan inode down we
5098                  * could have left over extents that were never meant to live,
5099                  * so we need to guarantee from this point on that everything
5100                  * will be consistent.
5101                  */
5102                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
5103                 btrfs_end_transaction(trans);
5104                 if (ret)
5105                         return ret;
5106
5107                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
5108                 truncate_setsize(inode, newsize);
5109
5110                 /* Disable nonlocked read DIO to avoid the end less truncate */
5111                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5112                 inode_dio_wait(inode);
5113                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5114
5115                 ret = btrfs_truncate(inode);
5116                 if (ret && inode->i_nlink) {
5117                         int err;
5118
5119                         /* To get a stable disk_i_size */
5120                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5121                         if (err) {
5122                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5123                                 return err;
5124                         }
5125
5126                         /*
5127                          * failed to truncate, disk_i_size is only adjusted down
5128                          * as we remove extents, so it should represent the true
5129                          * size of the inode, so reset the in memory size and
5130                          * delete our orphan entry.
5131                          */
5132                         trans = btrfs_join_transaction(root);
5133                         if (IS_ERR(trans)) {
5134                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5135                                 return ret;
5136                         }
5137                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5138                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
5139                         if (err)
5140                                 btrfs_abort_transaction(trans, err);
5141                         btrfs_end_transaction(trans);
5142                 }
5143         }
5144
5145         return ret;
5146 }
5147
5148 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5149 {
5150         struct inode *inode = d_inode(dentry);
5151         struct btrfs_root *root = BTRFS_I(inode)->root;
5152         int err;
5153
5154         if (btrfs_root_readonly(root))
5155                 return -EROFS;
5156
5157         err = setattr_prepare(dentry, attr);
5158         if (err)
5159                 return err;
5160
5161         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5162                 err = btrfs_setsize(inode, attr);
5163                 if (err)
5164                         return err;
5165         }
5166
5167         if (attr->ia_valid) {
5168                 setattr_copy(inode, attr);
5169                 inode_inc_iversion(inode);
5170                 err = btrfs_dirty_inode(inode);
5171
5172                 if (!err && attr->ia_valid & ATTR_MODE)
5173                         err = posix_acl_chmod(inode, inode->i_mode);
5174         }
5175
5176         return err;
5177 }
5178
5179 /*
5180  * While truncating the inode pages during eviction, we get the VFS calling
5181  * btrfs_invalidatepage() against each page of the inode. This is slow because
5182  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5183  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5184  * extent_state structures over and over, wasting lots of time.
5185  *
5186  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5187  * those expensive operations on a per page basis and do only the ordered io
5188  * finishing, while we release here the extent_map and extent_state structures,
5189  * without the excessive merging and splitting.
5190  */
5191 static void evict_inode_truncate_pages(struct inode *inode)
5192 {
5193         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5194         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5195         struct rb_node *node;
5196
5197         ASSERT(inode->i_state & I_FREEING);
5198         truncate_inode_pages_final(&inode->i_data);
5199
5200         write_lock(&map_tree->lock);
5201         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5202                 struct extent_map *em;
5203
5204                 node = rb_first(&map_tree->map);
5205                 em = rb_entry(node, struct extent_map, rb_node);
5206                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5207                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5208                 remove_extent_mapping(map_tree, em);
5209                 free_extent_map(em);
5210                 if (need_resched()) {
5211                         write_unlock(&map_tree->lock);
5212                         cond_resched();
5213                         write_lock(&map_tree->lock);
5214                 }
5215         }
5216         write_unlock(&map_tree->lock);
5217
5218         /*
5219          * Keep looping until we have no more ranges in the io tree.
5220          * We can have ongoing bios started by readpages (called from readahead)
5221          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5222          * still in progress (unlocked the pages in the bio but did not yet
5223          * unlocked the ranges in the io tree). Therefore this means some
5224          * ranges can still be locked and eviction started because before
5225          * submitting those bios, which are executed by a separate task (work
5226          * queue kthread), inode references (inode->i_count) were not taken
5227          * (which would be dropped in the end io callback of each bio).
5228          * Therefore here we effectively end up waiting for those bios and
5229          * anyone else holding locked ranges without having bumped the inode's
5230          * reference count - if we don't do it, when they access the inode's
5231          * io_tree to unlock a range it may be too late, leading to an
5232          * use-after-free issue.
5233          */
5234         spin_lock(&io_tree->lock);
5235         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5236                 struct extent_state *state;
5237                 struct extent_state *cached_state = NULL;
5238                 u64 start;
5239                 u64 end;
5240
5241                 node = rb_first(&io_tree->state);
5242                 state = rb_entry(node, struct extent_state, rb_node);
5243                 start = state->start;
5244                 end = state->end;
5245                 spin_unlock(&io_tree->lock);
5246
5247                 lock_extent_bits(io_tree, start, end, &cached_state);
5248
5249                 /*
5250                  * If still has DELALLOC flag, the extent didn't reach disk,
5251                  * and its reserved space won't be freed by delayed_ref.
5252                  * So we need to free its reserved space here.
5253                  * (Refer to comment in btrfs_invalidatepage, case 2)
5254                  *
5255                  * Note, end is the bytenr of last byte, so we need + 1 here.
5256                  */
5257                 if (state->state & EXTENT_DELALLOC)
5258                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5259
5260                 clear_extent_bit(io_tree, start, end,
5261                                  EXTENT_LOCKED | EXTENT_DIRTY |
5262                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5263                                  EXTENT_DEFRAG, 1, 1, &cached_state);
5264
5265                 cond_resched();
5266                 spin_lock(&io_tree->lock);
5267         }
5268         spin_unlock(&io_tree->lock);
5269 }
5270
5271 void btrfs_evict_inode(struct inode *inode)
5272 {
5273         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5274         struct btrfs_trans_handle *trans;
5275         struct btrfs_root *root = BTRFS_I(inode)->root;
5276         struct btrfs_block_rsv *rsv, *global_rsv;
5277         int steal_from_global = 0;
5278         u64 min_size;
5279         int ret;
5280
5281         trace_btrfs_inode_evict(inode);
5282
5283         if (!root) {
5284                 clear_inode(inode);
5285                 return;
5286         }
5287
5288         min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
5289
5290         evict_inode_truncate_pages(inode);
5291
5292         if (inode->i_nlink &&
5293             ((btrfs_root_refs(&root->root_item) != 0 &&
5294               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5295              btrfs_is_free_space_inode(BTRFS_I(inode))))
5296                 goto no_delete;
5297
5298         if (is_bad_inode(inode)) {
5299                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5300                 goto no_delete;
5301         }
5302         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5303         if (!special_file(inode->i_mode))
5304                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5305
5306         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5307
5308         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
5309                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5310                                  &BTRFS_I(inode)->runtime_flags));
5311                 goto no_delete;
5312         }
5313
5314         if (inode->i_nlink > 0) {
5315                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5316                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5317                 goto no_delete;
5318         }
5319
5320         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5321         if (ret) {
5322                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5323                 goto no_delete;
5324         }
5325
5326         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5327         if (!rsv) {
5328                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5329                 goto no_delete;
5330         }
5331         rsv->size = min_size;
5332         rsv->failfast = 1;
5333         global_rsv = &fs_info->global_block_rsv;
5334
5335         btrfs_i_size_write(BTRFS_I(inode), 0);
5336
5337         /*
5338          * This is a bit simpler than btrfs_truncate since we've already
5339          * reserved our space for our orphan item in the unlink, so we just
5340          * need to reserve some slack space in case we add bytes and update
5341          * inode item when doing the truncate.
5342          */
5343         while (1) {
5344                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5345                                              BTRFS_RESERVE_FLUSH_LIMIT);
5346
5347                 /*
5348                  * Try and steal from the global reserve since we will
5349                  * likely not use this space anyway, we want to try as
5350                  * hard as possible to get this to work.
5351                  */
5352                 if (ret)
5353                         steal_from_global++;
5354                 else
5355                         steal_from_global = 0;
5356                 ret = 0;
5357
5358                 /*
5359                  * steal_from_global == 0: we reserved stuff, hooray!
5360                  * steal_from_global == 1: we didn't reserve stuff, boo!
5361                  * steal_from_global == 2: we've committed, still not a lot of
5362                  * room but maybe we'll have room in the global reserve this
5363                  * time.
5364                  * steal_from_global == 3: abandon all hope!
5365                  */
5366                 if (steal_from_global > 2) {
5367                         btrfs_warn(fs_info,
5368                                    "Could not get space for a delete, will truncate on mount %d",
5369                                    ret);
5370                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5371                         btrfs_free_block_rsv(fs_info, rsv);
5372                         goto no_delete;
5373                 }
5374
5375                 trans = btrfs_join_transaction(root);
5376                 if (IS_ERR(trans)) {
5377                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5378                         btrfs_free_block_rsv(fs_info, rsv);
5379                         goto no_delete;
5380                 }
5381
5382                 /*
5383                  * We can't just steal from the global reserve, we need to make
5384                  * sure there is room to do it, if not we need to commit and try
5385                  * again.
5386                  */
5387                 if (steal_from_global) {
5388                         if (!btrfs_check_space_for_delayed_refs(trans, fs_info))
5389                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5390                                                               min_size, 0);
5391                         else
5392                                 ret = -ENOSPC;
5393                 }
5394
5395                 /*
5396                  * Couldn't steal from the global reserve, we have too much
5397                  * pending stuff built up, commit the transaction and try it
5398                  * again.
5399                  */
5400                 if (ret) {
5401                         ret = btrfs_commit_transaction(trans);
5402                         if (ret) {
5403                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5404                                 btrfs_free_block_rsv(fs_info, rsv);
5405                                 goto no_delete;
5406                         }
5407                         continue;
5408                 } else {
5409                         steal_from_global = 0;
5410                 }
5411
5412                 trans->block_rsv = rsv;
5413
5414                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5415                 if (ret != -ENOSPC && ret != -EAGAIN)
5416                         break;
5417
5418                 trans->block_rsv = &fs_info->trans_block_rsv;
5419                 btrfs_end_transaction(trans);
5420                 trans = NULL;
5421                 btrfs_btree_balance_dirty(fs_info);
5422         }
5423
5424         btrfs_free_block_rsv(fs_info, rsv);
5425
5426         /*
5427          * Errors here aren't a big deal, it just means we leave orphan items
5428          * in the tree.  They will be cleaned up on the next mount.
5429          */
5430         if (ret == 0) {
5431                 trans->block_rsv = root->orphan_block_rsv;
5432                 btrfs_orphan_del(trans, BTRFS_I(inode));
5433         } else {
5434                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5435         }
5436
5437         trans->block_rsv = &fs_info->trans_block_rsv;
5438         if (!(root == fs_info->tree_root ||
5439               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5440                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5441
5442         btrfs_end_transaction(trans);
5443         btrfs_btree_balance_dirty(fs_info);
5444 no_delete:
5445         btrfs_remove_delayed_node(BTRFS_I(inode));
5446         clear_inode(inode);
5447 }
5448
5449 /*
5450  * this returns the key found in the dir entry in the location pointer.
5451  * If no dir entries were found, location->objectid is 0.
5452  */
5453 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5454                                struct btrfs_key *location)
5455 {
5456         const char *name = dentry->d_name.name;
5457         int namelen = dentry->d_name.len;
5458         struct btrfs_dir_item *di;
5459         struct btrfs_path *path;
5460         struct btrfs_root *root = BTRFS_I(dir)->root;
5461         int ret = 0;
5462
5463         path = btrfs_alloc_path();
5464         if (!path)
5465                 return -ENOMEM;
5466
5467         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5468                         name, namelen, 0);
5469         if (IS_ERR(di))
5470                 ret = PTR_ERR(di);
5471
5472         if (IS_ERR_OR_NULL(di))
5473                 goto out_err;
5474
5475         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5476         if (location->type != BTRFS_INODE_ITEM_KEY &&
5477             location->type != BTRFS_ROOT_ITEM_KEY) {
5478                 btrfs_warn(root->fs_info,
5479 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5480                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5481                            location->objectid, location->type, location->offset);
5482                 goto out_err;
5483         }
5484 out:
5485         btrfs_free_path(path);
5486         return ret;
5487 out_err:
5488         location->objectid = 0;
5489         goto out;
5490 }
5491
5492 /*
5493  * when we hit a tree root in a directory, the btrfs part of the inode
5494  * needs to be changed to reflect the root directory of the tree root.  This
5495  * is kind of like crossing a mount point.
5496  */
5497 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5498                                     struct inode *dir,
5499                                     struct dentry *dentry,
5500                                     struct btrfs_key *location,
5501                                     struct btrfs_root **sub_root)
5502 {
5503         struct btrfs_path *path;
5504         struct btrfs_root *new_root;
5505         struct btrfs_root_ref *ref;
5506         struct extent_buffer *leaf;
5507         struct btrfs_key key;
5508         int ret;
5509         int err = 0;
5510
5511         path = btrfs_alloc_path();
5512         if (!path) {
5513                 err = -ENOMEM;
5514                 goto out;
5515         }
5516
5517         err = -ENOENT;
5518         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5519         key.type = BTRFS_ROOT_REF_KEY;
5520         key.offset = location->objectid;
5521
5522         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5523         if (ret) {
5524                 if (ret < 0)
5525                         err = ret;
5526                 goto out;
5527         }
5528
5529         leaf = path->nodes[0];
5530         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5531         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5532             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5533                 goto out;
5534
5535         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5536                                    (unsigned long)(ref + 1),
5537                                    dentry->d_name.len);
5538         if (ret)
5539                 goto out;
5540
5541         btrfs_release_path(path);
5542
5543         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5544         if (IS_ERR(new_root)) {
5545                 err = PTR_ERR(new_root);
5546                 goto out;
5547         }
5548
5549         *sub_root = new_root;
5550         location->objectid = btrfs_root_dirid(&new_root->root_item);
5551         location->type = BTRFS_INODE_ITEM_KEY;
5552         location->offset = 0;
5553         err = 0;
5554 out:
5555         btrfs_free_path(path);
5556         return err;
5557 }
5558
5559 static void inode_tree_add(struct inode *inode)
5560 {
5561         struct btrfs_root *root = BTRFS_I(inode)->root;
5562         struct btrfs_inode *entry;
5563         struct rb_node **p;
5564         struct rb_node *parent;
5565         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5566         u64 ino = btrfs_ino(BTRFS_I(inode));
5567
5568         if (inode_unhashed(inode))
5569                 return;
5570         parent = NULL;
5571         spin_lock(&root->inode_lock);
5572         p = &root->inode_tree.rb_node;
5573         while (*p) {
5574                 parent = *p;
5575                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5576
5577                 if (ino < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5578                         p = &parent->rb_left;
5579                 else if (ino > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5580                         p = &parent->rb_right;
5581                 else {
5582                         WARN_ON(!(entry->vfs_inode.i_state &
5583                                   (I_WILL_FREE | I_FREEING)));
5584                         rb_replace_node(parent, new, &root->inode_tree);
5585                         RB_CLEAR_NODE(parent);
5586                         spin_unlock(&root->inode_lock);
5587                         return;
5588                 }
5589         }
5590         rb_link_node(new, parent, p);
5591         rb_insert_color(new, &root->inode_tree);
5592         spin_unlock(&root->inode_lock);
5593 }
5594
5595 static void inode_tree_del(struct inode *inode)
5596 {
5597         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5598         struct btrfs_root *root = BTRFS_I(inode)->root;
5599         int empty = 0;
5600
5601         spin_lock(&root->inode_lock);
5602         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5603                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5604                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5605                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5606         }
5607         spin_unlock(&root->inode_lock);
5608
5609         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5610                 synchronize_srcu(&fs_info->subvol_srcu);
5611                 spin_lock(&root->inode_lock);
5612                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5613                 spin_unlock(&root->inode_lock);
5614                 if (empty)
5615                         btrfs_add_dead_root(root);
5616         }
5617 }
5618
5619 void btrfs_invalidate_inodes(struct btrfs_root *root)
5620 {
5621         struct btrfs_fs_info *fs_info = root->fs_info;
5622         struct rb_node *node;
5623         struct rb_node *prev;
5624         struct btrfs_inode *entry;
5625         struct inode *inode;
5626         u64 objectid = 0;
5627
5628         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
5629                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5630
5631         spin_lock(&root->inode_lock);
5632 again:
5633         node = root->inode_tree.rb_node;
5634         prev = NULL;
5635         while (node) {
5636                 prev = node;
5637                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5638
5639                 if (objectid < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5640                         node = node->rb_left;
5641                 else if (objectid > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5642                         node = node->rb_right;
5643                 else
5644                         break;
5645         }
5646         if (!node) {
5647                 while (prev) {
5648                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5649                         if (objectid <= btrfs_ino(BTRFS_I(&entry->vfs_inode))) {
5650                                 node = prev;
5651                                 break;
5652                         }
5653                         prev = rb_next(prev);
5654                 }
5655         }
5656         while (node) {
5657                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5658                 objectid = btrfs_ino(BTRFS_I(&entry->vfs_inode)) + 1;
5659                 inode = igrab(&entry->vfs_inode);
5660                 if (inode) {
5661                         spin_unlock(&root->inode_lock);
5662                         if (atomic_read(&inode->i_count) > 1)
5663                                 d_prune_aliases(inode);
5664                         /*
5665                          * btrfs_drop_inode will have it removed from
5666                          * the inode cache when its usage count
5667                          * hits zero.
5668                          */
5669                         iput(inode);
5670                         cond_resched();
5671                         spin_lock(&root->inode_lock);
5672                         goto again;
5673                 }
5674
5675                 if (cond_resched_lock(&root->inode_lock))
5676                         goto again;
5677
5678                 node = rb_next(node);
5679         }
5680         spin_unlock(&root->inode_lock);
5681 }
5682
5683 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5684 {
5685         struct btrfs_iget_args *args = p;
5686         inode->i_ino = args->location->objectid;
5687         memcpy(&BTRFS_I(inode)->location, args->location,
5688                sizeof(*args->location));
5689         BTRFS_I(inode)->root = args->root;
5690         return 0;
5691 }
5692
5693 static int btrfs_find_actor(struct inode *inode, void *opaque)
5694 {
5695         struct btrfs_iget_args *args = opaque;
5696         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5697                 args->root == BTRFS_I(inode)->root;
5698 }
5699
5700 static struct inode *btrfs_iget_locked(struct super_block *s,
5701                                        struct btrfs_key *location,
5702                                        struct btrfs_root *root)
5703 {
5704         struct inode *inode;
5705         struct btrfs_iget_args args;
5706         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5707
5708         args.location = location;
5709         args.root = root;
5710
5711         inode = iget5_locked(s, hashval, btrfs_find_actor,
5712                              btrfs_init_locked_inode,
5713                              (void *)&args);
5714         return inode;
5715 }
5716
5717 /* Get an inode object given its location and corresponding root.
5718  * Returns in *is_new if the inode was read from disk
5719  */
5720 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5721                          struct btrfs_root *root, int *new)
5722 {
5723         struct inode *inode;
5724
5725         inode = btrfs_iget_locked(s, location, root);
5726         if (!inode)
5727                 return ERR_PTR(-ENOMEM);
5728
5729         if (inode->i_state & I_NEW) {
5730                 int ret;
5731
5732                 ret = btrfs_read_locked_inode(inode);
5733                 if (!is_bad_inode(inode)) {
5734                         inode_tree_add(inode);
5735                         unlock_new_inode(inode);
5736                         if (new)
5737                                 *new = 1;
5738                 } else {
5739                         unlock_new_inode(inode);
5740                         iput(inode);
5741                         ASSERT(ret < 0);
5742                         inode = ERR_PTR(ret < 0 ? ret : -ESTALE);
5743                 }
5744         }
5745
5746         return inode;
5747 }
5748
5749 static struct inode *new_simple_dir(struct super_block *s,
5750                                     struct btrfs_key *key,
5751                                     struct btrfs_root *root)
5752 {
5753         struct inode *inode = new_inode(s);
5754
5755         if (!inode)
5756                 return ERR_PTR(-ENOMEM);
5757
5758         BTRFS_I(inode)->root = root;
5759         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5760         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5761
5762         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5763         inode->i_op = &btrfs_dir_ro_inode_operations;
5764         inode->i_opflags &= ~IOP_XATTR;
5765         inode->i_fop = &simple_dir_operations;
5766         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5767         inode->i_mtime = current_time(inode);
5768         inode->i_atime = inode->i_mtime;
5769         inode->i_ctime = inode->i_mtime;
5770         BTRFS_I(inode)->i_otime = inode->i_mtime;
5771
5772         return inode;
5773 }
5774
5775 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5776 {
5777         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5778         struct inode *inode;
5779         struct btrfs_root *root = BTRFS_I(dir)->root;
5780         struct btrfs_root *sub_root = root;
5781         struct btrfs_key location;
5782         int index;
5783         int ret = 0;
5784
5785         if (dentry->d_name.len > BTRFS_NAME_LEN)
5786                 return ERR_PTR(-ENAMETOOLONG);
5787
5788         ret = btrfs_inode_by_name(dir, dentry, &location);
5789         if (ret < 0)
5790                 return ERR_PTR(ret);
5791
5792         if (location.objectid == 0)
5793                 return ERR_PTR(-ENOENT);
5794
5795         if (location.type == BTRFS_INODE_ITEM_KEY) {
5796                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5797                 return inode;
5798         }
5799
5800         index = srcu_read_lock(&fs_info->subvol_srcu);
5801         ret = fixup_tree_root_location(fs_info, dir, dentry,
5802                                        &location, &sub_root);
5803         if (ret < 0) {
5804                 if (ret != -ENOENT)
5805                         inode = ERR_PTR(ret);
5806                 else
5807                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5808         } else {
5809                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5810         }
5811         srcu_read_unlock(&fs_info->subvol_srcu, index);
5812
5813         if (!IS_ERR(inode) && root != sub_root) {
5814                 down_read(&fs_info->cleanup_work_sem);
5815                 if (!sb_rdonly(inode->i_sb))
5816                         ret = btrfs_orphan_cleanup(sub_root);
5817                 up_read(&fs_info->cleanup_work_sem);
5818                 if (ret) {
5819                         iput(inode);
5820                         inode = ERR_PTR(ret);
5821                 }
5822         }
5823
5824         return inode;
5825 }
5826
5827 static int btrfs_dentry_delete(const struct dentry *dentry)
5828 {
5829         struct btrfs_root *root;
5830         struct inode *inode = d_inode(dentry);
5831
5832         if (!inode && !IS_ROOT(dentry))
5833                 inode = d_inode(dentry->d_parent);
5834
5835         if (inode) {
5836                 root = BTRFS_I(inode)->root;
5837                 if (btrfs_root_refs(&root->root_item) == 0)
5838                         return 1;
5839
5840                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5841                         return 1;
5842         }
5843         return 0;
5844 }
5845
5846 static void btrfs_dentry_release(struct dentry *dentry)
5847 {
5848         kfree(dentry->d_fsdata);
5849 }
5850
5851 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5852                                    unsigned int flags)
5853 {
5854         struct inode *inode;
5855
5856         inode = btrfs_lookup_dentry(dir, dentry);
5857         if (IS_ERR(inode)) {
5858                 if (PTR_ERR(inode) == -ENOENT)
5859                         inode = NULL;
5860                 else
5861                         return ERR_CAST(inode);
5862         }
5863
5864         return d_splice_alias(inode, dentry);
5865 }
5866
5867 unsigned char btrfs_filetype_table[] = {
5868         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5869 };
5870
5871 /*
5872  * All this infrastructure exists because dir_emit can fault, and we are holding
5873  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5874  * our information into that, and then dir_emit from the buffer.  This is
5875  * similar to what NFS does, only we don't keep the buffer around in pagecache
5876  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5877  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5878  * tree lock.
