fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/compat.h>
  34 #include <linux/bit_spinlock.h>
  35 #include <linux/xattr.h>
  36 #include <linux/posix_acl.h>
  37 #include <linux/falloc.h>
  38 #include <linux/slab.h>
  39 #include <linux/ratelimit.h>
  40 #include <linux/mount.h>
  41 #include <linux/btrfs.h>
  42 #include <linux/blkdev.h>
  43 #include <linux/posix_acl_xattr.h>
  44 #include <linux/uio.h>
  45 #include <linux/magic.h>
  46 #include "ctree.h"
  47 #include "disk-io.h"
  48 #include "transaction.h"
  49 #include "btrfs_inode.h"
  50 #include "print-tree.h"
  51 #include "ordered-data.h"
  52 #include "xattr.h"
  53 #include "tree-log.h"
  54 #include "volumes.h"
  55 #include "compression.h"
  56 #include "locking.h"
  57 #include "free-space-cache.h"
  58 #include "inode-map.h"
  59 #include "backref.h"
  60 #include "hash.h"
  61 #include "props.h"
  62 #include "qgroup.h"
  63 #include "dedupe.h"
  64
  65 struct btrfs_iget_args {
  66         struct btrfs_key *location;
  67         struct btrfs_root *root;
  68 };
  69
  70 struct btrfs_dio_data {
  71         u64 reserve;
  72         u64 unsubmitted_oe_range_start;
  73         u64 unsubmitted_oe_range_end;
  74         int overwrite;
  75 };
  76
  77 static const struct inode_operations btrfs_dir_inode_operations;
  78 static const struct inode_operations btrfs_symlink_inode_operations;
  79 static const struct inode_operations btrfs_dir_ro_inode_operations;
  80 static const struct inode_operations btrfs_special_inode_operations;
  81 static const struct inode_operations btrfs_file_inode_operations;
  82 static const struct address_space_operations btrfs_aops;
  83 static const struct address_space_operations btrfs_symlink_aops;
  84 static const struct file_operations btrfs_dir_file_operations;
  85 static const struct extent_io_ops btrfs_extent_io_ops;
  86
  87 static struct kmem_cache *btrfs_inode_cachep;
  88 struct kmem_cache *btrfs_trans_handle_cachep;
  89 struct kmem_cache *btrfs_path_cachep;
  90 struct kmem_cache *btrfs_free_space_cachep;
  91
  92 #define S_SHIFT 12
  93 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  94         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  95         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  96         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  97         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  98         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  99         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 100         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 101 };
 102
 103 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 104 static int btrfs_truncate(struct inode *inode);
 105 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 106 static noinline int cow_file_range(struct inode *inode,
 107                                    struct page *locked_page,
 108                                    u64 start, u64 end, u64 delalloc_end,
 109                                    int *page_started, unsigned long *nr_written,
 110                                    int unlock, struct btrfs_dedupe_hash *hash);
 111 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
 112                                        u64 orig_start, u64 block_start,
 113                                        u64 block_len, u64 orig_block_len,
 114                                        u64 ram_bytes, int compress_type,
 115                                        int type);
 116
 117 static void __endio_write_update_ordered(struct inode *inode,
 118                                          const u64 offset, const u64 bytes,
 119                                          const bool uptodate);
 120
 121 /*
 122  * Cleanup all submitted ordered extents in specified range to handle errors
 123  * from the fill_dellaloc() callback.
 124  *
 125  * NOTE: caller must ensure that when an error happens, it can not call
 126  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 127  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 128  * to be released, which we want to happen only when finishing the ordered
 129  * extent (btrfs_finish_ordered_io()). Also note that the caller of the
 130  * fill_delalloc() callback already does proper cleanup for the first page of
 131  * the range, that is, it invokes the callback writepage_end_io_hook() for the
 132  * range of the first page.
 133  */
 134 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 135                                                  const u64 offset,
 136                                                  const u64 bytes)
 137 {
 138         unsigned long index = offset >> PAGE_SHIFT;
 139         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 140         struct page *page;
 141
 142         while (index <= end_index) {
 143                 page = find_get_page(inode->i_mapping, index);
 144                 index++;
 145                 if (!page)
 146                         continue;
 147                 ClearPagePrivate2(page);
 148                 put_page(page);
 149         }
 150         return __endio_write_update_ordered(inode, offset + PAGE_SIZE,
 151                                             bytes - PAGE_SIZE, false);
 152 }
 153
 154 static int btrfs_dirty_inode(struct inode *inode);
 155
 156 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 157 void btrfs_test_inode_set_ops(struct inode *inode)
 158 {
 159         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 160 }
 161 #endif
 162
 163 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 164                                      struct inode *inode,  struct inode *dir,
 165                                      const struct qstr *qstr)
 166 {
 167         int err;
 168
 169         err = btrfs_init_acl(trans, inode, dir);
 170         if (!err)
 171                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 172         return err;
 173 }
 174
 175 /*
 176  * this does all the hard work for inserting an inline extent into
 177  * the btree.  The caller should have done a btrfs_drop_extents so that
 178  * no overlapping inline items exist in the btree
 179  */
 180 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 181                                 struct btrfs_path *path, int extent_inserted,
 182                                 struct btrfs_root *root, struct inode *inode,
 183                                 u64 start, size_t size, size_t compressed_size,
 184                                 int compress_type,
 185                                 struct page **compressed_pages)
 186 {
 187         struct extent_buffer *leaf;
 188         struct page *page = NULL;
 189         char *kaddr;
 190         unsigned long ptr;
 191         struct btrfs_file_extent_item *ei;
 192         int ret;
 193         size_t cur_size = size;
 194         unsigned long offset;
 195
 196         if (compressed_size && compressed_pages)
 197                 cur_size = compressed_size;
 198
 199         inode_add_bytes(inode, size);
 200
 201         if (!extent_inserted) {
 202                 struct btrfs_key key;
 203                 size_t datasize;
 204
 205                 key.objectid = btrfs_ino(BTRFS_I(inode));
 206                 key.offset = start;
 207                 key.type = BTRFS_EXTENT_DATA_KEY;
 208
 209                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 210                 path->leave_spinning = 1;
 211                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 212                                               datasize);
 213                 if (ret)
 214                         goto fail;
 215         }
 216         leaf = path->nodes[0];
 217         ei = btrfs_item_ptr(leaf, path->slots[0],
 218                             struct btrfs_file_extent_item);
 219         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 220         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 221         btrfs_set_file_extent_encryption(leaf, ei, 0);
 222         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 223         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 224         ptr = btrfs_file_extent_inline_start(ei);
 225
 226         if (compress_type != BTRFS_COMPRESS_NONE) {
 227                 struct page *cpage;
 228                 int i = 0;
 229                 while (compressed_size > 0) {
 230                         cpage = compressed_pages[i];
 231                         cur_size = min_t(unsigned long, compressed_size,
 232                                        PAGE_SIZE);
 233
 234                         kaddr = kmap_atomic(cpage);
 235                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 236                         kunmap_atomic(kaddr);
 237
 238                         i++;
 239                         ptr += cur_size;
 240                         compressed_size -= cur_size;
 241                 }
 242                 btrfs_set_file_extent_compression(leaf, ei,
 243                                                   compress_type);
 244         } else {
 245                 page = find_get_page(inode->i_mapping,
 246                                      start >> PAGE_SHIFT);
 247                 btrfs_set_file_extent_compression(leaf, ei, 0);
 248                 kaddr = kmap_atomic(page);
 249                 offset = start & (PAGE_SIZE - 1);
 250                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 251                 kunmap_atomic(kaddr);
 252                 put_page(page);
 253         }
 254         btrfs_mark_buffer_dirty(leaf);
 255         btrfs_release_path(path);
 256
 257         /*
 258          * we're an inline extent, so nobody can
 259          * extend the file past i_size without locking
 260          * a page we already have locked.
 261          *
 262          * We must do any isize and inode updates
 263          * before we unlock the pages.  Otherwise we
 264          * could end up racing with unlink.
 265          */
 266         BTRFS_I(inode)->disk_i_size = inode->i_size;
 267         ret = btrfs_update_inode(trans, root, inode);
 268
 269 fail:
 270         return ret;
 271 }
 272
 273
 274 /*
 275  * conditionally insert an inline extent into the file.  This
 276  * does the checks required to make sure the data is small enough
 277  * to fit as an inline extent.
 278  */
 279 static noinline int cow_file_range_inline(struct btrfs_root *root,
 280                                           struct inode *inode, u64 start,
 281                                           u64 end, size_t compressed_size,
 282                                           int compress_type,
 283                                           struct page **compressed_pages)
 284 {
 285         struct btrfs_fs_info *fs_info = root->fs_info;
 286         struct btrfs_trans_handle *trans;
 287         u64 isize = i_size_read(inode);
 288         u64 actual_end = min(end + 1, isize);
 289         u64 inline_len = actual_end - start;
 290         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 291         u64 data_len = inline_len;
 292         int ret;
 293         struct btrfs_path *path;
 294         int extent_inserted = 0;
 295         u32 extent_item_size;
 296
 297         if (compressed_size)
 298                 data_len = compressed_size;
 299
 300         if (start > 0 ||
 301             actual_end > fs_info->sectorsize ||
 302             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 303             (!compressed_size &&
 304             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 305             end + 1 < isize ||
 306             data_len > fs_info->max_inline) {
 307                 return 1;
 308         }
 309
 310         path = btrfs_alloc_path();
 311         if (!path)
 312                 return -ENOMEM;
 313
 314         trans = btrfs_join_transaction(root);
 315         if (IS_ERR(trans)) {
 316                 btrfs_free_path(path);
 317                 return PTR_ERR(trans);
 318         }
 319         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 320
 321         if (compressed_size && compressed_pages)
 322                 extent_item_size = btrfs_file_extent_calc_inline_size(
 323                    compressed_size);
 324         else
 325                 extent_item_size = btrfs_file_extent_calc_inline_size(
 326                     inline_len);
 327
 328         ret = __btrfs_drop_extents(trans, root, inode, path,
 329                                    start, aligned_end, NULL,
 330                                    1, 1, extent_item_size, &extent_inserted);
 331         if (ret) {
 332                 btrfs_abort_transaction(trans, ret);
 333                 goto out;
 334         }
 335
 336         if (isize > actual_end)
 337                 inline_len = min_t(u64, isize, actual_end);
 338         ret = insert_inline_extent(trans, path, extent_inserted,
 339                                    root, inode, start,
 340                                    inline_len, compressed_size,
 341                                    compress_type, compressed_pages);
 342         if (ret && ret != -ENOSPC) {
 343                 btrfs_abort_transaction(trans, ret);
 344                 goto out;
 345         } else if (ret == -ENOSPC) {
 346                 ret = 1;
 347                 goto out;
 348         }
 349
 350         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 351         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 352 out:
 353         /*
 354          * Don't forget to free the reserved space, as for inlined extent
 355          * it won't count as data extent, free them directly here.
 356          * And at reserve time, it's always aligned to page size, so
 357          * just free one page here.
 358          */
 359         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 360         btrfs_free_path(path);
 361         btrfs_end_transaction(trans);
 362         return ret;
 363 }
 364
 365 struct async_extent {
 366         u64 start;
 367         u64 ram_size;
 368         u64 compressed_size;
 369         struct page **pages;
 370         unsigned long nr_pages;
 371         int compress_type;
 372         struct list_head list;
 373 };
 374
 375 struct async_cow {
 376         struct inode *inode;
 377         struct btrfs_root *root;
 378         struct page *locked_page;
 379         u64 start;
 380         u64 end;
 381         unsigned int write_flags;
 382         struct list_head extents;
 383         struct btrfs_work work;
 384 };
 385
 386 static noinline int add_async_extent(struct async_cow *cow,
 387                                      u64 start, u64 ram_size,
 388                                      u64 compressed_size,
 389                                      struct page **pages,
 390                                      unsigned long nr_pages,
 391                                      int compress_type)
 392 {
 393         struct async_extent *async_extent;
 394
 395         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 396         BUG_ON(!async_extent); /* -ENOMEM */
 397         async_extent->start = start;
 398         async_extent->ram_size = ram_size;
 399         async_extent->compressed_size = compressed_size;
 400         async_extent->pages = pages;
 401         async_extent->nr_pages = nr_pages;
 402         async_extent->compress_type = compress_type;
 403         list_add_tail(&async_extent->list, &cow->extents);
 404         return 0;
 405 }
 406
 407 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 408 {
 409         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 410
 411         /* force compress */
 412         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 413                 return 1;
 414         /* defrag ioctl */
 415         if (BTRFS_I(inode)->defrag_compress)
 416                 return 1;
 417         /* bad compression ratios */
 418         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 419                 return 0;
 420         if (btrfs_test_opt(fs_info, COMPRESS) ||
 421             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 422             BTRFS_I(inode)->prop_compress)
 423                 return btrfs_compress_heuristic(inode, start, end);
 424         return 0;
 425 }
 426
 427 static inline void inode_should_defrag(struct btrfs_inode *inode,
 428                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 429 {
 430         /* If this is a small write inside eof, kick off a defrag */
 431         if (num_bytes < small_write &&
 432             (start > 0 || end + 1 < inode->disk_i_size))
 433                 btrfs_add_inode_defrag(NULL, inode);
 434 }
 435
 436 /*
 437  * we create compressed extents in two phases.  The first
 438  * phase compresses a range of pages that have already been
 439  * locked (both pages and state bits are locked).
 440  *
 441  * This is done inside an ordered work queue, and the compression
 442  * is spread across many cpus.  The actual IO submission is step
 443  * two, and the ordered work queue takes care of making sure that
 444  * happens in the same order things were put onto the queue by
 445  * writepages and friends.
 446  *
 447  * If this code finds it can't get good compression, it puts an
 448  * entry onto the work queue to write the uncompressed bytes.  This
 449  * makes sure that both compressed inodes and uncompressed inodes
 450  * are written in the same order that the flusher thread sent them
 451  * down.
 452  */
 453 static noinline void compress_file_range(struct inode *inode,
 454                                         struct page *locked_page,
 455                                         u64 start, u64 end,
 456                                         struct async_cow *async_cow,
 457                                         int *num_added)
 458 {
 459         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 460         struct btrfs_root *root = BTRFS_I(inode)->root;
 461         u64 blocksize = fs_info->sectorsize;
 462         u64 actual_end;
 463         u64 isize = i_size_read(inode);
 464         int ret = 0;
 465         struct page **pages = NULL;
 466         unsigned long nr_pages;
 467         unsigned long total_compressed = 0;
 468         unsigned long total_in = 0;
 469         int i;
 470         int will_compress;
 471         int compress_type = fs_info->compress_type;
 472         int redirty = 0;
 473
 474         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 475                         SZ_16K);
 476
 477         actual_end = min_t(u64, isize, end + 1);
 478 again:
 479         will_compress = 0;
 480         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 481         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 482         nr_pages = min_t(unsigned long, nr_pages,
 483                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 484
 485         /*
 486          * we don't want to send crud past the end of i_size through
 487          * compression, that's just a waste of CPU time.  So, if the
 488          * end of the file is before the start of our current
 489          * requested range of bytes, we bail out to the uncompressed
 490          * cleanup code that can deal with all of this.
 491          *
 492          * It isn't really the fastest way to fix things, but this is a
 493          * very uncommon corner.
 494          */
 495         if (actual_end <= start)
 496                 goto cleanup_and_bail_uncompressed;
 497
 498         total_compressed = actual_end - start;
 499
 500         /*
 501          * skip compression for a small file range(<=blocksize) that
 502          * isn't an inline extent, since it doesn't save disk space at all.
 503          */
 504         if (total_compressed <= blocksize &&
 505            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 506                 goto cleanup_and_bail_uncompressed;
 507
 508         total_compressed = min_t(unsigned long, total_compressed,
 509                         BTRFS_MAX_UNCOMPRESSED);
 510         total_in = 0;
 511         ret = 0;
 512
 513         /*
 514          * we do compression for mount -o compress and when the
 515          * inode has not been flagged as nocompress.  This flag can
 516          * change at any time if we discover bad compression ratios.
 517          */
 518         if (inode_need_compress(inode, start, end)) {
 519                 WARN_ON(pages);
 520                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 521                 if (!pages) {
 522                         /* just bail out to the uncompressed code */
 523                         goto cont;
 524                 }
 525
 526                 if (BTRFS_I(inode)->defrag_compress)
 527                         compress_type = BTRFS_I(inode)->defrag_compress;
 528                 else if (BTRFS_I(inode)->prop_compress)
 529                         compress_type = BTRFS_I(inode)->prop_compress;
 530
 531                 /*
 532                  * we need to call clear_page_dirty_for_io on each
 533                  * page in the range.  Otherwise applications with the file
 534                  * mmap'd can wander in and change the page contents while
 535                  * we are compressing them.
 536                  *
 537                  * If the compression fails for any reason, we set the pages
 538                  * dirty again later on.
 539                  */
 540                 extent_range_clear_dirty_for_io(inode, start, end);
 541                 redirty = 1;
 542
 543                 /* Compression level is applied here and only here */
 544                 ret = btrfs_compress_pages(
 545                         compress_type | (fs_info->compress_level << 4),
 546                                            inode->i_mapping, start,
 547                                            pages,
 548                                            &nr_pages,
 549                                            &total_in,
 550                                            &total_compressed);
 551
 552                 if (!ret) {
 553                         unsigned long offset = total_compressed &
 554                                 (PAGE_SIZE - 1);
 555                         struct page *page = pages[nr_pages - 1];
 556                         char *kaddr;
 557
 558                         /* zero the tail end of the last page, we might be
 559                          * sending it down to disk
 560                          */
 561                         if (offset) {
 562                                 kaddr = kmap_atomic(page);
 563                                 memset(kaddr + offset, 0,
 564                                        PAGE_SIZE - offset);
 565                                 kunmap_atomic(kaddr);
 566                         }
 567                         will_compress = 1;
 568                 }
 569         }
 570 cont:
 571         if (start == 0) {
 572                 /* lets try to make an inline extent */
 573                 if (ret || total_in < actual_end) {
 574                         /* we didn't compress the entire range, try
 575                          * to make an uncompressed inline extent.
 576                          */
 577                         ret = cow_file_range_inline(root, inode, start, end,
 578                                             0, BTRFS_COMPRESS_NONE, NULL);
 579                 } else {
 580                         /* try making a compressed inline extent */
 581                         ret = cow_file_range_inline(root, inode, start, end,
 582                                                     total_compressed,
 583                                                     compress_type, pages);
 584                 }
 585                 if (ret <= 0) {
 586                         unsigned long clear_flags = EXTENT_DELALLOC |
 587                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 588                                 EXTENT_DO_ACCOUNTING;
 589                         unsigned long page_error_op;
 590
 591                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 592
 593                         /*
 594                          * inline extent creation worked or returned error,
 595                          * we don't need to create any more async work items.
 596                          * Unlock and free up our temp pages.
 597                          *
 598                          * We use DO_ACCOUNTING here because we need the
 599                          * delalloc_release_metadata to be done _after_ we drop
 600                          * our outstanding extent for clearing delalloc for this
 601                          * range.
 602                          */
 603                         extent_clear_unlock_delalloc(inode, start, end, end,
 604                                                      NULL, clear_flags,
 605                                                      PAGE_UNLOCK |
 606                                                      PAGE_CLEAR_DIRTY |
 607                                                      PAGE_SET_WRITEBACK |
 608                                                      page_error_op |
 609                                                      PAGE_END_WRITEBACK);
 610                         goto free_pages_out;
 611                 }
 612         }
 613
 614         if (will_compress) {
 615                 /*
 616                  * we aren't doing an inline extent round the compressed size
 617                  * up to a block size boundary so the allocator does sane
 618                  * things
 619                  */
 620                 total_compressed = ALIGN(total_compressed, blocksize);
 621
 622                 /*
 623                  * one last check to make sure the compression is really a
 624                  * win, compare the page count read with the blocks on disk,
 625                  * compression must free at least one sector size
 626                  */
 627                 total_in = ALIGN(total_in, PAGE_SIZE);
 628                 if (total_compressed + blocksize <= total_in) {
 629                         *num_added += 1;
 630
 631                         /*
 632                          * The async work queues will take care of doing actual
 633                          * allocation on disk for these compressed pages, and
 634                          * will submit them to the elevator.
 635                          */
 636                         add_async_extent(async_cow, start, total_in,
 637                                         total_compressed, pages, nr_pages,
 638                                         compress_type);
 639
 640                         if (start + total_in < end) {
 641                                 start += total_in;
 642                                 pages = NULL;
 643                                 cond_resched();
 644                                 goto again;
 645                         }
 646                         return;
 647                 }
 648         }
 649         if (pages) {
 650                 /*
 651                  * the compression code ran but failed to make things smaller,
 652                  * free any pages it allocated and our page pointer array
 653                  */
 654                 for (i = 0; i < nr_pages; i++) {
 655                         WARN_ON(pages[i]->mapping);
 656                         put_page(pages[i]);
 657                 }
 658                 kfree(pages);
 659                 pages = NULL;
 660                 total_compressed = 0;
 661                 nr_pages = 0;
 662
 663                 /* flag the file so we don't compress in the future */
 664                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 665                     !(BTRFS_I(inode)->prop_compress)) {
 666                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 667                 }
 668         }
 669 cleanup_and_bail_uncompressed:
 670         /*
 671          * No compression, but we still need to write the pages in the file
 672          * we've been given so far.  redirty the locked page if it corresponds
 673          * to our extent and set things up for the async work queue to run
 674          * cow_file_range to do the normal delalloc dance.
 675          */
 676         if (page_offset(locked_page) >= start &&
 677             page_offset(locked_page) <= end)
 678                 __set_page_dirty_nobuffers(locked_page);
 679                 /* unlocked later on in the async handlers */
 680
 681         if (redirty)
 682                 extent_range_redirty_for_io(inode, start, end);
 683         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 684                          BTRFS_COMPRESS_NONE);
 685         *num_added += 1;
 686
 687         return;
 688
 689 free_pages_out:
 690         for (i = 0; i < nr_pages; i++) {
 691                 WARN_ON(pages[i]->mapping);
 692                 put_page(pages[i]);
 693         }
 694         kfree(pages);
 695 }
 696
 697 static void free_async_extent_pages(struct async_extent *async_extent)
 698 {
 699         int i;
 700
 701         if (!async_extent->pages)
 702                 return;
 703
 704         for (i = 0; i < async_extent->nr_pages; i++) {
 705                 WARN_ON(async_extent->pages[i]->mapping);
 706                 put_page(async_extent->pages[i]);
 707         }
 708         kfree(async_extent->pages);
 709         async_extent->nr_pages = 0;
 710         async_extent->pages = NULL;
 711 }
 712
 713 /*
 714  * phase two of compressed writeback.  This is the ordered portion
 715  * of the code, which only gets called in the order the work was
 716  * queued.  We walk all the async extents created by compress_file_range
 717  * and send them down to the disk.
 718  */
 719 static noinline void submit_compressed_extents(struct inode *inode,
 720                                               struct async_cow *async_cow)
 721 {
 722         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 723         struct async_extent *async_extent;
 724         u64 alloc_hint = 0;
 725         struct btrfs_key ins;
 726         struct extent_map *em;
 727         struct btrfs_root *root = BTRFS_I(inode)->root;
 728         struct extent_io_tree *io_tree;
 729         int ret = 0;
 730
 731 again:
 732         while (!list_empty(&async_cow->extents)) {
 733                 async_extent = list_entry(async_cow->extents.next,
 734                                           struct async_extent, list);
 735                 list_del(&async_extent->list);
 736
 737                 io_tree = &BTRFS_I(inode)->io_tree;
 738
 739 retry:
 740                 /* did the compression code fall back to uncompressed IO? */
 741                 if (!async_extent->pages) {
 742                         int page_started = 0;
 743                         unsigned long nr_written = 0;
 744
 745                         lock_extent(io_tree, async_extent->start,
 746                                          async_extent->start +
 747                                          async_extent->ram_size - 1);
 748
 749                         /* allocate blocks */
 750                         ret = cow_file_range(inode, async_cow->locked_page,
 751                                              async_extent->start,
 752                                              async_extent->start +
 753                                              async_extent->ram_size - 1,
 754                                              async_extent->start +
 755                                              async_extent->ram_size - 1,
 756                                              &page_started, &nr_written, 0,
 757                                              NULL);
 758
 759                         /* JDM XXX */
 760
 761                         /*
 762                          * if page_started, cow_file_range inserted an
 763                          * inline extent and took care of all the unlocking
 764                          * and IO for us.  Otherwise, we need to submit
 765                          * all those pages down to the drive.
