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