5879  */
5880 static int btrfs_opendir(struct inode *inode, struct file *file)
5881 {
5882         struct btrfs_file_private *private;
5883
5884         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5885         if (!private)
5886                 return -ENOMEM;
5887         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5888         if (!private->filldir_buf) {
5889                 kfree(private);
5890                 return -ENOMEM;
5891         }
5892         file->private_data = private;
5893         return 0;
5894 }
5895
5896 struct dir_entry {
5897         u64 ino;
5898         u64 offset;
5899         unsigned type;
5900         int name_len;
5901 };
5902
5903 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5904 {
5905         while (entries--) {
5906                 struct dir_entry *entry = addr;
5907                 char *name = (char *)(entry + 1);
5908
5909                 ctx->pos = entry->offset;
5910                 if (!dir_emit(ctx, name, entry->name_len, entry->ino,
5911                               entry->type))
5912                         return 1;
5913                 addr += sizeof(struct dir_entry) + entry->name_len;
5914                 ctx->pos++;
5915         }
5916         return 0;
5917 }
5918
5919 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5920 {
5921         struct inode *inode = file_inode(file);
5922         struct btrfs_root *root = BTRFS_I(inode)->root;
5923         struct btrfs_file_private *private = file->private_data;
5924         struct btrfs_dir_item *di;
5925         struct btrfs_key key;
5926         struct btrfs_key found_key;
5927         struct btrfs_path *path;
5928         void *addr;
5929         struct list_head ins_list;
5930         struct list_head del_list;
5931         int ret;
5932         struct extent_buffer *leaf;
5933         int slot;
5934         char *name_ptr;
5935         int name_len;
5936         int entries = 0;
5937         int total_len = 0;
5938         bool put = false;
5939         struct btrfs_key location;
5940
5941         if (!dir_emit_dots(file, ctx))
5942                 return 0;
5943
5944         path = btrfs_alloc_path();
5945         if (!path)
5946                 return -ENOMEM;
5947
5948         addr = private->filldir_buf;
5949         path->reada = READA_FORWARD;
5950
5951         INIT_LIST_HEAD(&ins_list);
5952         INIT_LIST_HEAD(&del_list);
5953         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5954
5955 again:
5956         key.type = BTRFS_DIR_INDEX_KEY;
5957         key.offset = ctx->pos;
5958         key.objectid = btrfs_ino(BTRFS_I(inode));
5959
5960         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5961         if (ret < 0)
5962                 goto err;
5963
5964         while (1) {
5965                 struct dir_entry *entry;
5966
5967                 leaf = path->nodes[0];
5968                 slot = path->slots[0];
5969                 if (slot >= btrfs_header_nritems(leaf)) {
5970                         ret = btrfs_next_leaf(root, path);
5971                         if (ret < 0)
5972                                 goto err;
5973                         else if (ret > 0)
5974                                 break;
5975                         continue;
5976                 }
5977
5978                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5979
5980                 if (found_key.objectid != key.objectid)
5981                         break;
5982                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
5983                         break;
5984                 if (found_key.offset < ctx->pos)
5985                         goto next;
5986                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
5987                         goto next;
5988                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5989                 name_len = btrfs_dir_name_len(leaf, di);
5990                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
5991                     PAGE_SIZE) {
5992                         btrfs_release_path(path);
5993                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5994                         if (ret)
5995                                 goto nopos;
5996                         addr = private->filldir_buf;
5997                         entries = 0;
5998                         total_len = 0;
5999                         goto again;
6000                 }
6001
6002                 entry = addr;
6003                 entry->name_len = name_len;
6004                 name_ptr = (char *)(entry + 1);
6005                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
6006                                    name_len);
6007                 entry->type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
6008                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
6009                 entry->ino = location.objectid;
6010                 entry->offset = found_key.offset;
6011                 entries++;
6012                 addr += sizeof(struct dir_entry) + name_len;
6013                 total_len += sizeof(struct dir_entry) + name_len;
6014 next:
6015                 path->slots[0]++;
6016         }
6017         btrfs_release_path(path);
6018
6019         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6020         if (ret)
6021                 goto nopos;
6022
6023         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
6024         if (ret)
6025                 goto nopos;
6026
6027         /*
6028          * Stop new entries from being returned after we return the last
6029          * entry.
6030          *
6031          * New directory entries are assigned a strictly increasing
6032          * offset.  This means that new entries created during readdir
6033          * are *guaranteed* to be seen in the future by that readdir.
6034          * This has broken buggy programs which operate on names as
6035          * they're returned by readdir.  Until we re-use freed offsets
6036          * we have this hack to stop new entries from being returned
6037          * under the assumption that they'll never reach this huge
6038          * offset.
6039          *
6040          * This is being careful not to overflow 32bit loff_t unless the
6041          * last entry requires it because doing so has broken 32bit apps
6042          * in the past.
6043          */
6044         if (ctx->pos >= INT_MAX)
6045                 ctx->pos = LLONG_MAX;
6046         else
6047                 ctx->pos = INT_MAX;
6048 nopos:
6049         ret = 0;
6050 err:
6051         if (put)
6052                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
6053         btrfs_free_path(path);
6054         return ret;
6055 }
6056
6057 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
6058 {
6059         struct btrfs_root *root = BTRFS_I(inode)->root;
6060         struct btrfs_trans_handle *trans;
6061         int ret = 0;
6062         bool nolock = false;
6063
6064         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6065                 return 0;
6066
6067         if (btrfs_fs_closing(root->fs_info) &&
6068                         btrfs_is_free_space_inode(BTRFS_I(inode)))
6069                 nolock = true;
6070
6071         if (wbc->sync_mode == WB_SYNC_ALL) {
6072                 if (nolock)
6073                         trans = btrfs_join_transaction_nolock(root);
6074                 else
6075                         trans = btrfs_join_transaction(root);
6076                 if (IS_ERR(trans))
6077                         return PTR_ERR(trans);
6078                 ret = btrfs_commit_transaction(trans);
6079         }
6080         return ret;
6081 }
6082
6083 /*
6084  * This is somewhat expensive, updating the tree every time the
6085  * inode changes.  But, it is most likely to find the inode in cache.
6086  * FIXME, needs more benchmarking...there are no reasons other than performance
6087  * to keep or drop this code.
6088  */
6089 static int btrfs_dirty_inode(struct inode *inode)
6090 {
6091         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6092         struct btrfs_root *root = BTRFS_I(inode)->root;
6093         struct btrfs_trans_handle *trans;
6094         int ret;
6095
6096         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6097                 return 0;
6098
6099         trans = btrfs_join_transaction(root);
6100         if (IS_ERR(trans))
6101                 return PTR_ERR(trans);
6102
6103         ret = btrfs_update_inode(trans, root, inode);
6104         if (ret && ret == -ENOSPC) {
6105                 /* whoops, lets try again with the full transaction */
6106                 btrfs_end_transaction(trans);
6107                 trans = btrfs_start_transaction(root, 1);
6108                 if (IS_ERR(trans))
6109                         return PTR_ERR(trans);
6110
6111                 ret = btrfs_update_inode(trans, root, inode);
6112         }
6113         btrfs_end_transaction(trans);
6114         if (BTRFS_I(inode)->delayed_node)
6115                 btrfs_balance_delayed_items(fs_info);
6116
6117         return ret;
6118 }
6119
6120 /*
6121  * This is a copy of file_update_time.  We need this so we can return error on
6122  * ENOSPC for updating the inode in the case of file write and mmap writes.
6123  */
6124 static int btrfs_update_time(struct inode *inode, struct timespec *now,
6125                              int flags)
6126 {
6127         struct btrfs_root *root = BTRFS_I(inode)->root;
6128
6129         if (btrfs_root_readonly(root))
6130                 return -EROFS;
6131
6132         if (flags & S_VERSION)
6133                 inode_inc_iversion(inode);
6134         if (flags & S_CTIME)
6135                 inode->i_ctime = *now;
6136         if (flags & S_MTIME)
6137                 inode->i_mtime = *now;
6138         if (flags & S_ATIME)
6139                 inode->i_atime = *now;
6140         return btrfs_dirty_inode(inode);
6141 }
6142
6143 /*
6144  * find the highest existing sequence number in a directory
6145  * and then set the in-memory index_cnt variable to reflect
6146  * free sequence numbers
6147  */
6148 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6149 {
6150         struct btrfs_root *root = inode->root;
6151         struct btrfs_key key, found_key;
6152         struct btrfs_path *path;
6153         struct extent_buffer *leaf;
6154         int ret;
6155
6156         key.objectid = btrfs_ino(inode);
6157         key.type = BTRFS_DIR_INDEX_KEY;
6158         key.offset = (u64)-1;
6159
6160         path = btrfs_alloc_path();
6161         if (!path)
6162                 return -ENOMEM;
6163
6164         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6165         if (ret < 0)
6166                 goto out;
6167         /* FIXME: we should be able to handle this */
6168         if (ret == 0)
6169                 goto out;
6170         ret = 0;
6171
6172         /*
6173          * MAGIC NUMBER EXPLANATION:
6174          * since we search a directory based on f_pos we have to start at 2
6175          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6176          * else has to start at 2
6177          */
6178         if (path->slots[0] == 0) {
6179                 inode->index_cnt = 2;
6180                 goto out;
6181         }
6182
6183         path->slots[0]--;
6184
6185         leaf = path->nodes[0];
6186         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6187
6188         if (found_key.objectid != btrfs_ino(inode) ||
6189             found_key.type != BTRFS_DIR_INDEX_KEY) {
6190                 inode->index_cnt = 2;
6191                 goto out;
6192         }
6193
6194         inode->index_cnt = found_key.offset + 1;
6195 out:
6196         btrfs_free_path(path);
6197         return ret;
6198 }
6199
6200 /*
6201  * helper to find a free sequence number in a given directory.  This current
6202  * code is very simple, later versions will do smarter things in the btree
6203  */
6204 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6205 {
6206         int ret = 0;
6207
6208         if (dir->index_cnt == (u64)-1) {
6209                 ret = btrfs_inode_delayed_dir_index_count(dir);
6210                 if (ret) {
6211                         ret = btrfs_set_inode_index_count(dir);
6212                         if (ret)
6213                                 return ret;
6214                 }
6215         }
6216
6217         *index = dir->index_cnt;
6218         dir->index_cnt++;
6219
6220         return ret;
6221 }
6222
6223 static int btrfs_insert_inode_locked(struct inode *inode)
6224 {
6225         struct btrfs_iget_args args;
6226         args.location = &BTRFS_I(inode)->location;
6227         args.root = BTRFS_I(inode)->root;
6228
6229         return insert_inode_locked4(inode,
6230                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6231                    btrfs_find_actor, &args);
6232 }
6233
6234 /*
6235  * Inherit flags from the parent inode.
6236  *
6237  * Currently only the compression flags and the cow flags are inherited.
6238  */
6239 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6240 {
6241         unsigned int flags;
6242
6243         if (!dir)
6244                 return;
6245
6246         flags = BTRFS_I(dir)->flags;
6247
6248         if (flags & BTRFS_INODE_NOCOMPRESS) {
6249                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6250                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6251         } else if (flags & BTRFS_INODE_COMPRESS) {
6252                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6253                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6254         }
6255
6256         if (flags & BTRFS_INODE_NODATACOW) {
6257                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6258                 if (S_ISREG(inode->i_mode))
6259                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6260         }
6261
6262         btrfs_update_iflags(inode);
6263 }
6264
6265 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6266                                      struct btrfs_root *root,
6267                                      struct inode *dir,
6268                                      const char *name, int name_len,
6269                                      u64 ref_objectid, u64 objectid,
6270                                      umode_t mode, u64 *index)
6271 {
6272         struct btrfs_fs_info *fs_info = root->fs_info;
6273         struct inode *inode;
6274         struct btrfs_inode_item *inode_item;
6275         struct btrfs_key *location;
6276         struct btrfs_path *path;
6277         struct btrfs_inode_ref *ref;
6278         struct btrfs_key key[2];
6279         u32 sizes[2];
6280         int nitems = name ? 2 : 1;
6281         unsigned long ptr;
6282         int ret;
6283
6284         path = btrfs_alloc_path();
6285         if (!path)
6286                 return ERR_PTR(-ENOMEM);
6287
6288         inode = new_inode(fs_info->sb);
6289         if (!inode) {
6290                 btrfs_free_path(path);
6291                 return ERR_PTR(-ENOMEM);
6292         }
6293
6294         /*
6295          * O_TMPFILE, set link count to 0, so that after this point,
6296          * we fill in an inode item with the correct link count.
6297          */
6298         if (!name)
6299                 set_nlink(inode, 0);
6300
6301         /*
6302          * we have to initialize this early, so we can reclaim the inode
6303          * number if we fail afterwards in this function.
6304          */
6305         inode->i_ino = objectid;
6306
6307         if (dir && name) {
6308                 trace_btrfs_inode_request(dir);
6309
6310                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6311                 if (ret) {
6312                         btrfs_free_path(path);
6313                         iput(inode);
6314                         return ERR_PTR(ret);
6315                 }
6316         } else if (dir) {
6317                 *index = 0;
6318         }
6319         /*
6320          * index_cnt is ignored for everything but a dir,
6321          * btrfs_set_inode_index_count has an explanation for the magic
6322          * number
6323          */
6324         BTRFS_I(inode)->index_cnt = 2;
6325         BTRFS_I(inode)->dir_index = *index;
6326         BTRFS_I(inode)->root = root;
6327         BTRFS_I(inode)->generation = trans->transid;
6328         inode->i_generation = BTRFS_I(inode)->generation;
6329
6330         /*
6331          * We could have gotten an inode number from somebody who was fsynced
6332          * and then removed in this same transaction, so let's just set full
6333          * sync since it will be a full sync anyway and this will blow away the
6334          * old info in the log.
6335          */
6336         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6337
6338         key[0].objectid = objectid;
6339         key[0].type = BTRFS_INODE_ITEM_KEY;
6340         key[0].offset = 0;
6341
6342         sizes[0] = sizeof(struct btrfs_inode_item);
6343
6344         if (name) {
6345                 /*
6346                  * Start new inodes with an inode_ref. This is slightly more
6347                  * efficient for small numbers of hard links since they will
6348                  * be packed into one item. Extended refs will kick in if we
6349                  * add more hard links than can fit in the ref item.
6350                  */
6351                 key[1].objectid = objectid;
6352                 key[1].type = BTRFS_INODE_REF_KEY;
6353                 key[1].offset = ref_objectid;
6354
6355                 sizes[1] = name_len + sizeof(*ref);
6356         }
6357
6358         location = &BTRFS_I(inode)->location;
6359         location->objectid = objectid;
6360         location->offset = 0;
6361         location->type = BTRFS_INODE_ITEM_KEY;
6362
6363         ret = btrfs_insert_inode_locked(inode);
6364         if (ret < 0)
6365                 goto fail;
6366
6367         path->leave_spinning = 1;
6368         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6369         if (ret != 0)
6370                 goto fail_unlock;
6371
6372         inode_init_owner(inode, dir, mode);
6373         inode_set_bytes(inode, 0);
6374
6375         inode->i_mtime = current_time(inode);
6376         inode->i_atime = inode->i_mtime;
6377         inode->i_ctime = inode->i_mtime;
6378         BTRFS_I(inode)->i_otime = inode->i_mtime;
6379
6380         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6381                                   struct btrfs_inode_item);
6382         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6383                              sizeof(*inode_item));
6384         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6385
6386         if (name) {
6387                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6388                                      struct btrfs_inode_ref);
6389                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6390                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6391                 ptr = (unsigned long)(ref + 1);
6392                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6393         }
6394
6395         btrfs_mark_buffer_dirty(path->nodes[0]);
6396         btrfs_free_path(path);
6397
6398         btrfs_inherit_iflags(inode, dir);
6399
6400         if (S_ISREG(mode)) {
6401                 if (btrfs_test_opt(fs_info, NODATASUM))
6402                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6403                 if (btrfs_test_opt(fs_info, NODATACOW))
6404                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6405                                 BTRFS_INODE_NODATASUM;
6406         }
6407
6408         inode_tree_add(inode);
6409
6410         trace_btrfs_inode_new(inode);
6411         btrfs_set_inode_last_trans(trans, inode);
6412
6413         btrfs_update_root_times(trans, root);
6414
6415         ret = btrfs_inode_inherit_props(trans, inode, dir);
6416         if (ret)
6417                 btrfs_err(fs_info,
6418                           "error inheriting props for ino %llu (root %llu): %d",
6419                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6420
6421         return inode;
6422
6423 fail_unlock:
6424         unlock_new_inode(inode);
6425 fail:
6426         if (dir && name)
6427                 BTRFS_I(dir)->index_cnt--;
6428         btrfs_free_path(path);
6429         iput(inode);
6430         return ERR_PTR(ret);
6431 }
6432
6433 static inline u8 btrfs_inode_type(struct inode *inode)
6434 {
6435         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6436 }
6437
6438 /*
6439  * utility function to add 'inode' into 'parent_inode' with
6440  * a give name and a given sequence number.
6441  * if 'add_backref' is true, also insert a backref from the
6442  * inode to the parent directory.
6443  */
6444 int btrfs_add_link(struct btrfs_trans_handle *trans,
6445                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6446                    const char *name, int name_len, int add_backref, u64 index)
6447 {
6448         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6449         int ret = 0;
6450         struct btrfs_key key;
6451         struct btrfs_root *root = parent_inode->root;
6452         u64 ino = btrfs_ino(inode);
6453         u64 parent_ino = btrfs_ino(parent_inode);
6454
6455         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6456                 memcpy(&key, &inode->root->root_key, sizeof(key));
6457         } else {
6458                 key.objectid = ino;
6459                 key.type = BTRFS_INODE_ITEM_KEY;
6460                 key.offset = 0;
6461         }
6462
6463         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6464                 ret = btrfs_add_root_ref(trans, fs_info, key.objectid,
6465                                          root->root_key.objectid, parent_ino,
6466                                          index, name, name_len);
6467         } else if (add_backref) {
6468                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6469                                              parent_ino, index);
6470         }
6471
6472         /* Nothing to clean up yet */
6473         if (ret)
6474                 return ret;
6475
6476         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6477                                     parent_inode, &key,
6478                                     btrfs_inode_type(&inode->vfs_inode), index);
6479         if (ret == -EEXIST || ret == -EOVERFLOW)
6480                 goto fail_dir_item;
6481         else if (ret) {
6482                 btrfs_abort_transaction(trans, ret);
6483                 return ret;
6484         }
6485
6486         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6487                            name_len * 2);
6488         inode_inc_iversion(&parent_inode->vfs_inode);
6489         parent_inode->vfs_inode.i_mtime = parent_inode->vfs_inode.i_ctime =
6490                 current_time(&parent_inode->vfs_inode);
6491         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6492         if (ret)
6493                 btrfs_abort_transaction(trans, ret);
6494         return ret;
6495
6496 fail_dir_item:
6497         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6498                 u64 local_index;
6499                 int err;
6500                 err = btrfs_del_root_ref(trans, fs_info, key.objectid,
6501                                          root->root_key.objectid, parent_ino,
6502                                          &local_index, name, name_len);
6503
6504         } else if (add_backref) {
6505                 u64 local_index;
6506                 int err;
6507
6508                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6509                                           ino, parent_ino, &local_index);
6510         }
6511         return ret;
6512 }
6513
6514 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6515                             struct btrfs_inode *dir, struct dentry *dentry,
6516                             struct btrfs_inode *inode, int backref, u64 index)
6517 {
6518         int err = btrfs_add_link(trans, dir, inode,
6519                                  dentry->d_name.name, dentry->d_name.len,
6520                                  backref, index);
6521         if (err > 0)
6522                 err = -EEXIST;
6523         return err;
6524 }
6525
6526 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6527                         umode_t mode, dev_t rdev)
6528 {
6529         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6530         struct btrfs_trans_handle *trans;
6531         struct btrfs_root *root = BTRFS_I(dir)->root;
6532         struct inode *inode = NULL;
6533         int err;
6534         int drop_inode = 0;
6535         u64 objectid;
6536         u64 index = 0;
6537
6538         /*
6539          * 2 for inode item and ref
6540          * 2 for dir items
6541          * 1 for xattr if selinux is on
6542          */
6543         trans = btrfs_start_transaction(root, 5);
6544         if (IS_ERR(trans))
6545                 return PTR_ERR(trans);
6546
6547         err = btrfs_find_free_ino(root, &objectid);
6548         if (err)
6549                 goto out_unlock;
6550
6551         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6552                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6553                         mode, &index);
6554         if (IS_ERR(inode)) {
6555                 err = PTR_ERR(inode);
6556                 goto out_unlock;
6557         }
6558
6559         /*
6560         * If the active LSM wants to access the inode during
6561         * d_instantiate it needs these. Smack checks to see
6562         * if the filesystem supports xattrs by looking at the
6563         * ops vector.