 766                          */
 767                         if (!page_started && !ret)
 768                                 extent_write_locked_range(io_tree,
 769                                                   inode, async_extent->start,
 770                                                   async_extent->start +
 771                                                   async_extent->ram_size - 1,
 772                                                   btrfs_get_extent,
 773                                                   WB_SYNC_ALL);
 774                         else if (ret)
 775                                 unlock_page(async_cow->locked_page);
 776                         kfree(async_extent);
 777                         cond_resched();
 778                         continue;
 779                 }
 780
 781                 lock_extent(io_tree, async_extent->start,
 782                             async_extent->start + async_extent->ram_size - 1);
 783
 784                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 785                                            async_extent->compressed_size,
 786                                            async_extent->compressed_size,
 787                                            0, alloc_hint, &ins, 1, 1);
 788                 if (ret) {
 789                         free_async_extent_pages(async_extent);
 790
 791                         if (ret == -ENOSPC) {
 792                                 unlock_extent(io_tree, async_extent->start,
 793                                               async_extent->start +
 794                                               async_extent->ram_size - 1);
 795
 796                                 /*
 797                                  * we need to redirty the pages if we decide to
 798                                  * fallback to uncompressed IO, otherwise we
 799                                  * will not submit these pages down to lower
 800                                  * layers.
 801                                  */
 802                                 extent_range_redirty_for_io(inode,
 803                                                 async_extent->start,
 804                                                 async_extent->start +
 805                                                 async_extent->ram_size - 1);
 806
 807                                 goto retry;
 808                         }
 809                         goto out_free;
 810                 }
 811                 /*
 812                  * here we're doing allocation and writeback of the
 813                  * compressed pages
 814                  */
 815                 em = create_io_em(inode, async_extent->start,
 816                                   async_extent->ram_size, /* len */
 817                                   async_extent->start, /* orig_start */
 818                                   ins.objectid, /* block_start */
 819                                   ins.offset, /* block_len */
 820                                   ins.offset, /* orig_block_len */
 821                                   async_extent->ram_size, /* ram_bytes */
 822                                   async_extent->compress_type,
 823                                   BTRFS_ORDERED_COMPRESSED);
 824                 if (IS_ERR(em))
 825                         /* ret value is not necessary due to void function */
 826                         goto out_free_reserve;
 827                 free_extent_map(em);
 828
 829                 ret = btrfs_add_ordered_extent_compress(inode,
 830                                                 async_extent->start,
 831                                                 ins.objectid,
 832                                                 async_extent->ram_size,
 833                                                 ins.offset,
 834                                                 BTRFS_ORDERED_COMPRESSED,
 835                                                 async_extent->compress_type);
 836                 if (ret) {
 837                         btrfs_drop_extent_cache(BTRFS_I(inode),
 838                                                 async_extent->start,
 839                                                 async_extent->start +
 840                                                 async_extent->ram_size - 1, 0);
 841                         goto out_free_reserve;
 842                 }
 843                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 844
 845                 /*
 846                  * clear dirty, set writeback and unlock the pages.
 847                  */
 848                 extent_clear_unlock_delalloc(inode, async_extent->start,
 849                                 async_extent->start +
 850                                 async_extent->ram_size - 1,
 851                                 async_extent->start +
 852                                 async_extent->ram_size - 1,
 853                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 854                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 855                                 PAGE_SET_WRITEBACK);
 856                 if (btrfs_submit_compressed_write(inode,
 857                                     async_extent->start,
 858                                     async_extent->ram_size,
 859                                     ins.objectid,
 860                                     ins.offset, async_extent->pages,
 861                                     async_extent->nr_pages,
 862                                     async_cow->write_flags)) {
 863                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 864                         struct page *p = async_extent->pages[0];
 865                         const u64 start = async_extent->start;
 866                         const u64 end = start + async_extent->ram_size - 1;
 867
 868                         p->mapping = inode->i_mapping;
 869                         tree->ops->writepage_end_io_hook(p, start, end,
 870                                                          NULL, 0);
 871                         p->mapping = NULL;
 872                         extent_clear_unlock_delalloc(inode, start, end, end,
 873                                                      NULL, 0,
 874                                                      PAGE_END_WRITEBACK |
 875                                                      PAGE_SET_ERROR);
 876                         free_async_extent_pages(async_extent);
 877                 }
 878                 alloc_hint = ins.objectid + ins.offset;
 879                 kfree(async_extent);
 880                 cond_resched();
 881         }
 882         return;
 883 out_free_reserve:
 884         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 885         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 886 out_free:
 887         extent_clear_unlock_delalloc(inode, async_extent->start,
 888                                      async_extent->start +
 889                                      async_extent->ram_size - 1,
 890                                      async_extent->start +
 891                                      async_extent->ram_size - 1,
 892                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 893                                      EXTENT_DELALLOC_NEW |
 894                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 895                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 896                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 897                                      PAGE_SET_ERROR);
 898         free_async_extent_pages(async_extent);
 899         kfree(async_extent);
 900         goto again;
 901 }
 902
 903 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 904                                       u64 num_bytes)
 905 {
 906         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 907         struct extent_map *em;
 908         u64 alloc_hint = 0;
 909
 910         read_lock(&em_tree->lock);
 911         em = search_extent_mapping(em_tree, start, num_bytes);
 912         if (em) {
 913                 /*
 914                  * if block start isn't an actual block number then find the
 915                  * first block in this inode and use that as a hint.  If that
 916                  * block is also bogus then just don't worry about it.
 917                  */
 918                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 919                         free_extent_map(em);
 920                         em = search_extent_mapping(em_tree, 0, 0);
 921                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 922                                 alloc_hint = em->block_start;
 923                         if (em)
 924                                 free_extent_map(em);
 925                 } else {
 926                         alloc_hint = em->block_start;
 927                         free_extent_map(em);
 928                 }
 929         }
 930         read_unlock(&em_tree->lock);
 931
 932         return alloc_hint;
 933 }
 934
 935 /*
 936  * when extent_io.c finds a delayed allocation range in the file,
 937  * the call backs end up in this code.  The basic idea is to
 938  * allocate extents on disk for the range, and create ordered data structs
 939  * in ram to track those extents.
 940  *
 941  * locked_page is the page that writepage had locked already.  We use
 942  * it to make sure we don't do extra locks or unlocks.
 943  *
 944  * *page_started is set to one if we unlock locked_page and do everything
 945  * required to start IO on it.  It may be clean and already done with
 946  * IO when we return.
 947  */
 948 static noinline int cow_file_range(struct inode *inode,
 949                                    struct page *locked_page,
 950                                    u64 start, u64 end, u64 delalloc_end,
 951                                    int *page_started, unsigned long *nr_written,
 952                                    int unlock, struct btrfs_dedupe_hash *hash)
 953 {
 954         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 955         struct btrfs_root *root = BTRFS_I(inode)->root;
 956         u64 alloc_hint = 0;
 957         u64 num_bytes;
 958         unsigned long ram_size;
 959         u64 disk_num_bytes;
 960         u64 cur_alloc_size = 0;
 961         u64 blocksize = fs_info->sectorsize;
 962         struct btrfs_key ins;
 963         struct extent_map *em;
 964         unsigned clear_bits;
 965         unsigned long page_ops;
 966         bool extent_reserved = false;
 967         int ret = 0;
 968
 969         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 970                 WARN_ON_ONCE(1);
 971                 ret = -EINVAL;
 972                 goto out_unlock;
 973         }
 974
 975         num_bytes = ALIGN(end - start + 1, blocksize);
 976         num_bytes = max(blocksize,  num_bytes);
 977         disk_num_bytes = num_bytes;
 978
 979         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 980
 981         if (start == 0) {
 982                 /* lets try to make an inline extent */
 983                 ret = cow_file_range_inline(root, inode, start, end, 0,
 984                                         BTRFS_COMPRESS_NONE, NULL);
 985                 if (ret == 0) {
 986                         /*
 987                          * We use DO_ACCOUNTING here because we need the
 988                          * delalloc_release_metadata to be run _after_ we drop
 989                          * our outstanding extent for clearing delalloc for this
 990                          * range.
 991                          */
 992                         extent_clear_unlock_delalloc(inode, start, end,
 993                                      delalloc_end, NULL,
 994                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 995                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 996                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
 997                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 998                                      PAGE_END_WRITEBACK);
 999                         *nr_written = *nr_written +
1000                              (end - start + PAGE_SIZE) / PAGE_SIZE;
1001                         *page_started = 1;
1002                         goto out;
1003                 } else if (ret < 0) {
1004                         goto out_unlock;
1005                 }
1006         }
1007
1008         BUG_ON(disk_num_bytes >
1009                btrfs_super_total_bytes(fs_info->super_copy));
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 (disk_num_bytes > 0) {
1016                 cur_alloc_size = disk_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 (disk_num_bytes < cur_alloc_size)
1081                         disk_num_bytes = 0;
1082                 else
1083                         disk_num_bytes -= cur_alloc_size;
1084                 num_bytes -= cur_alloc_size;
1085                 alloc_hint = ins.objectid + ins.offset;
1086                 start += cur_alloc_size;
1087                 extent_reserved = false;
1088
1089                 /*
1090                  * btrfs_reloc_clone_csums() error, since start is increased
1091                  * extent_clear_unlock_delalloc() at out_unlock label won't
1092                  * free metadata of current ordered extent, we're OK to exit.
1093                  */
1094                 if (ret)
1095                         goto out_unlock;
1096         }
1097 out:
1098         return ret;
1099
1100 out_drop_extent_cache:
1101         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1102 out_reserve:
1103         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1104         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1105 out_unlock:
1106         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1107                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1108         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1109                 PAGE_END_WRITEBACK;
1110         /*
1111          * If we reserved an extent for our delalloc range (or a subrange) and
1112          * failed to create the respective ordered extent, then it means that
1113          * when we reserved the extent we decremented the extent's size from
1114          * the data space_info's bytes_may_use counter and incremented the
1115          * space_info's bytes_reserved counter by the same amount. We must make
1116          * sure extent_clear_unlock_delalloc() does not try to decrement again
1117          * the data space_info's bytes_may_use counter, therefore we do not pass
1118          * it the flag EXTENT_CLEAR_DATA_RESV.
1119          */
1120         if (extent_reserved) {
1121                 extent_clear_unlock_delalloc(inode, start,
1122                                              start + cur_alloc_size,
1123                                              start + cur_alloc_size,
1124                                              locked_page,
1125                                              clear_bits,
1126                                              page_ops);
1127                 start += cur_alloc_size;
1128                 if (start >= end)
1129                         goto out;
1130         }
1131         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1132                                      locked_page,
1133                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1134                                      page_ops);
1135         goto out;
1136 }
1137
1138 /*
1139  * work queue call back to started compression on a file and pages
1140  */
1141 static noinline void async_cow_start(struct btrfs_work *work)
1142 {
1143         struct async_cow *async_cow;
1144         int num_added = 0;
1145         async_cow = container_of(work, struct async_cow, work);
1146
1147         compress_file_range(async_cow->inode, async_cow->locked_page,
1148                             async_cow->start, async_cow->end, async_cow,
1149                             &num_added);
1150         if (num_added == 0) {
1151                 btrfs_add_delayed_iput(async_cow->inode);
1152                 async_cow->inode = NULL;
1153         }
1154 }
1155
1156 /*
1157  * work queue call back to submit previously compressed pages
1158  */
1159 static noinline void async_cow_submit(struct btrfs_work *work)
1160 {
1161         struct btrfs_fs_info *fs_info;
1162         struct async_cow *async_cow;
1163         struct btrfs_root *root;
1164         unsigned long nr_pages;
1165
1166         async_cow = container_of(work, struct async_cow, work);
1167
1168         root = async_cow->root;
1169         fs_info = root->fs_info;
1170         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1171                 PAGE_SHIFT;
1172
1173         /*
1174          * atomic_sub_return implies a barrier for waitqueue_active
1175          */
1176         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1177             5 * SZ_1M &&
1178             waitqueue_active(&fs_info->async_submit_wait))
1179                 wake_up(&fs_info->async_submit_wait);
1180
1181         if (async_cow->inode)
1182                 submit_compressed_extents(async_cow->inode, async_cow);
1183 }
1184
1185 static noinline void async_cow_free(struct btrfs_work *work)
1186 {
1187         struct async_cow *async_cow;
1188         async_cow = container_of(work, struct async_cow, work);
1189         if (async_cow->inode)
1190                 btrfs_add_delayed_iput(async_cow->inode);
1191         kfree(async_cow);
1192 }
1193
1194 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1195                                 u64 start, u64 end, int *page_started,
1196                                 unsigned long *nr_written,
1197                                 unsigned int write_flags)
1198 {
1199         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1200         struct async_cow *async_cow;
1201         struct btrfs_root *root = BTRFS_I(inode)->root;
1202         unsigned long nr_pages;
1203         u64 cur_end;
1204
1205         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1206                          1, 0, NULL, GFP_NOFS);
1207         while (start < end) {
1208                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1209                 BUG_ON(!async_cow); /* -ENOMEM */
1210                 async_cow->inode = igrab(inode);
1211                 async_cow->root = root;
1212                 async_cow->locked_page = locked_page;
1213                 async_cow->start = start;
1214                 async_cow->write_flags = write_flags;
1215
1216                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1217                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1218                         cur_end = end;
1219                 else
1220                         cur_end = min(end, start + SZ_512K - 1);
1221
1222                 async_cow->end = cur_end;
1223                 INIT_LIST_HEAD(&async_cow->extents);
1224
1225                 btrfs_init_work(&async_cow->work,
1226                                 btrfs_delalloc_helper,
1227                                 async_cow_start, async_cow_submit,
1228                                 async_cow_free);
1229
1230                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1231                         PAGE_SHIFT;
1232                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1233
1234                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1235
1236                 *nr_written += nr_pages;
1237                 start = cur_end + 1;
1238         }
1239         *page_started = 1;
1240         return 0;
1241 }
1242
1243 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1244                                         u64 bytenr, u64 num_bytes)
1245 {
1246         int ret;
1247         struct btrfs_ordered_sum *sums;
1248         LIST_HEAD(list);
1249
1250         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1251                                        bytenr + num_bytes - 1, &list, 0);
1252         if (ret == 0 && list_empty(&list))
1253                 return 0;
1254
1255         while (!list_empty(&list)) {
1256                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1257                 list_del(&sums->list);
1258                 kfree(sums);
1259         }
1260         return 1;
1261 }
1262
1263 /*
1264  * when nowcow writeback call back.  This checks for snapshots or COW copies
1265  * of the extents that exist in the file, and COWs the file as required.
1266  *
1267  * If no cow copies or snapshots exist, we write directly to the existing
1268  * blocks on disk
1269  */
1270 static noinline int run_delalloc_nocow(struct inode *inode,
1271                                        struct page *locked_page,
1272                               u64 start, u64 end, int *page_started, int force,
1273                               unsigned long *nr_written)
1274 {
1275         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1276         struct btrfs_root *root = BTRFS_I(inode)->root;
1277         struct extent_buffer *leaf;
1278         struct btrfs_path *path;
1279         struct btrfs_file_extent_item *fi;
1280         struct btrfs_key found_key;
1281         struct extent_map *em;
1282         u64 cow_start;
1283         u64 cur_offset;
1284         u64 extent_end;
1285         u64 extent_offset;
1286         u64 disk_bytenr;
1287         u64 num_bytes;
1288         u64 disk_num_bytes;
1289         u64 ram_bytes;
1290         int extent_type;
1291         int ret, err;
1292         int type;
1293         int nocow;
1294         int check_prev = 1;
1295         bool nolock;
1296         u64 ino = btrfs_ino(BTRFS_I(inode));
1297
1298         path = btrfs_alloc_path();
1299         if (!path) {
1300                 extent_clear_unlock_delalloc(inode, start, end, end,
1301                                              locked_page,
1302                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1303                                              EXTENT_DO_ACCOUNTING |
1304                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1305                                              PAGE_CLEAR_DIRTY |
1306                                              PAGE_SET_WRITEBACK |
1307                                              PAGE_END_WRITEBACK);
1308                 return -ENOMEM;
1309         }
1310
1311         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1312
1313         cow_start = (u64)-1;
1314         cur_offset = start;
1315         while (1) {
1316                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1317                                                cur_offset, 0);
1318                 if (ret < 0)
1319                         goto error;
1320                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1321                         leaf = path->nodes[0];
1322                         btrfs_item_key_to_cpu(leaf, &found_key,
1323                                               path->slots[0] - 1);
1324                         if (found_key.objectid == ino &&
1325                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1326                                 path->slots[0]--;
1327                 }
1328                 check_prev = 0;
1329 next_slot:
1330                 leaf = path->nodes[0];
1331                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1332                         ret = btrfs_next_leaf(root, path);
1333                         if (ret < 0)
1334                                 goto error;
1335                         if (ret > 0)
1336                                 break;
1337                         leaf = path->nodes[0];
1338                 }
1339
1340                 nocow = 0;
1341                 disk_bytenr = 0;
1342                 num_bytes = 0;
1343                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1344
1345                 if (found_key.objectid > ino)
1346                         break;
1347                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1348                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1349                         path->slots[0]++;
1350                         goto next_slot;
1351                 }
1352                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1353                     found_key.offset > end)
1354                         break;
1355
1356                 if (found_key.offset > cur_offset) {
1357                         extent_end = found_key.offset;
1358                         extent_type = 0;
1359                         goto out_check;
1360                 }
1361
1362                 fi = btrfs_item_ptr(leaf, path->slots[0],
1363                                     struct btrfs_file_extent_item);
1364                 extent_type = btrfs_file_extent_type(leaf, fi);
1365
1366                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1367                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1368                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1369                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1370                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1371                         extent_end = found_key.offset +
1372                                 btrfs_file_extent_num_bytes(leaf, fi);
1373                         disk_num_bytes =
1374                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1375                         if (extent_end <= start) {
1376                                 path->slots[0]++;
1377                                 goto next_slot;
1378                         }
1379                         if (disk_bytenr == 0)
1380                                 goto out_check;
1381                         if (btrfs_file_extent_compression(leaf, fi) ||
1382                             btrfs_file_extent_encryption(leaf, fi) ||
1383                             btrfs_file_extent_other_encoding(leaf, fi))
1384                                 goto out_check;
1385                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1386                                 goto out_check;
1387                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1388                                 goto out_check;
1389                         if (btrfs_cross_ref_exist(root, ino,
1390                                                   found_key.offset -
1391                                                   extent_offset, disk_bytenr))
1392                                 goto out_check;
1393                         disk_bytenr += extent_offset;
1394                         disk_bytenr += cur_offset - found_key.offset;
1395                         num_bytes = min(end + 1, extent_end) - cur_offset;
1396                         /*
1397                          * if there are pending snapshots for this root,
1398                          * we fall into common COW way.
1399                          */
1400                         if (!nolock) {
1401                                 err = btrfs_start_write_no_snapshotting(root);
1402                                 if (!err)
1403                                         goto out_check;
1404                         }
1405                         /*
1406                          * force cow if csum exists in the range.
1407                          * this ensure that csum for a given extent are
1408                          * either valid or do not exist.
1409                          */
1410                         if (csum_exist_in_range(fs_info, disk_bytenr,
1411                                                 num_bytes)) {
1412                                 if (!nolock)
1413                                         btrfs_end_write_no_snapshotting(root);
1414                                 goto out_check;
1415                         }
1416                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr)) {
1417                                 if (!nolock)
1418                                         btrfs_end_write_no_snapshotting(root);
1419                                 goto out_check;
1420                         }
1421                         nocow = 1;
1422                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1423                         extent_end = found_key.offset +
1424                                 btrfs_file_extent_inline_len(leaf,
1425                                                      path->slots[0], fi);
1426                         extent_end = ALIGN(extent_end,
1427                                            fs_info->sectorsize);
1428                 } else {
1429                         BUG_ON(1);
1430                 }
1431 out_check:
1432                 if (extent_end <= start) {
1433                         path->slots[0]++;
1434                         if (!nolock && nocow)
1435                                 btrfs_end_write_no_snapshotting(root);
1436                         if (nocow)
1437                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1438                         goto next_slot;
1439                 }
1440                 if (!nocow) {
1441                         if (cow_start == (u64)-1)
1442                                 cow_start = cur_offset;
1443                         cur_offset = extent_end;
1444                         if (cur_offset > end)
1445                                 break;
1446                         path->slots[0]++;
1447                         goto next_slot;
1448                 }
1449
1450                 btrfs_release_path(path);
1451                 if (cow_start != (u64)-1) {
1452                         ret = cow_file_range(inode, locked_page,
1453                                              cow_start, found_key.offset - 1,
1454                                              end, page_started, nr_written, 1,
1455                                              NULL);
1456                         if (ret) {
1457                                 if (!nolock && nocow)
1458                                         btrfs_end_write_no_snapshotting(root);
1459                                 if (nocow)
1460                                         btrfs_dec_nocow_writers(fs_info,
1461                                                                 disk_bytenr);
1462                                 goto error;
1463                         }
1464                         cow_start = (u64)-1;
1465                 }
1466
1467                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1468                         u64 orig_start = found_key.offset - extent_offset;
1469
1470                         em = create_io_em(inode, cur_offset, num_bytes,
1471                                           orig_start,
1472                                           disk_bytenr, /* block_start */
1473                                           num_bytes, /* block_len */
1474                                           disk_num_bytes, /* orig_block_len */
1475                                           ram_bytes, BTRFS_COMPRESS_NONE,
1476                                           BTRFS_ORDERED_PREALLOC);
1477                         if (IS_ERR(em)) {
1478                                 if (!nolock && nocow)
1479                                         btrfs_end_write_no_snapshotting(root);
1480                                 if (nocow)
1481                                         btrfs_dec_nocow_writers(fs_info,
1482                                                                 disk_bytenr);
1483                                 ret = PTR_ERR(em);
1484                                 goto error;
1485                         }
1486                         free_extent_map(em);
1487                 }
1488
1489                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1490                         type = BTRFS_ORDERED_PREALLOC;
1491                 } else {
1492                         type = BTRFS_ORDERED_NOCOW;
1493                 }
1494
1495                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1496                                                num_bytes, num_bytes, type);
1497                 if (nocow)
1498                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1499                 BUG_ON(ret); /* -ENOMEM */
1500
1501                 if (root->root_key.objectid ==
1502                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1503                         /*
1504                          * Error handled later, as we must prevent
1505                          * extent_clear_unlock_delalloc() in error handler
1506                          * from freeing metadata of created ordered extent.
1507                          */
1508                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1509                                                       num_bytes);
1510
1511                 extent_clear_unlock_delalloc(inode, cur_offset,
1512                                              cur_offset + num_bytes - 1, end,
1513                                              locked_page, EXTENT_LOCKED |
1514                                              EXTENT_DELALLOC |
1515                                              EXTENT_CLEAR_DATA_RESV,
1516                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1517
1518                 if (!nolock && nocow)
1519                         btrfs_end_write_no_snapshotting(root);
1520                 cur_offset = extent_end;
1521
1522                 /*
1523                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1524                  * handler, as metadata for created ordered extent will only
1525                  * be freed by btrfs_finish_ordered_io().
1526                  */
1527                 if (ret)
1528                         goto error;
1529                 if (cur_offset > end)
1530                         break;
1531         }
1532         btrfs_release_path(path);
1533
1534         if (cur_offset <= end && cow_start == (u64)-1) {
1535                 cow_start = cur_offset;
1536                 cur_offset = end;
1537         }
1538
1539         if (cow_start != (u64)-1) {
1540                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1541                                      page_started, nr_written, 1, NULL);
1542                 if (ret)
1543                         goto error;
1544         }
1545
1546 error:
1547         if (ret && cur_offset < end)
1548                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1549                                              locked_page, EXTENT_LOCKED |
1550                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1551                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1552                                              PAGE_CLEAR_DIRTY |
1553                                              PAGE_SET_WRITEBACK |
1554                                              PAGE_END_WRITEBACK);
1555         btrfs_free_path(path);
1556         return ret;
1557 }
1558
1559 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1560 {
1561
1562         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1563             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1564                 return 0;
1565
1566         /*
1567          * @defrag_bytes is a hint value, no spinlock held here,
1568          * if is not zero, it means the file is defragging.
1569          * Force cow if given extent needs to be defragged.