6564         */
6565         inode->i_op = &btrfs_special_inode_operations;
6566         init_special_inode(inode, inode->i_mode, rdev);
6567
6568         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6569         if (err)
6570                 goto out_unlock_inode;
6571
6572         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6573                         0, index);
6574         if (err) {
6575                 goto out_unlock_inode;
6576         } else {
6577                 btrfs_update_inode(trans, root, inode);
6578                 unlock_new_inode(inode);
6579                 d_instantiate(dentry, inode);
6580         }
6581
6582 out_unlock:
6583         btrfs_end_transaction(trans);
6584         btrfs_btree_balance_dirty(fs_info);
6585         if (drop_inode) {
6586                 inode_dec_link_count(inode);
6587                 iput(inode);
6588         }
6589         return err;
6590
6591 out_unlock_inode:
6592         drop_inode = 1;
6593         unlock_new_inode(inode);
6594         goto out_unlock;
6595
6596 }
6597
6598 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6599                         umode_t mode, bool excl)
6600 {
6601         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6602         struct btrfs_trans_handle *trans;
6603         struct btrfs_root *root = BTRFS_I(dir)->root;
6604         struct inode *inode = NULL;
6605         int drop_inode_on_err = 0;
6606         int err;
6607         u64 objectid;
6608         u64 index = 0;
6609
6610         /*
6611          * 2 for inode item and ref
6612          * 2 for dir items
6613          * 1 for xattr if selinux is on
6614          */
6615         trans = btrfs_start_transaction(root, 5);
6616         if (IS_ERR(trans))
6617                 return PTR_ERR(trans);
6618
6619         err = btrfs_find_free_ino(root, &objectid);
6620         if (err)
6621                 goto out_unlock;
6622
6623         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6624                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6625                         mode, &index);
6626         if (IS_ERR(inode)) {
6627                 err = PTR_ERR(inode);
6628                 goto out_unlock;
6629         }
6630         drop_inode_on_err = 1;
6631         /*
6632         * If the active LSM wants to access the inode during
6633         * d_instantiate it needs these. Smack checks to see
6634         * if the filesystem supports xattrs by looking at the
6635         * ops vector.
6636         */
6637         inode->i_fop = &btrfs_file_operations;
6638         inode->i_op = &btrfs_file_inode_operations;
6639         inode->i_mapping->a_ops = &btrfs_aops;
6640
6641         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6642         if (err)
6643                 goto out_unlock_inode;
6644
6645         err = btrfs_update_inode(trans, root, inode);
6646         if (err)
6647                 goto out_unlock_inode;
6648
6649         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6650                         0, index);
6651         if (err)
6652                 goto out_unlock_inode;
6653
6654         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6655         unlock_new_inode(inode);
6656         d_instantiate(dentry, inode);
6657
6658 out_unlock:
6659         btrfs_end_transaction(trans);
6660         if (err && drop_inode_on_err) {
6661                 inode_dec_link_count(inode);
6662                 iput(inode);
6663         }
6664         btrfs_btree_balance_dirty(fs_info);
6665         return err;
6666
6667 out_unlock_inode:
6668         unlock_new_inode(inode);
6669         goto out_unlock;
6670
6671 }
6672
6673 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6674                       struct dentry *dentry)
6675 {
6676         struct btrfs_trans_handle *trans = NULL;
6677         struct btrfs_root *root = BTRFS_I(dir)->root;
6678         struct inode *inode = d_inode(old_dentry);
6679         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6680         u64 index;
6681         int err;
6682         int drop_inode = 0;
6683
6684         /* do not allow sys_link's with other subvols of the same device */
6685         if (root->objectid != BTRFS_I(inode)->root->objectid)
6686                 return -EXDEV;
6687
6688         if (inode->i_nlink >= BTRFS_LINK_MAX)
6689                 return -EMLINK;
6690
6691         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6692         if (err)
6693                 goto fail;
6694
6695         /*
6696          * 2 items for inode and inode ref
6697          * 2 items for dir items
6698          * 1 item for parent inode
6699          */
6700         trans = btrfs_start_transaction(root, 5);
6701         if (IS_ERR(trans)) {
6702                 err = PTR_ERR(trans);
6703                 trans = NULL;
6704                 goto fail;
6705         }
6706
6707         /* There are several dir indexes for this inode, clear the cache. */
6708         BTRFS_I(inode)->dir_index = 0ULL;
6709         inc_nlink(inode);
6710         inode_inc_iversion(inode);
6711         inode->i_ctime = current_time(inode);
6712         ihold(inode);
6713         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6714
6715         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6716                         1, index);
6717
6718         if (err) {
6719                 drop_inode = 1;
6720         } else {
6721                 struct dentry *parent = dentry->d_parent;
6722                 err = btrfs_update_inode(trans, root, inode);
6723                 if (err)
6724                         goto fail;
6725                 if (inode->i_nlink == 1) {
6726                         /*
6727                          * If new hard link count is 1, it's a file created
6728                          * with open(2) O_TMPFILE flag.
6729                          */
6730                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6731                         if (err)
6732                                 goto fail;
6733                 }
6734                 d_instantiate(dentry, inode);
6735                 btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent);
6736         }
6737
6738 fail:
6739         if (trans)
6740                 btrfs_end_transaction(trans);
6741         if (drop_inode) {
6742                 inode_dec_link_count(inode);
6743                 iput(inode);
6744         }
6745         btrfs_btree_balance_dirty(fs_info);
6746         return err;
6747 }
6748
6749 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6750 {
6751         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6752         struct inode *inode = NULL;
6753         struct btrfs_trans_handle *trans;
6754         struct btrfs_root *root = BTRFS_I(dir)->root;
6755         int err = 0;
6756         int drop_on_err = 0;
6757         u64 objectid = 0;
6758         u64 index = 0;
6759
6760         /*
6761          * 2 items for inode and ref
6762          * 2 items for dir items
6763          * 1 for xattr if selinux is on
6764          */
6765         trans = btrfs_start_transaction(root, 5);
6766         if (IS_ERR(trans))
6767                 return PTR_ERR(trans);
6768
6769         err = btrfs_find_free_ino(root, &objectid);
6770         if (err)
6771                 goto out_fail;
6772
6773         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6774                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6775                         S_IFDIR | mode, &index);
6776         if (IS_ERR(inode)) {
6777                 err = PTR_ERR(inode);
6778                 goto out_fail;
6779         }
6780
6781         drop_on_err = 1;
6782         /* these must be set before we unlock the inode */
6783         inode->i_op = &btrfs_dir_inode_operations;
6784         inode->i_fop = &btrfs_dir_file_operations;
6785
6786         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6787         if (err)
6788                 goto out_fail_inode;
6789
6790         btrfs_i_size_write(BTRFS_I(inode), 0);
6791         err = btrfs_update_inode(trans, root, inode);
6792         if (err)
6793                 goto out_fail_inode;
6794
6795         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6796                         dentry->d_name.name,
6797                         dentry->d_name.len, 0, index);
6798         if (err)
6799                 goto out_fail_inode;
6800
6801         d_instantiate(dentry, inode);
6802         /*
6803          * mkdir is special.  We're unlocking after we call d_instantiate
6804          * to avoid a race with nfsd calling d_instantiate.
6805          */
6806         unlock_new_inode(inode);
6807         drop_on_err = 0;
6808
6809 out_fail:
6810         btrfs_end_transaction(trans);
6811         if (drop_on_err) {
6812                 inode_dec_link_count(inode);
6813                 iput(inode);
6814         }
6815         btrfs_btree_balance_dirty(fs_info);
6816         return err;
6817
6818 out_fail_inode:
6819         unlock_new_inode(inode);
6820         goto out_fail;
6821 }
6822
6823 static noinline int uncompress_inline(struct btrfs_path *path,
6824                                       struct page *page,
6825                                       size_t pg_offset, u64 extent_offset,
6826                                       struct btrfs_file_extent_item *item)
6827 {
6828         int ret;
6829         struct extent_buffer *leaf = path->nodes[0];
6830         char *tmp;
6831         size_t max_size;
6832         unsigned long inline_size;
6833         unsigned long ptr;
6834         int compress_type;
6835
6836         WARN_ON(pg_offset != 0);
6837         compress_type = btrfs_file_extent_compression(leaf, item);
6838         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6839         inline_size = btrfs_file_extent_inline_item_len(leaf,
6840                                         btrfs_item_nr(path->slots[0]));
6841         tmp = kmalloc(inline_size, GFP_NOFS);
6842         if (!tmp)
6843                 return -ENOMEM;
6844         ptr = btrfs_file_extent_inline_start(item);
6845
6846         read_extent_buffer(leaf, tmp, ptr, inline_size);
6847
6848         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6849         ret = btrfs_decompress(compress_type, tmp, page,
6850                                extent_offset, inline_size, max_size);
6851
6852         /*
6853          * decompression code contains a memset to fill in any space between the end
6854          * of the uncompressed data and the end of max_size in case the decompressed
6855          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6856          * the end of an inline extent and the beginning of the next block, so we
6857          * cover that region here.
6858          */
6859
6860         if (max_size + pg_offset < PAGE_SIZE) {
6861                 char *map = kmap(page);
6862                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6863                 kunmap(page);
6864         }
6865         kfree(tmp);
6866         return ret;
6867 }
6868
6869 /*
6870  * a bit scary, this does extent mapping from logical file offset to the disk.
6871  * the ugly parts come from merging extents from the disk with the in-ram
6872  * representation.  This gets more complex because of the data=ordered code,
6873  * where the in-ram extents might be locked pending data=ordered completion.
6874  *
6875  * This also copies inline extents directly into the page.
6876  */
6877 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6878                 struct page *page,
6879             size_t pg_offset, u64 start, u64 len,
6880                 int create)
6881 {
6882         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6883         int ret;
6884         int err = 0;
6885         u64 extent_start = 0;
6886         u64 extent_end = 0;
6887         u64 objectid = btrfs_ino(inode);
6888         u32 found_type;
6889         struct btrfs_path *path = NULL;
6890         struct btrfs_root *root = inode->root;
6891         struct btrfs_file_extent_item *item;
6892         struct extent_buffer *leaf;
6893         struct btrfs_key found_key;
6894         struct extent_map *em = NULL;
6895         struct extent_map_tree *em_tree = &inode->extent_tree;
6896         struct extent_io_tree *io_tree = &inode->io_tree;
6897         const bool new_inline = !page || create;
6898
6899         read_lock(&em_tree->lock);
6900         em = lookup_extent_mapping(em_tree, start, len);
6901         if (em)
6902                 em->bdev = fs_info->fs_devices->latest_bdev;
6903         read_unlock(&em_tree->lock);
6904
6905         if (em) {
6906                 if (em->start > start || em->start + em->len <= start)
6907                         free_extent_map(em);
6908                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6909                         free_extent_map(em);
6910                 else
6911                         goto out;
6912         }
6913         em = alloc_extent_map();
6914         if (!em) {
6915                 err = -ENOMEM;
6916                 goto out;
6917         }
6918         em->bdev = fs_info->fs_devices->latest_bdev;
6919         em->start = EXTENT_MAP_HOLE;
6920         em->orig_start = EXTENT_MAP_HOLE;
6921         em->len = (u64)-1;
6922         em->block_len = (u64)-1;
6923
6924         if (!path) {
6925                 path = btrfs_alloc_path();
6926                 if (!path) {
6927                         err = -ENOMEM;
6928                         goto out;
6929                 }
6930                 /*
6931                  * Chances are we'll be called again, so go ahead and do
6932                  * readahead
6933                  */
6934                 path->reada = READA_FORWARD;
6935         }
6936
6937         ret = btrfs_lookup_file_extent(NULL, root, path, objectid, start, 0);
6938         if (ret < 0) {
6939                 err = ret;
6940                 goto out;
6941         }
6942
6943         if (ret != 0) {
6944                 if (path->slots[0] == 0)
6945                         goto not_found;
6946                 path->slots[0]--;
6947         }
6948
6949         leaf = path->nodes[0];
6950         item = btrfs_item_ptr(leaf, path->slots[0],
6951                               struct btrfs_file_extent_item);
6952         /* are we inside the extent that was found? */
6953         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6954         found_type = found_key.type;
6955         if (found_key.objectid != objectid ||
6956             found_type != BTRFS_EXTENT_DATA_KEY) {
6957                 /*
6958                  * If we backup past the first extent we want to move forward
6959                  * and see if there is an extent in front of us, otherwise we'll
6960                  * say there is a hole for our whole search range which can
6961                  * cause problems.
6962                  */
6963                 extent_end = start;
6964                 goto next;
6965         }
6966
6967         found_type = btrfs_file_extent_type(leaf, item);
6968         extent_start = found_key.offset;
6969         if (found_type == BTRFS_FILE_EXTENT_REG ||
6970             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6971                 extent_end = extent_start +
6972                        btrfs_file_extent_num_bytes(leaf, item);
6973
6974                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
6975                                                        extent_start);
6976         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6977                 size_t size;
6978                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6979                 extent_end = ALIGN(extent_start + size,
6980                                    fs_info->sectorsize);
6981
6982                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
6983                                                       path->slots[0],
6984                                                       extent_start);
6985         }
6986 next:
6987         if (start >= extent_end) {
6988                 path->slots[0]++;
6989                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6990                         ret = btrfs_next_leaf(root, path);
6991                         if (ret < 0) {
6992                                 err = ret;
6993                                 goto out;
6994                         }
6995                         if (ret > 0)
6996                                 goto not_found;
6997                         leaf = path->nodes[0];
6998                 }
6999                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7000                 if (found_key.objectid != objectid ||
7001                     found_key.type != BTRFS_EXTENT_DATA_KEY)
7002                         goto not_found;
7003                 if (start + len <= found_key.offset)
7004                         goto not_found;
7005                 if (start > found_key.offset)
7006                         goto next;
7007                 em->start = start;
7008                 em->orig_start = start;
7009                 em->len = found_key.offset - start;
7010                 goto not_found_em;
7011         }
7012
7013         btrfs_extent_item_to_extent_map(inode, path, item,
7014                         new_inline, em);
7015
7016         if (found_type == BTRFS_FILE_EXTENT_REG ||
7017             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7018                 goto insert;
7019         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
7020                 unsigned long ptr;
7021                 char *map;
7022                 size_t size;
7023                 size_t extent_offset;
7024                 size_t copy_size;
7025
7026                 if (new_inline)
7027                         goto out;
7028
7029                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
7030                 extent_offset = page_offset(page) + pg_offset - extent_start;
7031                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
7032                                   size - extent_offset);
7033                 em->start = extent_start + extent_offset;
7034                 em->len = ALIGN(copy_size, fs_info->sectorsize);
7035                 em->orig_block_len = em->len;
7036                 em->orig_start = em->start;
7037                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
7038                 if (!PageUptodate(page)) {
7039                         if (btrfs_file_extent_compression(leaf, item) !=
7040                             BTRFS_COMPRESS_NONE) {
7041                                 ret = uncompress_inline(path, page, pg_offset,
7042                                                         extent_offset, item);
7043                                 if (ret) {
7044                                         err = ret;
7045                                         goto out;
7046                                 }
7047                         } else {
7048                                 map = kmap(page);
7049                                 read_extent_buffer(leaf, map + pg_offset, ptr,
7050                                                    copy_size);
7051                                 if (pg_offset + copy_size < PAGE_SIZE) {
7052                                         memset(map + pg_offset + copy_size, 0,
7053                                                PAGE_SIZE - pg_offset -
7054                                                copy_size);
7055                                 }
7056                                 kunmap(page);
7057                         }
7058                         flush_dcache_page(page);
7059                 }
7060                 set_extent_uptodate(io_tree, em->start,
7061                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7062                 goto insert;
7063         }
7064 not_found:
7065         em->start = start;
7066         em->orig_start = start;
7067         em->len = len;
7068 not_found_em:
7069         em->block_start = EXTENT_MAP_HOLE;
7070 insert:
7071         btrfs_release_path(path);
7072         if (em->start > start || extent_map_end(em) <= start) {
7073                 btrfs_err(fs_info,
7074                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
7075                           em->start, em->len, start, len);
7076                 err = -EIO;
7077                 goto out;
7078         }
7079
7080         err = 0;
7081         write_lock(&em_tree->lock);
7082         err = btrfs_add_extent_mapping(em_tree, &em, start, len);
7083         write_unlock(&em_tree->lock);
7084 out:
7085
7086         trace_btrfs_get_extent(root, inode, em);
7087
7088         btrfs_free_path(path);
7089         if (err) {
7090                 free_extent_map(em);
7091                 return ERR_PTR(err);
7092         }
7093         BUG_ON(!em); /* Error is always set */
7094         return em;
7095 }
7096
7097 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7098                 struct page *page,
7099                 size_t pg_offset, u64 start, u64 len,
7100                 int create)
7101 {
7102         struct extent_map *em;
7103         struct extent_map *hole_em = NULL;
7104         u64 range_start = start;
7105         u64 end;
7106         u64 found;
7107         u64 found_end;
7108         int err = 0;
7109
7110         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7111         if (IS_ERR(em))
7112                 return em;
7113         /*
7114          * If our em maps to:
7115          * - a hole or
7116          * - a pre-alloc extent,
7117          * there might actually be delalloc bytes behind it.
7118          */
7119         if (em->block_start != EXTENT_MAP_HOLE &&
7120             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7121                 return em;
7122         else
7123                 hole_em = em;
7124
7125         /* check to see if we've wrapped (len == -1 or similar) */
7126         end = start + len;
7127         if (end < start)
7128                 end = (u64)-1;
7129         else
7130                 end -= 1;
7131
7132         em = NULL;
7133
7134         /* ok, we didn't find anything, lets look for delalloc */
7135         found = count_range_bits(&inode->io_tree, &range_start,
7136                                  end, len, EXTENT_DELALLOC, 1);
7137         found_end = range_start + found;
7138         if (found_end < range_start)
7139                 found_end = (u64)-1;
7140
7141         /*
7142          * we didn't find anything useful, return
7143          * the original results from get_extent()
7144          */
7145         if (range_start > end || found_end <= start) {
7146                 em = hole_em;
7147                 hole_em = NULL;
7148                 goto out;
7149         }
7150
7151         /* adjust the range_start to make sure it doesn't
7152          * go backwards from the start they passed in
7153          */
7154         range_start = max(start, range_start);
7155         found = found_end - range_start;
7156
7157         if (found > 0) {
7158                 u64 hole_start = start;
7159                 u64 hole_len = len;
7160
7161                 em = alloc_extent_map();
7162                 if (!em) {
7163                         err = -ENOMEM;
7164                         goto out;
7165                 }
7166                 /*
7167                  * when btrfs_get_extent can't find anything it
7168                  * returns one huge hole
7169                  *
7170                  * make sure what it found really fits our range, and
7171                  * adjust to make sure it is based on the start from
7172                  * the caller
7173                  */
7174                 if (hole_em) {
7175                         u64 calc_end = extent_map_end(hole_em);
7176
7177                         if (calc_end <= start || (hole_em->start > end)) {
7178                                 free_extent_map(hole_em);
7179                                 hole_em = NULL;
7180                         } else {
7181                                 hole_start = max(hole_em->start, start);
7182                                 hole_len = calc_end - hole_start;
7183                         }
7184                 }
7185                 em->bdev = NULL;
7186                 if (hole_em && range_start > hole_start) {
7187                         /* our hole starts before our delalloc, so we
7188                          * have to return just the parts of the hole
7189                          * that go until  the delalloc starts
7190                          */
7191                         em->len = min(hole_len,
7192                                       range_start - hole_start);
7193                         em->start = hole_start;
7194                         em->orig_start = hole_start;
7195                         /*
7196                          * don't adjust block start at all,
7197                          * it is fixed at EXTENT_MAP_HOLE
7198                          */
7199                         em->block_start = hole_em->block_start;
7200                         em->block_len = hole_len;
7201                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7202                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7203                 } else {
7204                         em->start = range_start;
7205                         em->len = found;
7206                         em->orig_start = range_start;
7207                         em->block_start = EXTENT_MAP_DELALLOC;
7208                         em->block_len = found;
7209                 }
7210         } else {
7211                 return hole_em;
7212         }
7213 out:
7214
7215         free_extent_map(hole_em);
7216         if (err) {
7217                 free_extent_map(em);
7218                 return ERR_PTR(err);
7219         }
7220         return em;
7221 }
7222
7223 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7224                                                   const u64 start,
7225                                                   const u64 len,
7226                                                   const u64 orig_start,
7227                                                   const u64 block_start,
7228                                                   const u64 block_len,
7229                                                   const u64 orig_block_len,
7230                                                   const u64 ram_bytes,
7231                                                   const int type)
7232 {
7233         struct extent_map *em = NULL;
7234         int ret;
7235
7236         if (type != BTRFS_ORDERED_NOCOW) {
7237                 em = create_io_em(inode, start, len, orig_start,
7238                                   block_start, block_len, orig_block_len,
7239                                   ram_bytes,
7240                                   BTRFS_COMPRESS_NONE, /* compress_type */
7241                                   type);
7242                 if (IS_ERR(em))
7243                         goto out;
7244         }
7245         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7246                                            len, block_len, type);
7247         if (ret) {
7248                 if (em) {
7249                         free_extent_map(em);
7250                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7251                                                 start + len - 1, 0);
7252                 }
7253                 em = ERR_PTR(ret);
7254         }
7255  out:
7256
7257         return em;
7258 }
7259
7260 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7261                                                   u64 start, u64 len)
7262 {
7263         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7264         struct btrfs_root *root = BTRFS_I(inode)->root;
7265         struct extent_map *em;
7266         struct btrfs_key ins;
7267         u64 alloc_hint;
7268         int ret;
7269
7270         alloc_hint = get_extent_allocation_hint(inode, start, len);
7271         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7272                                    0, alloc_hint, &ins, 1, 1);
7273         if (ret)
7274                 return ERR_PTR(ret);
7275
7276         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7277                                      ins.objectid, ins.offset, ins.offset,
7278                                      ins.offset, BTRFS_ORDERED_REGULAR);
7279         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7280         if (IS_ERR(em))
7281                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7282                                            ins.offset, 1);
7283
7284         return em;
7285 }
7286
7287 /*
7288  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7289  * block must be cow'd
7290  */
7291 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7292                               u64 *orig_start, u64 *orig_block_len,
7293                               u64 *ram_bytes)
7294 {
7295         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7296         struct btrfs_path *path;
7297         int ret;
7298         struct extent_buffer *leaf;
7299         struct btrfs_root *root = BTRFS_I(inode)->root;
7300         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7301         struct btrfs_file_extent_item *fi;
7302         struct btrfs_key key;
7303         u64 disk_bytenr;
7304         u64 backref_offset;
7305         u64 extent_end;
7306         u64 num_bytes;
7307         int slot;
7308         int found_type;
7309         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7310
7311         path = btrfs_alloc_path();
7312         if (!path)
7313                 return -ENOMEM;
7314
7315         ret = btrfs_lookup_file_extent(NULL, root, path,
7316                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7317         if (ret < 0)
7318                 goto out;
7319
7320         slot = path->slots[0];
7321         if (ret == 1) {
7322                 if (slot == 0) {
7323                         /* can't find the item, must cow */
7324                         ret = 0;
7325                         goto out;
7326                 }
7327                 slot--;
7328         }
7329         ret = 0;
7330         leaf = path->nodes[0];
7331         btrfs_item_key_to_cpu(leaf, &key, slot);
7332         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7333             key.type != BTRFS_EXTENT_DATA_KEY) {
7334                 /* not our file or wrong item type, must cow */
7335                 goto out;
7336         }
7337
7338         if (key.offset > offset) {
7339                 /* Wrong offset, must cow */
7340                 goto out;
7341         }
7342
7343         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7344         found_type = btrfs_file_extent_type(leaf, fi);
7345         if (found_type != BTRFS_FILE_EXTENT_REG &&
7346             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7347                 /* not a regular extent, must cow */
7348                 goto out;
7349         }
7350
7351         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7352                 goto out;
7353
7354         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7355         if (extent_end <= offset)
7356                 goto out;
7357
7358         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7359         if (disk_bytenr == 0)
7360                 goto out;
7361
7362         if (btrfs_file_extent_compression(leaf, fi) ||
7363             btrfs_file_extent_encryption(leaf, fi) ||
7364             btrfs_file_extent_other_encoding(leaf, fi))
7365                 goto out;
7366
7367         backref_offset = btrfs_file_extent_offset(leaf, fi);
7368
7369         if (orig_start) {
7370                 *orig_start = key.offset - backref_offset;
7371                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7372                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7373         }
7374
7375         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7376                 goto out;
7377
7378         num_bytes = min(offset + *len, extent_end) - offset;
7379         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7380                 u64 range_end;
7381
7382                 range_end = round_up(offset + num_bytes,
7383                                      root->fs_info->sectorsize) - 1;
7384                 ret = test_range_bit(io_tree, offset, range_end,
7385                                      EXTENT_DELALLOC, 0, NULL);
7386                 if (ret) {
7387                         ret = -EAGAIN;
7388                         goto out;
7389                 }
7390         }
7391
7392         btrfs_release_path(path);
7393
7394         /*
7395          * look for other files referencing this extent, if we
7396          * find any we must cow
7397          */
7398
7399         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7400                                     key.offset - backref_offset, disk_bytenr);
7401         if (ret) {
7402                 ret = 0;
7403                 goto out;
7404         }
7405
7406         /*
7407          * adjust disk_bytenr and num_bytes to cover just the bytes
7408          * in this extent we are about to write.  If there
7409          * are any csums in that range we have to cow in order
7410          * to keep the csums correct
7411          */
7412         disk_bytenr += backref_offset;
7413         disk_bytenr += offset - key.offset;
7414         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7415                 goto out;
7416         /*
7417          * all of the above have passed, it is safe to overwrite this extent
7418          * without cow
7419          */
7420         *len = num_bytes;
7421         ret = 1;
7422 out:
7423         btrfs_free_path(path);
7424         return ret;
7425 }
7426
7427 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7428 {
7429         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7430         bool found = false;
7431         void **pagep = NULL;
7432         struct page *page = NULL;
7433         unsigned long start_idx;
7434         unsigned long end_idx;
7435
7436         start_idx = start >> PAGE_SHIFT;
7437
7438         /*
7439          * end is the last byte in the last page.  end == start is legal
7440          */
7441         end_idx = end >> PAGE_SHIFT;
7442
7443         rcu_read_lock();
7444
7445         /* Most of the code in this while loop is lifted from
7446          * find_get_page.  It's been modified to begin searching from a
7447          * page and return just the first page found in that range.  If the
7448          * found idx is less than or equal to the end idx then we know that
7449          * a page exists.  If no pages are found or if those pages are
7450          * outside of the range then we're fine (yay!) */
7451         while (page == NULL &&
7452                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7453                 page = radix_tree_deref_slot(pagep);
7454                 if (unlikely(!page))
7455                         break;
7456
7457                 if (radix_tree_exception(page)) {
7458                         if (radix_tree_deref_retry(page)) {
7459                                 page = NULL;
7460                                 continue;
7461                         }
7462                         /*
7463                          * Otherwise, shmem/tmpfs must be storing a swap entry
7464                          * here as an exceptional entry: so return it without
7465                          * attempting to raise page count.
7466                          */
7467                         page = NULL;
7468                         break; /* TODO: Is this relevant for this use case? */
7469                 }
7470
7471                 if (!page_cache_get_speculative(page)) {
7472                         page = NULL;
7473                         continue;
7474                 }
7475
7476                 /*
7477                  * Has the page moved?
7478                  * This is part of the lockless pagecache protocol. See
7479                  * include/linux/pagemap.h for details.
7480                  */
7481                 if (unlikely(page != *pagep)) {
7482                         put_page(page);
7483                         page = NULL;
7484                 }
7485         }
7486
7487         if (page) {
7488                 if (page->index <= end_idx)
7489                         found = true;
7490                 put_page(page);
7491         }
7492
7493         rcu_read_unlock();
7494         return found;
7495 }
7496
7497 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7498                               struct extent_state **cached_state, int writing)
7499 {
7500         struct btrfs_ordered_extent *ordered;
7501         int ret = 0;
7502
7503         while (1) {
7504                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7505                                  cached_state);
7506                 /*
7507                  * We're concerned with the entire range that we're going to be
7508                  * doing DIO to, so we need to make sure there's no ordered
7509                  * extents in this range.
7510                  */
7511                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7512                                                      lockend - lockstart + 1);
7513
7514                 /*
7515                  * We need to make sure there are no buffered pages in this
7516                  * range either, we could have raced between the invalidate in
7517                  * generic_file_direct_write and locking the extent.  The
7518                  * invalidate needs to happen so that reads after a write do not
7519                  * get stale data.
7520                  */
7521                 if (!ordered &&
7522                     (!writing ||
7523                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7524                         break;
7525
7526                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7527                                      cached_state);
7528
7529                 if (ordered) {
7530                         /*
7531                          * If we are doing a DIO read and the ordered extent we
7532                          * found is for a buffered write, we can not wait for it
7533                          * to complete and retry, because if we do so we can
7534                          * deadlock with concurrent buffered writes on page
7535                          * locks. This happens only if our DIO read covers more
7536                          * than one extent map, if at this point has already
7537                          * created an ordered extent for a previous extent map
7538                          * and locked its range in the inode's io tree, and a
7539                          * concurrent write against that previous extent map's
7540                          * range and this range started (we unlock the ranges
7541                          * in the io tree only when the bios complete and
7542                          * buffered writes always lock pages before attempting
7543                          * to lock range in the io tree).
7544                          */
7545                         if (writing ||
7546                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7547                                 btrfs_start_ordered_extent(inode, ordered, 1);
7548                         else
7549                                 ret = -ENOTBLK;
7550                         btrfs_put_ordered_extent(ordered);
7551                 } else {
7552                         /*
7553                          * We could trigger writeback for this range (and wait
7554                          * for it to complete) and then invalidate the pages for
7555                          * this range (through invalidate_inode_pages2_range()),
7556                          * but that can lead us to a deadlock with a concurrent
7557                          * call to readpages() (a buffered read or a defrag call
7558                          * triggered a readahead) on a page lock due to an
7559                          * ordered dio extent we created before but did not have
7560                          * yet a corresponding bio submitted (whence it can not
7561                          * complete), which makes readpages() wait for that
7562                          * ordered extent to complete while holding a lock on
7563                          * that page.
7564                          */
7565                         ret = -ENOTBLK;
7566                 }
7567
7568                 if (ret)
7569                         break;
7570
7571                 cond_resched();
7572         }
7573
7574         return ret;
7575 }
7576
7577 /* The callers of this must take lock_extent() */
7578 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7579                                        u64 orig_start, u64 block_start,
7580                                        u64 block_len, u64 orig_block_len,
7581                                        u64 ram_bytes, int compress_type,
7582                                        int type)
7583 {
7584         struct extent_map_tree *em_tree;
7585         struct extent_map *em;
7586         struct btrfs_root *root = BTRFS_I(inode)->root;
7587         int ret;
7588
7589         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7590                type == BTRFS_ORDERED_COMPRESSED ||
7591                type == BTRFS_ORDERED_NOCOW ||
7592                type == BTRFS_ORDERED_REGULAR);
7593
7594         em_tree = &BTRFS_I(inode)->extent_tree;
7595         em = alloc_extent_map();
7596         if (!em)
7597                 return ERR_PTR(-ENOMEM);
7598
7599         em->start = start;
7600         em->orig_start = orig_start;
7601         em->len = len;
7602         em->block_len = block_len;
7603         em->block_start = block_start;
7604         em->bdev = root->fs_info->fs_devices->latest_bdev;
7605         em->orig_block_len = orig_block_len;
7606         em->ram_bytes = ram_bytes;
7607         em->generation = -1;
7608         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7609         if (type == BTRFS_ORDERED_PREALLOC) {
7610                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7611         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7612                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7613                 em->compress_type = compress_type;
7614         }
7615
7616         do {
7617                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7618                                 em->start + em->len - 1, 0);
7619                 write_lock(&em_tree->lock);
7620                 ret = add_extent_mapping(em_tree, em, 1);
7621                 write_unlock(&em_tree->lock);
7622                 /*
7623                  * The caller has taken lock_extent(), who could race with us
7624                  * to add em?
7625                  */
7626         } while (ret == -EEXIST);
7627
7628         if (ret) {
7629                 free_extent_map(em);
7630                 return ERR_PTR(ret);
7631         }
7632
7633         /* em got 2 refs now, callers needs to do free_extent_map once. */
7634         return em;
7635 }
7636
7637 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7638                                    struct buffer_head *bh_result, int create)
7639 {
7640         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7641         struct extent_map *em;
7642         struct extent_state *cached_state = NULL;
7643         struct btrfs_dio_data *dio_data = NULL;
7644         u64 start = iblock << inode->i_blkbits;
7645         u64 lockstart, lockend;
7646         u64 len = bh_result->b_size;
7647         int unlock_bits = EXTENT_LOCKED;
7648         int ret = 0;
7649
7650         if (create)
7651                 unlock_bits |= EXTENT_DIRTY;
7652         else
7653                 len = min_t(u64, len, fs_info->sectorsize);
7654
7655         lockstart = start;
7656         lockend = start + len - 1;
7657
7658         if (current->journal_info) {
7659                 /*
7660                  * Need to pull our outstanding extents and set journal_info to NULL so
7661                  * that anything that needs to check if there's a transaction doesn't get
7662                  * confused.
7663                  */
7664                 dio_data = current->journal_info;
7665                 current->journal_info = NULL;
7666         }
7667
7668         /*
7669          * If this errors out it's because we couldn't invalidate pagecache for
7670          * this range and we need to fallback to buffered.
7671          */
7672         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7673                                create)) {
7674                 ret = -ENOTBLK;
7675                 goto err;
7676         }
7677
7678         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7679         if (IS_ERR(em)) {
7680                 ret = PTR_ERR(em);
7681                 goto unlock_err;
7682         }
7683
7684         /*
7685          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7686          * io.  INLINE is special, and we could probably kludge it in here, but
7687          * it's still buffered so for safety lets just fall back to the generic
7688          * buffered path.
7689          *
7690          * For COMPRESSED we _have_ to read the entire extent in so we can
7691          * decompress it, so there will be buffering required no matter what we
7692          * do, so go ahead and fallback to buffered.
7693          *
7694          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7695          * to buffered IO.  Don't blame me, this is the price we pay for using
7696          * the generic code.
7697          */
7698         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7699             em->block_start == EXTENT_MAP_INLINE) {
7700                 free_extent_map(em);
7701                 ret = -ENOTBLK;
7702                 goto unlock_err;
7703         }
7704
7705         /* Just a good old fashioned hole, return */
7706         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7707                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7708                 free_extent_map(em);
7709                 goto unlock_err;
7710         }
7711
7712         /*
7713          * We don't allocate a new extent in the following cases
7714          *
7715          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7716          * existing extent.
7717          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7718          * just use the extent.
7719          *
7720          */
7721         if (!create) {
7722                 len = min(len, em->len - (start - em->start));
7723                 lockstart = start + len;
7724                 goto unlock;
7725         }
7726
7727         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7728             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7729              em->block_start != EXTENT_MAP_HOLE)) {
7730                 int type;
7731                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7732
7733                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7734                         type = BTRFS_ORDERED_PREALLOC;
7735                 else
7736                         type = BTRFS_ORDERED_NOCOW;
7737                 len = min(len, em->len - (start - em->start));
7738                 block_start = em->block_start + (start - em->start);
7739
7740                 if (can_nocow_extent(inode, start, &len, &orig_start,
7741                                      &orig_block_len, &ram_bytes) == 1 &&
7742                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7743                         struct extent_map *em2;
7744
7745                         em2 = btrfs_create_dio_extent(inode, start, len,
7746                                                       orig_start, block_start,
7747                                                       len, orig_block_len,
7748                                                       ram_bytes, type);
7749                         btrfs_dec_nocow_writers(fs_info, block_start);
7750                         if (type == BTRFS_ORDERED_PREALLOC) {
7751                                 free_extent_map(em);
7752                                 em = em2;
7753                         }
7754                         if (em2 && IS_ERR(em2)) {
7755                                 ret = PTR_ERR(em2);
7756                                 goto unlock_err;
7757                         }
7758                         /*
7759                          * For inode marked NODATACOW or extent marked PREALLOC,
7760                          * use the existing or preallocated extent, so does not
7761                          * need to adjust btrfs_space_info's bytes_may_use.
7762                          */
7763                         btrfs_free_reserved_data_space_noquota(inode,
7764                                         start, len);
7765                         goto unlock;
7766                 }
7767         }
7768
7769         /*
7770          * this will cow the extent, reset the len in case we changed
7771          * it above
7772          */
7773         len = bh_result->b_size;
7774         free_extent_map(em);
7775         em = btrfs_new_extent_direct(inode, start, len);
7776         if (IS_ERR(em)) {
7777                 ret = PTR_ERR(em);
7778                 goto unlock_err;
7779         }
7780         len = min(len, em->len - (start - em->start));
7781 unlock:
7782         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7783                 inode->i_blkbits;
7784         bh_result->b_size = len;
7785         bh_result->b_bdev = em->bdev;
7786         set_buffer_mapped(bh_result);
7787         if (create) {
7788                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7789                         set_buffer_new(bh_result);
7790
7791                 /*
7792                  * Need to update the i_size under the extent lock so buffered
7793                  * readers will get the updated i_size when we unlock.
7794                  */
7795                 if (!dio_data->overwrite && start + len > i_size_read(inode))
7796                         i_size_write(inode, start + len);
7797
7798                 WARN_ON(dio_data->reserve < len);
7799                 dio_data->reserve -= len;
7800                 dio_data->unsubmitted_oe_range_end = start + len;
7801                 current->journal_info = dio_data;
7802         }
7803
7804         /*
7805          * In the case of write we need to clear and unlock the entire range,
7806          * in the case of read we need to unlock only the end area that we
7807          * aren't using if there is any left over space.
7808          */
7809         if (lockstart < lockend) {
7810                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7811                                  lockend, unlock_bits, 1, 0,
7812                                  &cached_state);
7813         } else {
7814                 free_extent_state(cached_state);
7815         }
7816
7817         free_extent_map(em);
7818
7819         return 0;
7820
7821 unlock_err:
7822         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7823                          unlock_bits, 1, 0, &cached_state);
7824 err:
7825         if (dio_data)
7826                 current->journal_info = dio_data;
7827         return ret;
7828 }
7829
7830 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7831                                                  struct bio *bio,
7832                                                  int mirror_num)
7833 {
7834         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7835         blk_status_t ret;
7836
7837         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7838
7839         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7840         if (ret)
7841                 return ret;
7842
7843         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
7844
7845         return ret;
7846 }
7847
7848 static int btrfs_check_dio_repairable(struct inode *inode,
7849                                       struct bio *failed_bio,
7850                                       struct io_failure_record *failrec,
7851                                       int failed_mirror)
7852 {
7853         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7854         int num_copies;
7855
7856         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7857         if (num_copies == 1) {
7858                 /*
7859                  * we only have a single copy of the data, so don't bother with
7860                  * all the retry and error correction code that follows. no
7861                  * matter what the error is, it is very likely to persist.