1570          */
1571         if (BTRFS_I(inode)->defrag_bytes &&
1572             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1573                            EXTENT_DEFRAG, 0, NULL))
1574                 return 1;
1575
1576         return 0;
1577 }
1578
1579 /*
1580  * extent_io.c call back to do delayed allocation processing
1581  */
1582 static int run_delalloc_range(void *private_data, struct page *locked_page,
1583                               u64 start, u64 end, int *page_started,
1584                               unsigned long *nr_written,
1585                               struct writeback_control *wbc)
1586 {
1587         struct inode *inode = private_data;
1588         int ret;
1589         int force_cow = need_force_cow(inode, start, end);
1590         unsigned int write_flags = wbc_to_write_flags(wbc);
1591
1592         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1593                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1594                                          page_started, 1, nr_written);
1595         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1596                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1597                                          page_started, 0, nr_written);
1598         } else if (!inode_need_compress(inode, start, end)) {
1599                 ret = cow_file_range(inode, locked_page, start, end, end,
1600                                       page_started, nr_written, 1, NULL);
1601         } else {
1602                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1603                         &BTRFS_I(inode)->runtime_flags);
1604                 ret = cow_file_range_async(inode, locked_page, start, end,
1605                                            page_started, nr_written,
1606                                            write_flags);
1607         }
1608         if (ret)
1609                 btrfs_cleanup_ordered_extents(inode, start, end - start + 1);
1610         return ret;
1611 }
1612
1613 static void btrfs_split_extent_hook(void *private_data,
1614                                     struct extent_state *orig, u64 split)
1615 {
1616         struct inode *inode = private_data;
1617         u64 size;
1618
1619         /* not delalloc, ignore it */
1620         if (!(orig->state & EXTENT_DELALLOC))
1621                 return;
1622
1623         size = orig->end - orig->start + 1;
1624         if (size > BTRFS_MAX_EXTENT_SIZE) {
1625                 u32 num_extents;
1626                 u64 new_size;
1627
1628                 /*
1629                  * See the explanation in btrfs_merge_extent_hook, the same
1630                  * applies here, just in reverse.
1631                  */
1632                 new_size = orig->end - split + 1;
1633                 num_extents = count_max_extents(new_size);
1634                 new_size = split - orig->start;
1635                 num_extents += count_max_extents(new_size);
1636                 if (count_max_extents(size) >= num_extents)
1637                         return;
1638         }
1639
1640         spin_lock(&BTRFS_I(inode)->lock);
1641         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1642         spin_unlock(&BTRFS_I(inode)->lock);
1643 }
1644
1645 /*
1646  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1647  * extents so we can keep track of new extents that are just merged onto old
1648  * extents, such as when we are doing sequential writes, so we can properly
1649  * account for the metadata space we'll need.
1650  */
1651 static void btrfs_merge_extent_hook(void *private_data,
1652                                     struct extent_state *new,
1653                                     struct extent_state *other)
1654 {
1655         struct inode *inode = private_data;
1656         u64 new_size, old_size;
1657         u32 num_extents;
1658
1659         /* not delalloc, ignore it */
1660         if (!(other->state & EXTENT_DELALLOC))
1661                 return;
1662
1663         if (new->start > other->start)
1664                 new_size = new->end - other->start + 1;
1665         else
1666                 new_size = other->end - new->start + 1;
1667
1668         /* we're not bigger than the max, unreserve the space and go */
1669         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1670                 spin_lock(&BTRFS_I(inode)->lock);
1671                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1672                 spin_unlock(&BTRFS_I(inode)->lock);
1673                 return;
1674         }
1675
1676         /*
1677          * We have to add up either side to figure out how many extents were
1678          * accounted for before we merged into one big extent.  If the number of
1679          * extents we accounted for is <= the amount we need for the new range
1680          * then we can return, otherwise drop.  Think of it like this
1681          *
1682          * [ 4k][MAX_SIZE]
1683          *
1684          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1685          * need 2 outstanding extents, on one side we have 1 and the other side
1686          * we have 1 so they are == and we can return.  But in this case
1687          *
1688          * [MAX_SIZE+4k][MAX_SIZE+4k]
1689          *
1690          * Each range on their own accounts for 2 extents, but merged together
1691          * they are only 3 extents worth of accounting, so we need to drop in
1692          * this case.
1693          */
1694         old_size = other->end - other->start + 1;
1695         num_extents = count_max_extents(old_size);
1696         old_size = new->end - new->start + 1;
1697         num_extents += count_max_extents(old_size);
1698         if (count_max_extents(new_size) >= num_extents)
1699                 return;
1700
1701         spin_lock(&BTRFS_I(inode)->lock);
1702         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1703         spin_unlock(&BTRFS_I(inode)->lock);
1704 }
1705
1706 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1707                                       struct inode *inode)
1708 {
1709         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1710
1711         spin_lock(&root->delalloc_lock);
1712         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1713                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1714                               &root->delalloc_inodes);
1715                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1716                         &BTRFS_I(inode)->runtime_flags);
1717                 root->nr_delalloc_inodes++;
1718                 if (root->nr_delalloc_inodes == 1) {
1719                         spin_lock(&fs_info->delalloc_root_lock);
1720                         BUG_ON(!list_empty(&root->delalloc_root));
1721                         list_add_tail(&root->delalloc_root,
1722                                       &fs_info->delalloc_roots);
1723                         spin_unlock(&fs_info->delalloc_root_lock);
1724                 }
1725         }
1726         spin_unlock(&root->delalloc_lock);
1727 }
1728
1729 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1730                                      struct btrfs_inode *inode)
1731 {
1732         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1733
1734         spin_lock(&root->delalloc_lock);
1735         if (!list_empty(&inode->delalloc_inodes)) {
1736                 list_del_init(&inode->delalloc_inodes);
1737                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1738                           &inode->runtime_flags);
1739                 root->nr_delalloc_inodes--;
1740                 if (!root->nr_delalloc_inodes) {
1741                         spin_lock(&fs_info->delalloc_root_lock);
1742                         BUG_ON(list_empty(&root->delalloc_root));
1743                         list_del_init(&root->delalloc_root);
1744                         spin_unlock(&fs_info->delalloc_root_lock);
1745                 }
1746         }
1747         spin_unlock(&root->delalloc_lock);
1748 }
1749
1750 /*
1751  * extent_io.c set_bit_hook, used to track delayed allocation
1752  * bytes in this file, and to maintain the list of inodes that
1753  * have pending delalloc work to be done.
1754  */
1755 static void btrfs_set_bit_hook(void *private_data,
1756                                struct extent_state *state, unsigned *bits)
1757 {
1758         struct inode *inode = private_data;
1759
1760         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1761
1762         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1763                 WARN_ON(1);
1764         /*
1765          * set_bit and clear bit hooks normally require _irqsave/restore
1766          * but in this case, we are only testing for the DELALLOC
1767          * bit, which is only set or cleared with irqs on
1768          */
1769         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1770                 struct btrfs_root *root = BTRFS_I(inode)->root;
1771                 u64 len = state->end + 1 - state->start;
1772                 u32 num_extents = count_max_extents(len);
1773                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1774
1775                 spin_lock(&BTRFS_I(inode)->lock);
1776                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1777                 spin_unlock(&BTRFS_I(inode)->lock);
1778
1779                 /* For sanity tests */
1780                 if (btrfs_is_testing(fs_info))
1781                         return;
1782
1783                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1784                                          fs_info->delalloc_batch);
1785                 spin_lock(&BTRFS_I(inode)->lock);
1786                 BTRFS_I(inode)->delalloc_bytes += len;
1787                 if (*bits & EXTENT_DEFRAG)
1788                         BTRFS_I(inode)->defrag_bytes += len;
1789                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1790                                          &BTRFS_I(inode)->runtime_flags))
1791                         btrfs_add_delalloc_inodes(root, inode);
1792                 spin_unlock(&BTRFS_I(inode)->lock);
1793         }
1794
1795         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1796             (*bits & EXTENT_DELALLOC_NEW)) {
1797                 spin_lock(&BTRFS_I(inode)->lock);
1798                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1799                         state->start;
1800                 spin_unlock(&BTRFS_I(inode)->lock);
1801         }
1802 }
1803
1804 /*
1805  * extent_io.c clear_bit_hook, see set_bit_hook for why
1806  */
1807 static void btrfs_clear_bit_hook(void *private_data,
1808                                  struct extent_state *state,
1809                                  unsigned *bits)
1810 {
1811         struct btrfs_inode *inode = BTRFS_I((struct inode *)private_data);
1812         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1813         u64 len = state->end + 1 - state->start;
1814         u32 num_extents = count_max_extents(len);
1815
1816         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1817                 spin_lock(&inode->lock);
1818                 inode->defrag_bytes -= len;
1819                 spin_unlock(&inode->lock);
1820         }
1821
1822         /*
1823          * set_bit and clear bit hooks normally require _irqsave/restore
1824          * but in this case, we are only testing for the DELALLOC
1825          * bit, which is only set or cleared with irqs on
1826          */
1827         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1828                 struct btrfs_root *root = inode->root;
1829                 bool do_list = !btrfs_is_free_space_inode(inode);
1830
1831                 spin_lock(&inode->lock);
1832                 btrfs_mod_outstanding_extents(inode, -num_extents);
1833                 spin_unlock(&inode->lock);
1834
1835                 /*
1836                  * We don't reserve metadata space for space cache inodes so we
1837                  * don't need to call dellalloc_release_metadata if there is an
1838                  * error.
1839                  */
1840                 if (*bits & EXTENT_CLEAR_META_RESV &&
1841                     root != fs_info->tree_root)
1842                         btrfs_delalloc_release_metadata(inode, len);
1843
1844                 /* For sanity tests. */
1845                 if (btrfs_is_testing(fs_info))
1846                         return;
1847
1848                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1849                     do_list && !(state->state & EXTENT_NORESERVE) &&
1850                     (*bits & EXTENT_CLEAR_DATA_RESV))
1851                         btrfs_free_reserved_data_space_noquota(
1852                                         &inode->vfs_inode,
1853                                         state->start, len);
1854
1855                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1856                                          fs_info->delalloc_batch);
1857                 spin_lock(&inode->lock);
1858                 inode->delalloc_bytes -= len;
1859                 if (do_list && inode->delalloc_bytes == 0 &&
1860                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1861                                         &inode->runtime_flags))
1862                         btrfs_del_delalloc_inode(root, inode);
1863                 spin_unlock(&inode->lock);
1864         }
1865
1866         if ((state->state & EXTENT_DELALLOC_NEW) &&
1867             (*bits & EXTENT_DELALLOC_NEW)) {
1868                 spin_lock(&inode->lock);
1869                 ASSERT(inode->new_delalloc_bytes >= len);
1870                 inode->new_delalloc_bytes -= len;
1871                 spin_unlock(&inode->lock);
1872         }
1873 }
1874
1875 /*
1876  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1877  * we don't create bios that span stripes or chunks
1878  *
1879  * return 1 if page cannot be merged to bio
1880  * return 0 if page can be merged to bio
1881  * return error otherwise
1882  */
1883 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1884                          size_t size, struct bio *bio,
1885                          unsigned long bio_flags)
1886 {
1887         struct inode *inode = page->mapping->host;
1888         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1889         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1890         u64 length = 0;
1891         u64 map_length;
1892         int ret;
1893
1894         if (bio_flags & EXTENT_BIO_COMPRESSED)
1895                 return 0;
1896
1897         length = bio->bi_iter.bi_size;
1898         map_length = length;
1899         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1900                               NULL, 0);
1901         if (ret < 0)
1902                 return ret;
1903         if (map_length < length + size)
1904                 return 1;
1905         return 0;
1906 }
1907
1908 /*
1909  * in order to insert checksums into the metadata in large chunks,
1910  * we wait until bio submission time.   All the pages in the bio are
1911  * checksummed and sums are attached onto the ordered extent record.
1912  *
1913  * At IO completion time the cums attached on the ordered extent record
1914  * are inserted into the btree
1915  */
1916 static blk_status_t __btrfs_submit_bio_start(void *private_data, struct bio *bio,
1917                                     int mirror_num, unsigned long bio_flags,
1918                                     u64 bio_offset)
1919 {
1920         struct inode *inode = private_data;
1921         blk_status_t ret = 0;
1922
1923         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1924         BUG_ON(ret); /* -ENOMEM */
1925         return 0;
1926 }
1927
1928 /*
1929  * in order to insert checksums into the metadata in large chunks,
1930  * we wait until bio submission time.   All the pages in the bio are
1931  * checksummed and sums are attached onto the ordered extent record.
1932  *
1933  * At IO completion time the cums attached on the ordered extent record
1934  * are inserted into the btree
1935  */
1936 static blk_status_t __btrfs_submit_bio_done(void *private_data, struct bio *bio,
1937                           int mirror_num, unsigned long bio_flags,
1938                           u64 bio_offset)
1939 {
1940         struct inode *inode = private_data;
1941         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1942         blk_status_t ret;
1943
1944         ret = btrfs_map_bio(fs_info, bio, mirror_num, 1);
1945         if (ret) {
1946                 bio->bi_status = ret;
1947                 bio_endio(bio);
1948         }
1949         return ret;
1950 }
1951
1952 /*
1953  * extent_io.c submission hook. This does the right thing for csum calculation
1954  * on write, or reading the csums from the tree before a read
1955  */
1956 static blk_status_t btrfs_submit_bio_hook(void *private_data, struct bio *bio,
1957                                  int mirror_num, unsigned long bio_flags,
1958                                  u64 bio_offset)
1959 {
1960         struct inode *inode = private_data;
1961         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1962         struct btrfs_root *root = BTRFS_I(inode)->root;
1963         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1964         blk_status_t ret = 0;
1965         int skip_sum;
1966         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1967
1968         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1969
1970         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
1971                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1972
1973         if (bio_op(bio) != REQ_OP_WRITE) {
1974                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
1975                 if (ret)
1976                         goto out;
1977
1978                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1979                         ret = btrfs_submit_compressed_read(inode, bio,
1980                                                            mirror_num,
1981                                                            bio_flags);
1982                         goto out;
1983                 } else if (!skip_sum) {
1984                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
1985                         if (ret)
1986                                 goto out;
1987                 }
1988                 goto mapit;
1989         } else if (async && !skip_sum) {
1990                 /* csum items have already been cloned */
1991                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1992                         goto mapit;
1993                 /* we're doing a write, do the async checksumming */
1994                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
1995                                           bio_offset, inode,
1996                                           __btrfs_submit_bio_start,
1997                                           __btrfs_submit_bio_done);
1998                 goto out;
1999         } else if (!skip_sum) {
2000                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2001                 if (ret)
2002                         goto out;
2003         }
2004
2005 mapit:
2006         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
2007
2008 out:
2009         if (ret) {
2010                 bio->bi_status = ret;
2011                 bio_endio(bio);
2012         }
2013         return ret;
2014 }
2015
2016 /*
2017  * given a list of ordered sums record them in the inode.  This happens
2018  * at IO completion time based on sums calculated at bio submission time.
2019  */
2020 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2021                              struct inode *inode, struct list_head *list)
2022 {
2023         struct btrfs_ordered_sum *sum;
2024
2025         list_for_each_entry(sum, list, list) {
2026                 trans->adding_csums = 1;
2027                 btrfs_csum_file_blocks(trans,
2028                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2029                 trans->adding_csums = 0;
2030         }
2031         return 0;
2032 }
2033
2034 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2035                               unsigned int extra_bits,
2036                               struct extent_state **cached_state, int dedupe)
2037 {
2038         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
2039         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2040                                    extra_bits, cached_state);
2041 }
2042
2043 /* see btrfs_writepage_start_hook for details on why this is required */
2044 struct btrfs_writepage_fixup {
2045         struct page *page;
2046         struct btrfs_work work;
2047 };
2048
2049 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2050 {
2051         struct btrfs_writepage_fixup *fixup;
2052         struct btrfs_ordered_extent *ordered;
2053         struct extent_state *cached_state = NULL;
2054         struct extent_changeset *data_reserved = NULL;
2055         struct page *page;
2056         struct inode *inode;
2057         u64 page_start;
2058         u64 page_end;
2059         int ret;
2060
2061         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2062         page = fixup->page;
2063 again:
2064         lock_page(page);
2065         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2066                 ClearPageChecked(page);
2067                 goto out_page;
2068         }
2069
2070         inode = page->mapping->host;
2071         page_start = page_offset(page);
2072         page_end = page_offset(page) + PAGE_SIZE - 1;
2073
2074         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2075                          &cached_state);
2076
2077         /* already ordered? We're done */
2078         if (PagePrivate2(page))
2079                 goto out;
2080
2081         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2082                                         PAGE_SIZE);
2083         if (ordered) {
2084                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2085                                      page_end, &cached_state, GFP_NOFS);
2086                 unlock_page(page);
2087                 btrfs_start_ordered_extent(inode, ordered, 1);
2088                 btrfs_put_ordered_extent(ordered);
2089                 goto again;
2090         }
2091
2092         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2093                                            PAGE_SIZE);
2094         if (ret) {
2095                 mapping_set_error(page->mapping, ret);
2096                 end_extent_writepage(page, ret, page_start, page_end);
2097                 ClearPageChecked(page);
2098                 goto out;
2099          }
2100
2101         btrfs_set_extent_delalloc(inode, page_start, page_end, 0, &cached_state,
2102                                   0);
2103         ClearPageChecked(page);
2104         set_page_dirty(page);
2105         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2106 out:
2107         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2108                              &cached_state, GFP_NOFS);
2109 out_page:
2110         unlock_page(page);
2111         put_page(page);
2112         kfree(fixup);
2113         extent_changeset_free(data_reserved);
2114 }
2115
2116 /*
2117  * There are a few paths in the higher layers of the kernel that directly
2118  * set the page dirty bit without asking the filesystem if it is a
2119  * good idea.  This causes problems because we want to make sure COW
2120  * properly happens and the data=ordered rules are followed.
2121  *
2122  * In our case any range that doesn't have the ORDERED bit set
2123  * hasn't been properly setup for IO.  We kick off an async process
2124  * to fix it up.  The async helper will wait for ordered extents, set
2125  * the delalloc bit and make it safe to write the page.
2126  */
2127 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2128 {
2129         struct inode *inode = page->mapping->host;
2130         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2131         struct btrfs_writepage_fixup *fixup;
2132
2133         /* this page is properly in the ordered list */
2134         if (TestClearPagePrivate2(page))
2135                 return 0;
2136
2137         if (PageChecked(page))
2138                 return -EAGAIN;
2139
2140         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2141         if (!fixup)
2142                 return -EAGAIN;
2143
2144         SetPageChecked(page);
2145         get_page(page);
2146         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2147                         btrfs_writepage_fixup_worker, NULL, NULL);
2148         fixup->page = page;
2149         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2150         return -EBUSY;
2151 }
2152
2153 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2154                                        struct inode *inode, u64 file_pos,
2155                                        u64 disk_bytenr, u64 disk_num_bytes,
2156                                        u64 num_bytes, u64 ram_bytes,
2157                                        u8 compression, u8 encryption,
2158                                        u16 other_encoding, int extent_type)
2159 {
2160         struct btrfs_root *root = BTRFS_I(inode)->root;
2161         struct btrfs_file_extent_item *fi;
2162         struct btrfs_path *path;
2163         struct extent_buffer *leaf;
2164         struct btrfs_key ins;
2165         u64 qg_released;
2166         int extent_inserted = 0;
2167         int ret;
2168
2169         path = btrfs_alloc_path();
2170         if (!path)
2171                 return -ENOMEM;
2172
2173         /*
2174          * we may be replacing one extent in the tree with another.
2175          * The new extent is pinned in the extent map, and we don't want
2176          * to drop it from the cache until it is completely in the btree.
2177          *
2178          * So, tell btrfs_drop_extents to leave this extent in the cache.
2179          * the caller is expected to unpin it and allow it to be merged
2180          * with the others.
2181          */
2182         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2183                                    file_pos + num_bytes, NULL, 0,
2184                                    1, sizeof(*fi), &extent_inserted);
2185         if (ret)
2186                 goto out;
2187
2188         if (!extent_inserted) {
2189                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2190                 ins.offset = file_pos;
2191                 ins.type = BTRFS_EXTENT_DATA_KEY;
2192
2193                 path->leave_spinning = 1;
2194                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2195                                               sizeof(*fi));
2196                 if (ret)
2197                         goto out;
2198         }
2199         leaf = path->nodes[0];
2200         fi = btrfs_item_ptr(leaf, path->slots[0],
2201                             struct btrfs_file_extent_item);
2202         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2203         btrfs_set_file_extent_type(leaf, fi, extent_type);
2204         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2205         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2206         btrfs_set_file_extent_offset(leaf, fi, 0);
2207         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2208         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2209         btrfs_set_file_extent_compression(leaf, fi, compression);
2210         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2211         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2212
2213         btrfs_mark_buffer_dirty(leaf);
2214         btrfs_release_path(path);
2215
2216         inode_add_bytes(inode, num_bytes);
2217
2218         ins.objectid = disk_bytenr;
2219         ins.offset = disk_num_bytes;
2220         ins.type = BTRFS_EXTENT_ITEM_KEY;
2221
2222         /*
2223          * Release the reserved range from inode dirty range map, as it is
2224          * already moved into delayed_ref_head
2225          */
2226         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2227         if (ret < 0)
2228                 goto out;
2229         qg_released = ret;
2230         ret = btrfs_alloc_reserved_file_extent(trans, root,
2231                                                btrfs_ino(BTRFS_I(inode)),
2232                                                file_pos, qg_released, &ins);
2233 out:
2234         btrfs_free_path(path);
2235
2236         return ret;
2237 }
2238
2239 /* snapshot-aware defrag */
2240 struct sa_defrag_extent_backref {
2241         struct rb_node node;
2242         struct old_sa_defrag_extent *old;
2243         u64 root_id;
2244         u64 inum;
2245         u64 file_pos;
2246         u64 extent_offset;
2247         u64 num_bytes;
2248         u64 generation;
2249 };
2250
2251 struct old_sa_defrag_extent {
2252         struct list_head list;
2253         struct new_sa_defrag_extent *new;
2254
2255         u64 extent_offset;
2256         u64 bytenr;
2257         u64 offset;
2258         u64 len;
2259         int count;
2260 };
2261
2262 struct new_sa_defrag_extent {
2263         struct rb_root root;
2264         struct list_head head;
2265         struct btrfs_path *path;
2266         struct inode *inode;
2267         u64 file_pos;
2268         u64 len;
2269         u64 bytenr;
2270         u64 disk_len;
2271         u8 compress_type;
2272 };
2273
2274 static int backref_comp(struct sa_defrag_extent_backref *b1,
2275                         struct sa_defrag_extent_backref *b2)
2276 {
2277         if (b1->root_id < b2->root_id)
2278                 return -1;
2279         else if (b1->root_id > b2->root_id)
2280                 return 1;
2281
2282         if (b1->inum < b2->inum)
2283                 return -1;
2284         else if (b1->inum > b2->inum)
2285                 return 1;
2286
2287         if (b1->file_pos < b2->file_pos)
2288                 return -1;
2289         else if (b1->file_pos > b2->file_pos)
2290                 return 1;
2291
2292         /*
2293          * [------------------------------] ===> (a range of space)
2294          *     |<--->|   |<---->| =============> (fs/file tree A)
2295          * |<---------------------------->| ===> (fs/file tree B)
2296          *
2297          * A range of space can refer to two file extents in one tree while
2298          * refer to only one file extent in another tree.
2299          *
2300          * So we may process a disk offset more than one time(two extents in A)
2301          * and locate at the same extent(one extent in B), then insert two same
2302          * backrefs(both refer to the extent in B).
2303          */
2304         return 0;
2305 }
2306
2307 static void backref_insert(struct rb_root *root,
2308                            struct sa_defrag_extent_backref *backref)
2309 {
2310         struct rb_node **p = &root->rb_node;
2311         struct rb_node *parent = NULL;
2312         struct sa_defrag_extent_backref *entry;
2313         int ret;
2314
2315         while (*p) {
2316                 parent = *p;
2317                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2318
2319                 ret = backref_comp(backref, entry);
2320                 if (ret < 0)
2321                         p = &(*p)->rb_left;
2322                 else
2323                         p = &(*p)->rb_right;
2324         }
2325
2326         rb_link_node(&backref->node, parent, p);
2327         rb_insert_color(&backref->node, root);
2328 }
2329
2330 /*
2331  * Note the backref might has changed, and in this case we just return 0.