7862                  */
7863                 btrfs_debug(fs_info,
7864                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7865                         num_copies, failrec->this_mirror, failed_mirror);
7866                 return 0;
7867         }
7868
7869         failrec->failed_mirror = failed_mirror;
7870         failrec->this_mirror++;
7871         if (failrec->this_mirror == failed_mirror)
7872                 failrec->this_mirror++;
7873
7874         if (failrec->this_mirror > num_copies) {
7875                 btrfs_debug(fs_info,
7876                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7877                         num_copies, failrec->this_mirror, failed_mirror);
7878                 return 0;
7879         }
7880
7881         return 1;
7882 }
7883
7884 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
7885                                    struct page *page, unsigned int pgoff,
7886                                    u64 start, u64 end, int failed_mirror,
7887                                    bio_end_io_t *repair_endio, void *repair_arg)
7888 {
7889         struct io_failure_record *failrec;
7890         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7891         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
7892         struct bio *bio;
7893         int isector;
7894         unsigned int read_mode = 0;
7895         int segs;
7896         int ret;
7897         blk_status_t status;
7898
7899         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7900
7901         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7902         if (ret)
7903                 return errno_to_blk_status(ret);
7904
7905         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7906                                          failed_mirror);
7907         if (!ret) {
7908                 free_io_failure(failure_tree, io_tree, failrec);
7909                 return BLK_STS_IOERR;
7910         }
7911
7912         segs = bio_segments(failed_bio);
7913         if (segs > 1 ||
7914             (failed_bio->bi_io_vec->bv_len > btrfs_inode_sectorsize(inode)))
7915                 read_mode |= REQ_FAILFAST_DEV;
7916
7917         isector = start - btrfs_io_bio(failed_bio)->logical;
7918         isector >>= inode->i_sb->s_blocksize_bits;
7919         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7920                                 pgoff, isector, repair_endio, repair_arg);
7921         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
7922
7923         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7924                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
7925                     read_mode, failrec->this_mirror, failrec->in_validation);
7926
7927         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7928         if (status) {
7929                 free_io_failure(failure_tree, io_tree, failrec);
7930                 bio_put(bio);
7931         }
7932
7933         return status;
7934 }
7935
7936 struct btrfs_retry_complete {
7937         struct completion done;
7938         struct inode *inode;
7939         u64 start;
7940         int uptodate;
7941 };
7942
7943 static void btrfs_retry_endio_nocsum(struct bio *bio)
7944 {
7945         struct btrfs_retry_complete *done = bio->bi_private;
7946         struct inode *inode = done->inode;
7947         struct bio_vec *bvec;
7948         struct extent_io_tree *io_tree, *failure_tree;
7949         int i;
7950
7951         if (bio->bi_status)
7952                 goto end;
7953
7954         ASSERT(bio->bi_vcnt == 1);
7955         io_tree = &BTRFS_I(inode)->io_tree;
7956         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7957         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(inode));
7958
7959         done->uptodate = 1;
7960         ASSERT(!bio_flagged(bio, BIO_CLONED));
7961         bio_for_each_segment_all(bvec, bio, i)
7962                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
7963                                  io_tree, done->start, bvec->bv_page,
7964                                  btrfs_ino(BTRFS_I(inode)), 0);
7965 end:
7966         complete(&done->done);
7967         bio_put(bio);
7968 }
7969
7970 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
7971                                                 struct btrfs_io_bio *io_bio)
7972 {
7973         struct btrfs_fs_info *fs_info;
7974         struct bio_vec bvec;
7975         struct bvec_iter iter;
7976         struct btrfs_retry_complete done;
7977         u64 start;
7978         unsigned int pgoff;
7979         u32 sectorsize;
7980         int nr_sectors;
7981         blk_status_t ret;
7982         blk_status_t err = BLK_STS_OK;
7983
7984         fs_info = BTRFS_I(inode)->root->fs_info;
7985         sectorsize = fs_info->sectorsize;
7986
7987         start = io_bio->logical;
7988         done.inode = inode;
7989         io_bio->bio.bi_iter = io_bio->iter;
7990
7991         bio_for_each_segment(bvec, &io_bio->bio, iter) {
7992                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
7993                 pgoff = bvec.bv_offset;
7994
7995 next_block_or_try_again:
7996                 done.uptodate = 0;
7997                 done.start = start;
7998                 init_completion(&done.done);
7999
8000                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8001                                 pgoff, start, start + sectorsize - 1,
8002                                 io_bio->mirror_num,
8003                                 btrfs_retry_endio_nocsum, &done);
8004                 if (ret) {
8005                         err = ret;
8006                         goto next;
8007                 }
8008
8009                 wait_for_completion_io(&done.done);
8010
8011                 if (!done.uptodate) {
8012                         /* We might have another mirror, so try again */
8013                         goto next_block_or_try_again;
8014                 }
8015
8016 next:
8017                 start += sectorsize;
8018
8019                 nr_sectors--;
8020                 if (nr_sectors) {
8021                         pgoff += sectorsize;
8022                         ASSERT(pgoff < PAGE_SIZE);
8023                         goto next_block_or_try_again;
8024                 }
8025         }
8026
8027         return err;
8028 }
8029
8030 static void btrfs_retry_endio(struct bio *bio)
8031 {
8032         struct btrfs_retry_complete *done = bio->bi_private;
8033         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8034         struct extent_io_tree *io_tree, *failure_tree;
8035         struct inode *inode = done->inode;
8036         struct bio_vec *bvec;
8037         int uptodate;
8038         int ret;
8039         int i;
8040
8041         if (bio->bi_status)
8042                 goto end;
8043
8044         uptodate = 1;
8045
8046         ASSERT(bio->bi_vcnt == 1);
8047         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(done->inode));
8048
8049         io_tree = &BTRFS_I(inode)->io_tree;
8050         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8051
8052         ASSERT(!bio_flagged(bio, BIO_CLONED));
8053         bio_for_each_segment_all(bvec, bio, i) {
8054                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
8055                                              bvec->bv_offset, done->start,
8056                                              bvec->bv_len);
8057                 if (!ret)
8058                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
8059                                          failure_tree, io_tree, done->start,
8060                                          bvec->bv_page,
8061                                          btrfs_ino(BTRFS_I(inode)),
8062                                          bvec->bv_offset);
8063                 else
8064                         uptodate = 0;
8065         }
8066
8067         done->uptodate = uptodate;
8068 end:
8069         complete(&done->done);
8070         bio_put(bio);
8071 }
8072
8073 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
8074                 struct btrfs_io_bio *io_bio, blk_status_t err)
8075 {
8076         struct btrfs_fs_info *fs_info;
8077         struct bio_vec bvec;
8078         struct bvec_iter iter;
8079         struct btrfs_retry_complete done;
8080         u64 start;
8081         u64 offset = 0;
8082         u32 sectorsize;
8083         int nr_sectors;
8084         unsigned int pgoff;
8085         int csum_pos;
8086         bool uptodate = (err == 0);
8087         int ret;
8088         blk_status_t status;
8089
8090         fs_info = BTRFS_I(inode)->root->fs_info;
8091         sectorsize = fs_info->sectorsize;
8092
8093         err = BLK_STS_OK;
8094         start = io_bio->logical;
8095         done.inode = inode;
8096         io_bio->bio.bi_iter = io_bio->iter;
8097
8098         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8099                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8100
8101                 pgoff = bvec.bv_offset;
8102 next_block:
8103                 if (uptodate) {
8104                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8105                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
8106                                         bvec.bv_page, pgoff, start, sectorsize);
8107                         if (likely(!ret))
8108                                 goto next;
8109                 }
8110 try_again:
8111                 done.uptodate = 0;
8112                 done.start = start;
8113                 init_completion(&done.done);
8114
8115                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8116                                         pgoff, start, start + sectorsize - 1,
8117                                         io_bio->mirror_num, btrfs_retry_endio,
8118                                         &done);
8119                 if (status) {
8120                         err = status;
8121                         goto next;
8122                 }
8123
8124                 wait_for_completion_io(&done.done);
8125
8126                 if (!done.uptodate) {
8127                         /* We might have another mirror, so try again */
8128                         goto try_again;
8129                 }
8130 next:
8131                 offset += sectorsize;
8132                 start += sectorsize;
8133
8134                 ASSERT(nr_sectors);
8135
8136                 nr_sectors--;
8137                 if (nr_sectors) {
8138                         pgoff += sectorsize;
8139                         ASSERT(pgoff < PAGE_SIZE);
8140                         goto next_block;
8141                 }
8142         }
8143
8144         return err;
8145 }
8146
8147 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8148                 struct btrfs_io_bio *io_bio, blk_status_t err)
8149 {
8150         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8151
8152         if (skip_csum) {
8153                 if (unlikely(err))
8154                         return __btrfs_correct_data_nocsum(inode, io_bio);
8155                 else
8156                         return BLK_STS_OK;
8157         } else {
8158                 return __btrfs_subio_endio_read(inode, io_bio, err);
8159         }
8160 }
8161
8162 static void btrfs_endio_direct_read(struct bio *bio)
8163 {
8164         struct btrfs_dio_private *dip = bio->bi_private;
8165         struct inode *inode = dip->inode;
8166         struct bio *dio_bio;
8167         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8168         blk_status_t err = bio->bi_status;
8169
8170         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8171                 err = btrfs_subio_endio_read(inode, io_bio, err);
8172
8173         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8174                       dip->logical_offset + dip->bytes - 1);
8175         dio_bio = dip->dio_bio;
8176
8177         kfree(dip);
8178
8179         dio_bio->bi_status = err;
8180         dio_end_io(dio_bio);
8181
8182         if (io_bio->end_io)
8183                 io_bio->end_io(io_bio, blk_status_to_errno(err));
8184         bio_put(bio);
8185 }
8186
8187 static void __endio_write_update_ordered(struct inode *inode,
8188                                          const u64 offset, const u64 bytes,
8189                                          const bool uptodate)
8190 {
8191         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8192         struct btrfs_ordered_extent *ordered = NULL;
8193         struct btrfs_workqueue *wq;
8194         btrfs_work_func_t func;
8195         u64 ordered_offset = offset;
8196         u64 ordered_bytes = bytes;
8197         u64 last_offset;
8198         int ret;
8199
8200         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
8201                 wq = fs_info->endio_freespace_worker;
8202                 func = btrfs_freespace_write_helper;
8203         } else {
8204                 wq = fs_info->endio_write_workers;
8205                 func = btrfs_endio_write_helper;
8206         }
8207
8208 again:
8209         last_offset = ordered_offset;
8210         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8211                                                    &ordered_offset,
8212                                                    ordered_bytes,
8213                                                    uptodate);
8214         if (!ret)
8215                 goto out_test;
8216
8217         btrfs_init_work(&ordered->work, func, finish_ordered_fn, NULL, NULL);
8218         btrfs_queue_work(wq, &ordered->work);
8219 out_test:
8220         /*
8221          * If btrfs_dec_test_ordered_pending does not find any ordered extent
8222          * in the range, we can exit.
8223          */
8224         if (ordered_offset == last_offset)
8225                 return;
8226         /*
8227          * our bio might span multiple ordered extents.  If we haven't
8228          * completed the accounting for the whole dio, go back and try again
8229          */
8230         if (ordered_offset < offset + bytes) {
8231                 ordered_bytes = offset + bytes - ordered_offset;
8232                 ordered = NULL;
8233                 goto again;
8234         }
8235 }
8236
8237 static void btrfs_endio_direct_write(struct bio *bio)
8238 {
8239         struct btrfs_dio_private *dip = bio->bi_private;
8240         struct bio *dio_bio = dip->dio_bio;
8241
8242         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8243                                      dip->bytes, !bio->bi_status);
8244
8245         kfree(dip);
8246
8247         dio_bio->bi_status = bio->bi_status;
8248         dio_end_io(dio_bio);
8249         bio_put(bio);
8250 }
8251
8252 static blk_status_t __btrfs_submit_bio_start_direct_io(void *private_data,
8253                                     struct bio *bio, int mirror_num,
8254                                     unsigned long bio_flags, u64 offset)
8255 {
8256         struct inode *inode = private_data;
8257         blk_status_t ret;
8258         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8259         BUG_ON(ret); /* -ENOMEM */
8260         return 0;
8261 }
8262
8263 static void btrfs_end_dio_bio(struct bio *bio)
8264 {
8265         struct btrfs_dio_private *dip = bio->bi_private;
8266         blk_status_t err = bio->bi_status;
8267
8268         if (err)
8269                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8270                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8271                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8272                            bio->bi_opf,
8273                            (unsigned long long)bio->bi_iter.bi_sector,
8274                            bio->bi_iter.bi_size, err);
8275
8276         if (dip->subio_endio)
8277                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8278
8279         if (err) {
8280                 dip->errors = 1;
8281
8282                 /*
8283                  * before atomic variable goto zero, we must make sure
8284                  * dip->errors is perceived to be set.
8285                  */
8286                 smp_mb__before_atomic();
8287         }
8288
8289         /* if there are more bios still pending for this dio, just exit */
8290         if (!atomic_dec_and_test(&dip->pending_bios))
8291                 goto out;
8292
8293         if (dip->errors) {
8294                 bio_io_error(dip->orig_bio);
8295         } else {
8296                 dip->dio_bio->bi_status = BLK_STS_OK;
8297                 bio_endio(dip->orig_bio);
8298         }
8299 out:
8300         bio_put(bio);
8301 }
8302
8303 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8304                                                  struct btrfs_dio_private *dip,
8305                                                  struct bio *bio,
8306                                                  u64 file_offset)
8307 {
8308         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8309         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8310         blk_status_t ret;
8311
8312         /*
8313          * We load all the csum data we need when we submit
8314          * the first bio to reduce the csum tree search and
8315          * contention.
8316          */
8317         if (dip->logical_offset == file_offset) {
8318                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8319                                                 file_offset);
8320                 if (ret)
8321                         return ret;
8322         }
8323
8324         if (bio == dip->orig_bio)
8325                 return 0;
8326
8327         file_offset -= dip->logical_offset;
8328         file_offset >>= inode->i_sb->s_blocksize_bits;
8329         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8330
8331         return 0;
8332 }
8333
8334 static inline blk_status_t
8335 __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode, u64 file_offset,
8336                        int async_submit)
8337 {
8338         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8339         struct btrfs_dio_private *dip = bio->bi_private;
8340         bool write = bio_op(bio) == REQ_OP_WRITE;
8341         blk_status_t ret;
8342
8343         /* Check btrfs_submit_bio_hook() for rules about async submit. */
8344         if (async_submit)
8345                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8346
8347         if (!write) {
8348                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8349                 if (ret)
8350                         goto err;
8351         }
8352
8353         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8354                 goto map;
8355
8356         if (write && async_submit) {
8357                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8358                                           file_offset, inode,
8359                                           __btrfs_submit_bio_start_direct_io,
8360                                           __btrfs_submit_bio_done);
8361                 goto err;
8362         } else if (write) {
8363                 /*
8364                  * If we aren't doing async submit, calculate the csum of the
8365                  * bio now.
8366                  */
8367                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8368                 if (ret)
8369                         goto err;
8370         } else {
8371                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8372                                                      file_offset);
8373                 if (ret)
8374                         goto err;
8375         }
8376 map:
8377         ret = btrfs_map_bio(fs_info, bio, 0, 0);
8378 err:
8379         return ret;
8380 }
8381
8382 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8383 {
8384         struct inode *inode = dip->inode;
8385         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8386         struct bio *bio;
8387         struct bio *orig_bio = dip->orig_bio;
8388         u64 start_sector = orig_bio->bi_iter.bi_sector;
8389         u64 file_offset = dip->logical_offset;
8390         u64 map_length;
8391         int async_submit = 0;
8392         u64 submit_len;
8393         int clone_offset = 0;
8394         int clone_len;
8395         int ret;
8396         blk_status_t status;
8397
8398         map_length = orig_bio->bi_iter.bi_size;
8399         submit_len = map_length;
8400         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio), start_sector << 9,
8401                               &map_length, NULL, 0);
8402         if (ret)
8403                 return -EIO;
8404
8405         if (map_length >= submit_len) {
8406                 bio = orig_bio;
8407                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8408                 goto submit;
8409         }
8410
8411         /* async crcs make it difficult to collect full stripe writes. */
8412         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8413                 async_submit = 0;
8414         else
8415                 async_submit = 1;
8416
8417         /* bio split */
8418         ASSERT(map_length <= INT_MAX);
8419         atomic_inc(&dip->pending_bios);
8420         do {
8421                 clone_len = min_t(int, submit_len, map_length);
8422
8423                 /*
8424                  * This will never fail as it's passing GPF_NOFS and
8425                  * the allocation is backed by btrfs_bioset.
8426                  */
8427                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8428                                               clone_len);
8429                 bio->bi_private = dip;
8430                 bio->bi_end_io = btrfs_end_dio_bio;
8431                 btrfs_io_bio(bio)->logical = file_offset;
8432
8433                 ASSERT(submit_len >= clone_len);
8434                 submit_len -= clone_len;
8435                 if (submit_len == 0)
8436                         break;
8437
8438                 /*
8439                  * Increase the count before we submit the bio so we know
8440                  * the end IO handler won't happen before we increase the
8441                  * count. Otherwise, the dip might get freed before we're
8442                  * done setting it up.
8443                  */
8444                 atomic_inc(&dip->pending_bios);
8445
8446                 status = __btrfs_submit_dio_bio(bio, inode, file_offset,
8447                                                 async_submit);
8448                 if (status) {
8449                         bio_put(bio);
8450                         atomic_dec(&dip->pending_bios);
8451                         goto out_err;
8452                 }
8453
8454                 clone_offset += clone_len;
8455                 start_sector += clone_len >> 9;
8456                 file_offset += clone_len;
8457
8458                 map_length = submit_len;
8459                 ret = btrfs_map_block(fs_info, btrfs_op(orig_bio),
8460                                       start_sector << 9, &map_length, NULL, 0);
8461                 if (ret)
8462                         goto out_err;
8463         } while (submit_len > 0);
8464
8465 submit:
8466         status = __btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8467         if (!status)
8468                 return 0;
8469
8470         bio_put(bio);
8471 out_err:
8472         dip->errors = 1;
8473         /*
8474          * before atomic variable goto zero, we must
8475          * make sure dip->errors is perceived to be set.
8476          */
8477         smp_mb__before_atomic();
8478         if (atomic_dec_and_test(&dip->pending_bios))
8479                 bio_io_error(dip->orig_bio);
8480
8481         /* bio_end_io() will handle error, so we needn't return it */
8482         return 0;
8483 }
8484
8485 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8486                                 loff_t file_offset)
8487 {
8488         struct btrfs_dio_private *dip = NULL;
8489         struct bio *bio = NULL;
8490         struct btrfs_io_bio *io_bio;
8491         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8492         int ret = 0;
8493
8494         bio = btrfs_bio_clone(dio_bio);
8495
8496         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8497         if (!dip) {
8498                 ret = -ENOMEM;
8499                 goto free_ordered;
8500         }
8501
8502         dip->private = dio_bio->bi_private;
8503         dip->inode = inode;
8504         dip->logical_offset = file_offset;
8505         dip->bytes = dio_bio->bi_iter.bi_size;
8506         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8507         bio->bi_private = dip;
8508         dip->orig_bio = bio;
8509         dip->dio_bio = dio_bio;
8510         atomic_set(&dip->pending_bios, 0);
8511         io_bio = btrfs_io_bio(bio);
8512         io_bio->logical = file_offset;
8513
8514         if (write) {
8515                 bio->bi_end_io = btrfs_endio_direct_write;
8516         } else {
8517                 bio->bi_end_io = btrfs_endio_direct_read;
8518                 dip->subio_endio = btrfs_subio_endio_read;
8519         }
8520
8521         /*
8522          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8523          * even if we fail to submit a bio, because in such case we do the
8524          * corresponding error handling below and it must not be done a second
8525          * time by btrfs_direct_IO().
8526          */
8527         if (write) {
8528                 struct btrfs_dio_data *dio_data = current->journal_info;
8529
8530                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8531                         dip->bytes;
8532                 dio_data->unsubmitted_oe_range_start =
8533                         dio_data->unsubmitted_oe_range_end;
8534         }
8535
8536         ret = btrfs_submit_direct_hook(dip);
8537         if (!ret)
8538                 return;
8539
8540         if (io_bio->end_io)
8541                 io_bio->end_io(io_bio, ret);
8542
8543 free_ordered:
8544         /*
8545          * If we arrived here it means either we failed to submit the dip
8546          * or we either failed to clone the dio_bio or failed to allocate the
8547          * dip. If we cloned the dio_bio and allocated the dip, we can just
8548          * call bio_endio against our io_bio so that we get proper resource
8549          * cleanup if we fail to submit the dip, otherwise, we must do the
8550          * same as btrfs_endio_direct_[write|read] because we can't call these
8551          * callbacks - they require an allocated dip and a clone of dio_bio.
8552          */
8553         if (bio && dip) {
8554                 bio_io_error(bio);
8555                 /*
8556                  * The end io callbacks free our dip, do the final put on bio
8557                  * and all the cleanup and final put for dio_bio (through
8558                  * dio_end_io()).
8559                  */
8560                 dip = NULL;
8561                 bio = NULL;
8562         } else {
8563                 if (write)
8564                         __endio_write_update_ordered(inode,
8565                                                 file_offset,
8566                                                 dio_bio->bi_iter.bi_size,
8567                                                 false);
8568                 else
8569                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8570                               file_offset + dio_bio->bi_iter.bi_size - 1);
8571
8572                 dio_bio->bi_status = BLK_STS_IOERR;
8573                 /*
8574                  * Releases and cleans up our dio_bio, no need to bio_put()
8575                  * nor bio_endio()/bio_io_error() against dio_bio.
8576                  */
8577                 dio_end_io(dio_bio);
8578         }
8579         if (bio)
8580                 bio_put(bio);
8581         kfree(dip);
8582 }
8583
8584 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8585                                const struct iov_iter *iter, loff_t offset)
8586 {
8587         int seg;
8588         int i;
8589         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8590         ssize_t retval = -EINVAL;
8591
8592         if (offset & blocksize_mask)
8593                 goto out;
8594
8595         if (iov_iter_alignment(iter) & blocksize_mask)
8596                 goto out;
8597
8598         /* If this is a write we don't need to check anymore */
8599         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8600                 return 0;
8601         /*
8602          * Check to make sure we don't have duplicate iov_base's in this
8603          * iovec, if so return EINVAL, otherwise we'll get csum errors
8604          * when reading back.
8605          */
8606         for (seg = 0; seg < iter->nr_segs; seg++) {
8607                 for (i = seg + 1; i < iter->nr_segs; i++) {
8608                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8609                                 goto out;
8610                 }
8611         }
8612         retval = 0;
8613 out:
8614         return retval;
8615 }
8616
8617 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8618 {
8619         struct file *file = iocb->ki_filp;
8620         struct inode *inode = file->f_mapping->host;
8621         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8622         struct btrfs_dio_data dio_data = { 0 };
8623         struct extent_changeset *data_reserved = NULL;
8624         loff_t offset = iocb->ki_pos;
8625         size_t count = 0;
8626         int flags = 0;
8627         bool wakeup = true;
8628         bool relock = false;
8629         ssize_t ret;
8630
8631         if (check_direct_IO(fs_info, iter, offset))
8632                 return 0;
8633
8634         inode_dio_begin(inode);
8635
8636         /*
8637          * The generic stuff only does filemap_write_and_wait_range, which
8638          * isn't enough if we've written compressed pages to this area, so
8639          * we need to flush the dirty pages again to make absolutely sure
8640          * that any outstanding dirty pages are on disk.
8641          */
8642         count = iov_iter_count(iter);
8643         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8644                      &BTRFS_I(inode)->runtime_flags))
8645                 filemap_fdatawrite_range(inode->i_mapping, offset,
8646                                          offset + count - 1);
8647
8648         if (iov_iter_rw(iter) == WRITE) {
8649                 /*
8650                  * If the write DIO is beyond the EOF, we need update
8651                  * the isize, but it is protected by i_mutex. So we can
8652                  * not unlock the i_mutex at this case.
8653                  */
8654                 if (offset + count <= inode->i_size) {
8655                         dio_data.overwrite = 1;
8656                         inode_unlock(inode);
8657                         relock = true;
8658                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8659                         ret = -EAGAIN;
8660                         goto out;
8661                 }
8662                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8663                                                    offset, count);
8664                 if (ret)
8665                         goto out;
8666
8667                 /*
8668                  * We need to know how many extents we reserved so that we can
8669                  * do the accounting properly if we go over the number we
8670                  * originally calculated.  Abuse current->journal_info for this.
8671                  */
8672                 dio_data.reserve = round_up(count,
8673                                             fs_info->sectorsize);
8674                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8675                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8676                 current->journal_info = &dio_data;
8677                 down_read(&BTRFS_I(inode)->dio_sem);
8678         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8679                                      &BTRFS_I(inode)->runtime_flags)) {
8680                 inode_dio_end(inode);
8681                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8682                 wakeup = false;
8683         }
8684
8685         ret = __blockdev_direct_IO(iocb, inode,
8686                                    fs_info->fs_devices->latest_bdev,
8687                                    iter, btrfs_get_blocks_direct, NULL,
8688                                    btrfs_submit_direct, flags);
8689         if (iov_iter_rw(iter) == WRITE) {
8690                 up_read(&BTRFS_I(inode)->dio_sem);
8691                 current->journal_info = NULL;
8692                 if (ret < 0 && ret != -EIOCBQUEUED) {
8693                         if (dio_data.reserve)
8694                                 btrfs_delalloc_release_space(inode, data_reserved,
8695                                         offset, dio_data.reserve);
8696                         /*
8697                          * On error we might have left some ordered extents
8698                          * without submitting corresponding bios for them, so
8699                          * cleanup them up to avoid other tasks getting them
8700                          * and waiting for them to complete forever.
8701                          */
8702                         if (dio_data.unsubmitted_oe_range_start <
8703                             dio_data.unsubmitted_oe_range_end)
8704                                 __endio_write_update_ordered(inode,
8705                                         dio_data.unsubmitted_oe_range_start,
8706                                         dio_data.unsubmitted_oe_range_end -
8707                                         dio_data.unsubmitted_oe_range_start,
8708                                         false);
8709                 } else if (ret >= 0 && (size_t)ret < count)
8710                         btrfs_delalloc_release_space(inode, data_reserved,
8711                                         offset, count - (size_t)ret);
8712                 btrfs_delalloc_release_extents(BTRFS_I(inode), count);
8713         }
8714 out:
8715         if (wakeup)
8716                 inode_dio_end(inode);
8717         if (relock)
8718                 inode_lock(inode);
8719
8720         extent_changeset_free(data_reserved);
8721         return ret;
8722 }
8723
8724 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8725
8726 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8727                 __u64 start, __u64 len)
8728 {
8729         int     ret;
8730
8731         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8732         if (ret)
8733                 return ret;
8734
8735         return extent_fiemap(inode, fieinfo, start, len);
8736 }
8737
8738 int btrfs_readpage(struct file *file, struct page *page)
8739 {
8740         struct extent_io_tree *tree;
8741         tree = &BTRFS_I(page->mapping->host)->io_tree;
8742         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8743 }
8744
8745 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8746 {
8747         struct inode *inode = page->mapping->host;
8748         int ret;
8749
8750         if (current->flags & PF_MEMALLOC) {
8751                 redirty_page_for_writepage(wbc, page);
8752                 unlock_page(page);
8753                 return 0;
8754         }
8755
8756         /*
8757          * If we are under memory pressure we will call this directly from the
8758          * VM, we need to make sure we have the inode referenced for the ordered
8759          * extent.  If not just return like we didn't do anything.
8760          */
8761         if (!igrab(inode)) {
8762                 redirty_page_for_writepage(wbc, page);
8763                 return AOP_WRITEPAGE_ACTIVATE;
8764         }
8765         ret = extent_write_full_page(page, wbc);
8766         btrfs_add_delayed_iput(inode);
8767         return ret;
8768 }
8769
8770 static int btrfs_writepages(struct address_space *mapping,
8771                             struct writeback_control *wbc)
8772 {
8773         struct extent_io_tree *tree;
8774
8775         tree = &BTRFS_I(mapping->host)->io_tree;
8776         return extent_writepages(tree, mapping, wbc);
8777 }
8778
8779 static int
8780 btrfs_readpages(struct file *file, struct address_space *mapping,
8781                 struct list_head *pages, unsigned nr_pages)
8782 {
8783         struct extent_io_tree *tree;
8784         tree = &BTRFS_I(mapping->host)->io_tree;
8785         return extent_readpages(tree, mapping, pages, nr_pages);
8786 }
8787 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8788 {
8789         struct extent_io_tree *tree;
8790         struct extent_map_tree *map;
8791         int ret;
8792
8793         tree = &BTRFS_I(page->mapping->host)->io_tree;
8794         map = &BTRFS_I(page->mapping->host)->extent_tree;
8795         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8796         if (ret == 1) {
8797                 ClearPagePrivate(page);
8798                 set_page_private(page, 0);
8799                 put_page(page);
8800         }
8801         return ret;
8802 }
8803
8804 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8805 {
8806         if (PageWriteback(page) || PageDirty(page))
8807                 return 0;
8808         return __btrfs_releasepage(page, gfp_flags);
8809 }
8810
8811 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8812                                  unsigned int length)
8813 {
8814         struct inode *inode = page->mapping->host;
8815         struct extent_io_tree *tree;
8816         struct btrfs_ordered_extent *ordered;
8817         struct extent_state *cached_state = NULL;
8818         u64 page_start = page_offset(page);
8819         u64 page_end = page_start + PAGE_SIZE - 1;
8820         u64 start;
8821         u64 end;
8822         int inode_evicting = inode->i_state & I_FREEING;
8823
8824         /*
8825          * we have the page locked, so new writeback can't start,
8826          * and the dirty bit won't be cleared while we are here.
8827          *
8828          * Wait for IO on this page so that we can safely clear
8829          * the PagePrivate2 bit and do ordered accounting
8830          */
8831         wait_on_page_writeback(page);
8832
8833         tree = &BTRFS_I(inode)->io_tree;
8834         if (offset) {
8835                 btrfs_releasepage(page, GFP_NOFS);
8836                 return;
8837         }
8838
8839         if (!inode_evicting)
8840                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8841 again:
8842         start = page_start;
8843         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8844                                         page_end - start + 1);
8845         if (ordered) {
8846                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8847                 /*
8848                  * IO on this page will never be started, so we need
8849                  * to account for any ordered extents now
8850                  */
8851                 if (!inode_evicting)
8852                         clear_extent_bit(tree, start, end,
8853                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8854                                          EXTENT_DELALLOC_NEW |
8855                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8856                                          EXTENT_DEFRAG, 1, 0, &cached_state);
8857                 /*
8858                  * whoever cleared the private bit is responsible
8859                  * for the finish_ordered_io
8860                  */
8861                 if (TestClearPagePrivate2(page)) {
8862                         struct btrfs_ordered_inode_tree *tree;
8863                         u64 new_len;
8864
8865                         tree = &BTRFS_I(inode)->ordered_tree;
8866
8867                         spin_lock_irq(&tree->lock);
8868                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8869                         new_len = start - ordered->file_offset;
8870                         if (new_len < ordered->truncated_len)
8871                                 ordered->truncated_len = new_len;
8872                         spin_unlock_irq(&tree->lock);
8873
8874                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8875                                                            start,
8876                                                            end - start + 1, 1))
8877                                 btrfs_finish_ordered_io(ordered);
8878                 }
8879                 btrfs_put_ordered_extent(ordered);
8880                 if (!inode_evicting) {
8881                         cached_state = NULL;
8882                         lock_extent_bits(tree, start, end,
8883                                          &cached_state);
8884                 }
8885
8886                 start = end + 1;
8887                 if (start < page_end)
8888                         goto again;
8889         }
8890
8891         /*
8892          * Qgroup reserved space handler
8893          * Page here will be either
8894          * 1) Already written to disk
8895          *    In this case, its reserved space is released from data rsv map
8896          *    and will be freed by delayed_ref handler finally.
8897          *    So even we call qgroup_free_data(), it won't decrease reserved
8898          *    space.
8899          * 2) Not written to disk
8900          *    This means the reserved space should be freed here. However,
8901          *    if a truncate invalidates the page (by clearing PageDirty)
8902          *    and the page is accounted for while allocating extent
8903          *    in btrfs_check_data_free_space() we let delayed_ref to
8904          *    free the entire extent.
8905          */
8906         if (PageDirty(page))
8907                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
8908         if (!inode_evicting) {
8909                 clear_extent_bit(tree, page_start, page_end,
8910                                  EXTENT_LOCKED | EXTENT_DIRTY |
8911                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8912                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
8913                                  &cached_state);
8914
8915                 __btrfs_releasepage(page, GFP_NOFS);
8916         }
8917
8918         ClearPageChecked(page);
8919         if (PagePrivate(page)) {
8920                 ClearPagePrivate(page);
8921                 set_page_private(page, 0);
8922                 put_page(page);
8923         }
8924 }
8925
8926 /*
8927  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8928  * called from a page fault handler when a page is first dirtied. Hence we must
8929  * be careful to check for EOF conditions here. We set the page up correctly
8930  * for a written page which means we get ENOSPC checking when writing into
8931  * holes and correct delalloc and unwritten extent mapping on filesystems that
8932  * support these features.
8933  *
8934  * We are not allowed to take the i_mutex here so we have to play games to
8935  * protect against truncate races as the page could now be beyond EOF.  Because
8936  * vmtruncate() writes the inode size before removing pages, once we have the
8937  * page lock we can determine safely if the page is beyond EOF. If it is not
8938  * beyond EOF, then the page is guaranteed safe against truncation until we
8939  * unlock the page.
8940  */
8941 int btrfs_page_mkwrite(struct vm_fault *vmf)
8942 {
8943         struct page *page = vmf->page;
8944         struct inode *inode = file_inode(vmf->vma->vm_file);
8945         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8946         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8947         struct btrfs_ordered_extent *ordered;
8948         struct extent_state *cached_state = NULL;
8949         struct extent_changeset *data_reserved = NULL;
8950         char *kaddr;
8951         unsigned long zero_start;
8952         loff_t size;
8953         int ret;
8954         int reserved = 0;
8955         u64 reserved_space;
8956         u64 page_start;
8957         u64 page_end;
8958         u64 end;
8959
8960         reserved_space = PAGE_SIZE;
8961
8962         sb_start_pagefault(inode->i_sb);
8963         page_start = page_offset(page);
8964         page_end = page_start + PAGE_SIZE - 1;
8965         end = page_end;
8966
8967         /*
8968          * Reserving delalloc space after obtaining the page lock can lead to
8969          * deadlock. For example, if a dirty page is locked by this function
8970          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8971          * dirty page write out, then the btrfs_writepage() function could
8972          * end up waiting indefinitely to get a lock on the page currently
8973          * being processed by btrfs_page_mkwrite() function.
8974          */
8975         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
8976                                            reserved_space);
8977         if (!ret) {
8978                 ret = file_update_time(vmf->vma->vm_file);
8979                 reserved = 1;
8980         }
8981         if (ret) {
8982                 if (ret == -ENOMEM)
8983                         ret = VM_FAULT_OOM;
8984                 else /* -ENOSPC, -EIO, etc */
8985                         ret = VM_FAULT_SIGBUS;
8986                 if (reserved)
8987                         goto out;
8988                 goto out_noreserve;
8989         }
8990
8991         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8992 again:
8993         lock_page(page);
8994         size = i_size_read(inode);
8995
8996         if ((page->mapping != inode->i_mapping) ||
8997             (page_start >= size)) {
8998                 /* page got truncated out from underneath us */
8999                 goto out_unlock;
9000         }
9001         wait_on_page_writeback(page);
9002
9003         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
9004         set_page_extent_mapped(page);
9005
9006         /*
9007          * we can't set the delalloc bits if there are pending ordered
9008          * extents.  Drop our locks and wait for them to finish
9009          */
9010         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
9011                         PAGE_SIZE);
9012         if (ordered) {
9013                 unlock_extent_cached(io_tree, page_start, page_end,
9014                                      &cached_state);
9015                 unlock_page(page);
9016                 btrfs_start_ordered_extent(inode, ordered, 1);
9017                 btrfs_put_ordered_extent(ordered);
9018                 goto again;
9019         }
9020
9021         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9022                 reserved_space = round_up(size - page_start,
9023                                           fs_info->sectorsize);
9024                 if (reserved_space < PAGE_SIZE) {
9025                         end = page_start + reserved_space - 1;
9026                         btrfs_delalloc_release_space(inode, data_reserved,
9027                                         page_start, PAGE_SIZE - reserved_space);
9028                 }
9029         }
9030
9031         /*
9032          * page_mkwrite gets called when the page is firstly dirtied after it's
9033          * faulted in, but write(2) could also dirty a page and set delalloc
9034          * bits, thus in this case for space account reason, we still need to
9035          * clear any delalloc bits within this page range since we have to
9036          * reserve data&meta space before lock_page() (see above comments).
9037          */
9038         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9039                           EXTENT_DIRTY | EXTENT_DELALLOC |
9040                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9041                           0, 0, &cached_state);
9042
9043         ret = btrfs_set_extent_delalloc(inode, page_start, end, 0,
9044                                         &cached_state, 0);
9045         if (ret) {
9046                 unlock_extent_cached(io_tree, page_start, page_end,
9047                                      &cached_state);
9048                 ret = VM_FAULT_SIGBUS;
9049                 goto out_unlock;
9050         }
9051         ret = 0;
9052
9053         /* page is wholly or partially inside EOF */
9054         if (page_start + PAGE_SIZE > size)
9055                 zero_start = size & ~PAGE_MASK;
9056         else
9057                 zero_start = PAGE_SIZE;
9058
9059         if (zero_start != PAGE_SIZE) {
9060                 kaddr = kmap(page);
9061                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9062                 flush_dcache_page(page);
9063                 kunmap(page);
9064         }
9065         ClearPageChecked(page);
9066         set_page_dirty(page);
9067         SetPageUptodate(page);
9068
9069         BTRFS_I(inode)->last_trans = fs_info->generation;
9070         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9071         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9072
9073         unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
9074
9075 out_unlock:
9076         if (!ret) {
9077                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9078                 sb_end_pagefault(inode->i_sb);
9079                 extent_changeset_free(data_reserved);
9080                 return VM_FAULT_LOCKED;
9081         }
9082         unlock_page(page);
9083 out:
9084         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9085         btrfs_delalloc_release_space(inode, data_reserved, page_start,
9086                                      reserved_space);
9087 out_noreserve:
9088         sb_end_pagefault(inode->i_sb);
9089         extent_changeset_free(data_reserved);
9090         return ret;
9091 }
9092
9093 static int btrfs_truncate(struct inode *inode)
9094 {
9095         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9096         struct btrfs_root *root = BTRFS_I(inode)->root;
9097         struct btrfs_block_rsv *rsv;
9098         int ret = 0;
9099         int err = 0;
9100         struct btrfs_trans_handle *trans;
9101         u64 mask = fs_info->sectorsize - 1;
9102         u64 min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
9103
9104         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9105                                        (u64)-1);
9106         if (ret)
9107                 return ret;
9108
9109         /*
9110          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9111          * 3 things going on here
9112          *
9113          * 1) We need to reserve space for our orphan item and the space to
9114          * delete our orphan item.  Lord knows we don't want to have a dangling
9115          * orphan item because we didn't reserve space to remove it.
9116          *
9117          * 2) We need to reserve space to update our inode.
9118          *
9119          * 3) We need to have something to cache all the space that is going to
9120          * be free'd up by the truncate operation, but also have some slack
9121          * space reserved in case it uses space during the truncate (thank you
9122          * very much snapshotting).
9123          *
9124          * And we need these to all be separate.  The fact is we can use a lot of
9125          * space doing the truncate, and we have no earthly idea how much space
9126          * we will use, so we need the truncate reservation to be separate so it
9127          * doesn't end up using space reserved for updating the inode or
9128          * removing the orphan item.  We also need to be able to stop the
9129          * transaction and start a new one, which means we need to be able to
9130          * update the inode several times, and we have no idea of knowing how
9131          * many times that will be, so we can't just reserve 1 item for the
9132          * entirety of the operation, so that has to be done separately as well.
9133          * Then there is the orphan item, which does indeed need to be held on
9134          * to for the whole operation, and we need nobody to touch this reserved
9135          * space except the orphan code.
9136          *
9137          * So that leaves us with
9138          *
9139          * 1) root->orphan_block_rsv - for the orphan deletion.
9140          * 2) rsv - for the truncate reservation, which we will steal from the
9141          * transaction reservation.
9142          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9143          * updating the inode.
9144          */
9145         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9146         if (!rsv)
9147                 return -ENOMEM;
9148         rsv->size = min_size;
9149         rsv->failfast = 1;
9150
9151         /*
9152          * 1 for the truncate slack space
9153          * 1 for updating the inode.
9154          */
9155         trans = btrfs_start_transaction(root, 2);
9156         if (IS_ERR(trans)) {
9157                 err = PTR_ERR(trans);
9158                 goto out;
9159         }
9160
9161         /* Migrate the slack space for the truncate to our reserve */
9162         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9163                                       min_size, 0);
9164         BUG_ON(ret);
9165
9166         /*
9167          * So if we truncate and then write and fsync we normally would just
9168          * write the extents that changed, which is a problem if we need to
9169          * first truncate that entire inode.  So set this flag so we write out
9170          * all of the extents in the inode to the sync log so we're completely
9171          * safe.
9172          */
9173         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9174         trans->block_rsv = rsv;
9175
9176         while (1) {
9177                 ret = btrfs_truncate_inode_items(trans, root, inode,
9178                                                  inode->i_size,
9179                                                  BTRFS_EXTENT_DATA_KEY);
9180                 trans->block_rsv = &fs_info->trans_block_rsv;
9181                 if (ret != -ENOSPC && ret != -EAGAIN) {
9182                         err = ret;
9183                         break;
9184                 }
9185
9186                 ret = btrfs_update_inode(trans, root, inode);
9187                 if (ret) {
9188                         err = ret;
9189                         break;
9190                 }
9191
9192                 btrfs_end_transaction(trans);
9193                 btrfs_btree_balance_dirty(fs_info);
9194
9195                 trans = btrfs_start_transaction(root, 2);
9196                 if (IS_ERR(trans)) {
9197                         ret = err = PTR_ERR(trans);
9198                         trans = NULL;
9199                         break;
9200                 }
9201
9202                 btrfs_block_rsv_release(fs_info, rsv, -1);
9203                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9204                                               rsv, min_size, 0);
9205                 BUG_ON(ret);    /* shouldn't happen */
9206                 trans->block_rsv = rsv;
9207         }
9208
9209         /*
9210          * We can't call btrfs_truncate_block inside a trans handle as we could
9211          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
9212          * we've truncated everything except the last little bit, and can do
9213          * btrfs_truncate_block and then update the disk_i_size.
9214          */
9215         if (ret == NEED_TRUNCATE_BLOCK) {
9216                 btrfs_end_transaction(trans);
9217                 btrfs_btree_balance_dirty(fs_info);
9218
9219                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
9220                 if (ret)
9221                         goto out;
9222                 trans = btrfs_start_transaction(root, 1);
9223                 if (IS_ERR(trans)) {
9224                         ret = PTR_ERR(trans);
9225                         goto out;
9226                 }
9227                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
9228         }
9229
9230         if (ret == 0 && inode->i_nlink > 0) {
9231                 trans->block_rsv = root->orphan_block_rsv;
9232                 ret = btrfs_orphan_del(trans, BTRFS_I(inode));
9233                 if (ret)
9234                         err = ret;
9235         }
9236
9237         if (trans) {
9238                 trans->block_rsv = &fs_info->trans_block_rsv;
9239                 ret = btrfs_update_inode(trans, root, inode);
9240                 if (ret && !err)
9241                         err = ret;
9242
9243                 ret = btrfs_end_transaction(trans);
9244                 btrfs_btree_balance_dirty(fs_info);
9245         }
9246 out:
9247         btrfs_free_block_rsv(fs_info, rsv);
9248
9249         if (ret && !err)
9250                 err = ret;
9251
9252         return err;
9253 }
9254
9255 /*
9256  * create a new subvolume directory/inode (helper for the ioctl).
9257  */
9258 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9259                              struct btrfs_root *new_root,
9260                              struct btrfs_root *parent_root,
9261                              u64 new_dirid)
9262 {
9263         struct inode *inode;
9264         int err;
9265         u64 index = 0;
9266
9267         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9268                                 new_dirid, new_dirid,
9269                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9270                                 &index);
9271         if (IS_ERR(inode))
9272                 return PTR_ERR(inode);
9273         inode->i_op = &btrfs_dir_inode_operations;
9274         inode->i_fop = &btrfs_dir_file_operations;
9275
9276         set_nlink(inode, 1);
9277         btrfs_i_size_write(BTRFS_I(inode), 0);
9278         unlock_new_inode(inode);
9279
9280         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9281         if (err)
9282                 btrfs_err(new_root->fs_info,
9283                           "error inheriting subvolume %llu properties: %d",
9284                           new_root->root_key.objectid, err);
9285
9286         err = btrfs_update_inode(trans, new_root, inode);
9287
9288         iput(inode);
9289         return err;
9290 }
9291
9292 struct inode *btrfs_alloc_inode(struct super_block *sb)
9293 {
9294         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
9295         struct btrfs_inode *ei;
9296         struct inode *inode;
9297
9298         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_KERNEL);
9299         if (!ei)
9300                 return NULL;
9301
9302         ei->root = NULL;
9303         ei->generation = 0;
9304         ei->last_trans = 0;
9305         ei->last_sub_trans = 0;
9306         ei->logged_trans = 0;
9307         ei->delalloc_bytes = 0;
9308         ei->new_delalloc_bytes = 0;
9309         ei->defrag_bytes = 0;
9310         ei->disk_i_size = 0;
9311         ei->flags = 0;
9312         ei->csum_bytes = 0;
9313         ei->index_cnt = (u64)-1;
9314         ei->dir_index = 0;
9315         ei->last_unlink_trans = 0;
9316         ei->last_log_commit = 0;
9317         ei->delayed_iput_count = 0;
9318
9319         spin_lock_init(&ei->lock);
9320         ei->outstanding_extents = 0;
9321         if (sb->s_magic != BTRFS_TEST_MAGIC)
9322                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
9323                                               BTRFS_BLOCK_RSV_DELALLOC);
9324         ei->runtime_flags = 0;
9325         ei->prop_compress = BTRFS_COMPRESS_NONE;
9326         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9327
9328         ei->delayed_node = NULL;
9329
9330         ei->i_otime.tv_sec = 0;
9331         ei->i_otime.tv_nsec = 0;
9332
9333         inode = &ei->vfs_inode;
9334         extent_map_tree_init(&ei->extent_tree);
9335         extent_io_tree_init(&ei->io_tree, inode);
9336         extent_io_tree_init(&ei->io_failure_tree, inode);
9337         ei->io_tree.track_uptodate = 1;
9338         ei->io_failure_tree.track_uptodate = 1;
9339         atomic_set(&ei->sync_writers, 0);
9340         mutex_init(&ei->log_mutex);
9341         mutex_init(&ei->delalloc_mutex);
9342         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9343         INIT_LIST_HEAD(&ei->delalloc_inodes);
9344         INIT_LIST_HEAD(&ei->delayed_iput);
9345         RB_CLEAR_NODE(&ei->rb_node);
9346         init_rwsem(&ei->dio_sem);
9347
9348         return inode;
9349 }
9350
9351 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9352 void btrfs_test_destroy_inode(struct inode *inode)
9353 {
9354         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9355         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9356 }
9357 #endif
9358
9359 static void btrfs_i_callback(struct rcu_head *head)
9360 {
9361         struct inode *inode = container_of(head, struct inode, i_rcu);
9362         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9363 }
9364
9365 void btrfs_destroy_inode(struct inode *inode)
9366 {
9367         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9368         struct btrfs_ordered_extent *ordered;
9369         struct btrfs_root *root = BTRFS_I(inode)->root;
9370
9371         WARN_ON(!hlist_empty(&inode->i_dentry));
9372         WARN_ON(inode->i_data.nrpages);
9373         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
9374         WARN_ON(BTRFS_I(inode)->block_rsv.size);
9375         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9376         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9377         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9378         WARN_ON(BTRFS_I(inode)->csum_bytes);
9379         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9380
9381         /*
9382          * This can happen where we create an inode, but somebody else also
9383          * created the same inode and we need to destroy the one we already
9384          * created.