2332  */
2333 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2334                                        void *ctx)
2335 {
2336         struct btrfs_file_extent_item *extent;
2337         struct old_sa_defrag_extent *old = ctx;
2338         struct new_sa_defrag_extent *new = old->new;
2339         struct btrfs_path *path = new->path;
2340         struct btrfs_key key;
2341         struct btrfs_root *root;
2342         struct sa_defrag_extent_backref *backref;
2343         struct extent_buffer *leaf;
2344         struct inode *inode = new->inode;
2345         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2346         int slot;
2347         int ret;
2348         u64 extent_offset;
2349         u64 num_bytes;
2350
2351         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2352             inum == btrfs_ino(BTRFS_I(inode)))
2353                 return 0;
2354
2355         key.objectid = root_id;
2356         key.type = BTRFS_ROOT_ITEM_KEY;
2357         key.offset = (u64)-1;
2358
2359         root = btrfs_read_fs_root_no_name(fs_info, &key);
2360         if (IS_ERR(root)) {
2361                 if (PTR_ERR(root) == -ENOENT)
2362                         return 0;
2363                 WARN_ON(1);
2364                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2365                          inum, offset, root_id);
2366                 return PTR_ERR(root);
2367         }
2368
2369         key.objectid = inum;
2370         key.type = BTRFS_EXTENT_DATA_KEY;
2371         if (offset > (u64)-1 << 32)
2372                 key.offset = 0;
2373         else
2374                 key.offset = offset;
2375
2376         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2377         if (WARN_ON(ret < 0))
2378                 return ret;
2379         ret = 0;
2380
2381         while (1) {
2382                 cond_resched();
2383
2384                 leaf = path->nodes[0];
2385                 slot = path->slots[0];
2386
2387                 if (slot >= btrfs_header_nritems(leaf)) {
2388                         ret = btrfs_next_leaf(root, path);
2389                         if (ret < 0) {
2390                                 goto out;
2391                         } else if (ret > 0) {
2392                                 ret = 0;
2393                                 goto out;
2394                         }
2395                         continue;
2396                 }
2397
2398                 path->slots[0]++;
2399
2400                 btrfs_item_key_to_cpu(leaf, &key, slot);
2401
2402                 if (key.objectid > inum)
2403                         goto out;
2404
2405                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2406                         continue;
2407
2408                 extent = btrfs_item_ptr(leaf, slot,
2409                                         struct btrfs_file_extent_item);
2410
2411                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2412                         continue;
2413
2414                 /*
2415                  * 'offset' refers to the exact key.offset,
2416                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2417                  * (key.offset - extent_offset).
2418                  */
2419                 if (key.offset != offset)
2420                         continue;
2421
2422                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2423                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2424
2425                 if (extent_offset >= old->extent_offset + old->offset +
2426                     old->len || extent_offset + num_bytes <=
2427                     old->extent_offset + old->offset)
2428                         continue;
2429                 break;
2430         }
2431
2432         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2433         if (!backref) {
2434                 ret = -ENOENT;
2435                 goto out;
2436         }
2437
2438         backref->root_id = root_id;
2439         backref->inum = inum;
2440         backref->file_pos = offset;
2441         backref->num_bytes = num_bytes;
2442         backref->extent_offset = extent_offset;
2443         backref->generation = btrfs_file_extent_generation(leaf, extent);
2444         backref->old = old;
2445         backref_insert(&new->root, backref);
2446         old->count++;
2447 out:
2448         btrfs_release_path(path);
2449         WARN_ON(ret);
2450         return ret;
2451 }
2452
2453 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2454                                    struct new_sa_defrag_extent *new)
2455 {
2456         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2457         struct old_sa_defrag_extent *old, *tmp;
2458         int ret;
2459
2460         new->path = path;
2461
2462         list_for_each_entry_safe(old, tmp, &new->head, list) {
2463                 ret = iterate_inodes_from_logical(old->bytenr +
2464                                                   old->extent_offset, fs_info,
2465                                                   path, record_one_backref,
2466                                                   old, false);
2467                 if (ret < 0 && ret != -ENOENT)
2468                         return false;
2469
2470                 /* no backref to be processed for this extent */
2471                 if (!old->count) {
2472                         list_del(&old->list);
2473                         kfree(old);
2474                 }
2475         }
2476
2477         if (list_empty(&new->head))
2478                 return false;
2479
2480         return true;
2481 }
2482
2483 static int relink_is_mergable(struct extent_buffer *leaf,
2484                               struct btrfs_file_extent_item *fi,
2485                               struct new_sa_defrag_extent *new)
2486 {
2487         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2488                 return 0;
2489
2490         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2491                 return 0;
2492
2493         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2494                 return 0;
2495
2496         if (btrfs_file_extent_encryption(leaf, fi) ||
2497             btrfs_file_extent_other_encoding(leaf, fi))
2498                 return 0;
2499
2500         return 1;
2501 }
2502
2503 /*
2504  * Note the backref might has changed, and in this case we just return 0.
2505  */
2506 static noinline int relink_extent_backref(struct btrfs_path *path,
2507                                  struct sa_defrag_extent_backref *prev,
2508                                  struct sa_defrag_extent_backref *backref)
2509 {
2510         struct btrfs_file_extent_item *extent;
2511         struct btrfs_file_extent_item *item;
2512         struct btrfs_ordered_extent *ordered;
2513         struct btrfs_trans_handle *trans;
2514         struct btrfs_root *root;
2515         struct btrfs_key key;
2516         struct extent_buffer *leaf;
2517         struct old_sa_defrag_extent *old = backref->old;
2518         struct new_sa_defrag_extent *new = old->new;
2519         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2520         struct inode *inode;
2521         struct extent_state *cached = NULL;
2522         int ret = 0;
2523         u64 start;
2524         u64 len;
2525         u64 lock_start;
2526         u64 lock_end;
2527         bool merge = false;
2528         int index;
2529
2530         if (prev && prev->root_id == backref->root_id &&
2531             prev->inum == backref->inum &&
2532             prev->file_pos + prev->num_bytes == backref->file_pos)
2533                 merge = true;
2534
2535         /* step 1: get root */
2536         key.objectid = backref->root_id;
2537         key.type = BTRFS_ROOT_ITEM_KEY;
2538         key.offset = (u64)-1;
2539
2540         index = srcu_read_lock(&fs_info->subvol_srcu);
2541
2542         root = btrfs_read_fs_root_no_name(fs_info, &key);
2543         if (IS_ERR(root)) {
2544                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2545                 if (PTR_ERR(root) == -ENOENT)
2546                         return 0;
2547                 return PTR_ERR(root);
2548         }
2549
2550         if (btrfs_root_readonly(root)) {
2551                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2552                 return 0;
2553         }
2554
2555         /* step 2: get inode */
2556         key.objectid = backref->inum;
2557         key.type = BTRFS_INODE_ITEM_KEY;
2558         key.offset = 0;
2559
2560         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2561         if (IS_ERR(inode)) {
2562                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2563                 return 0;
2564         }
2565
2566         srcu_read_unlock(&fs_info->subvol_srcu, index);
2567
2568         /* step 3: relink backref */
2569         lock_start = backref->file_pos;
2570         lock_end = backref->file_pos + backref->num_bytes - 1;
2571         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2572                          &cached);
2573
2574         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2575         if (ordered) {
2576                 btrfs_put_ordered_extent(ordered);
2577                 goto out_unlock;
2578         }
2579
2580         trans = btrfs_join_transaction(root);
2581         if (IS_ERR(trans)) {
2582                 ret = PTR_ERR(trans);
2583                 goto out_unlock;
2584         }
2585
2586         key.objectid = backref->inum;
2587         key.type = BTRFS_EXTENT_DATA_KEY;
2588         key.offset = backref->file_pos;
2589
2590         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2591         if (ret < 0) {
2592                 goto out_free_path;
2593         } else if (ret > 0) {
2594                 ret = 0;
2595                 goto out_free_path;
2596         }
2597
2598         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2599                                 struct btrfs_file_extent_item);
2600
2601         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2602             backref->generation)
2603                 goto out_free_path;
2604
2605         btrfs_release_path(path);
2606
2607         start = backref->file_pos;
2608         if (backref->extent_offset < old->extent_offset + old->offset)
2609                 start += old->extent_offset + old->offset -
2610                          backref->extent_offset;
2611
2612         len = min(backref->extent_offset + backref->num_bytes,
2613                   old->extent_offset + old->offset + old->len);
2614         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2615
2616         ret = btrfs_drop_extents(trans, root, inode, start,
2617                                  start + len, 1);
2618         if (ret)
2619                 goto out_free_path;
2620 again:
2621         key.objectid = btrfs_ino(BTRFS_I(inode));
2622         key.type = BTRFS_EXTENT_DATA_KEY;
2623         key.offset = start;
2624
2625         path->leave_spinning = 1;
2626         if (merge) {
2627                 struct btrfs_file_extent_item *fi;
2628                 u64 extent_len;
2629                 struct btrfs_key found_key;
2630
2631                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2632                 if (ret < 0)
2633                         goto out_free_path;
2634
2635                 path->slots[0]--;
2636                 leaf = path->nodes[0];
2637                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2638
2639                 fi = btrfs_item_ptr(leaf, path->slots[0],
2640                                     struct btrfs_file_extent_item);
2641                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2642
2643                 if (extent_len + found_key.offset == start &&
2644                     relink_is_mergable(leaf, fi, new)) {
2645                         btrfs_set_file_extent_num_bytes(leaf, fi,
2646                                                         extent_len + len);
2647                         btrfs_mark_buffer_dirty(leaf);
2648                         inode_add_bytes(inode, len);
2649
2650                         ret = 1;
2651                         goto out_free_path;
2652                 } else {
2653                         merge = false;
2654                         btrfs_release_path(path);
2655                         goto again;
2656                 }
2657         }
2658
2659         ret = btrfs_insert_empty_item(trans, root, path, &key,
2660                                         sizeof(*extent));
2661         if (ret) {
2662                 btrfs_abort_transaction(trans, ret);
2663                 goto out_free_path;
2664         }
2665
2666         leaf = path->nodes[0];
2667         item = btrfs_item_ptr(leaf, path->slots[0],
2668                                 struct btrfs_file_extent_item);
2669         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2670         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2671         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2672         btrfs_set_file_extent_num_bytes(leaf, item, len);
2673         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2674         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2675         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2676         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2677         btrfs_set_file_extent_encryption(leaf, item, 0);
2678         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2679
2680         btrfs_mark_buffer_dirty(leaf);
2681         inode_add_bytes(inode, len);
2682         btrfs_release_path(path);
2683
2684         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2685                         new->disk_len, 0,
2686                         backref->root_id, backref->inum,
2687                         new->file_pos); /* start - extent_offset */
2688         if (ret) {
2689                 btrfs_abort_transaction(trans, ret);
2690                 goto out_free_path;
2691         }
2692
2693         ret = 1;
2694 out_free_path:
2695         btrfs_release_path(path);
2696         path->leave_spinning = 0;
2697         btrfs_end_transaction(trans);
2698 out_unlock:
2699         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2700                              &cached, GFP_NOFS);
2701         iput(inode);
2702         return ret;
2703 }
2704
2705 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2706 {
2707         struct old_sa_defrag_extent *old, *tmp;
2708
2709         if (!new)
2710                 return;
2711
2712         list_for_each_entry_safe(old, tmp, &new->head, list) {
2713                 kfree(old);
2714         }
2715         kfree(new);
2716 }
2717
2718 static void relink_file_extents(struct new_sa_defrag_extent *new)
2719 {
2720         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2721         struct btrfs_path *path;
2722         struct sa_defrag_extent_backref *backref;
2723         struct sa_defrag_extent_backref *prev = NULL;
2724         struct inode *inode;
2725         struct btrfs_root *root;
2726         struct rb_node *node;
2727         int ret;
2728
2729         inode = new->inode;
2730         root = BTRFS_I(inode)->root;
2731
2732         path = btrfs_alloc_path();
2733         if (!path)
2734                 return;
2735
2736         if (!record_extent_backrefs(path, new)) {
2737                 btrfs_free_path(path);
2738                 goto out;
2739         }
2740         btrfs_release_path(path);
2741
2742         while (1) {
2743                 node = rb_first(&new->root);
2744                 if (!node)
2745                         break;
2746                 rb_erase(node, &new->root);
2747
2748                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2749
2750                 ret = relink_extent_backref(path, prev, backref);
2751                 WARN_ON(ret < 0);
2752
2753                 kfree(prev);
2754
2755                 if (ret == 1)
2756                         prev = backref;
2757                 else
2758                         prev = NULL;
2759                 cond_resched();
2760         }
2761         kfree(prev);
2762
2763         btrfs_free_path(path);
2764 out:
2765         free_sa_defrag_extent(new);
2766
2767         atomic_dec(&fs_info->defrag_running);
2768         wake_up(&fs_info->transaction_wait);
2769 }
2770
2771 static struct new_sa_defrag_extent *
2772 record_old_file_extents(struct inode *inode,
2773                         struct btrfs_ordered_extent *ordered)
2774 {
2775         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2776         struct btrfs_root *root = BTRFS_I(inode)->root;
2777         struct btrfs_path *path;
2778         struct btrfs_key key;
2779         struct old_sa_defrag_extent *old;
2780         struct new_sa_defrag_extent *new;
2781         int ret;
2782
2783         new = kmalloc(sizeof(*new), GFP_NOFS);
2784         if (!new)
2785                 return NULL;
2786
2787         new->inode = inode;
2788         new->file_pos = ordered->file_offset;
2789         new->len = ordered->len;
2790         new->bytenr = ordered->start;
2791         new->disk_len = ordered->disk_len;
2792         new->compress_type = ordered->compress_type;
2793         new->root = RB_ROOT;
2794         INIT_LIST_HEAD(&new->head);
2795
2796         path = btrfs_alloc_path();
2797         if (!path)
2798                 goto out_kfree;
2799
2800         key.objectid = btrfs_ino(BTRFS_I(inode));
2801         key.type = BTRFS_EXTENT_DATA_KEY;
2802         key.offset = new->file_pos;
2803
2804         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2805         if (ret < 0)
2806                 goto out_free_path;
2807         if (ret > 0 && path->slots[0] > 0)
2808                 path->slots[0]--;
2809
2810         /* find out all the old extents for the file range */
2811         while (1) {
2812                 struct btrfs_file_extent_item *extent;
2813                 struct extent_buffer *l;
2814                 int slot;
2815                 u64 num_bytes;
2816                 u64 offset;
2817                 u64 end;
2818                 u64 disk_bytenr;
2819                 u64 extent_offset;
2820
2821                 l = path->nodes[0];
2822                 slot = path->slots[0];
2823
2824                 if (slot >= btrfs_header_nritems(l)) {
2825                         ret = btrfs_next_leaf(root, path);
2826                         if (ret < 0)
2827                                 goto out_free_path;
2828                         else if (ret > 0)
2829                                 break;
2830                         continue;
2831                 }
2832
2833                 btrfs_item_key_to_cpu(l, &key, slot);
2834
2835                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2836                         break;
2837                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2838                         break;
2839                 if (key.offset >= new->file_pos + new->len)
2840                         break;
2841
2842                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2843
2844                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2845                 if (key.offset + num_bytes < new->file_pos)
2846                         goto next;
2847
2848                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2849                 if (!disk_bytenr)
2850                         goto next;
2851
2852                 extent_offset = btrfs_file_extent_offset(l, extent);
2853
2854                 old = kmalloc(sizeof(*old), GFP_NOFS);
2855                 if (!old)
2856                         goto out_free_path;
2857
2858                 offset = max(new->file_pos, key.offset);
2859                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2860
2861                 old->bytenr = disk_bytenr;
2862                 old->extent_offset = extent_offset;
2863                 old->offset = offset - key.offset;
2864                 old->len = end - offset;
2865                 old->new = new;
2866                 old->count = 0;
2867                 list_add_tail(&old->list, &new->head);
2868 next:
2869                 path->slots[0]++;
2870                 cond_resched();
2871         }
2872
2873         btrfs_free_path(path);
2874         atomic_inc(&fs_info->defrag_running);
2875
2876         return new;
2877
2878 out_free_path:
2879         btrfs_free_path(path);
2880 out_kfree:
2881         free_sa_defrag_extent(new);
2882         return NULL;
2883 }
2884
2885 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2886                                          u64 start, u64 len)
2887 {
2888         struct btrfs_block_group_cache *cache;
2889
2890         cache = btrfs_lookup_block_group(fs_info, start);
2891         ASSERT(cache);
2892
2893         spin_lock(&cache->lock);
2894         cache->delalloc_bytes -= len;
2895         spin_unlock(&cache->lock);
2896
2897         btrfs_put_block_group(cache);
2898 }
2899
2900 /* as ordered data IO finishes, this gets called so we can finish
2901  * an ordered extent if the range of bytes in the file it covers are
2902  * fully written.
2903  */
2904 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2905 {
2906         struct inode *inode = ordered_extent->inode;
2907         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2908         struct btrfs_root *root = BTRFS_I(inode)->root;
2909         struct btrfs_trans_handle *trans = NULL;
2910         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2911         struct extent_state *cached_state = NULL;
2912         struct new_sa_defrag_extent *new = NULL;
2913         int compress_type = 0;
2914         int ret = 0;
2915         u64 logical_len = ordered_extent->len;
2916         bool nolock;
2917         bool truncated = false;
2918         bool range_locked = false;
2919         bool clear_new_delalloc_bytes = false;
2920
2921         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2922             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2923             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2924                 clear_new_delalloc_bytes = true;
2925
2926         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2927
2928         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2929                 ret = -EIO;
2930                 goto out;
2931         }
2932
2933         btrfs_free_io_failure_record(BTRFS_I(inode),
2934                         ordered_extent->file_offset,
2935                         ordered_extent->file_offset +
2936                         ordered_extent->len - 1);
2937
2938         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2939                 truncated = true;
2940                 logical_len = ordered_extent->truncated_len;
2941                 /* Truncated the entire extent, don't bother adding */
2942                 if (!logical_len)
2943                         goto out;
2944         }
2945
2946         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2947                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2948
2949                 /*
2950                  * For mwrite(mmap + memset to write) case, we still reserve
2951                  * space for NOCOW range.
2952                  * As NOCOW won't cause a new delayed ref, just free the space
2953                  */
2954                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
2955                                        ordered_extent->len);
2956                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2957                 if (nolock)
2958                         trans = btrfs_join_transaction_nolock(root);
2959                 else
2960                         trans = btrfs_join_transaction(root);
2961                 if (IS_ERR(trans)) {
2962                         ret = PTR_ERR(trans);
2963                         trans = NULL;
2964                         goto out;
2965                 }
2966                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
2967                 ret = btrfs_update_inode_fallback(trans, root, inode);
2968                 if (ret) /* -ENOMEM or corruption */
2969                         btrfs_abort_transaction(trans, ret);
2970                 goto out;
2971         }
2972
2973         range_locked = true;
2974         lock_extent_bits(io_tree, ordered_extent->file_offset,
2975                          ordered_extent->file_offset + ordered_extent->len - 1,
2976                          &cached_state);
2977
2978         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2979                         ordered_extent->file_offset + ordered_extent->len - 1,
2980                         EXTENT_DEFRAG, 0, cached_state);
2981         if (ret) {
2982                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2983                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2984                         /* the inode is shared */
2985                         new = record_old_file_extents(inode, ordered_extent);
2986
2987                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2988                         ordered_extent->file_offset + ordered_extent->len - 1,
2989                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2990         }
2991
2992         if (nolock)
2993                 trans = btrfs_join_transaction_nolock(root);
2994         else
2995                 trans = btrfs_join_transaction(root);
2996         if (IS_ERR(trans)) {
2997                 ret = PTR_ERR(trans);
2998                 trans = NULL;
2999                 goto out;
3000         }
3001
3002         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3003
3004         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3005                 compress_type = ordered_extent->compress_type;
3006         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3007                 BUG_ON(compress_type);
3008                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3009                                        ordered_extent->len);
3010                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3011                                                 ordered_extent->file_offset,
3012                                                 ordered_extent->file_offset +
3013                                                 logical_len);
3014         } else {
3015                 BUG_ON(root == fs_info->tree_root);
3016                 ret = insert_reserved_file_extent(trans, inode,
3017                                                 ordered_extent->file_offset,
3018                                                 ordered_extent->start,
3019                                                 ordered_extent->disk_len,
3020                                                 logical_len, logical_len,
3021                                                 compress_type, 0, 0,
3022                                                 BTRFS_FILE_EXTENT_REG);
3023                 if (!ret)
3024                         btrfs_release_delalloc_bytes(fs_info,
3025                                                      ordered_extent->start,
3026                                                      ordered_extent->disk_len);
3027         }
3028         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3029                            ordered_extent->file_offset, ordered_extent->len,
3030                            trans->transid);
3031         if (ret < 0) {
3032                 btrfs_abort_transaction(trans, ret);
3033                 goto out;
3034         }
3035
3036         add_pending_csums(trans, inode, &ordered_extent->list);
3037
3038         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3039         ret = btrfs_update_inode_fallback(trans, root, inode);
3040         if (ret) { /* -ENOMEM or corruption */
3041                 btrfs_abort_transaction(trans, ret);
3042                 goto out;
3043         }
3044         ret = 0;
3045 out:
3046         if (range_locked || clear_new_delalloc_bytes) {
3047                 unsigned int clear_bits = 0;
3048
3049                 if (range_locked)
3050                         clear_bits |= EXTENT_LOCKED;
3051                 if (clear_new_delalloc_bytes)
3052                         clear_bits |= EXTENT_DELALLOC_NEW;
3053                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3054                                  ordered_extent->file_offset,
3055                                  ordered_extent->file_offset +
3056                                  ordered_extent->len - 1,
3057                                  clear_bits,
3058                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3059                                  0, &cached_state, GFP_NOFS);
3060         }
3061
3062         if (trans)
3063                 btrfs_end_transaction(trans);
3064
3065         if (ret || truncated) {
3066                 u64 start, end;
3067
3068                 if (truncated)
3069                         start = ordered_extent->file_offset + logical_len;
3070                 else
3071                         start = ordered_extent->file_offset;
3072                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3073                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
3074
3075                 /* Drop the cache for the part of the extent we didn't write. */
3076                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3077
3078                 /*
3079                  * If the ordered extent had an IOERR or something else went
3080                  * wrong we need to return the space for this ordered extent
3081                  * back to the allocator.  We only free the extent in the
3082                  * truncated case if we didn't write out the extent at all.
3083                  */
3084                 if ((ret || !logical_len) &&
3085                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3086                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3087                         btrfs_free_reserved_extent(fs_info,
3088                                                    ordered_extent->start,
3089                                                    ordered_extent->disk_len, 1);
3090         }
3091
3092
3093         /*
3094          * This needs to be done to make sure anybody waiting knows we are done
3095          * updating everything for this ordered extent.
3096          */
3097         btrfs_remove_ordered_extent(inode, ordered_extent);
3098
3099         /* for snapshot-aware defrag */
3100         if (new) {
3101                 if (ret) {
3102                         free_sa_defrag_extent(new);
3103                         atomic_dec(&fs_info->defrag_running);
3104                 } else {
3105                         relink_file_extents(new);
3106                 }
3107         }
3108
3109         /* once for us */
3110         btrfs_put_ordered_extent(ordered_extent);
3111         /* once for the tree */
3112         btrfs_put_ordered_extent(ordered_extent);
3113
3114         return ret;
3115 }
3116
3117 static void finish_ordered_fn(struct btrfs_work *work)
3118 {
3119         struct btrfs_ordered_extent *ordered_extent;
3120         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3121         btrfs_finish_ordered_io(ordered_extent);
3122 }
3123
3124 static void btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3125                                 struct extent_state *state, int uptodate)
3126 {
3127         struct inode *inode = page->mapping->host;
3128         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3129         struct btrfs_ordered_extent *ordered_extent = NULL;
3130         struct btrfs_workqueue *wq;
3131         btrfs_work_func_t func;
3132
3133         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3134
3135         ClearPagePrivate2(page);
3136         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3137                                             end - start + 1, uptodate))
3138                 return;
3139
3140         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
3141                 wq = fs_info->endio_freespace_worker;
3142                 func = btrfs_freespace_write_helper;
3143         } else {
3144                 wq = fs_info->endio_write_workers;
3145                 func = btrfs_endio_write_helper;
3146         }
3147
3148         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3149                         NULL);
3150         btrfs_queue_work(wq, &ordered_extent->work);
3151 }
3152
3153 static int __readpage_endio_check(struct inode *inode,
3154                                   struct btrfs_io_bio *io_bio,
3155                                   int icsum, struct page *page,
3156                                   int pgoff, u64 start, size_t len)
3157 {
3158         char *kaddr;
3159         u32 csum_expected;
3160         u32 csum = ~(u32)0;
3161
3162         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3163
3164         kaddr = kmap_atomic(page);
3165         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3166         btrfs_csum_final(csum, (u8 *)&csum);
3167         if (csum != csum_expected)
3168                 goto zeroit;
3169
3170         kunmap_atomic(kaddr);
3171         return 0;
3172 zeroit:
3173         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3174                                     io_bio->mirror_num);
3175         memset(kaddr + pgoff, 1, len);
3176         flush_dcache_page(page);
3177         kunmap_atomic(kaddr);
3178         return -EIO;
3179 }
3180
3181 /*
3182  * when reads are done, we need to check csums to verify the data is correct
3183  * if there's a match, we allow the bio to finish.  If not, the code in
3184  * extent_io.c will try to find good copies for us.