9385          */
9386         if (!root)
9387                 goto free;
9388
9389         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9390                      &BTRFS_I(inode)->runtime_flags)) {
9391                 btrfs_info(fs_info, "inode %llu still on the orphan list",
9392                            btrfs_ino(BTRFS_I(inode)));
9393                 atomic_dec(&root->orphan_inodes);
9394         }
9395
9396         while (1) {
9397                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9398                 if (!ordered)
9399                         break;
9400                 else {
9401                         btrfs_err(fs_info,
9402                                   "found ordered extent %llu %llu on inode cleanup",
9403                                   ordered->file_offset, ordered->len);
9404                         btrfs_remove_ordered_extent(inode, ordered);
9405                         btrfs_put_ordered_extent(ordered);
9406                         btrfs_put_ordered_extent(ordered);
9407                 }
9408         }
9409         btrfs_qgroup_check_reserved_leak(inode);
9410         inode_tree_del(inode);
9411         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9412 free:
9413         call_rcu(&inode->i_rcu, btrfs_i_callback);
9414 }
9415
9416 int btrfs_drop_inode(struct inode *inode)
9417 {
9418         struct btrfs_root *root = BTRFS_I(inode)->root;
9419
9420         if (root == NULL)
9421                 return 1;
9422
9423         /* the snap/subvol tree is on deleting */
9424         if (btrfs_root_refs(&root->root_item) == 0)
9425                 return 1;
9426         else
9427                 return generic_drop_inode(inode);
9428 }
9429
9430 static void init_once(void *foo)
9431 {
9432         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9433
9434         inode_init_once(&ei->vfs_inode);
9435 }
9436
9437 void btrfs_destroy_cachep(void)
9438 {
9439         /*
9440          * Make sure all delayed rcu free inodes are flushed before we
9441          * destroy cache.
9442          */
9443         rcu_barrier();
9444         kmem_cache_destroy(btrfs_inode_cachep);
9445         kmem_cache_destroy(btrfs_trans_handle_cachep);
9446         kmem_cache_destroy(btrfs_path_cachep);
9447         kmem_cache_destroy(btrfs_free_space_cachep);
9448 }
9449
9450 int __init btrfs_init_cachep(void)
9451 {
9452         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9453                         sizeof(struct btrfs_inode), 0,
9454                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9455                         init_once);
9456         if (!btrfs_inode_cachep)
9457                 goto fail;
9458
9459         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9460                         sizeof(struct btrfs_trans_handle), 0,
9461                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9462         if (!btrfs_trans_handle_cachep)
9463                 goto fail;
9464
9465         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9466                         sizeof(struct btrfs_path), 0,
9467                         SLAB_MEM_SPREAD, NULL);
9468         if (!btrfs_path_cachep)
9469                 goto fail;
9470
9471         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9472                         sizeof(struct btrfs_free_space), 0,
9473                         SLAB_MEM_SPREAD, NULL);
9474         if (!btrfs_free_space_cachep)
9475                 goto fail;
9476
9477         return 0;
9478 fail:
9479         btrfs_destroy_cachep();
9480         return -ENOMEM;
9481 }
9482
9483 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9484                          u32 request_mask, unsigned int flags)
9485 {
9486         u64 delalloc_bytes;
9487         struct inode *inode = d_inode(path->dentry);
9488         u32 blocksize = inode->i_sb->s_blocksize;
9489         u32 bi_flags = BTRFS_I(inode)->flags;
9490
9491         stat->result_mask |= STATX_BTIME;
9492         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9493         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9494         if (bi_flags & BTRFS_INODE_APPEND)
9495                 stat->attributes |= STATX_ATTR_APPEND;
9496         if (bi_flags & BTRFS_INODE_COMPRESS)
9497                 stat->attributes |= STATX_ATTR_COMPRESSED;
9498         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9499                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9500         if (bi_flags & BTRFS_INODE_NODUMP)
9501                 stat->attributes |= STATX_ATTR_NODUMP;
9502
9503         stat->attributes_mask |= (STATX_ATTR_APPEND |
9504                                   STATX_ATTR_COMPRESSED |
9505                                   STATX_ATTR_IMMUTABLE |
9506                                   STATX_ATTR_NODUMP);
9507
9508         generic_fillattr(inode, stat);
9509         stat->dev = BTRFS_I(inode)->root->anon_dev;
9510
9511         spin_lock(&BTRFS_I(inode)->lock);
9512         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9513         spin_unlock(&BTRFS_I(inode)->lock);
9514         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9515                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9516         return 0;
9517 }
9518
9519 static int btrfs_rename_exchange(struct inode *old_dir,
9520                               struct dentry *old_dentry,
9521                               struct inode *new_dir,
9522                               struct dentry *new_dentry)
9523 {
9524         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9525         struct btrfs_trans_handle *trans;
9526         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9527         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9528         struct inode *new_inode = new_dentry->d_inode;
9529         struct inode *old_inode = old_dentry->d_inode;
9530         struct timespec ctime = current_time(old_inode);
9531         struct dentry *parent;
9532         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9533         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9534         u64 old_idx = 0;
9535         u64 new_idx = 0;
9536         u64 root_objectid;
9537         int ret;
9538         bool root_log_pinned = false;
9539         bool dest_log_pinned = false;
9540
9541         /* we only allow rename subvolume link between subvolumes */
9542         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9543                 return -EXDEV;
9544
9545         /* close the race window with snapshot create/destroy ioctl */
9546         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9547                 down_read(&fs_info->subvol_sem);
9548         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9549                 down_read(&fs_info->subvol_sem);
9550
9551         /*
9552          * We want to reserve the absolute worst case amount of items.  So if
9553          * both inodes are subvols and we need to unlink them then that would
9554          * require 4 item modifications, but if they are both normal inodes it
9555          * would require 5 item modifications, so we'll assume their normal
9556          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9557          * should cover the worst case number of items we'll modify.
9558          */
9559         trans = btrfs_start_transaction(root, 12);
9560         if (IS_ERR(trans)) {
9561                 ret = PTR_ERR(trans);
9562                 goto out_notrans;
9563         }
9564
9565         /*
9566          * We need to find a free sequence number both in the source and
9567          * in the destination directory for the exchange.
9568          */
9569         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9570         if (ret)
9571                 goto out_fail;
9572         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9573         if (ret)
9574                 goto out_fail;
9575
9576         BTRFS_I(old_inode)->dir_index = 0ULL;
9577         BTRFS_I(new_inode)->dir_index = 0ULL;
9578
9579         /* Reference for the source. */
9580         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9581                 /* force full log commit if subvolume involved. */
9582                 btrfs_set_log_full_commit(fs_info, trans);
9583         } else {
9584                 btrfs_pin_log_trans(root);
9585                 root_log_pinned = true;
9586                 ret = btrfs_insert_inode_ref(trans, dest,
9587                                              new_dentry->d_name.name,
9588                                              new_dentry->d_name.len,
9589                                              old_ino,
9590                                              btrfs_ino(BTRFS_I(new_dir)),
9591                                              old_idx);
9592                 if (ret)
9593                         goto out_fail;
9594         }
9595
9596         /* And now for the dest. */
9597         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9598                 /* force full log commit if subvolume involved. */
9599                 btrfs_set_log_full_commit(fs_info, trans);
9600         } else {
9601                 btrfs_pin_log_trans(dest);
9602                 dest_log_pinned = true;
9603                 ret = btrfs_insert_inode_ref(trans, root,
9604                                              old_dentry->d_name.name,
9605                                              old_dentry->d_name.len,
9606                                              new_ino,
9607                                              btrfs_ino(BTRFS_I(old_dir)),
9608                                              new_idx);
9609                 if (ret)
9610                         goto out_fail;
9611         }
9612
9613         /* Update inode version and ctime/mtime. */
9614         inode_inc_iversion(old_dir);
9615         inode_inc_iversion(new_dir);
9616         inode_inc_iversion(old_inode);
9617         inode_inc_iversion(new_inode);
9618         old_dir->i_ctime = old_dir->i_mtime = ctime;
9619         new_dir->i_ctime = new_dir->i_mtime = ctime;
9620         old_inode->i_ctime = ctime;
9621         new_inode->i_ctime = ctime;
9622
9623         if (old_dentry->d_parent != new_dentry->d_parent) {
9624                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9625                                 BTRFS_I(old_inode), 1);
9626                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9627                                 BTRFS_I(new_inode), 1);
9628         }
9629
9630         /* src is a subvolume */
9631         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9632                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9633                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9634                                           root_objectid,
9635                                           old_dentry->d_name.name,
9636                                           old_dentry->d_name.len);
9637         } else { /* src is an inode */
9638                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9639                                            BTRFS_I(old_dentry->d_inode),
9640                                            old_dentry->d_name.name,
9641                                            old_dentry->d_name.len);
9642                 if (!ret)
9643                         ret = btrfs_update_inode(trans, root, old_inode);
9644         }
9645         if (ret) {
9646                 btrfs_abort_transaction(trans, ret);
9647                 goto out_fail;
9648         }
9649
9650         /* dest is a subvolume */
9651         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9652                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9653                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9654                                           root_objectid,
9655                                           new_dentry->d_name.name,
9656                                           new_dentry->d_name.len);
9657         } else { /* dest is an inode */
9658                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9659                                            BTRFS_I(new_dentry->d_inode),
9660                                            new_dentry->d_name.name,
9661                                            new_dentry->d_name.len);
9662                 if (!ret)
9663                         ret = btrfs_update_inode(trans, dest, new_inode);
9664         }
9665         if (ret) {
9666                 btrfs_abort_transaction(trans, ret);
9667                 goto out_fail;
9668         }
9669
9670         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9671                              new_dentry->d_name.name,
9672                              new_dentry->d_name.len, 0, old_idx);
9673         if (ret) {
9674                 btrfs_abort_transaction(trans, ret);
9675                 goto out_fail;
9676         }
9677
9678         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9679                              old_dentry->d_name.name,
9680                              old_dentry->d_name.len, 0, new_idx);
9681         if (ret) {
9682                 btrfs_abort_transaction(trans, ret);
9683                 goto out_fail;
9684         }
9685
9686         if (old_inode->i_nlink == 1)
9687                 BTRFS_I(old_inode)->dir_index = old_idx;
9688         if (new_inode->i_nlink == 1)
9689                 BTRFS_I(new_inode)->dir_index = new_idx;
9690
9691         if (root_log_pinned) {
9692                 parent = new_dentry->d_parent;
9693                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9694                                 parent);
9695                 btrfs_end_log_trans(root);
9696                 root_log_pinned = false;
9697         }
9698         if (dest_log_pinned) {
9699                 parent = old_dentry->d_parent;
9700                 btrfs_log_new_name(trans, BTRFS_I(new_inode), BTRFS_I(new_dir),
9701                                 parent);
9702                 btrfs_end_log_trans(dest);
9703                 dest_log_pinned = false;
9704         }
9705 out_fail:
9706         /*
9707          * If we have pinned a log and an error happened, we unpin tasks
9708          * trying to sync the log and force them to fallback to a transaction
9709          * commit if the log currently contains any of the inodes involved in
9710          * this rename operation (to ensure we do not persist a log with an
9711          * inconsistent state for any of these inodes or leading to any
9712          * inconsistencies when replayed). If the transaction was aborted, the
9713          * abortion reason is propagated to userspace when attempting to commit
9714          * the transaction. If the log does not contain any of these inodes, we
9715          * allow the tasks to sync it.
9716          */
9717         if (ret && (root_log_pinned || dest_log_pinned)) {
9718                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9719                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9720                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9721                     (new_inode &&
9722                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9723                         btrfs_set_log_full_commit(fs_info, trans);
9724
9725                 if (root_log_pinned) {
9726                         btrfs_end_log_trans(root);
9727                         root_log_pinned = false;
9728                 }
9729                 if (dest_log_pinned) {
9730                         btrfs_end_log_trans(dest);
9731                         dest_log_pinned = false;
9732                 }
9733         }
9734         ret = btrfs_end_transaction(trans);
9735 out_notrans:
9736         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9737                 up_read(&fs_info->subvol_sem);
9738         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9739                 up_read(&fs_info->subvol_sem);
9740
9741         return ret;
9742 }
9743
9744 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9745                                      struct btrfs_root *root,
9746                                      struct inode *dir,
9747                                      struct dentry *dentry)
9748 {
9749         int ret;
9750         struct inode *inode;
9751         u64 objectid;
9752         u64 index;
9753
9754         ret = btrfs_find_free_ino(root, &objectid);
9755         if (ret)
9756                 return ret;
9757
9758         inode = btrfs_new_inode(trans, root, dir,
9759                                 dentry->d_name.name,
9760                                 dentry->d_name.len,
9761                                 btrfs_ino(BTRFS_I(dir)),
9762                                 objectid,
9763                                 S_IFCHR | WHITEOUT_MODE,
9764                                 &index);
9765
9766         if (IS_ERR(inode)) {
9767                 ret = PTR_ERR(inode);
9768                 return ret;
9769         }
9770
9771         inode->i_op = &btrfs_special_inode_operations;
9772         init_special_inode(inode, inode->i_mode,
9773                 WHITEOUT_DEV);
9774
9775         ret = btrfs_init_inode_security(trans, inode, dir,
9776                                 &dentry->d_name);
9777         if (ret)
9778                 goto out;
9779
9780         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9781                                 BTRFS_I(inode), 0, index);
9782         if (ret)
9783                 goto out;
9784
9785         ret = btrfs_update_inode(trans, root, inode);
9786 out:
9787         unlock_new_inode(inode);
9788         if (ret)
9789                 inode_dec_link_count(inode);
9790         iput(inode);
9791
9792         return ret;
9793 }
9794
9795 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9796                            struct inode *new_dir, struct dentry *new_dentry,
9797                            unsigned int flags)
9798 {
9799         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9800         struct btrfs_trans_handle *trans;
9801         unsigned int trans_num_items;
9802         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9803         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9804         struct inode *new_inode = d_inode(new_dentry);
9805         struct inode *old_inode = d_inode(old_dentry);
9806         u64 index = 0;
9807         u64 root_objectid;
9808         int ret;
9809         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9810         bool log_pinned = false;
9811
9812         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9813                 return -EPERM;
9814
9815         /* we only allow rename subvolume link between subvolumes */
9816         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9817                 return -EXDEV;
9818
9819         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9820             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9821                 return -ENOTEMPTY;
9822
9823         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9824             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9825                 return -ENOTEMPTY;
9826
9827
9828         /* check for collisions, even if the  name isn't there */
9829         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9830                              new_dentry->d_name.name,
9831                              new_dentry->d_name.len);
9832
9833         if (ret) {
9834                 if (ret == -EEXIST) {
9835                         /* we shouldn't get
9836                          * eexist without a new_inode */
9837                         if (WARN_ON(!new_inode)) {
9838                                 return ret;
9839                         }
9840                 } else {
9841                         /* maybe -EOVERFLOW */
9842                         return ret;
9843                 }
9844         }
9845         ret = 0;
9846
9847         /*
9848          * we're using rename to replace one file with another.  Start IO on it
9849          * now so  we don't add too much work to the end of the transaction
9850          */
9851         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9852                 filemap_flush(old_inode->i_mapping);
9853
9854         /* close the racy window with snapshot create/destroy ioctl */
9855         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9856                 down_read(&fs_info->subvol_sem);
9857         /*
9858          * We want to reserve the absolute worst case amount of items.  So if
9859          * both inodes are subvols and we need to unlink them then that would
9860          * require 4 item modifications, but if they are both normal inodes it
9861          * would require 5 item modifications, so we'll assume they are normal
9862          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9863          * should cover the worst case number of items we'll modify.
9864          * If our rename has the whiteout flag, we need more 5 units for the
9865          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9866          * when selinux is enabled).
9867          */
9868         trans_num_items = 11;
9869         if (flags & RENAME_WHITEOUT)
9870                 trans_num_items += 5;
9871         trans = btrfs_start_transaction(root, trans_num_items);
9872         if (IS_ERR(trans)) {
9873                 ret = PTR_ERR(trans);
9874                 goto out_notrans;
9875         }
9876
9877         if (dest != root)
9878                 btrfs_record_root_in_trans(trans, dest);
9879
9880         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9881         if (ret)
9882                 goto out_fail;
9883
9884         BTRFS_I(old_inode)->dir_index = 0ULL;
9885         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9886                 /* force full log commit if subvolume involved. */
9887                 btrfs_set_log_full_commit(fs_info, trans);
9888         } else {
9889                 btrfs_pin_log_trans(root);
9890                 log_pinned = true;
9891                 ret = btrfs_insert_inode_ref(trans, dest,
9892                                              new_dentry->d_name.name,
9893                                              new_dentry->d_name.len,
9894                                              old_ino,
9895                                              btrfs_ino(BTRFS_I(new_dir)), index);
9896                 if (ret)
9897                         goto out_fail;
9898         }
9899
9900         inode_inc_iversion(old_dir);
9901         inode_inc_iversion(new_dir);
9902         inode_inc_iversion(old_inode);
9903         old_dir->i_ctime = old_dir->i_mtime =
9904         new_dir->i_ctime = new_dir->i_mtime =
9905         old_inode->i_ctime = current_time(old_dir);
9906
9907         if (old_dentry->d_parent != new_dentry->d_parent)
9908                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9909                                 BTRFS_I(old_inode), 1);
9910
9911         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9912                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9913                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9914                                         old_dentry->d_name.name,
9915                                         old_dentry->d_name.len);
9916         } else {
9917                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9918                                         BTRFS_I(d_inode(old_dentry)),
9919                                         old_dentry->d_name.name,
9920                                         old_dentry->d_name.len);
9921                 if (!ret)
9922                         ret = btrfs_update_inode(trans, root, old_inode);
9923         }
9924         if (ret) {
9925                 btrfs_abort_transaction(trans, ret);
9926                 goto out_fail;
9927         }
9928
9929         if (new_inode) {
9930                 inode_inc_iversion(new_inode);
9931                 new_inode->i_ctime = current_time(new_inode);
9932                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
9933                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9934                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9935                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9936                                                 root_objectid,
9937                                                 new_dentry->d_name.name,
9938                                                 new_dentry->d_name.len);
9939                         BUG_ON(new_inode->i_nlink == 0);
9940                 } else {
9941                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9942                                                  BTRFS_I(d_inode(new_dentry)),
9943                                                  new_dentry->d_name.name,
9944                                                  new_dentry->d_name.len);
9945                 }
9946                 if (!ret && new_inode->i_nlink == 0)
9947                         ret = btrfs_orphan_add(trans,
9948                                         BTRFS_I(d_inode(new_dentry)));
9949                 if (ret) {
9950                         btrfs_abort_transaction(trans, ret);
9951                         goto out_fail;
9952                 }
9953         }
9954
9955         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9956                              new_dentry->d_name.name,
9957                              new_dentry->d_name.len, 0, index);
9958         if (ret) {
9959                 btrfs_abort_transaction(trans, ret);
9960                 goto out_fail;
9961         }
9962
9963         if (old_inode->i_nlink == 1)
9964                 BTRFS_I(old_inode)->dir_index = index;
9965
9966         if (log_pinned) {
9967                 struct dentry *parent = new_dentry->d_parent;
9968
9969                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9970                                 parent);
9971                 btrfs_end_log_trans(root);
9972                 log_pinned = false;
9973         }
9974
9975         if (flags & RENAME_WHITEOUT) {
9976                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9977                                                 old_dentry);
9978
9979                 if (ret) {
9980                         btrfs_abort_transaction(trans, ret);
9981                         goto out_fail;
9982                 }
9983         }
9984 out_fail:
9985         /*
9986          * If we have pinned the log and an error happened, we unpin tasks
9987          * trying to sync the log and force them to fallback to a transaction
9988          * commit if the log currently contains any of the inodes involved in
9989          * this rename operation (to ensure we do not persist a log with an
9990          * inconsistent state for any of these inodes or leading to any
9991          * inconsistencies when replayed). If the transaction was aborted, the
9992          * abortion reason is propagated to userspace when attempting to commit
9993          * the transaction. If the log does not contain any of these inodes, we
9994          * allow the tasks to sync it.
9995          */
9996         if (ret && log_pinned) {
9997                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9998                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9999                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
10000                     (new_inode &&
10001                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
10002                         btrfs_set_log_full_commit(fs_info, trans);
10003
10004                 btrfs_end_log_trans(root);
10005                 log_pinned = false;
10006         }
10007         btrfs_end_transaction(trans);
10008 out_notrans:
10009         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
10010                 up_read(&fs_info->subvol_sem);
10011
10012         return ret;
10013 }
10014
10015 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
10016                          struct inode *new_dir, struct dentry *new_dentry,
10017                          unsigned int flags)
10018 {
10019         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
10020                 return -EINVAL;
10021
10022         if (flags & RENAME_EXCHANGE)
10023                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
10024                                           new_dentry);
10025
10026         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
10027 }
10028
10029 static void btrfs_run_delalloc_work(struct btrfs_work *work)
10030 {
10031         struct btrfs_delalloc_work *delalloc_work;
10032         struct inode *inode;
10033
10034         delalloc_work = container_of(work, struct btrfs_delalloc_work,
10035                                      work);
10036         inode = delalloc_work->inode;
10037         filemap_flush(inode->i_mapping);
10038         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
10039                                 &BTRFS_I(inode)->runtime_flags))
10040                 filemap_flush(inode->i_mapping);
10041
10042         if (delalloc_work->delay_iput)
10043                 btrfs_add_delayed_iput(inode);
10044         else
10045                 iput(inode);
10046         complete(&delalloc_work->completion);
10047 }
10048
10049 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
10050                                                     int delay_iput)
10051 {
10052         struct btrfs_delalloc_work *work;
10053
10054         work = kmalloc(sizeof(*work), GFP_NOFS);
10055         if (!work)
10056                 return NULL;
10057
10058         init_completion(&work->completion);
10059         INIT_LIST_HEAD(&work->list);
10060         work->inode = inode;
10061         work->delay_iput = delay_iput;
10062         WARN_ON_ONCE(!inode);
10063         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
10064                         btrfs_run_delalloc_work, NULL, NULL);
10065
10066         return work;
10067 }
10068
10069 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
10070 {
10071         wait_for_completion(&work->completion);
10072         kfree(work);
10073 }
10074
10075 /*
10076  * some fairly slow code that needs optimization. This walks the list
10077  * of all the inodes with pending delalloc and forces them to disk.