3185  */
3186 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3187                                       u64 phy_offset, struct page *page,
3188                                       u64 start, u64 end, int mirror)
3189 {
3190         size_t offset = start - page_offset(page);
3191         struct inode *inode = page->mapping->host;
3192         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3193         struct btrfs_root *root = BTRFS_I(inode)->root;
3194
3195         if (PageChecked(page)) {
3196                 ClearPageChecked(page);
3197                 return 0;
3198         }
3199
3200         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3201                 return 0;
3202
3203         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3204             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3205                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3206                 return 0;
3207         }
3208
3209         phy_offset >>= inode->i_sb->s_blocksize_bits;
3210         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3211                                       start, (size_t)(end - start + 1));
3212 }
3213
3214 void btrfs_add_delayed_iput(struct inode *inode)
3215 {
3216         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3217         struct btrfs_inode *binode = BTRFS_I(inode);
3218
3219         if (atomic_add_unless(&inode->i_count, -1, 1))
3220                 return;
3221
3222         spin_lock(&fs_info->delayed_iput_lock);
3223         if (binode->delayed_iput_count == 0) {
3224                 ASSERT(list_empty(&binode->delayed_iput));
3225                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3226         } else {
3227                 binode->delayed_iput_count++;
3228         }
3229         spin_unlock(&fs_info->delayed_iput_lock);
3230 }
3231
3232 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3233 {
3234
3235         spin_lock(&fs_info->delayed_iput_lock);
3236         while (!list_empty(&fs_info->delayed_iputs)) {
3237                 struct btrfs_inode *inode;
3238
3239                 inode = list_first_entry(&fs_info->delayed_iputs,
3240                                 struct btrfs_inode, delayed_iput);
3241                 if (inode->delayed_iput_count) {
3242                         inode->delayed_iput_count--;
3243                         list_move_tail(&inode->delayed_iput,
3244                                         &fs_info->delayed_iputs);
3245                 } else {
3246                         list_del_init(&inode->delayed_iput);
3247                 }
3248                 spin_unlock(&fs_info->delayed_iput_lock);
3249                 iput(&inode->vfs_inode);
3250                 spin_lock(&fs_info->delayed_iput_lock);
3251         }
3252         spin_unlock(&fs_info->delayed_iput_lock);
3253 }
3254
3255 /*
3256  * This is called in transaction commit time. If there are no orphan
3257  * files in the subvolume, it removes orphan item and frees block_rsv
3258  * structure.
3259  */
3260 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3261                               struct btrfs_root *root)
3262 {
3263         struct btrfs_fs_info *fs_info = root->fs_info;
3264         struct btrfs_block_rsv *block_rsv;
3265         int ret;
3266
3267         if (atomic_read(&root->orphan_inodes) ||
3268             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3269                 return;
3270
3271         spin_lock(&root->orphan_lock);
3272         if (atomic_read(&root->orphan_inodes)) {
3273                 spin_unlock(&root->orphan_lock);
3274                 return;
3275         }
3276
3277         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3278                 spin_unlock(&root->orphan_lock);
3279                 return;
3280         }
3281
3282         block_rsv = root->orphan_block_rsv;
3283         root->orphan_block_rsv = NULL;
3284         spin_unlock(&root->orphan_lock);
3285
3286         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3287             btrfs_root_refs(&root->root_item) > 0) {
3288                 ret = btrfs_del_orphan_item(trans, fs_info->tree_root,
3289                                             root->root_key.objectid);
3290                 if (ret)
3291                         btrfs_abort_transaction(trans, ret);
3292                 else
3293                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3294                                   &root->state);
3295         }
3296
3297         if (block_rsv) {
3298                 WARN_ON(block_rsv->size > 0);
3299                 btrfs_free_block_rsv(fs_info, block_rsv);
3300         }
3301 }
3302
3303 /*
3304  * This creates an orphan entry for the given inode in case something goes
3305  * wrong in the middle of an unlink/truncate.
3306  *
3307  * NOTE: caller of this function should reserve 5 units of metadata for
3308  *       this function.
3309  */
3310 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3311                 struct btrfs_inode *inode)
3312 {
3313         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
3314         struct btrfs_root *root = inode->root;
3315         struct btrfs_block_rsv *block_rsv = NULL;
3316         int reserve = 0;
3317         int insert = 0;
3318         int ret;
3319
3320         if (!root->orphan_block_rsv) {
3321                 block_rsv = btrfs_alloc_block_rsv(fs_info,
3322                                                   BTRFS_BLOCK_RSV_TEMP);
3323                 if (!block_rsv)
3324                         return -ENOMEM;
3325         }
3326
3327         spin_lock(&root->orphan_lock);
3328         if (!root->orphan_block_rsv) {
3329                 root->orphan_block_rsv = block_rsv;
3330         } else if (block_rsv) {
3331                 btrfs_free_block_rsv(fs_info, block_rsv);
3332                 block_rsv = NULL;
3333         }
3334
3335         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3336                               &inode->runtime_flags)) {
3337 #if 0
3338                 /*
3339                  * For proper ENOSPC handling, we should do orphan
3340                  * cleanup when mounting. But this introduces backward
3341                  * compatibility issue.
3342                  */
3343                 if (!xchg(&root->orphan_item_inserted, 1))
3344                         insert = 2;
3345                 else
3346                         insert = 1;
3347 #endif
3348                 insert = 1;
3349                 atomic_inc(&root->orphan_inodes);
3350         }
3351
3352         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3353                               &inode->runtime_flags))
3354                 reserve = 1;
3355         spin_unlock(&root->orphan_lock);
3356
3357         /* grab metadata reservation from transaction handle */
3358         if (reserve) {
3359                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3360                 ASSERT(!ret);
3361                 if (ret) {
3362                         atomic_dec(&root->orphan_inodes);
3363                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3364                                   &inode->runtime_flags);
3365                         if (insert)
3366                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3367                                           &inode->runtime_flags);
3368                         return ret;
3369                 }
3370         }
3371
3372         /* insert an orphan item to track this unlinked/truncated file */
3373         if (insert >= 1) {
3374                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3375                 if (ret) {
3376                         atomic_dec(&root->orphan_inodes);
3377                         if (reserve) {
3378                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3379                                           &inode->runtime_flags);
3380                                 btrfs_orphan_release_metadata(inode);
3381                         }
3382                         if (ret != -EEXIST) {
3383                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3384                                           &inode->runtime_flags);
3385                                 btrfs_abort_transaction(trans, ret);
3386                                 return ret;
3387                         }
3388                 }
3389                 ret = 0;
3390         }
3391
3392         /* insert an orphan item to track subvolume contains orphan files */
3393         if (insert >= 2) {
3394                 ret = btrfs_insert_orphan_item(trans, fs_info->tree_root,
3395                                                root->root_key.objectid);
3396                 if (ret && ret != -EEXIST) {
3397                         btrfs_abort_transaction(trans, ret);
3398                         return ret;
3399                 }
3400         }
3401         return 0;
3402 }
3403
3404 /*
3405  * We have done the truncate/delete so we can go ahead and remove the orphan
3406  * item for this particular inode.
3407  */
3408 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3409                             struct btrfs_inode *inode)
3410 {
3411         struct btrfs_root *root = inode->root;
3412         int delete_item = 0;
3413         int release_rsv = 0;
3414         int ret = 0;
3415
3416         spin_lock(&root->orphan_lock);
3417         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3418                                &inode->runtime_flags))
3419                 delete_item = 1;
3420
3421         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3422                                &inode->runtime_flags))
3423                 release_rsv = 1;
3424         spin_unlock(&root->orphan_lock);
3425
3426         if (delete_item) {
3427                 atomic_dec(&root->orphan_inodes);
3428                 if (trans)
3429                         ret = btrfs_del_orphan_item(trans, root,
3430                                                     btrfs_ino(inode));
3431         }
3432
3433         if (release_rsv)
3434                 btrfs_orphan_release_metadata(inode);
3435
3436         return ret;
3437 }
3438
3439 /*
3440  * this cleans up any orphans that may be left on the list from the last use
3441  * of this root.
3442  */
3443 int btrfs_orphan_cleanup(struct btrfs_root *root)
3444 {
3445         struct btrfs_fs_info *fs_info = root->fs_info;
3446         struct btrfs_path *path;
3447         struct extent_buffer *leaf;
3448         struct btrfs_key key, found_key;
3449         struct btrfs_trans_handle *trans;
3450         struct inode *inode;
3451         u64 last_objectid = 0;
3452         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3453
3454         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3455                 return 0;
3456
3457         path = btrfs_alloc_path();
3458         if (!path) {
3459                 ret = -ENOMEM;
3460                 goto out;
3461         }
3462         path->reada = READA_BACK;
3463
3464         key.objectid = BTRFS_ORPHAN_OBJECTID;
3465         key.type = BTRFS_ORPHAN_ITEM_KEY;
3466         key.offset = (u64)-1;
3467
3468         while (1) {
3469                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3470                 if (ret < 0)
3471                         goto out;
3472
3473                 /*
3474                  * if ret == 0 means we found what we were searching for, which
3475                  * is weird, but possible, so only screw with path if we didn't
3476                  * find the key and see if we have stuff that matches
3477                  */
3478                 if (ret > 0) {
3479                         ret = 0;
3480                         if (path->slots[0] == 0)
3481                                 break;
3482                         path->slots[0]--;
3483                 }
3484
3485                 /* pull out the item */
3486                 leaf = path->nodes[0];
3487                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3488
3489                 /* make sure the item matches what we want */
3490                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3491                         break;
3492                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3493                         break;
3494
3495                 /* release the path since we're done with it */
3496                 btrfs_release_path(path);
3497
3498                 /*
3499                  * this is where we are basically btrfs_lookup, without the
3500                  * crossing root thing.  we store the inode number in the
3501                  * offset of the orphan item.
3502                  */
3503
3504                 if (found_key.offset == last_objectid) {
3505                         btrfs_err(fs_info,
3506                                   "Error removing orphan entry, stopping orphan cleanup");
3507                         ret = -EINVAL;
3508                         goto out;
3509                 }
3510
3511                 last_objectid = found_key.offset;
3512
3513                 found_key.objectid = found_key.offset;
3514                 found_key.type = BTRFS_INODE_ITEM_KEY;
3515                 found_key.offset = 0;
3516                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3517                 ret = PTR_ERR_OR_ZERO(inode);
3518                 if (ret && ret != -ENOENT)
3519                         goto out;
3520
3521                 if (ret == -ENOENT && root == fs_info->tree_root) {
3522                         struct btrfs_root *dead_root;
3523                         struct btrfs_fs_info *fs_info = root->fs_info;
3524                         int is_dead_root = 0;
3525
3526                         /*
3527                          * this is an orphan in the tree root. Currently these
3528                          * could come from 2 sources:
3529                          *  a) a snapshot deletion in progress
3530                          *  b) a free space cache inode
3531                          * We need to distinguish those two, as the snapshot
3532                          * orphan must not get deleted.
3533                          * find_dead_roots already ran before us, so if this
3534                          * is a snapshot deletion, we should find the root
3535                          * in the dead_roots list
3536                          */
3537                         spin_lock(&fs_info->trans_lock);
3538                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3539                                             root_list) {
3540                                 if (dead_root->root_key.objectid ==
3541                                     found_key.objectid) {
3542                                         is_dead_root = 1;
3543                                         break;
3544                                 }
3545                         }
3546                         spin_unlock(&fs_info->trans_lock);
3547                         if (is_dead_root) {
3548                                 /* prevent this orphan from being found again */
3549                                 key.offset = found_key.objectid - 1;
3550                                 continue;
3551                         }
3552                 }
3553                 /*
3554                  * Inode is already gone but the orphan item is still there,
3555                  * kill the orphan item.
3556                  */
3557                 if (ret == -ENOENT) {
3558                         trans = btrfs_start_transaction(root, 1);
3559                         if (IS_ERR(trans)) {
3560                                 ret = PTR_ERR(trans);
3561                                 goto out;
3562                         }
3563                         btrfs_debug(fs_info, "auto deleting %Lu",
3564                                     found_key.objectid);
3565                         ret = btrfs_del_orphan_item(trans, root,
3566                                                     found_key.objectid);
3567                         btrfs_end_transaction(trans);
3568                         if (ret)
3569                                 goto out;
3570                         continue;
3571                 }
3572
3573                 /*
3574                  * add this inode to the orphan list so btrfs_orphan_del does
3575                  * the proper thing when we hit it
3576                  */
3577                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3578                         &BTRFS_I(inode)->runtime_flags);
3579                 atomic_inc(&root->orphan_inodes);
3580
3581                 /* if we have links, this was a truncate, lets do that */
3582                 if (inode->i_nlink) {
3583                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3584                                 iput(inode);
3585                                 continue;
3586                         }
3587                         nr_truncate++;
3588
3589                         /* 1 for the orphan item deletion. */
3590                         trans = btrfs_start_transaction(root, 1);
3591                         if (IS_ERR(trans)) {
3592                                 iput(inode);
3593                                 ret = PTR_ERR(trans);
3594                                 goto out;
3595                         }
3596                         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
3597                         btrfs_end_transaction(trans);
3598                         if (ret) {
3599                                 iput(inode);
3600                                 goto out;
3601                         }
3602
3603                         ret = btrfs_truncate(inode);
3604                         if (ret)
3605                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
3606                 } else {
3607                         nr_unlink++;
3608                 }
3609
3610                 /* this will do delete_inode and everything for us */
3611                 iput(inode);
3612                 if (ret)
3613                         goto out;
3614         }
3615         /* release the path since we're done with it */
3616         btrfs_release_path(path);
3617
3618         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3619
3620         if (root->orphan_block_rsv)
3621                 btrfs_block_rsv_release(fs_info, root->orphan_block_rsv,
3622                                         (u64)-1);
3623
3624         if (root->orphan_block_rsv ||
3625             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3626                 trans = btrfs_join_transaction(root);
3627                 if (!IS_ERR(trans))
3628                         btrfs_end_transaction(trans);
3629         }
3630
3631         if (nr_unlink)
3632                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3633         if (nr_truncate)
3634                 btrfs_debug(fs_info, "truncated %d orphans", nr_truncate);
3635
3636 out:
3637         if (ret)
3638                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3639         btrfs_free_path(path);
3640         return ret;
3641 }
3642
3643 /*
3644  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3645  * don't find any xattrs, we know there can't be any acls.
3646  *
3647  * slot is the slot the inode is in, objectid is the objectid of the inode
3648  */
3649 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3650                                           int slot, u64 objectid,
3651                                           int *first_xattr_slot)
3652 {
3653         u32 nritems = btrfs_header_nritems(leaf);
3654         struct btrfs_key found_key;
3655         static u64 xattr_access = 0;
3656         static u64 xattr_default = 0;
3657         int scanned = 0;
3658
3659         if (!xattr_access) {
3660                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3661                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3662                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3663                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3664         }
3665
3666         slot++;
3667         *first_xattr_slot = -1;
3668         while (slot < nritems) {
3669                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3670
3671                 /* we found a different objectid, there must not be acls */
3672                 if (found_key.objectid != objectid)
3673                         return 0;
3674
3675                 /* we found an xattr, assume we've got an acl */
3676                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3677                         if (*first_xattr_slot == -1)
3678                                 *first_xattr_slot = slot;
3679                         if (found_key.offset == xattr_access ||
3680                             found_key.offset == xattr_default)
3681                                 return 1;
3682                 }
3683
3684                 /*
3685                  * we found a key greater than an xattr key, there can't
3686                  * be any acls later on
3687                  */
3688                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3689                         return 0;
3690
3691                 slot++;
3692                 scanned++;
3693
3694                 /*
3695                  * it goes inode, inode backrefs, xattrs, extents,
3696                  * so if there are a ton of hard links to an inode there can
3697                  * be a lot of backrefs.  Don't waste time searching too hard,
3698                  * this is just an optimization
3699                  */
3700                 if (scanned >= 8)
3701                         break;
3702         }
3703         /* we hit the end of the leaf before we found an xattr or
3704          * something larger than an xattr.  We have to assume the inode
3705          * has acls
3706          */
3707         if (*first_xattr_slot == -1)
3708                 *first_xattr_slot = slot;
3709         return 1;
3710 }
3711
3712 /*
3713  * read an inode from the btree into the in-memory inode
3714  */
3715 static int btrfs_read_locked_inode(struct inode *inode)
3716 {
3717         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3718         struct btrfs_path *path;
3719         struct extent_buffer *leaf;
3720         struct btrfs_inode_item *inode_item;
3721         struct btrfs_root *root = BTRFS_I(inode)->root;
3722         struct btrfs_key location;
3723         unsigned long ptr;
3724         int maybe_acls;
3725         u32 rdev;
3726         int ret;
3727         bool filled = false;
3728         int first_xattr_slot;
3729
3730         ret = btrfs_fill_inode(inode, &rdev);
3731         if (!ret)
3732                 filled = true;
3733
3734         path = btrfs_alloc_path();
3735         if (!path) {
3736                 ret = -ENOMEM;
3737                 goto make_bad;
3738         }
3739
3740         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3741
3742         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3743         if (ret) {
3744                 if (ret > 0)
3745                         ret = -ENOENT;
3746                 goto make_bad;
3747         }
3748
3749         leaf = path->nodes[0];
3750
3751         if (filled)
3752                 goto cache_index;
3753
3754         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3755                                     struct btrfs_inode_item);
3756         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3757         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3758         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3759         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3760         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3761
3762         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3763         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3764
3765         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3766         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3767
3768         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3769         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3770
3771         BTRFS_I(inode)->i_otime.tv_sec =
3772                 btrfs_timespec_sec(leaf, &inode_item->otime);
3773         BTRFS_I(inode)->i_otime.tv_nsec =
3774                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3775
3776         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3777         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3778         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3779
3780         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3781         inode->i_generation = BTRFS_I(inode)->generation;
3782         inode->i_rdev = 0;
3783         rdev = btrfs_inode_rdev(leaf, inode_item);
3784
3785         BTRFS_I(inode)->index_cnt = (u64)-1;
3786         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3787
3788 cache_index:
3789         /*
3790          * If we were modified in the current generation and evicted from memory
3791          * and then re-read we need to do a full sync since we don't have any
3792          * idea about which extents were modified before we were evicted from
3793          * cache.
3794          *
3795          * This is required for both inode re-read from disk and delayed inode
3796          * in delayed_nodes_tree.
3797          */
3798         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3799                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3800                         &BTRFS_I(inode)->runtime_flags);
3801
3802         /*
3803          * We don't persist the id of the transaction where an unlink operation
3804          * against the inode was last made. So here we assume the inode might
3805          * have been evicted, and therefore the exact value of last_unlink_trans
3806          * lost, and set it to last_trans to avoid metadata inconsistencies
3807          * between the inode and its parent if the inode is fsync'ed and the log
3808          * replayed. For example, in the scenario:
3809          *
3810          * touch mydir/foo
3811          * ln mydir/foo mydir/bar
3812          * sync
3813          * unlink mydir/bar
3814          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3815          * xfs_io -c fsync mydir/foo
3816          * <power failure>
3817          * mount fs, triggers fsync log replay
3818          *
3819          * We must make sure that when we fsync our inode foo we also log its
3820          * parent inode, otherwise after log replay the parent still has the
3821          * dentry with the "bar" name but our inode foo has a link count of 1
3822          * and doesn't have an inode ref with the name "bar" anymore.
3823          *
3824          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3825          * but it guarantees correctness at the expense of occasional full
3826          * transaction commits on fsync if our inode is a directory, or if our
3827          * inode is not a directory, logging its parent unnecessarily.
3828          */
3829         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3830
3831         path->slots[0]++;
3832         if (inode->i_nlink != 1 ||
3833             path->slots[0] >= btrfs_header_nritems(leaf))
3834                 goto cache_acl;
3835
3836         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3837         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3838                 goto cache_acl;
3839
3840         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3841         if (location.type == BTRFS_INODE_REF_KEY) {
3842                 struct btrfs_inode_ref *ref;
3843
3844                 ref = (struct btrfs_inode_ref *)ptr;
3845                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3846         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3847                 struct btrfs_inode_extref *extref;
3848
3849                 extref = (struct btrfs_inode_extref *)ptr;
3850                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3851                                                                      extref);
3852         }
3853 cache_acl:
3854         /*
3855          * try to precache a NULL acl entry for files that don't have
3856          * any xattrs or acls
3857          */
3858         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3859                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3860         if (first_xattr_slot != -1) {
3861                 path->slots[0] = first_xattr_slot;
3862                 ret = btrfs_load_inode_props(inode, path);
3863                 if (ret)
3864                         btrfs_err(fs_info,
3865                                   "error loading props for ino %llu (root %llu): %d",
3866                                   btrfs_ino(BTRFS_I(inode)),
3867                                   root->root_key.objectid, ret);
3868         }
3869         btrfs_free_path(path);
3870
3871         if (!maybe_acls)
3872                 cache_no_acl(inode);
3873
3874         switch (inode->i_mode & S_IFMT) {
3875         case S_IFREG:
3876                 inode->i_mapping->a_ops = &btrfs_aops;
3877                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3878                 inode->i_fop = &btrfs_file_operations;
3879                 inode->i_op = &btrfs_file_inode_operations;
3880                 break;
3881         case S_IFDIR:
3882                 inode->i_fop = &btrfs_dir_file_operations;
3883                 inode->i_op = &btrfs_dir_inode_operations;
3884                 break;
3885         case S_IFLNK:
3886                 inode->i_op = &btrfs_symlink_inode_operations;
3887                 inode_nohighmem(inode);
3888                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3889                 break;
3890         default:
3891                 inode->i_op = &btrfs_special_inode_operations;
3892                 init_special_inode(inode, inode->i_mode, rdev);
3893                 break;
3894         }
3895
3896         btrfs_update_iflags(inode);
3897         return 0;
3898
3899 make_bad:
3900         btrfs_free_path(path);
3901         make_bad_inode(inode);
3902         return ret;
3903 }
3904
3905 /*
3906  * given a leaf and an inode, copy the inode fields into the leaf
3907  */
3908 static void fill_inode_item(struct btrfs_trans_handle *trans,
3909                             struct extent_buffer *leaf,
3910                             struct btrfs_inode_item *item,
3911                             struct inode *inode)
3912 {
3913         struct btrfs_map_token token;
3914
3915         btrfs_init_map_token(&token);
3916
3917         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3918         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3919         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3920                                    &token);
3921         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3922         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3923
3924         btrfs_set_token_timespec_sec(leaf, &item->atime,
3925                                      inode->i_atime.tv_sec, &token);
3926         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3927                                       inode->i_atime.tv_nsec, &token);
3928
3929         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3930                                      inode->i_mtime.tv_sec, &token);
3931         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3932                                       inode->i_mtime.tv_nsec, &token);
3933
3934         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3935                                      inode->i_ctime.tv_sec, &token);
3936         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3937                                       inode->i_ctime.tv_nsec, &token);
3938
3939         btrfs_set_token_timespec_sec(leaf, &item->otime,
3940                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3941         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3942                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3943
3944         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3945                                      &token);
3946         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3947                                          &token);
3948         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3949         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3950         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3951         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3952         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3953 }
3954
3955 /*
3956  * copy everything in the in-memory inode into the btree.
3957  */
3958 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3959                                 struct btrfs_root *root, struct inode *inode)
3960 {
3961         struct btrfs_inode_item *inode_item;
3962         struct btrfs_path *path;
3963         struct extent_buffer *leaf;
3964         int ret;
3965
3966         path = btrfs_alloc_path();
3967         if (!path)
3968                 return -ENOMEM;
3969
3970         path->leave_spinning = 1;
3971         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3972                                  1);
3973         if (ret) {
3974                 if (ret > 0)
3975                         ret = -ENOENT;
3976                 goto failed;
3977         }
3978
3979         leaf = path->nodes[0];
3980         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3981                                     struct btrfs_inode_item);
3982
3983         fill_inode_item(trans, leaf, inode_item, inode);
3984         btrfs_mark_buffer_dirty(leaf);
3985         btrfs_set_inode_last_trans(trans, inode);
3986         ret = 0;
3987 failed:
3988         btrfs_free_path(path);
3989         return ret;
3990 }
3991
3992 /*
3993  * copy everything in the in-memory inode into the btree.
3994  */
3995 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3996                                 struct btrfs_root *root, struct inode *inode)
3997 {
3998         struct btrfs_fs_info *fs_info = root->fs_info;
3999         int ret;
4000
4001         /*
4002          * If the inode is a free space inode, we can deadlock during commit
4003          * if we put it into the delayed code.