10078  */
10079 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
10080                                    int nr)
10081 {
10082         struct btrfs_inode *binode;
10083         struct inode *inode;
10084         struct btrfs_delalloc_work *work, *next;
10085         struct list_head works;
10086         struct list_head splice;
10087         int ret = 0;
10088
10089         INIT_LIST_HEAD(&works);
10090         INIT_LIST_HEAD(&splice);
10091
10092         mutex_lock(&root->delalloc_mutex);
10093         spin_lock(&root->delalloc_lock);
10094         list_splice_init(&root->delalloc_inodes, &splice);
10095         while (!list_empty(&splice)) {
10096                 binode = list_entry(splice.next, struct btrfs_inode,
10097                                     delalloc_inodes);
10098
10099                 list_move_tail(&binode->delalloc_inodes,
10100                                &root->delalloc_inodes);
10101                 inode = igrab(&binode->vfs_inode);
10102                 if (!inode) {
10103                         cond_resched_lock(&root->delalloc_lock);
10104                         continue;
10105                 }
10106                 spin_unlock(&root->delalloc_lock);
10107
10108                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10109                 if (!work) {
10110                         if (delay_iput)
10111                                 btrfs_add_delayed_iput(inode);
10112                         else
10113                                 iput(inode);
10114                         ret = -ENOMEM;
10115                         goto out;
10116                 }
10117                 list_add_tail(&work->list, &works);
10118                 btrfs_queue_work(root->fs_info->flush_workers,
10119                                  &work->work);
10120                 ret++;
10121                 if (nr != -1 && ret >= nr)
10122                         goto out;
10123                 cond_resched();
10124                 spin_lock(&root->delalloc_lock);
10125         }
10126         spin_unlock(&root->delalloc_lock);
10127
10128 out:
10129         list_for_each_entry_safe(work, next, &works, list) {
10130                 list_del_init(&work->list);
10131                 btrfs_wait_and_free_delalloc_work(work);
10132         }
10133
10134         if (!list_empty_careful(&splice)) {
10135                 spin_lock(&root->delalloc_lock);
10136                 list_splice_tail(&splice, &root->delalloc_inodes);
10137                 spin_unlock(&root->delalloc_lock);
10138         }
10139         mutex_unlock(&root->delalloc_mutex);
10140         return ret;
10141 }
10142
10143 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10144 {
10145         struct btrfs_fs_info *fs_info = root->fs_info;
10146         int ret;
10147
10148         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10149                 return -EROFS;
10150
10151         ret = __start_delalloc_inodes(root, delay_iput, -1);
10152         if (ret > 0)
10153                 ret = 0;
10154         return ret;
10155 }
10156
10157 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10158                                int nr)
10159 {
10160         struct btrfs_root *root;
10161         struct list_head splice;
10162         int ret;
10163
10164         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10165                 return -EROFS;
10166
10167         INIT_LIST_HEAD(&splice);
10168
10169         mutex_lock(&fs_info->delalloc_root_mutex);
10170         spin_lock(&fs_info->delalloc_root_lock);
10171         list_splice_init(&fs_info->delalloc_roots, &splice);
10172         while (!list_empty(&splice) && nr) {
10173                 root = list_first_entry(&splice, struct btrfs_root,
10174                                         delalloc_root);
10175                 root = btrfs_grab_fs_root(root);
10176                 BUG_ON(!root);
10177                 list_move_tail(&root->delalloc_root,
10178                                &fs_info->delalloc_roots);
10179                 spin_unlock(&fs_info->delalloc_root_lock);
10180
10181                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10182                 btrfs_put_fs_root(root);
10183                 if (ret < 0)
10184                         goto out;
10185
10186                 if (nr != -1) {
10187                         nr -= ret;
10188                         WARN_ON(nr < 0);
10189                 }
10190                 spin_lock(&fs_info->delalloc_root_lock);
10191         }
10192         spin_unlock(&fs_info->delalloc_root_lock);
10193
10194         ret = 0;
10195 out:
10196         if (!list_empty_careful(&splice)) {
10197                 spin_lock(&fs_info->delalloc_root_lock);
10198                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10199                 spin_unlock(&fs_info->delalloc_root_lock);
10200         }
10201         mutex_unlock(&fs_info->delalloc_root_mutex);
10202         return ret;
10203 }
10204
10205 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10206                          const char *symname)
10207 {
10208         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10209         struct btrfs_trans_handle *trans;
10210         struct btrfs_root *root = BTRFS_I(dir)->root;
10211         struct btrfs_path *path;
10212         struct btrfs_key key;
10213         struct inode *inode = NULL;
10214         int err;
10215         int drop_inode = 0;
10216         u64 objectid;
10217         u64 index = 0;
10218         int name_len;
10219         int datasize;
10220         unsigned long ptr;
10221         struct btrfs_file_extent_item *ei;
10222         struct extent_buffer *leaf;
10223
10224         name_len = strlen(symname);
10225         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10226                 return -ENAMETOOLONG;
10227
10228         /*
10229          * 2 items for inode item and ref
10230          * 2 items for dir items
10231          * 1 item for updating parent inode item
10232          * 1 item for the inline extent item
10233          * 1 item for xattr if selinux is on
10234          */
10235         trans = btrfs_start_transaction(root, 7);
10236         if (IS_ERR(trans))
10237                 return PTR_ERR(trans);
10238
10239         err = btrfs_find_free_ino(root, &objectid);
10240         if (err)
10241                 goto out_unlock;
10242
10243         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10244                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10245                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10246         if (IS_ERR(inode)) {
10247                 err = PTR_ERR(inode);
10248                 goto out_unlock;
10249         }
10250
10251         /*
10252         * If the active LSM wants to access the inode during
10253         * d_instantiate it needs these. Smack checks to see
10254         * if the filesystem supports xattrs by looking at the
10255         * ops vector.
10256         */
10257         inode->i_fop = &btrfs_file_operations;
10258         inode->i_op = &btrfs_file_inode_operations;
10259         inode->i_mapping->a_ops = &btrfs_aops;
10260         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10261
10262         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10263         if (err)
10264                 goto out_unlock_inode;
10265
10266         path = btrfs_alloc_path();
10267         if (!path) {
10268                 err = -ENOMEM;
10269                 goto out_unlock_inode;
10270         }
10271         key.objectid = btrfs_ino(BTRFS_I(inode));
10272         key.offset = 0;
10273         key.type = BTRFS_EXTENT_DATA_KEY;
10274         datasize = btrfs_file_extent_calc_inline_size(name_len);
10275         err = btrfs_insert_empty_item(trans, root, path, &key,
10276                                       datasize);
10277         if (err) {
10278                 btrfs_free_path(path);
10279                 goto out_unlock_inode;
10280         }
10281         leaf = path->nodes[0];
10282         ei = btrfs_item_ptr(leaf, path->slots[0],
10283                             struct btrfs_file_extent_item);
10284         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10285         btrfs_set_file_extent_type(leaf, ei,
10286                                    BTRFS_FILE_EXTENT_INLINE);
10287         btrfs_set_file_extent_encryption(leaf, ei, 0);
10288         btrfs_set_file_extent_compression(leaf, ei, 0);
10289         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10290         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10291
10292         ptr = btrfs_file_extent_inline_start(ei);
10293         write_extent_buffer(leaf, symname, ptr, name_len);
10294         btrfs_mark_buffer_dirty(leaf);
10295         btrfs_free_path(path);
10296
10297         inode->i_op = &btrfs_symlink_inode_operations;
10298         inode_nohighmem(inode);
10299         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10300         inode_set_bytes(inode, name_len);
10301         btrfs_i_size_write(BTRFS_I(inode), name_len);
10302         err = btrfs_update_inode(trans, root, inode);
10303         /*
10304          * Last step, add directory indexes for our symlink inode. This is the
10305          * last step to avoid extra cleanup of these indexes if an error happens
10306          * elsewhere above.
10307          */
10308         if (!err)
10309                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10310                                 BTRFS_I(inode), 0, index);
10311         if (err) {
10312                 drop_inode = 1;
10313                 goto out_unlock_inode;
10314         }
10315
10316         unlock_new_inode(inode);
10317         d_instantiate(dentry, inode);
10318
10319 out_unlock:
10320         btrfs_end_transaction(trans);
10321         if (drop_inode) {
10322                 inode_dec_link_count(inode);
10323                 iput(inode);
10324         }
10325         btrfs_btree_balance_dirty(fs_info);
10326         return err;
10327
10328 out_unlock_inode:
10329         drop_inode = 1;
10330         unlock_new_inode(inode);
10331         goto out_unlock;
10332 }
10333
10334 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10335                                        u64 start, u64 num_bytes, u64 min_size,
10336                                        loff_t actual_len, u64 *alloc_hint,
10337                                        struct btrfs_trans_handle *trans)
10338 {
10339         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10340         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10341         struct extent_map *em;
10342         struct btrfs_root *root = BTRFS_I(inode)->root;
10343         struct btrfs_key ins;
10344         u64 cur_offset = start;
10345         u64 i_size;
10346         u64 cur_bytes;
10347         u64 last_alloc = (u64)-1;
10348         int ret = 0;
10349         bool own_trans = true;
10350         u64 end = start + num_bytes - 1;
10351
10352         if (trans)
10353                 own_trans = false;
10354         while (num_bytes > 0) {
10355                 if (own_trans) {
10356                         trans = btrfs_start_transaction(root, 3);
10357                         if (IS_ERR(trans)) {
10358                                 ret = PTR_ERR(trans);
10359                                 break;
10360                         }
10361                 }
10362
10363                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10364                 cur_bytes = max(cur_bytes, min_size);
10365                 /*
10366                  * If we are severely fragmented we could end up with really
10367                  * small allocations, so if the allocator is returning small
10368                  * chunks lets make its job easier by only searching for those
10369                  * sized chunks.
10370                  */
10371                 cur_bytes = min(cur_bytes, last_alloc);
10372                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10373                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10374                 if (ret) {
10375                         if (own_trans)
10376                                 btrfs_end_transaction(trans);
10377                         break;
10378                 }
10379                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10380
10381                 last_alloc = ins.offset;
10382                 ret = insert_reserved_file_extent(trans, inode,
10383                                                   cur_offset, ins.objectid,
10384                                                   ins.offset, ins.offset,
10385                                                   ins.offset, 0, 0, 0,
10386                                                   BTRFS_FILE_EXTENT_PREALLOC);
10387                 if (ret) {
10388                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10389                                                    ins.offset, 0);
10390                         btrfs_abort_transaction(trans, ret);
10391                         if (own_trans)
10392                                 btrfs_end_transaction(trans);
10393                         break;
10394                 }
10395
10396                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10397                                         cur_offset + ins.offset -1, 0);
10398
10399                 em = alloc_extent_map();
10400                 if (!em) {
10401                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10402                                 &BTRFS_I(inode)->runtime_flags);
10403                         goto next;
10404                 }
10405
10406                 em->start = cur_offset;
10407                 em->orig_start = cur_offset;
10408                 em->len = ins.offset;
10409                 em->block_start = ins.objectid;
10410                 em->block_len = ins.offset;
10411                 em->orig_block_len = ins.offset;
10412                 em->ram_bytes = ins.offset;
10413                 em->bdev = fs_info->fs_devices->latest_bdev;
10414                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10415                 em->generation = trans->transid;
10416
10417                 while (1) {
10418                         write_lock(&em_tree->lock);
10419                         ret = add_extent_mapping(em_tree, em, 1);
10420                         write_unlock(&em_tree->lock);
10421                         if (ret != -EEXIST)
10422                                 break;
10423                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10424                                                 cur_offset + ins.offset - 1,
10425                                                 0);
10426                 }
10427                 free_extent_map(em);
10428 next:
10429                 num_bytes -= ins.offset;
10430                 cur_offset += ins.offset;
10431                 *alloc_hint = ins.objectid + ins.offset;
10432
10433                 inode_inc_iversion(inode);
10434                 inode->i_ctime = current_time(inode);
10435                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10436                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10437                     (actual_len > inode->i_size) &&
10438                     (cur_offset > inode->i_size)) {
10439                         if (cur_offset > actual_len)
10440                                 i_size = actual_len;
10441                         else
10442                                 i_size = cur_offset;
10443                         i_size_write(inode, i_size);
10444                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10445                 }
10446
10447                 ret = btrfs_update_inode(trans, root, inode);
10448
10449                 if (ret) {
10450                         btrfs_abort_transaction(trans, ret);
10451                         if (own_trans)
10452                                 btrfs_end_transaction(trans);
10453                         break;
10454                 }
10455
10456                 if (own_trans)
10457                         btrfs_end_transaction(trans);
10458         }
10459         if (cur_offset < end)
10460                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10461                         end - cur_offset + 1);
10462         return ret;
10463 }
10464
10465 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10466                               u64 start, u64 num_bytes, u64 min_size,
10467                               loff_t actual_len, u64 *alloc_hint)
10468 {
10469         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10470                                            min_size, actual_len, alloc_hint,
10471                                            NULL);
10472 }
10473
10474 int btrfs_prealloc_file_range_trans(struct inode *inode,
10475                                     struct btrfs_trans_handle *trans, int mode,
10476                                     u64 start, u64 num_bytes, u64 min_size,
10477                                     loff_t actual_len, u64 *alloc_hint)
10478 {
10479         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10480                                            min_size, actual_len, alloc_hint, trans);
10481 }
10482
10483 static int btrfs_set_page_dirty(struct page *page)
10484 {
10485         return __set_page_dirty_nobuffers(page);
10486 }
10487
10488 static int btrfs_permission(struct inode *inode, int mask)
10489 {
10490         struct btrfs_root *root = BTRFS_I(inode)->root;
10491         umode_t mode = inode->i_mode;
10492
10493         if (mask & MAY_WRITE &&
10494             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10495                 if (btrfs_root_readonly(root))
10496                         return -EROFS;
10497                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10498                         return -EACCES;
10499         }
10500         return generic_permission(inode, mask);
10501 }
10502
10503 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10504 {
10505         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10506         struct btrfs_trans_handle *trans;
10507         struct btrfs_root *root = BTRFS_I(dir)->root;
10508         struct inode *inode = NULL;
10509         u64 objectid;
10510         u64 index;
10511         int ret = 0;
10512
10513         /*
10514          * 5 units required for adding orphan entry
10515          */
10516         trans = btrfs_start_transaction(root, 5);
10517         if (IS_ERR(trans))
10518                 return PTR_ERR(trans);
10519
10520         ret = btrfs_find_free_ino(root, &objectid);
10521         if (ret)
10522                 goto out;
10523
10524         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10525                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10526         if (IS_ERR(inode)) {
10527                 ret = PTR_ERR(inode);
10528                 inode = NULL;
10529                 goto out;
10530         }
10531
10532         inode->i_fop = &btrfs_file_operations;
10533         inode->i_op = &btrfs_file_inode_operations;
10534
10535         inode->i_mapping->a_ops = &btrfs_aops;
10536         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10537
10538         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10539         if (ret)
10540                 goto out_inode;
10541
10542         ret = btrfs_update_inode(trans, root, inode);
10543         if (ret)
10544                 goto out_inode;
10545         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10546         if (ret)
10547                 goto out_inode;
10548
10549         /*
10550          * We set number of links to 0 in btrfs_new_inode(), and here we set
10551          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10552          * through:
10553          *
10554          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10555          */
10556         set_nlink(inode, 1);
10557         unlock_new_inode(inode);
10558         d_tmpfile(dentry, inode);
10559         mark_inode_dirty(inode);
10560
10561 out:
10562         btrfs_end_transaction(trans);
10563         if (ret)
10564                 iput(inode);
10565         btrfs_btree_balance_dirty(fs_info);
10566         return ret;
10567
10568 out_inode:
10569         unlock_new_inode(inode);
10570         goto out;
10571
10572 }
10573
10574 __attribute__((const))
10575 static int btrfs_readpage_io_failed_hook(struct page *page, int failed_mirror)
10576 {
10577         return -EAGAIN;
10578 }
10579
10580 static struct btrfs_fs_info *iotree_fs_info(void *private_data)
10581 {
10582         struct inode *inode = private_data;
10583         return btrfs_sb(inode->i_sb);
10584 }
10585
10586 static void btrfs_check_extent_io_range(void *private_data, const char *caller,
10587                                         u64 start, u64 end)
10588 {
10589         struct inode *inode = private_data;
10590         u64 isize;
10591
10592         isize = i_size_read(inode);
10593         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
10594                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
10595                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
10596                         caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
10597         }
10598 }
10599
10600 void btrfs_set_range_writeback(void *private_data, u64 start, u64 end)
10601 {
10602         struct inode *inode = private_data;
10603         unsigned long index = start >> PAGE_SHIFT;
10604         unsigned long end_index = end >> PAGE_SHIFT;
10605         struct page *page;
10606
10607         while (index <= end_index) {
10608                 page = find_get_page(inode->i_mapping, index);
10609                 ASSERT(page); /* Pages should be in the extent_io_tree */
10610                 set_page_writeback(page);
10611                 put_page(page);
10612                 index++;
10613         }
10614 }
10615
10616 static const struct inode_operations btrfs_dir_inode_operations = {
10617         .getattr        = btrfs_getattr,
10618         .lookup         = btrfs_lookup,
10619         .create         = btrfs_create,
10620         .unlink         = btrfs_unlink,
10621         .link           = btrfs_link,
10622         .mkdir          = btrfs_mkdir,
10623         .rmdir          = btrfs_rmdir,
10624         .rename         = btrfs_rename2,
10625         .symlink        = btrfs_symlink,
10626         .setattr        = btrfs_setattr,
10627         .mknod          = btrfs_mknod,
10628         .listxattr      = btrfs_listxattr,
10629         .permission     = btrfs_permission,
10630         .get_acl        = btrfs_get_acl,
10631         .set_acl        = btrfs_set_acl,
10632         .update_time    = btrfs_update_time,
10633         .tmpfile        = btrfs_tmpfile,
10634 };
10635 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10636         .lookup         = btrfs_lookup,
10637         .permission     = btrfs_permission,
10638         .update_time    = btrfs_update_time,
10639 };
10640
10641 static const struct file_operations btrfs_dir_file_operations = {
10642         .llseek         = generic_file_llseek,
10643         .read           = generic_read_dir,
10644         .iterate_shared = btrfs_real_readdir,
10645         .open           = btrfs_opendir,
10646         .unlocked_ioctl = btrfs_ioctl,
10647 #ifdef CONFIG_COMPAT
10648         .compat_ioctl   = btrfs_compat_ioctl,
10649 #endif
10650         .release        = btrfs_release_file,
10651         .fsync          = btrfs_sync_file,
10652 };
10653
10654 static const struct extent_io_ops btrfs_extent_io_ops = {
10655         /* mandatory callbacks */
10656         .submit_bio_hook = btrfs_submit_bio_hook,
10657         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10658         .merge_bio_hook = btrfs_merge_bio_hook,
10659         .readpage_io_failed_hook = btrfs_readpage_io_failed_hook,
10660         .tree_fs_info = iotree_fs_info,
10661         .set_range_writeback = btrfs_set_range_writeback,
10662
10663         /* optional callbacks */
10664         .fill_delalloc = run_delalloc_range,
10665         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10666         .writepage_start_hook = btrfs_writepage_start_hook,
10667         .set_bit_hook = btrfs_set_bit_hook,
10668         .clear_bit_hook = btrfs_clear_bit_hook,
10669         .merge_extent_hook = btrfs_merge_extent_hook,
10670         .split_extent_hook = btrfs_split_extent_hook,
10671         .check_extent_io_range = btrfs_check_extent_io_range,
10672 };
10673
10674 /*
10675  * btrfs doesn't support the bmap operation because swapfiles
10676  * use bmap to make a mapping of extents in the file.  They assume
10677  * these extents won't change over the life of the file and they
10678  * use the bmap result to do IO directly to the drive.
10679  *
10680  * the btrfs bmap call would return logical addresses that aren't
10681  * suitable for IO and they also will change frequently as COW
10682  * operations happen.  So, swapfile + btrfs == corruption.
10683  *
10684  * For now we're avoiding this by dropping bmap.
10685  */
10686 static const struct address_space_operations btrfs_aops = {
10687         .readpage       = btrfs_readpage,
10688         .writepage      = btrfs_writepage,
10689         .writepages     = btrfs_writepages,
10690         .readpages      = btrfs_readpages,
10691         .direct_IO      = btrfs_direct_IO,
10692         .invalidatepage = btrfs_invalidatepage,
10693         .releasepage    = btrfs_releasepage,
10694         .set_page_dirty = btrfs_set_page_dirty,
10695         .error_remove_page = generic_error_remove_page,
10696 };
10697
10698 static const struct address_space_operations btrfs_symlink_aops = {
10699         .readpage       = btrfs_readpage,
10700         .writepage      = btrfs_writepage,
10701         .invalidatepage = btrfs_invalidatepage,
10702         .releasepage    = btrfs_releasepage,
10703 };
10704
10705 static const struct inode_operations btrfs_file_inode_operations = {
10706         .getattr        = btrfs_getattr,
10707         .setattr        = btrfs_setattr,
10708         .listxattr      = btrfs_listxattr,
10709         .permission     = btrfs_permission,
10710         .fiemap         = btrfs_fiemap,
10711         .get_acl        = btrfs_get_acl,
10712         .set_acl        = btrfs_set_acl,
10713         .update_time    = btrfs_update_time,
10714 };
10715 static const struct inode_operations btrfs_special_inode_operations = {
10716         .getattr        = btrfs_getattr,
10717         .setattr        = btrfs_setattr,
10718         .permission     = btrfs_permission,
10719         .listxattr      = btrfs_listxattr,
10720         .get_acl        = btrfs_get_acl,
10721         .set_acl        = btrfs_set_acl,
10722         .update_time    = btrfs_update_time,
10723 };
10724 static const struct inode_operations btrfs_symlink_inode_operations = {
10725         .get_link       = page_get_link,
10726         .getattr        = btrfs_getattr,
10727         .setattr        = btrfs_setattr,
10728         .permission     = btrfs_permission,
10729         .listxattr      = btrfs_listxattr,
10730         .update_time    = btrfs_update_time,
10731 };
10732
10733 const struct dentry_operations btrfs_dentry_operations = {
10734         .d_delete       = btrfs_dentry_delete,
10735         .d_release      = btrfs_dentry_release,
10736 };