4004          *
4005          * The data relocation inode should also be directly updated
4006          * without delay
4007          */
4008         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
4009             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
4010             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
4011                 btrfs_update_root_times(trans, root);
4012
4013                 ret = btrfs_delayed_update_inode(trans, root, inode);
4014                 if (!ret)
4015                         btrfs_set_inode_last_trans(trans, inode);
4016                 return ret;
4017         }
4018
4019         return btrfs_update_inode_item(trans, root, inode);
4020 }
4021
4022 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
4023                                          struct btrfs_root *root,
4024                                          struct inode *inode)
4025 {
4026         int ret;
4027
4028         ret = btrfs_update_inode(trans, root, inode);
4029         if (ret == -ENOSPC)
4030                 return btrfs_update_inode_item(trans, root, inode);
4031         return ret;
4032 }
4033
4034 /*
4035  * unlink helper that gets used here in inode.c and in the tree logging
4036  * recovery code.  It remove a link in a directory with a given name, and
4037  * also drops the back refs in the inode to the directory
4038  */
4039 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4040                                 struct btrfs_root *root,
4041                                 struct btrfs_inode *dir,
4042                                 struct btrfs_inode *inode,
4043                                 const char *name, int name_len)
4044 {
4045         struct btrfs_fs_info *fs_info = root->fs_info;
4046         struct btrfs_path *path;
4047         int ret = 0;
4048         struct extent_buffer *leaf;
4049         struct btrfs_dir_item *di;
4050         struct btrfs_key key;
4051         u64 index;
4052         u64 ino = btrfs_ino(inode);
4053         u64 dir_ino = btrfs_ino(dir);
4054
4055         path = btrfs_alloc_path();
4056         if (!path) {
4057                 ret = -ENOMEM;
4058                 goto out;
4059         }
4060
4061         path->leave_spinning = 1;
4062         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4063                                     name, name_len, -1);
4064         if (IS_ERR(di)) {
4065                 ret = PTR_ERR(di);
4066                 goto err;
4067         }
4068         if (!di) {
4069                 ret = -ENOENT;
4070                 goto err;
4071         }
4072         leaf = path->nodes[0];
4073         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4074         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4075         if (ret)
4076                 goto err;
4077         btrfs_release_path(path);
4078
4079         /*
4080          * If we don't have dir index, we have to get it by looking up
4081          * the inode ref, since we get the inode ref, remove it directly,
4082          * it is unnecessary to do delayed deletion.
4083          *
4084          * But if we have dir index, needn't search inode ref to get it.
4085          * Since the inode ref is close to the inode item, it is better
4086          * that we delay to delete it, and just do this deletion when
4087          * we update the inode item.
4088          */
4089         if (inode->dir_index) {
4090                 ret = btrfs_delayed_delete_inode_ref(inode);
4091                 if (!ret) {
4092                         index = inode->dir_index;
4093                         goto skip_backref;
4094                 }
4095         }
4096
4097         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4098                                   dir_ino, &index);
4099         if (ret) {
4100                 btrfs_info(fs_info,
4101                         "failed to delete reference to %.*s, inode %llu parent %llu",
4102                         name_len, name, ino, dir_ino);
4103                 btrfs_abort_transaction(trans, ret);
4104                 goto err;
4105         }
4106 skip_backref:
4107         ret = btrfs_delete_delayed_dir_index(trans, fs_info, dir, index);
4108         if (ret) {
4109                 btrfs_abort_transaction(trans, ret);
4110                 goto err;
4111         }
4112
4113         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
4114                         dir_ino);
4115         if (ret != 0 && ret != -ENOENT) {
4116                 btrfs_abort_transaction(trans, ret);
4117                 goto err;
4118         }
4119
4120         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
4121                         index);
4122         if (ret == -ENOENT)
4123                 ret = 0;
4124         else if (ret)
4125                 btrfs_abort_transaction(trans, ret);
4126 err:
4127         btrfs_free_path(path);
4128         if (ret)
4129                 goto out;
4130
4131         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
4132         inode_inc_iversion(&inode->vfs_inode);
4133         inode_inc_iversion(&dir->vfs_inode);
4134         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4135                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4136         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4137 out:
4138         return ret;
4139 }
4140
4141 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4142                        struct btrfs_root *root,
4143                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4144                        const char *name, int name_len)
4145 {
4146         int ret;
4147         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4148         if (!ret) {
4149                 drop_nlink(&inode->vfs_inode);
4150                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4151         }
4152         return ret;
4153 }
4154
4155 /*
4156  * helper to start transaction for unlink and rmdir.
4157  *
4158  * unlink and rmdir are special in btrfs, they do not always free space, so
4159  * if we cannot make our reservations the normal way try and see if there is
4160  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4161  * allow the unlink to occur.
4162  */
4163 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4164 {
4165         struct btrfs_root *root = BTRFS_I(dir)->root;
4166
4167         /*
4168          * 1 for the possible orphan item
4169          * 1 for the dir item
4170          * 1 for the dir index
4171          * 1 for the inode ref
4172          * 1 for the inode
4173          */
4174         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4175 }
4176
4177 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4178 {
4179         struct btrfs_root *root = BTRFS_I(dir)->root;
4180         struct btrfs_trans_handle *trans;
4181         struct inode *inode = d_inode(dentry);
4182         int ret;
4183
4184         trans = __unlink_start_trans(dir);
4185         if (IS_ERR(trans))
4186                 return PTR_ERR(trans);
4187
4188         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4189                         0);
4190
4191         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4192                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4193                         dentry->d_name.len);
4194         if (ret)
4195                 goto out;
4196
4197         if (inode->i_nlink == 0) {
4198                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4199                 if (ret)
4200                         goto out;
4201         }
4202
4203 out:
4204         btrfs_end_transaction(trans);
4205         btrfs_btree_balance_dirty(root->fs_info);
4206         return ret;
4207 }
4208
4209 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4210                         struct btrfs_root *root,
4211                         struct inode *dir, u64 objectid,
4212                         const char *name, int name_len)
4213 {
4214         struct btrfs_fs_info *fs_info = root->fs_info;
4215         struct btrfs_path *path;
4216         struct extent_buffer *leaf;
4217         struct btrfs_dir_item *di;
4218         struct btrfs_key key;
4219         u64 index;
4220         int ret;
4221         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4222
4223         path = btrfs_alloc_path();
4224         if (!path)
4225                 return -ENOMEM;
4226
4227         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4228                                    name, name_len, -1);
4229         if (IS_ERR_OR_NULL(di)) {
4230                 if (!di)
4231                         ret = -ENOENT;
4232                 else
4233                         ret = PTR_ERR(di);
4234                 goto out;
4235         }
4236
4237         leaf = path->nodes[0];
4238         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4239         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4240         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4241         if (ret) {
4242                 btrfs_abort_transaction(trans, ret);
4243                 goto out;
4244         }
4245         btrfs_release_path(path);
4246
4247         ret = btrfs_del_root_ref(trans, fs_info, objectid,
4248                                  root->root_key.objectid, dir_ino,
4249                                  &index, name, name_len);
4250         if (ret < 0) {
4251                 if (ret != -ENOENT) {
4252                         btrfs_abort_transaction(trans, ret);
4253                         goto out;
4254                 }
4255                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4256                                                  name, name_len);
4257                 if (IS_ERR_OR_NULL(di)) {
4258                         if (!di)
4259                                 ret = -ENOENT;
4260                         else
4261                                 ret = PTR_ERR(di);
4262                         btrfs_abort_transaction(trans, ret);
4263                         goto out;
4264                 }
4265
4266                 leaf = path->nodes[0];
4267                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4268                 btrfs_release_path(path);
4269                 index = key.offset;
4270         }
4271         btrfs_release_path(path);
4272
4273         ret = btrfs_delete_delayed_dir_index(trans, fs_info, BTRFS_I(dir), index);
4274         if (ret) {
4275                 btrfs_abort_transaction(trans, ret);
4276                 goto out;
4277         }
4278
4279         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4280         inode_inc_iversion(dir);
4281         dir->i_mtime = dir->i_ctime = current_time(dir);
4282         ret = btrfs_update_inode_fallback(trans, root, dir);
4283         if (ret)
4284                 btrfs_abort_transaction(trans, ret);
4285 out:
4286         btrfs_free_path(path);
4287         return ret;
4288 }
4289
4290 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4291 {
4292         struct inode *inode = d_inode(dentry);
4293         int err = 0;
4294         struct btrfs_root *root = BTRFS_I(dir)->root;
4295         struct btrfs_trans_handle *trans;
4296         u64 last_unlink_trans;
4297
4298         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4299                 return -ENOTEMPTY;
4300         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4301                 return -EPERM;
4302
4303         trans = __unlink_start_trans(dir);
4304         if (IS_ERR(trans))
4305                 return PTR_ERR(trans);
4306
4307         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4308                 err = btrfs_unlink_subvol(trans, root, dir,
4309                                           BTRFS_I(inode)->location.objectid,
4310                                           dentry->d_name.name,
4311                                           dentry->d_name.len);
4312                 goto out;
4313         }
4314
4315         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4316         if (err)
4317                 goto out;
4318
4319         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4320
4321         /* now the directory is empty */
4322         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4323                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4324                         dentry->d_name.len);
4325         if (!err) {
4326                 btrfs_i_size_write(BTRFS_I(inode), 0);
4327                 /*
4328                  * Propagate the last_unlink_trans value of the deleted dir to
4329                  * its parent directory. This is to prevent an unrecoverable
4330                  * log tree in the case we do something like this:
4331                  * 1) create dir foo
4332                  * 2) create snapshot under dir foo
4333                  * 3) delete the snapshot
4334                  * 4) rmdir foo
4335                  * 5) mkdir foo
4336                  * 6) fsync foo or some file inside foo
4337                  */
4338                 if (last_unlink_trans >= trans->transid)
4339                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4340         }
4341 out:
4342         btrfs_end_transaction(trans);
4343         btrfs_btree_balance_dirty(root->fs_info);
4344
4345         return err;
4346 }
4347
4348 static int truncate_space_check(struct btrfs_trans_handle *trans,
4349                                 struct btrfs_root *root,
4350                                 u64 bytes_deleted)
4351 {
4352         struct btrfs_fs_info *fs_info = root->fs_info;
4353         int ret;
4354
4355         /*
4356          * This is only used to apply pressure to the enospc system, we don't
4357          * intend to use this reservation at all.
4358          */
4359         bytes_deleted = btrfs_csum_bytes_to_leaves(fs_info, bytes_deleted);
4360         bytes_deleted *= fs_info->nodesize;
4361         ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
4362                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4363         if (!ret) {
4364                 trace_btrfs_space_reservation(fs_info, "transaction",
4365                                               trans->transid,
4366                                               bytes_deleted, 1);
4367                 trans->bytes_reserved += bytes_deleted;
4368         }
4369         return ret;
4370
4371 }
4372
4373 /*
4374  * Return this if we need to call truncate_block for the last bit of the
4375  * truncate.
4376  */
4377 #define NEED_TRUNCATE_BLOCK 1
4378
4379 /*
4380  * this can truncate away extent items, csum items and directory items.
4381  * It starts at a high offset and removes keys until it can't find
4382  * any higher than new_size
4383  *
4384  * csum items that cross the new i_size are truncated to the new size
4385  * as well.
4386  *
4387  * min_type is the minimum key type to truncate down to.  If set to 0, this
4388  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4389  */
4390 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4391                                struct btrfs_root *root,
4392                                struct inode *inode,
4393                                u64 new_size, u32 min_type)
4394 {
4395         struct btrfs_fs_info *fs_info = root->fs_info;
4396         struct btrfs_path *path;
4397         struct extent_buffer *leaf;
4398         struct btrfs_file_extent_item *fi;
4399         struct btrfs_key key;
4400         struct btrfs_key found_key;
4401         u64 extent_start = 0;
4402         u64 extent_num_bytes = 0;
4403         u64 extent_offset = 0;
4404         u64 item_end = 0;
4405         u64 last_size = new_size;
4406         u32 found_type = (u8)-1;
4407         int found_extent;
4408         int del_item;
4409         int pending_del_nr = 0;
4410         int pending_del_slot = 0;
4411         int extent_type = -1;
4412         int ret;
4413         int err = 0;
4414         u64 ino = btrfs_ino(BTRFS_I(inode));
4415         u64 bytes_deleted = 0;
4416         bool be_nice = false;
4417         bool should_throttle = false;
4418         bool should_end = false;
4419
4420         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4421
4422         /*
4423          * for non-free space inodes and ref cows, we want to back off from
4424          * time to time
4425          */
4426         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4427             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4428                 be_nice = true;
4429
4430         path = btrfs_alloc_path();
4431         if (!path)
4432                 return -ENOMEM;
4433         path->reada = READA_BACK;
4434
4435         /*
4436          * We want to drop from the next block forward in case this new size is
4437          * not block aligned since we will be keeping the last block of the
4438          * extent just the way it is.
4439          */
4440         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4441             root == fs_info->tree_root)
4442                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4443                                         fs_info->sectorsize),
4444                                         (u64)-1, 0);
4445
4446         /*
4447          * This function is also used to drop the items in the log tree before
4448          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4449          * it is used to drop the loged items. So we shouldn't kill the delayed
4450          * items.
4451          */
4452         if (min_type == 0 && root == BTRFS_I(inode)->root)
4453                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4454
4455         key.objectid = ino;
4456         key.offset = (u64)-1;
4457         key.type = (u8)-1;
4458
4459 search_again:
4460         /*
4461          * with a 16K leaf size and 128MB extents, you can actually queue
4462          * up a huge file in a single leaf.  Most of the time that
4463          * bytes_deleted is > 0, it will be huge by the time we get here
4464          */
4465         if (be_nice && bytes_deleted > SZ_32M) {
4466                 if (btrfs_should_end_transaction(trans)) {
4467                         err = -EAGAIN;
4468                         goto error;
4469                 }
4470         }
4471
4472
4473         path->leave_spinning = 1;
4474         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4475         if (ret < 0) {
4476                 err = ret;
4477                 goto out;
4478         }
4479
4480         if (ret > 0) {
4481                 /* there are no items in the tree for us to truncate, we're
4482                  * done
4483                  */
4484                 if (path->slots[0] == 0)
4485                         goto out;
4486                 path->slots[0]--;
4487         }
4488
4489         while (1) {
4490                 fi = NULL;
4491                 leaf = path->nodes[0];
4492                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4493                 found_type = found_key.type;
4494
4495                 if (found_key.objectid != ino)
4496                         break;
4497
4498                 if (found_type < min_type)
4499                         break;
4500
4501                 item_end = found_key.offset;
4502                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4503                         fi = btrfs_item_ptr(leaf, path->slots[0],
4504                                             struct btrfs_file_extent_item);
4505                         extent_type = btrfs_file_extent_type(leaf, fi);
4506                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4507                                 item_end +=
4508                                     btrfs_file_extent_num_bytes(leaf, fi);
4509
4510                                 trace_btrfs_truncate_show_fi_regular(
4511                                         BTRFS_I(inode), leaf, fi,
4512                                         found_key.offset);
4513                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4514                                 item_end += btrfs_file_extent_inline_len(leaf,
4515                                                          path->slots[0], fi);
4516
4517                                 trace_btrfs_truncate_show_fi_inline(
4518                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4519                                         found_key.offset);
4520                         }
4521                         item_end--;
4522                 }
4523                 if (found_type > min_type) {
4524                         del_item = 1;
4525                 } else {
4526                         if (item_end < new_size)
4527                                 break;
4528                         if (found_key.offset >= new_size)
4529                                 del_item = 1;
4530                         else
4531                                 del_item = 0;
4532                 }
4533                 found_extent = 0;
4534                 /* FIXME, shrink the extent if the ref count is only 1 */
4535                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4536                         goto delete;
4537
4538                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4539                         u64 num_dec;
4540                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4541                         if (!del_item) {
4542                                 u64 orig_num_bytes =
4543                                         btrfs_file_extent_num_bytes(leaf, fi);
4544                                 extent_num_bytes = ALIGN(new_size -
4545                                                 found_key.offset,
4546                                                 fs_info->sectorsize);
4547                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4548                                                          extent_num_bytes);
4549                                 num_dec = (orig_num_bytes -
4550                                            extent_num_bytes);
4551                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4552                                              &root->state) &&
4553                                     extent_start != 0)
4554                                         inode_sub_bytes(inode, num_dec);
4555                                 btrfs_mark_buffer_dirty(leaf);
4556                         } else {
4557                                 extent_num_bytes =
4558                                         btrfs_file_extent_disk_num_bytes(leaf,
4559                                                                          fi);
4560                                 extent_offset = found_key.offset -
4561                                         btrfs_file_extent_offset(leaf, fi);
4562
4563                                 /* FIXME blocksize != 4096 */
4564                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4565                                 if (extent_start != 0) {
4566                                         found_extent = 1;
4567                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4568                                                      &root->state))
4569                                                 inode_sub_bytes(inode, num_dec);
4570                                 }
4571                         }
4572                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4573                         /*
4574                          * we can't truncate inline items that have had
4575                          * special encodings
4576                          */
4577                         if (!del_item &&
4578                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4579                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4580                             btrfs_file_extent_compression(leaf, fi) == 0) {
4581                                 u32 size = (u32)(new_size - found_key.offset);
4582
4583                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4584                                 size = btrfs_file_extent_calc_inline_size(size);
4585                                 btrfs_truncate_item(root->fs_info, path, size, 1);
4586                         } else if (!del_item) {
4587                                 /*
4588                                  * We have to bail so the last_size is set to
4589                                  * just before this extent.
4590                                  */
4591                                 err = NEED_TRUNCATE_BLOCK;
4592                                 break;
4593                         }
4594
4595                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4596                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4597                 }
4598 delete:
4599                 if (del_item)
4600                         last_size = found_key.offset;
4601                 else
4602                         last_size = new_size;
4603                 if (del_item) {
4604                         if (!pending_del_nr) {
4605                                 /* no pending yet, add ourselves */
4606                                 pending_del_slot = path->slots[0];
4607                                 pending_del_nr = 1;
4608                         } else if (pending_del_nr &&
4609                                    path->slots[0] + 1 == pending_del_slot) {
4610                                 /* hop on the pending chunk */
4611                                 pending_del_nr++;
4612                                 pending_del_slot = path->slots[0];
4613                         } else {
4614                                 BUG();
4615                         }
4616                 } else {
4617                         break;
4618                 }
4619                 should_throttle = false;
4620
4621                 if (found_extent &&
4622                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4623                      root == fs_info->tree_root)) {
4624                         btrfs_set_path_blocking(path);
4625                         bytes_deleted += extent_num_bytes;
4626                         ret = btrfs_free_extent(trans, root, extent_start,
4627                                                 extent_num_bytes, 0,
4628                                                 btrfs_header_owner(leaf),
4629                                                 ino, extent_offset);
4630                         BUG_ON(ret);
4631                         if (btrfs_should_throttle_delayed_refs(trans, fs_info))
4632                                 btrfs_async_run_delayed_refs(fs_info,
4633                                         trans->delayed_ref_updates * 2,
4634                                         trans->transid, 0);
4635                         if (be_nice) {
4636                                 if (truncate_space_check(trans, root,
4637                                                          extent_num_bytes)) {
4638                                         should_end = true;
4639                                 }
4640                                 if (btrfs_should_throttle_delayed_refs(trans,
4641                                                                        fs_info))
4642                                         should_throttle = true;
4643                         }
4644                 }
4645
4646                 if (found_type == BTRFS_INODE_ITEM_KEY)
4647                         break;
4648
4649                 if (path->slots[0] == 0 ||
4650                     path->slots[0] != pending_del_slot ||
4651                     should_throttle || should_end) {
4652                         if (pending_del_nr) {
4653                                 ret = btrfs_del_items(trans, root, path,
4654                                                 pending_del_slot,
4655                                                 pending_del_nr);
4656                                 if (ret) {
4657                                         btrfs_abort_transaction(trans, ret);
4658                                         goto error;
4659                                 }
4660                                 pending_del_nr = 0;
4661                         }
4662                         btrfs_release_path(path);
4663                         if (should_throttle) {
4664                                 unsigned long updates = trans->delayed_ref_updates;
4665                                 if (updates) {
4666                                         trans->delayed_ref_updates = 0;
4667                                         ret = btrfs_run_delayed_refs(trans,
4668                                                                    fs_info,
4669                                                                    updates * 2);
4670                                         if (ret && !err)
4671                                                 err = ret;
4672                                 }
4673                         }
4674                         /*
4675                          * if we failed to refill our space rsv, bail out
4676                          * and let the transaction restart
4677                          */
4678                         if (should_end) {
4679                                 err = -EAGAIN;
4680                                 goto error;
4681                         }
4682                         goto search_again;
4683                 } else {
4684                         path->slots[0]--;
4685                 }
4686         }
4687 out:
4688         if (pending_del_nr) {
4689                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4690                                       pending_del_nr);
4691                 if (ret)
4692                         btrfs_abort_transaction(trans, ret);
4693         }
4694 error:
4695         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4696                 ASSERT(last_size >= new_size);
4697                 if (!err && last_size > new_size)
4698                         last_size = new_size;
4699                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4700         }
4701
4702         btrfs_free_path(path);
4703
4704         if (be_nice && bytes_deleted > SZ_32M) {
4705                 unsigned long updates = trans->delayed_ref_updates;
4706                 if (updates) {
4707                         trans->delayed_ref_updates = 0;
4708                         ret = btrfs_run_delayed_refs(trans, fs_info,
4709                                                      updates * 2);
4710                         if (ret && !err)
4711                                 err = ret;
4712                 }
4713         }
4714         return err;
4715 }
4716
4717 /*
4718  * btrfs_truncate_block - read, zero a chunk and write a block
4719  * @inode - inode that we're zeroing
4720  * @from - the offset to start zeroing
4721  * @len - the length to zero, 0 to zero the entire range respective to the
4722  *      offset
4723  * @front - zero up to the offset instead of from the offset on
4724  *
4725  * This will find the block for the "from" offset and cow the block and zero the
4726  * part we want to zero.  This is used with truncate and hole punching.
4727  */
4728 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4729                         int front)
4730 {
4731         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4732         struct address_space *mapping = inode->i_mapping;
4733         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4734         struct btrfs_ordered_extent *ordered;
4735         struct extent_state *cached_state = NULL;
4736         struct extent_changeset *data_reserved = NULL;
4737         char *kaddr;
4738         u32 blocksize = fs_info->sectorsize;
4739         pgoff_t index = from >> PAGE_SHIFT;
4740         unsigned offset = from & (blocksize - 1);
4741         struct page *page;
4742         gfp_t mask = btrfs_alloc_write_mask(mapping);
4743         int ret = 0;
4744         u64 block_start;
4745         u64 block_end;
4746
4747         if ((offset & (blocksize - 1)) == 0 &&
4748             (!len || ((len & (blocksize - 1)) == 0)))
4749                 goto out;
4750
4751         block_start = round_down(from, blocksize);
4752         block_end = block_start + blocksize - 1;
4753
4754         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4755                                            block_start, blocksize);
4756         if (ret)
4757                 goto out;
4758
4759 again:
4760         page = find_or_create_page(mapping, index, mask);
4761         if (!page) {
4762                 btrfs_delalloc_release_space(inode, data_reserved,
4763                                              block_start, blocksize);
4764                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4765                 ret = -ENOMEM;
4766                 goto out;
4767         }
4768
4769         if (!PageUptodate(page)) {
4770                 ret = btrfs_readpage(NULL, page);
4771                 lock_page(page);
4772                 if (page->mapping != mapping) {
4773                         unlock_page(page);
4774                         put_page(page);
4775                         goto again;
4776                 }
4777                 if (!PageUptodate(page)) {
4778                         ret = -EIO;
4779                         goto out_unlock;
4780                 }
4781         }
4782         wait_on_page_writeback(page);
4783
4784         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4785         set_page_extent_mapped(page);
4786
4787         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4788         if (ordered) {
4789                 unlock_extent_cached(io_tree, block_start, block_end,
4790                                      &cached_state, GFP_NOFS);
4791                 unlock_page(page);
4792                 put_page(page);
4793                 btrfs_start_ordered_extent(inode, ordered, 1);
4794                 btrfs_put_ordered_extent(ordered);
4795                 goto again;
4796         }
4797
4798         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4799                           EXTENT_DIRTY | EXTENT_DELALLOC |
4800                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4801                           0, 0, &cached_state, GFP_NOFS);
4802
4803         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
4804                                         &cached_state, 0);
4805         if (ret) {
4806                 unlock_extent_cached(io_tree, block_start, block_end,
4807                                      &cached_state, GFP_NOFS);
4808                 goto out_unlock;
4809         }
4810
4811         if (offset != blocksize) {
4812                 if (!len)
4813                         len = blocksize - offset;
4814                 kaddr = kmap(page);
4815                 if (front)
4816                         memset(kaddr + (block_start - page_offset(page)),
4817                                 0, offset);
4818                 else
4819                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4820                                 0, len);
4821                 flush_dcache_page(page);
4822                 kunmap(page);
4823         }
4824         ClearPageChecked(page);
4825         set_page_dirty(page);
4826         unlock_extent_cached(io_tree, block_start, block_end, &cached_state,
4827                              GFP_NOFS);
4828
4829 out_unlock:
4830         if (ret)
4831                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
4832                                              blocksize);
4833         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4834         unlock_page(page);
4835         put_page(page);
4836 out:
4837         extent_changeset_free(data_reserved);
4838         return ret;
4839 }
4840
4841 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4842                              u64 offset, u64 len)
4843 {
4844         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4845         struct btrfs_trans_handle *trans;
4846         int ret;
4847
4848         /*
4849          * Still need to make sure the inode looks like it's been updated so
4850          * that any holes get logged if we fsync.
4851          */
4852         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4853                 BTRFS_I(inode)->last_trans = fs_info->generation;
4854                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4855                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4856                 return 0;
4857         }
4858
4859         /*
4860          * 1 - for the one we're dropping
4861          * 1 - for the one we're adding
4862          * 1 - for updating the inode.
4863          */
4864         trans = btrfs_start_transaction(root, 3);
4865         if (IS_ERR(trans))
4866                 return PTR_ERR(trans);
4867
4868         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4869         if (ret) {
4870                 btrfs_abort_transaction(trans, ret);
4871                 btrfs_end_transaction(trans);
4872                 return ret;
4873         }
4874
4875         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4876                         offset, 0, 0, len, 0, len, 0, 0, 0);
4877         if (ret)
4878                 btrfs_abort_transaction(trans, ret);
4879         else
4880                 btrfs_update_inode(trans, root, inode);
4881         btrfs_end_transaction(trans);
4882         return ret;
4883 }
4884
4885 /*
4886  * This function puts in dummy file extents for the area we're creating a hole
4887  * for.  So if we are truncating this file to a larger size we need to insert
4888  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4889  * the range between oldsize and size
4890  */
4891 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4892 {
4893         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4894         struct btrfs_root *root = BTRFS_I(inode)->root;
4895         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4896         struct extent_map *em = NULL;
4897         struct extent_state *cached_state = NULL;
4898         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4899         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
4900         u64 block_end = ALIGN(size, fs_info->sectorsize);
4901         u64 last_byte;
4902         u64 cur_offset;
4903         u64 hole_size;
4904         int err = 0;
4905
4906         /*
4907          * If our size started in the middle of a block we need to zero out the
4908          * rest of the block before we expand the i_size, otherwise we could
4909          * expose stale data.
4910          */
4911         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4912         if (err)
4913                 return err;
4914
4915         if (size <= hole_start)
4916                 return 0;
4917
4918         while (1) {
4919                 struct btrfs_ordered_extent *ordered;
4920
4921                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4922                                  &cached_state);
4923                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), hole_start,
4924                                                      block_end - hole_start);
4925                 if (!ordered)
4926                         break;
4927                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4928                                      &cached_state, GFP_NOFS);
4929                 btrfs_start_ordered_extent(inode, ordered, 1);
4930                 btrfs_put_ordered_extent(ordered);
4931         }
4932
4933         cur_offset = hole_start;
4934         while (1) {
4935                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
4936                                 block_end - cur_offset, 0);
4937                 if (IS_ERR(em)) {
4938                         err = PTR_ERR(em);
4939                         em = NULL;
4940                         break;
4941                 }
4942                 last_byte = min(extent_map_end(em), block_end);
4943                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
4944                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4945                         struct extent_map *hole_em;
4946                         hole_size = last_byte - cur_offset;
4947
4948                         err = maybe_insert_hole(root, inode, cur_offset,
4949                                                 hole_size);
4950                         if (err)
4951                                 break;
4952                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
4953                                                 cur_offset + hole_size - 1, 0);
4954                         hole_em = alloc_extent_map();
4955                         if (!hole_em) {
4956                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4957                                         &BTRFS_I(inode)->runtime_flags);
4958                                 goto next;
4959                         }
4960                         hole_em->start = cur_offset;
4961                         hole_em->len = hole_size;
4962                         hole_em->orig_start = cur_offset;
4963
4964                         hole_em->block_start = EXTENT_MAP_HOLE;
4965                         hole_em->block_len = 0;
4966                         hole_em->orig_block_len = 0;
4967                         hole_em->ram_bytes = hole_size;
4968                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
4969                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4970                         hole_em->generation = fs_info->generation;
4971
4972                         while (1) {
4973                                 write_lock(&em_tree->lock);
4974                                 err = add_extent_mapping(em_tree, hole_em, 1);
4975                                 write_unlock(&em_tree->lock);
4976                                 if (err != -EEXIST)
4977                                         break;
4978                                 btrfs_drop_extent_cache(BTRFS_I(inode),
4979                                                         cur_offset,
4980                                                         cur_offset +
4981                                                         hole_size - 1, 0);
4982                         }
4983                         free_extent_map(hole_em);
4984                 }
4985 next:
4986                 free_extent_map(em);
4987                 em = NULL;
4988                 cur_offset = last_byte;
4989                 if (cur_offset >= block_end)
4990                         break;
4991         }
4992         free_extent_map(em);
4993         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4994                              GFP_NOFS);
4995         return err;
4996 }
4997
4998 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4999 {
5000         struct btrfs_root *root = BTRFS_I(inode)->root;
5001         struct btrfs_trans_handle *trans;
5002         loff_t oldsize = i_size_read(inode);
5003         loff_t newsize = attr->ia_size;
5004         int mask = attr->ia_valid;
5005         int ret;
5006
5007         /*
5008          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5009          * special case where we need to update the times despite not having
5010          * these flags set.  For all other operations the VFS set these flags
5011          * explicitly if it wants a timestamp update.
5012          */
5013         if (newsize != oldsize) {
5014                 inode_inc_iversion(inode);
5015                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5016                         inode->i_ctime = inode->i_mtime =
5017                                 current_time(inode);
5018         }
5019
5020         if (newsize > oldsize) {
5021                 /*
5022                  * Don't do an expanding truncate while snapshotting is ongoing.
5023                  * This is to ensure the snapshot captures a fully consistent
5024                  * state of this file - if the snapshot captures this expanding
5025                  * truncation, it must capture all writes that happened before
5026                  * this truncation.
5027                  */
5028                 btrfs_wait_for_snapshot_creation(root);
5029                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5030                 if (ret) {
5031                         btrfs_end_write_no_snapshotting(root);
5032                         return ret;
5033                 }
5034
5035                 trans = btrfs_start_transaction(root, 1);
5036                 if (IS_ERR(trans)) {
5037                         btrfs_end_write_no_snapshotting(root);
5038                         return PTR_ERR(trans);
5039                 }
5040
5041                 i_size_write(inode, newsize);
5042                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5043                 pagecache_isize_extended(inode, oldsize, newsize);
5044                 ret = btrfs_update_inode(trans, root, inode);
5045                 btrfs_end_write_no_snapshotting(root);
5046                 btrfs_end_transaction(trans);
5047         } else {
5048
5049                 /*
5050                  * We're truncating a file that used to have good data down to
5051                  * zero. Make sure it gets into the ordered flush list so that
5052                  * any new writes get down to disk quickly.
5053                  */
5054                 if (newsize == 0)
5055                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5056                                 &BTRFS_I(inode)->runtime_flags);
5057
5058                 /*
5059                  * 1 for the orphan item we're going to add
5060                  * 1 for the orphan item deletion.
5061                  */
5062                 trans = btrfs_start_transaction(root, 2);
5063                 if (IS_ERR(trans))
5064                         return PTR_ERR(trans);
5065
5066                 /*
5067                  * We need to do this in case we fail at _any_ point during the
5068                  * actual truncate.  Once we do the truncate_setsize we could
5069                  * invalidate pages which forces any outstanding ordered io to
5070                  * be instantly completed which will give us extents that need
5071                  * to be truncated.  If we fail to get an orphan inode down we
5072                  * could have left over extents that were never meant to live,
5073                  * so we need to guarantee from this point on that everything
5074                  * will be consistent.
5075                  */
5076                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
5077                 btrfs_end_transaction(trans);
5078                 if (ret)
5079                         return ret;
5080
5081                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
5082                 truncate_setsize(inode, newsize);
5083
5084                 /* Disable nonlocked read DIO to avoid the end less truncate */
5085                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5086                 inode_dio_wait(inode);
5087                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5088
5089                 ret = btrfs_truncate(inode);
5090                 if (ret && inode->i_nlink) {
5091                         int err;
5092
5093                         /* To get a stable disk_i_size */
5094                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5095                         if (err) {
5096                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5097                                 return err;
5098                         }
5099
5100                         /*
5101                          * failed to truncate, disk_i_size is only adjusted down
5102                          * as we remove extents, so it should represent the true
5103                          * size of the inode, so reset the in memory size and
5104                          * delete our orphan entry.
5105                          */
5106                         trans = btrfs_join_transaction(root);
5107                         if (IS_ERR(trans)) {
5108                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5109                                 return ret;
5110                         }
5111                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5112                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
5113                         if (err)
5114                                 btrfs_abort_transaction(trans, err);
5115                         btrfs_end_transaction(trans);
5116                 }
5117         }
5118
5119         return ret;
5120 }
5121
5122 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5123 {
5124         struct inode *inode = d_inode(dentry);
5125         struct btrfs_root *root = BTRFS_I(inode)->root;
5126         int err;
5127
5128         if (btrfs_root_readonly(root))
5129                 return -EROFS;
5130
5131         err = setattr_prepare(dentry, attr);
5132         if (err)
5133                 return err;
5134
5135         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5136                 err = btrfs_setsize(inode, attr);
5137                 if (err)
5138                         return err;
5139         }
5140
5141         if (attr->ia_valid) {
5142                 setattr_copy(inode, attr);
5143                 inode_inc_iversion(inode);
5144                 err = btrfs_dirty_inode(inode);
5145
5146                 if (!err && attr->ia_valid & ATTR_MODE)
5147                         err = posix_acl_chmod(inode, inode->i_mode);
5148         }
5149
5150         return err;
5151 }
5152
5153 /*
5154  * While truncating the inode pages during eviction, we get the VFS calling
5155  * btrfs_invalidatepage() against each page of the inode. This is slow because
5156  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5157  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5158  * extent_state structures over and over, wasting lots of time.
5159  *
5160  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5161  * those expensive operations on a per page basis and do only the ordered io
5162  * finishing, while we release here the extent_map and extent_state structures,
5163  * without the excessive merging and splitting.
5164  */
5165 static void evict_inode_truncate_pages(struct inode *inode)
5166 {
5167         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5168         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5169         struct rb_node *node;
5170
5171         ASSERT(inode->i_state & I_FREEING);
5172         truncate_inode_pages_final(&inode->i_data);
5173
5174         write_lock(&map_tree->lock);
5175         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5176                 struct extent_map *em;
5177
5178                 node = rb_first(&map_tree->map);
5179                 em = rb_entry(node, struct extent_map, rb_node);
5180                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5181                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5182                 remove_extent_mapping(map_tree, em);
5183                 free_extent_map(em);
5184                 if (need_resched()) {
5185                         write_unlock(&map_tree->lock);
5186                         cond_resched();
5187                         write_lock(&map_tree->lock);
5188                 }
5189         }
5190         write_unlock(&map_tree->lock);
5191
5192         /*
5193          * Keep looping until we have no more ranges in the io tree.
5194          * We can have ongoing bios started by readpages (called from readahead)
5195          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5196          * still in progress (unlocked the pages in the bio but did not yet
5197          * unlocked the ranges in the io tree). Therefore this means some
5198          * ranges can still be locked and eviction started because before
5199          * submitting those bios, which are executed by a separate task (work
5200          * queue kthread), inode references (inode->i_count) were not taken
5201          * (which would be dropped in the end io callback of each bio).
5202          * Therefore here we effectively end up waiting for those bios and
5203          * anyone else holding locked ranges without having bumped the inode's
5204          * reference count - if we don't do it, when they access the inode's
5205          * io_tree to unlock a range it may be too late, leading to an
5206          * use-after-free issue.
5207          */
5208         spin_lock(&io_tree->lock);
5209         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5210                 struct extent_state *state;
5211                 struct extent_state *cached_state = NULL;
5212                 u64 start;
5213                 u64 end;
5214
5215                 node = rb_first(&io_tree->state);
5216                 state = rb_entry(node, struct extent_state, rb_node);
5217                 start = state->start;
5218                 end = state->end;
5219                 spin_unlock(&io_tree->lock);
5220
5221                 lock_extent_bits(io_tree, start, end, &cached_state);
5222
5223                 /*
5224                  * If still has DELALLOC flag, the extent didn't reach disk,
5225                  * and its reserved space won't be freed by delayed_ref.
5226                  * So we need to free its reserved space here.
5227                  * (Refer to comment in btrfs_invalidatepage, case 2)
5228                  *
5229                  * Note, end is the bytenr of last byte, so we need + 1 here.
5230                  */
5231                 if (state->state & EXTENT_DELALLOC)
5232                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5233
5234                 clear_extent_bit(io_tree, start, end,
5235                                  EXTENT_LOCKED | EXTENT_DIRTY |
5236                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5237                                  EXTENT_DEFRAG, 1, 1,
5238                                  &cached_state, GFP_NOFS);
5239
5240                 cond_resched();
5241                 spin_lock(&io_tree->lock);
5242         }
5243         spin_unlock(&io_tree->lock);
5244 }
5245
5246 void btrfs_evict_inode(struct inode *inode)
5247 {
5248         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5249         struct btrfs_trans_handle *trans;
5250         struct btrfs_root *root = BTRFS_I(inode)->root;
5251         struct btrfs_block_rsv *rsv, *global_rsv;
5252         int steal_from_global = 0;
5253         u64 min_size;
5254         int ret;
5255
5256         trace_btrfs_inode_evict(inode);
5257
5258         if (!root) {
5259                 kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
5260                 return;
5261         }
5262
5263         min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
5264
5265         evict_inode_truncate_pages(inode);
5266
5267         if (inode->i_nlink &&
5268             ((btrfs_root_refs(&root->root_item) != 0 &&
5269               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5270              btrfs_is_free_space_inode(BTRFS_I(inode))))
5271                 goto no_delete;
5272
5273         if (is_bad_inode(inode)) {
5274                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5275                 goto no_delete;
5276         }
5277         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5278         if (!special_file(inode->i_mode))
5279                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5280
5281         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5282
5283         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
5284                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5285                                  &BTRFS_I(inode)->runtime_flags));
5286                 goto no_delete;
5287         }
5288
5289         if (inode->i_nlink > 0) {
5290                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5291                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5292                 goto no_delete;
5293         }
5294
5295         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5296         if (ret) {
5297                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5298                 goto no_delete;
5299         }
5300
5301         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5302         if (!rsv) {
5303                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5304                 goto no_delete;
5305         }
5306         rsv->size = min_size;
5307         rsv->failfast = 1;
5308         global_rsv = &fs_info->global_block_rsv;
5309
5310         btrfs_i_size_write(BTRFS_I(inode), 0);
5311
5312         /*
5313          * This is a bit simpler than btrfs_truncate since we've already
5314          * reserved our space for our orphan item in the unlink, so we just
5315          * need to reserve some slack space in case we add bytes and update
5316          * inode item when doing the truncate.
5317          */
5318         while (1) {
5319                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5320                                              BTRFS_RESERVE_FLUSH_LIMIT);
5321
5322                 /*
5323                  * Try and steal from the global reserve since we will
5324                  * likely not use this space anyway, we want to try as
5325                  * hard as possible to get this to work.
5326                  */
5327                 if (ret)
5328                         steal_from_global++;
5329                 else
5330                         steal_from_global = 0;
5331                 ret = 0;
5332
5333                 /*
5334                  * steal_from_global == 0: we reserved stuff, hooray!
5335                  * steal_from_global == 1: we didn't reserve stuff, boo!
5336                  * steal_from_global == 2: we've committed, still not a lot of
5337                  * room but maybe we'll have room in the global reserve this
5338                  * time.
5339                  * steal_from_global == 3: abandon all hope!
5340                  */
5341                 if (steal_from_global > 2) {
5342                         btrfs_warn(fs_info,
5343                                    "Could not get space for a delete, will truncate on mount %d",
5344                                    ret);
5345                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5346                         btrfs_free_block_rsv(fs_info, rsv);
5347                         goto no_delete;
5348                 }
5349
5350                 trans = btrfs_join_transaction(root);
5351                 if (IS_ERR(trans)) {
5352                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5353                         btrfs_free_block_rsv(fs_info, rsv);
5354                         goto no_delete;
5355                 }
5356
5357                 /*
5358                  * We can't just steal from the global reserve, we need to make
5359                  * sure there is room to do it, if not we need to commit and try
5360                  * again.
5361                  */
5362                 if (steal_from_global) {
5363                         if (!btrfs_check_space_for_delayed_refs(trans, fs_info))
5364                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5365                                                               min_size, 0);
5366                         else
5367                                 ret = -ENOSPC;
5368                 }
5369
5370                 /*
5371                  * Couldn't steal from the global reserve, we have too much
5372                  * pending stuff built up, commit the transaction and try it
5373                  * again.
5374                  */
5375                 if (ret) {
5376                         ret = btrfs_commit_transaction(trans);
5377                         if (ret) {
5378                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5379                                 btrfs_free_block_rsv(fs_info, rsv);
5380                                 goto no_delete;
5381                         }
5382                         continue;
5383                 } else {
5384                         steal_from_global = 0;
5385                 }
5386
5387                 trans->block_rsv = rsv;
5388
5389                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5390                 if (ret != -ENOSPC && ret != -EAGAIN)
5391                         break;
5392
5393                 trans->block_rsv = &fs_info->trans_block_rsv;
5394                 btrfs_end_transaction(trans);
5395                 trans = NULL;
5396                 btrfs_btree_balance_dirty(fs_info);
5397         }
5398
5399         btrfs_free_block_rsv(fs_info, rsv);
5400
5401         /*
5402          * Errors here aren't a big deal, it just means we leave orphan items
5403          * in the tree.  They will be cleaned up on the next mount.
5404          */
5405         if (ret == 0) {
5406                 trans->block_rsv = root->orphan_block_rsv;
5407                 btrfs_orphan_del(trans, BTRFS_I(inode));
5408         } else {
5409                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5410         }
5411
5412         trans->block_rsv = &fs_info->trans_block_rsv;
5413         if (!(root == fs_info->tree_root ||
5414               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5415                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5416
5417         btrfs_end_transaction(trans);
5418         btrfs_btree_balance_dirty(fs_info);
5419 no_delete:
5420         btrfs_remove_delayed_node(BTRFS_I(inode));
5421         clear_inode(inode);
5422 }
5423
5424 /*
5425  * this returns the key found in the dir entry in the location pointer.
5426  * If no dir entries were found, location->objectid is 0.
5427  */
5428 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5429                                struct btrfs_key *location)
5430 {
5431         const char *name = dentry->d_name.name;
5432         int namelen = dentry->d_name.len;
5433         struct btrfs_dir_item *di;
5434         struct btrfs_path *path;
5435         struct btrfs_root *root = BTRFS_I(dir)->root;
5436         int ret = 0;
5437
5438         path = btrfs_alloc_path();
5439         if (!path)
5440                 return -ENOMEM;
5441
5442         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5443                         name, namelen, 0);
5444         if (IS_ERR(di))
5445                 ret = PTR_ERR(di);
5446
5447         if (IS_ERR_OR_NULL(di))
5448                 goto out_err;
5449
5450         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5451         if (location->type != BTRFS_INODE_ITEM_KEY &&
5452             location->type != BTRFS_ROOT_ITEM_KEY) {
5453                 btrfs_warn(root->fs_info,
5454 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5455                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5456                            location->objectid, location->type, location->offset);
5457                 goto out_err;
5458         }
5459 out:
5460         btrfs_free_path(path);
5461         return ret;
5462 out_err:
5463         location->objectid = 0;
5464         goto out;
5465 }
5466
5467 /*
5468  * when we hit a tree root in a directory, the btrfs part of the inode
5469  * needs to be changed to reflect the root directory of the tree root.  This
5470  * is kind of like crossing a mount point.
5471  */
5472 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5473                                     struct inode *dir,
5474                                     struct dentry *dentry,
5475                                     struct btrfs_key *location,
5476                                     struct btrfs_root **sub_root)
5477 {
5478         struct btrfs_path *path;
5479         struct btrfs_root *new_root;
5480         struct btrfs_root_ref *ref;
5481         struct extent_buffer *leaf;
5482         struct btrfs_key key;
5483         int ret;
5484         int err = 0;
5485
5486         path = btrfs_alloc_path();
5487         if (!path) {
5488                 err = -ENOMEM;
5489                 goto out;
5490         }
5491
5492         err = -ENOENT;
5493         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5494         key.type = BTRFS_ROOT_REF_KEY;
5495         key.offset = location->objectid;
5496
5497         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5498         if (ret) {
5499                 if (ret < 0)
5500                         err = ret;
5501                 goto out;
5502         }
5503
5504         leaf = path->nodes[0];
5505         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5506         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5507             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5508                 goto out;
5509
5510         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5511                                    (unsigned long)(ref + 1),
5512                                    dentry->d_name.len);
5513         if (ret)
5514                 goto out;
5515
5516         btrfs_release_path(path);
5517
5518         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5519         if (IS_ERR(new_root)) {
5520                 err = PTR_ERR(new_root);
5521                 goto out;
5522         }
5523
5524         *sub_root = new_root;
5525         location->objectid = btrfs_root_dirid(&new_root->root_item);
5526         location->type = BTRFS_INODE_ITEM_KEY;
5527         location->offset = 0;
5528         err = 0;
5529 out:
5530         btrfs_free_path(path);
5531         return err;
5532 }
5533
5534 static void inode_tree_add(struct inode *inode)
5535 {
5536         struct btrfs_root *root = BTRFS_I(inode)->root;
5537         struct btrfs_inode *entry;
5538         struct rb_node **p;
5539         struct rb_node *parent;
5540         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5541         u64 ino = btrfs_ino(BTRFS_I(inode));
5542
5543         if (inode_unhashed(inode))
5544                 return;
5545         parent = NULL;
5546         spin_lock(&root->inode_lock);
5547         p = &root->inode_tree.rb_node;
5548         while (*p) {
5549                 parent = *p;
5550                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5551
5552                 if (ino < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5553                         p = &parent->rb_left;
5554                 else if (ino > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5555                         p = &parent->rb_right;
5556                 else {
5557                         WARN_ON(!(entry->vfs_inode.i_state &
5558                                   (I_WILL_FREE | I_FREEING)));
5559                         rb_replace_node(parent, new, &root->inode_tree);
5560                         RB_CLEAR_NODE(parent);
5561                         spin_unlock(&root->inode_lock);
5562                         return;
5563                 }
5564         }
5565         rb_link_node(new, parent, p);
5566         rb_insert_color(new, &root->inode_tree);
5567         spin_unlock(&root->inode_lock);
5568 }
5569
5570 static void inode_tree_del(struct inode *inode)
5571 {
5572         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5573         struct btrfs_root *root = BTRFS_I(inode)->root;
5574         int empty = 0;
5575
5576         spin_lock(&root->inode_lock);
5577         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5578                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5579                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5580                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5581         }
5582         spin_unlock(&root->inode_lock);
5583
5584         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5585                 synchronize_srcu(&fs_info->subvol_srcu);
5586                 spin_lock(&root->inode_lock);
5587                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5588                 spin_unlock(&root->inode_lock);
5589                 if (empty)
5590                         btrfs_add_dead_root(root);
5591         }
5592 }
5593
5594 void btrfs_invalidate_inodes(struct btrfs_root *root)
5595 {
5596         struct btrfs_fs_info *fs_info = root->fs_info;
5597         struct rb_node *node;
5598         struct rb_node *prev;
5599         struct btrfs_inode *entry;
5600         struct inode *inode;
5601         u64 objectid = 0;
5602
5603         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
5604                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5605
5606         spin_lock(&root->inode_lock);
5607 again:
5608         node = root->inode_tree.rb_node;
5609         prev = NULL;
5610         while (node) {
5611                 prev = node;
5612                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5613
5614                 if (objectid < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5615                         node = node->rb_left;
5616                 else if (objectid > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5617                         node = node->rb_right;
5618                 else
5619                         break;
5620         }
5621         if (!node) {
5622                 while (prev) {
5623                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5624                         if (objectid <= btrfs_ino(BTRFS_I(&entry->vfs_inode))) {
5625                                 node = prev;
5626                                 break;
5627                         }
5628                         prev = rb_next(prev);
5629                 }
5630         }
5631         while (node) {
5632                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5633                 objectid = btrfs_ino(BTRFS_I(&entry->vfs_inode)) + 1;
5634                 inode = igrab(&entry->vfs_inode);
5635                 if (inode) {
5636                         spin_unlock(&root->inode_lock);
5637                         if (atomic_read(&inode->i_count) > 1)
5638                                 d_prune_aliases(inode);
5639                         /*
5640                          * btrfs_drop_inode will have it removed from
5641                          * the inode cache when its usage count
5642                          * hits zero.
5643                          */
5644                         iput(inode);
5645                         cond_resched();
5646                         spin_lock(&root->inode_lock);
5647                         goto again;
5648                 }
5649
5650                 if (cond_resched_lock(&root->inode_lock))
5651                         goto again;
5652
5653                 node = rb_next(node);
5654         }
5655         spin_unlock(&root->inode_lock);
5656 }
5657
5658 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5659 {
5660         struct btrfs_iget_args *args = p;
5661         inode->i_ino = args->location->objectid;
5662         memcpy(&BTRFS_I(inode)->location, args->location,
5663                sizeof(*args->location));
5664         BTRFS_I(inode)->root = args->root;
5665         return 0;
5666 }
5667
5668 static int btrfs_find_actor(struct inode *inode, void *opaque)
5669 {
5670         struct btrfs_iget_args *args = opaque;
5671         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5672                 args->root == BTRFS_I(inode)->root;
5673 }
5674
5675 static struct inode *btrfs_iget_locked(struct super_block *s,
5676                                        struct btrfs_key *location,
5677                                        struct btrfs_root *root)
5678 {
5679         struct inode *inode;
5680         struct btrfs_iget_args args;
5681         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5682
5683         args.location = location;
5684         args.root = root;
5685
5686         inode = iget5_locked(s, hashval, btrfs_find_actor,
5687                              btrfs_init_locked_inode,
5688                              (void *)&args);
5689         return inode;
5690 }
5691
5692 /* Get an inode object given its location and corresponding root.
5693  * Returns in *is_new if the inode was read from disk
5694  */
5695 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5696                          struct btrfs_root *root, int *new)
5697 {
5698         struct inode *inode;
5699
5700         inode = btrfs_iget_locked(s, location, root);
5701         if (!inode)
5702                 return ERR_PTR(-ENOMEM);
5703
5704         if (inode->i_state & I_NEW) {
5705                 int ret;
5706
5707                 ret = btrfs_read_locked_inode(inode);
5708                 if (!is_bad_inode(inode)) {
5709                         inode_tree_add(inode);
5710                         unlock_new_inode(inode);
5711                         if (new)
5712                                 *new = 1;
5713                 } else {
5714                         unlock_new_inode(inode);
5715                         iput(inode);
5716                         ASSERT(ret < 0);
5717                         inode = ERR_PTR(ret < 0 ? ret : -ESTALE);
5718                 }
5719         }
5720
5721         return inode;
5722 }
5723
5724 static struct inode *new_simple_dir(struct super_block *s,
5725                                     struct btrfs_key *key,
5726                                     struct btrfs_root *root)
5727 {
5728         struct inode *inode = new_inode(s);
5729
5730         if (!inode)
5731                 return ERR_PTR(-ENOMEM);
5732
5733         BTRFS_I(inode)->root = root;
5734         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5735         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5736
5737         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5738         inode->i_op = &btrfs_dir_ro_inode_operations;
5739         inode->i_opflags &= ~IOP_XATTR;
5740         inode->i_fop = &simple_dir_operations;
5741         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5742         inode->i_mtime = current_time(inode);
5743         inode->i_atime = inode->i_mtime;
5744         inode->i_ctime = inode->i_mtime;
5745         BTRFS_I(inode)->i_otime = inode->i_mtime;
5746
5747         return inode;
5748 }
5749
5750 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5751 {
5752         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5753         struct inode *inode;
5754         struct btrfs_root *root = BTRFS_I(dir)->root;
5755         struct btrfs_root *sub_root = root;
5756         struct btrfs_key location;
5757         int index;
5758         int ret = 0;
5759
5760         if (dentry->d_name.len > BTRFS_NAME_LEN)
5761                 return ERR_PTR(-ENAMETOOLONG);
5762
5763         ret = btrfs_inode_by_name(dir, dentry, &location);
5764         if (ret < 0)
5765                 return ERR_PTR(ret);
5766
5767         if (location.objectid == 0)
5768                 return ERR_PTR(-ENOENT);
5769
5770         if (location.type == BTRFS_INODE_ITEM_KEY) {
5771                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5772                 return inode;
5773         }
5774
5775         index = srcu_read_lock(&fs_info->subvol_srcu);
5776         ret = fixup_tree_root_location(fs_info, dir, dentry,
5777                                        &location, &sub_root);
5778         if (ret < 0) {
5779                 if (ret != -ENOENT)
5780                         inode = ERR_PTR(ret);
5781                 else
5782                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5783         } else {
5784                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5785         }
5786         srcu_read_unlock(&fs_info->subvol_srcu, index);
5787
5788         if (!IS_ERR(inode) && root != sub_root) {
5789                 down_read(&fs_info->cleanup_work_sem);
5790                 if (!sb_rdonly(inode->i_sb))
5791                         ret = btrfs_orphan_cleanup(sub_root);
5792                 up_read(&fs_info->cleanup_work_sem);
5793                 if (ret) {
5794                         iput(inode);
5795                         inode = ERR_PTR(ret);
5796                 }
5797         }
5798
5799         return inode;
5800 }
5801
5802 static int btrfs_dentry_delete(const struct dentry *dentry)
5803 {
5804         struct btrfs_root *root;
5805         struct inode *inode = d_inode(dentry);
5806
5807         if (!inode && !IS_ROOT(dentry))
5808                 inode = d_inode(dentry->d_parent);
5809
5810         if (inode) {
5811                 root = BTRFS_I(inode)->root;
5812                 if (btrfs_root_refs(&root->root_item) == 0)
5813                         return 1;
5814
5815                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5816                         return 1;
5817         }
5818         return 0;
5819 }
5820
5821 static void btrfs_dentry_release(struct dentry *dentry)
5822 {
5823         kfree(dentry->d_fsdata);
5824 }
5825
5826 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5827                                    unsigned int flags)
5828 {
5829         struct inode *inode;
5830
5831         inode = btrfs_lookup_dentry(dir, dentry);
5832         if (IS_ERR(inode)) {
5833                 if (PTR_ERR(inode) == -ENOENT)
5834                         inode = NULL;
5835                 else
5836                         return ERR_CAST(inode);
5837         }
5838
5839         return d_splice_alias(inode, dentry);
5840 }
5841
5842 unsigned char btrfs_filetype_table[] = {
5843         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5844 };
5845
5846 /*
5847  * All this infrastructure exists because dir_emit can fault, and we are holding
5848  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5849  * our information into that, and then dir_emit from the buffer.  This is
5850  * similar to what NFS does, only we don't keep the buffer around in pagecache
5851  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5852  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5853  * tree lock.
5854  */
5855 static int btrfs_opendir(struct inode *inode, struct file *file)
5856 {
5857         struct btrfs_file_private *private;
5858
5859         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5860         if (!private)
5861                 return -ENOMEM;
5862         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5863         if (!private->filldir_buf) {
5864                 kfree(private);
5865                 return -ENOMEM;
5866         }
5867         file->private_data = private;
5868         return 0;
5869 }
5870
5871 struct dir_entry {
5872         u64 ino;
5873         u64 offset;
5874         unsigned type;
5875         int name_len;
5876 };
5877
5878 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5879 {
5880         while (entries--) {
5881                 struct dir_entry *entry = addr;
5882                 char *name = (char *)(entry + 1);
5883
5884                 ctx->pos = entry->offset;
5885                 if (!dir_emit(ctx, name, entry->name_len, entry->ino,
5886                               entry->type))
5887                         return 1;
5888                 addr += sizeof(struct dir_entry) + entry->name_len;
5889                 ctx->pos++;
5890         }
5891         return 0;
5892 }
5893
5894 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5895 {
5896         struct inode *inode = file_inode(file);
5897         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5898         struct btrfs_root *root = BTRFS_I(inode)->root;
5899         struct btrfs_file_private *private = file->private_data;
5900         struct btrfs_dir_item *di;
5901         struct btrfs_key key;
5902         struct btrfs_key found_key;
5903         struct btrfs_path *path;
5904         void *addr;
5905         struct list_head ins_list;
5906         struct list_head del_list;
5907         int ret;
5908         struct extent_buffer *leaf;
5909         int slot;
5910         char *name_ptr;
5911         int name_len;
5912         int entries = 0;
5913         int total_len = 0;
5914         bool put = false;
5915         struct btrfs_key location;
5916
5917         if (!dir_emit_dots(file, ctx))
5918                 return 0;
5919
5920         path = btrfs_alloc_path();
5921         if (!path)
5922                 return -ENOMEM;
5923
5924         addr = private->filldir_buf;
5925         path->reada = READA_FORWARD;
5926
5927         INIT_LIST_HEAD(&ins_list);
5928         INIT_LIST_HEAD(&del_list);
5929         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5930
5931 again:
5932         key.type = BTRFS_DIR_INDEX_KEY;
5933         key.offset = ctx->pos;
5934         key.objectid = btrfs_ino(BTRFS_I(inode));
5935
5936         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5937         if (ret < 0)
5938                 goto err;
5939
5940         while (1) {
5941                 struct dir_entry *entry;
5942
5943                 leaf = path->nodes[0];
5944                 slot = path->slots[0];
5945                 if (slot >= btrfs_header_nritems(leaf)) {
5946                         ret = btrfs_next_leaf(root, path);
5947                         if (ret < 0)
5948                                 goto err;
5949                         else if (ret > 0)
5950                                 break;
5951                         continue;
5952                 }
5953
5954                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5955
5956                 if (found_key.objectid != key.objectid)
5957                         break;
5958                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
5959                         break;
5960                 if (found_key.offset < ctx->pos)
5961                         goto next;
5962                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
5963                         goto next;
5964                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5965                 if (verify_dir_item(fs_info, leaf, slot, di))
5966                         goto next;
5967
5968                 name_len = btrfs_dir_name_len(leaf, di);
5969                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
5970                     PAGE_SIZE) {
5971                         btrfs_release_path(path);
5972                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5973                         if (ret)
5974                                 goto nopos;
5975                         addr = private->filldir_buf;
5976                         entries = 0;
5977                         total_len = 0;
5978                         goto again;
5979                 }
5980
5981                 entry = addr;
5982                 entry->name_len = name_len;
5983                 name_ptr = (char *)(entry + 1);
5984                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
5985                                    name_len);
5986                 entry->type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5987                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
5988                 entry->ino = location.objectid;
5989                 entry->offset = found_key.offset;
5990                 entries++;
5991                 addr += sizeof(struct dir_entry) + name_len;
5992                 total_len += sizeof(struct dir_entry) + name_len;
5993 next:
5994                 path->slots[0]++;
5995         }
5996         btrfs_release_path(path);
5997
5998         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5999         if (ret)
6000                 goto nopos;
6001
6002         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
6003         if (ret)
6004                 goto nopos;
6005
6006         /*
6007          * Stop new entries from being returned after we return the last
6008          * entry.
6009          *
6010          * New directory entries are assigned a strictly increasing
6011          * offset.  This means that new entries created during readdir
6012          * are *guaranteed* to be seen in the future by that readdir.
6013          * This has broken buggy programs which operate on names as
6014          * they're returned by readdir.  Until we re-use freed offsets
6015          * we have this hack to stop new entries from being returned
6016          * under the assumption that they'll never reach this huge
6017          * offset.
6018          *
6019          * This is being careful not to overflow 32bit loff_t unless the
6020          * last entry requires it because doing so has broken 32bit apps
6021          * in the past.
6022          */
6023         if (ctx->pos >= INT_MAX)
6024                 ctx->pos = LLONG_MAX;
6025         else
6026                 ctx->pos = INT_MAX;
6027 nopos:
6028         ret = 0;
6029 err:
6030         if (put)
6031                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
6032         btrfs_free_path(path);
6033         return ret;
6034 }
6035
6036 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
6037 {
6038         struct btrfs_root *root = BTRFS_I(inode)->root;
6039         struct btrfs_trans_handle *trans;
6040         int ret = 0;
6041         bool nolock = false;
6042
6043         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6044                 return 0;
6045
6046         if (btrfs_fs_closing(root->fs_info) &&
6047                         btrfs_is_free_space_inode(BTRFS_I(inode)))
6048                 nolock = true;
6049
6050         if (wbc->sync_mode == WB_SYNC_ALL) {
6051                 if (nolock)
6052                         trans = btrfs_join_transaction_nolock(root);
6053                 else
6054                         trans = btrfs_join_transaction(root);
6055                 if (IS_ERR(trans))
6056                         return PTR_ERR(trans);
6057                 ret = btrfs_commit_transaction(trans);
6058         }
6059         return ret;
6060 }
6061
6062 /*
6063  * This is somewhat expensive, updating the tree every time the
6064  * inode changes.  But, it is most likely to find the inode in cache.
6065  * FIXME, needs more benchmarking...there are no reasons other than performance
6066  * to keep or drop this code.
6067  */
6068 static int btrfs_dirty_inode(struct inode *inode)
6069 {
6070         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6071         struct btrfs_root *root = BTRFS_I(inode)->root;
6072         struct btrfs_trans_handle *trans;
6073         int ret;
6074
6075         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6076                 return 0;
6077
6078         trans = btrfs_join_transaction(root);
6079         if (IS_ERR(trans))
6080                 return PTR_ERR(trans);
6081
6082         ret = btrfs_update_inode(trans, root, inode);
6083         if (ret && ret == -ENOSPC) {
6084                 /* whoops, lets try again with the full transaction */
6085                 btrfs_end_transaction(trans);
6086                 trans = btrfs_start_transaction(root, 1);
6087                 if (IS_ERR(trans))
6088                         return PTR_ERR(trans);
6089
6090                 ret = btrfs_update_inode(trans, root, inode);
6091         }
6092         btrfs_end_transaction(trans);
6093         if (BTRFS_I(inode)->delayed_node)
6094                 btrfs_balance_delayed_items(fs_info);
6095
6096         return ret;
6097 }
6098
6099 /*
6100  * This is a copy of file_update_time.  We need this so we can return error on
6101  * ENOSPC for updating the inode in the case of file write and mmap writes.
6102  */
6103 static int btrfs_update_time(struct inode *inode, struct timespec *now,
6104                              int flags)
6105 {
6106         struct btrfs_root *root = BTRFS_I(inode)->root;
6107
6108         if (btrfs_root_readonly(root))
6109                 return -EROFS;
6110
6111         if (flags & S_VERSION)
6112                 inode_inc_iversion(inode);
6113         if (flags & S_CTIME)
6114                 inode->i_ctime = *now;
6115         if (flags & S_MTIME)
6116                 inode->i_mtime = *now;
6117         if (flags & S_ATIME)
6118                 inode->i_atime = *now;
6119         return btrfs_dirty_inode(inode);
6120 }
6121
6122 /*
6123  * find the highest existing sequence number in a directory
6124  * and then set the in-memory index_cnt variable to reflect
6125  * free sequence numbers
6126  */
6127 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6128 {
6129         struct btrfs_root *root = inode->root;
6130         struct btrfs_key key, found_key;
6131         struct btrfs_path *path;
6132         struct extent_buffer *leaf;
6133         int ret;
6134
6135         key.objectid = btrfs_ino(inode);
6136         key.type = BTRFS_DIR_INDEX_KEY;
6137         key.offset = (u64)-1;
6138
6139         path = btrfs_alloc_path();
6140         if (!path)
6141                 return -ENOMEM;
6142
6143         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6144         if (ret < 0)
6145                 goto out;
6146         /* FIXME: we should be able to handle this */
6147         if (ret == 0)
6148                 goto out;
6149         ret = 0;
6150
6151         /*
6152          * MAGIC NUMBER EXPLANATION:
6153          * since we search a directory based on f_pos we have to start at 2
6154          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6155          * else has to start at 2
6156          */
6157         if (path->slots[0] == 0) {
6158                 inode->index_cnt = 2;
6159                 goto out;
6160         }
6161
6162         path->slots[0]--;
6163
6164         leaf = path->nodes[0];
6165         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6166
6167         if (found_key.objectid != btrfs_ino(inode) ||
6168             found_key.type != BTRFS_DIR_INDEX_KEY) {
6169                 inode->index_cnt = 2;
6170                 goto out;
6171         }
6172
6173         inode->index_cnt = found_key.offset + 1;
6174 out:
6175         btrfs_free_path(path);
6176         return ret;
6177 }
6178
6179 /*
6180  * helper to find a free sequence number in a given directory.  This current
6181  * code is very simple, later versions will do smarter things in the btree
6182  */
6183 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6184 {
6185         int ret = 0;
6186
6187         if (dir->index_cnt == (u64)-1) {
6188                 ret = btrfs_inode_delayed_dir_index_count(dir);
6189                 if (ret) {
6190                         ret = btrfs_set_inode_index_count(dir);
6191                         if (ret)
6192                                 return ret;
6193                 }
6194         }
6195
6196         *index = dir->index_cnt;
6197         dir->index_cnt++;
6198
6199         return ret;
6200 }
6201
6202 static int btrfs_insert_inode_locked(struct inode *inode)
6203 {
6204         struct btrfs_iget_args args;
6205         args.location = &BTRFS_I(inode)->location;
6206         args.root = BTRFS_I(inode)->root;
6207
6208         return insert_inode_locked4(inode,
6209                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6210                    btrfs_find_actor, &args);
6211 }
6212
6213 /*
6214  * Inherit flags from the parent inode.
6215  *
6216  * Currently only the compression flags and the cow flags are inherited.
6217  */
6218 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6219 {
6220         unsigned int flags;
6221
6222         if (!dir)
6223                 return;
6224
6225         flags = BTRFS_I(dir)->flags;
6226
6227         if (flags & BTRFS_INODE_NOCOMPRESS) {
6228                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6229                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6230         } else if (flags & BTRFS_INODE_COMPRESS) {
6231                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6232                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6233         }
6234
6235         if (flags & BTRFS_INODE_NODATACOW) {
6236                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6237                 if (S_ISREG(inode->i_mode))
6238                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6239         }
6240
6241         btrfs_update_iflags(inode);
6242 }
6243
6244 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6245                                      struct btrfs_root *root,
6246                                      struct inode *dir,
6247                                      const char *name, int name_len,
6248                                      u64 ref_objectid, u64 objectid,
6249                                      umode_t mode, u64 *index)
6250 {
6251         struct btrfs_fs_info *fs_info = root->fs_info;
6252         struct inode *inode;
6253         struct btrfs_inode_item *inode_item;
6254         struct btrfs_key *location;
6255         struct btrfs_path *path;
6256         struct btrfs_inode_ref *ref;
6257         struct btrfs_key key[2];
6258         u32 sizes[2];
6259         int nitems = name ? 2 : 1;
6260         unsigned long ptr;
6261         int ret;
6262
6263         path = btrfs_alloc_path();
6264         if (!path)
6265                 return ERR_PTR(-ENOMEM);
6266
6267         inode = new_inode(fs_info->sb);
6268         if (!inode) {
6269                 btrfs_free_path(path);
6270                 return ERR_PTR(-ENOMEM);
6271         }
6272
6273         /*
6274          * O_TMPFILE, set link count to 0, so that after this point,
6275          * we fill in an inode item with the correct link count.
6276          */
6277         if (!name)
6278                 set_nlink(inode, 0);
6279
6280         /*
6281          * we have to initialize this early, so we can reclaim the inode
6282          * number if we fail afterwards in this function.
6283          */
6284         inode->i_ino = objectid;
6285
6286         if (dir && name) {
6287                 trace_btrfs_inode_request(dir);
6288
6289                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6290                 if (ret) {
6291                         btrfs_free_path(path);
6292                         iput(inode);
6293                         return ERR_PTR(ret);
6294                 }
6295         } else if (dir) {
6296                 *index = 0;
6297         }
6298         /*
6299          * index_cnt is ignored for everything but a dir,
6300          * btrfs_get_inode_index_count has an explanation for the magic
6301          * number
6302          */
6303         BTRFS_I(inode)->index_cnt = 2;
6304         BTRFS_I(inode)->dir_index = *index;
6305         BTRFS_I(inode)->root = root;
6306         BTRFS_I(inode)->generation = trans->transid;
6307         inode->i_generation = BTRFS_I(inode)->generation;
6308
6309         /*
6310          * We could have gotten an inode number from somebody who was fsynced
6311          * and then removed in this same transaction, so let's just set full
6312          * sync since it will be a full sync anyway and this will blow away the
6313          * old info in the log.
6314          */
6315         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6316
6317         key[0].objectid = objectid;
6318         key[0].type = BTRFS_INODE_ITEM_KEY;
6319         key[0].offset = 0;
6320
6321         sizes[0] = sizeof(struct btrfs_inode_item);
6322
6323         if (name) {
6324                 /*
6325                  * Start new inodes with an inode_ref. This is slightly more
6326                  * efficient for small numbers of hard links since they will
6327                  * be packed into one item. Extended refs will kick in if we
6328                  * add more hard links than can fit in the ref item.
6329                  */
6330                 key[1].objectid = objectid;
6331                 key[1].type = BTRFS_INODE_REF_KEY;
6332                 key[1].offset = ref_objectid;
6333
6334                 sizes[1] = name_len + sizeof(*ref);
6335         }
6336
6337         location = &BTRFS_I(inode)->location;
6338         location->objectid = objectid;
6339         location->offset = 0;
6340         location->type = BTRFS_INODE_ITEM_KEY;
6341
6342         ret = btrfs_insert_inode_locked(inode);
6343         if (ret < 0)
6344                 goto fail;
6345
6346         path->leave_spinning = 1;
6347         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6348         if (ret != 0)
6349                 goto fail_unlock;
6350
6351         inode_init_owner(inode, dir, mode);
6352         inode_set_bytes(inode, 0);
6353
6354         inode->i_mtime = current_time(inode);
6355         inode->i_atime = inode->i_mtime;
6356         inode->i_ctime = inode->i_mtime;
6357         BTRFS_I(inode)->i_otime = inode->i_mtime;
6358
6359         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6360                                   struct btrfs_inode_item);
6361         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6362                              sizeof(*inode_item));
6363         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6364
6365         if (name) {
6366                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6367                                      struct btrfs_inode_ref);
6368                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6369                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6370                 ptr = (unsigned long)(ref + 1);
6371                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6372         }
6373
6374         btrfs_mark_buffer_dirty(path->nodes[0]);
6375         btrfs_free_path(path);
6376
6377         btrfs_inherit_iflags(inode, dir);
6378
6379         if (S_ISREG(mode)) {
6380                 if (btrfs_test_opt(fs_info, NODATASUM))
6381                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6382                 if (btrfs_test_opt(fs_info, NODATACOW))
6383                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6384                                 BTRFS_INODE_NODATASUM;
6385         }
6386
6387         inode_tree_add(inode);
6388
6389         trace_btrfs_inode_new(inode);
6390         btrfs_set_inode_last_trans(trans, inode);
6391
6392         btrfs_update_root_times(trans, root);
6393
6394         ret = btrfs_inode_inherit_props(trans, inode, dir);
6395         if (ret)
6396                 btrfs_err(fs_info,
6397                           "error inheriting props for ino %llu (root %llu): %d",
6398                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6399
6400         return inode;
6401
6402 fail_unlock:
6403         unlock_new_inode(inode);
6404 fail:
6405         if (dir && name)
6406                 BTRFS_I(dir)->index_cnt--;
6407         btrfs_free_path(path);
6408         iput(inode);
6409         return ERR_PTR(ret);
6410 }
6411
6412 static inline u8 btrfs_inode_type(struct inode *inode)
6413 {
6414         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6415 }
6416
6417 /*
6418  * utility function to add 'inode' into 'parent_inode' with
6419  * a give name and a given sequence number.
6420  * if 'add_backref' is true, also insert a backref from the
6421  * inode to the parent directory.
6422  */
6423 int btrfs_add_link(struct btrfs_trans_handle *trans,
6424                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6425                    const char *name, int name_len, int add_backref, u64 index)
6426 {
6427         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6428         int ret = 0;
6429         struct btrfs_key key;
6430         struct btrfs_root *root = parent_inode->root;
6431         u64 ino = btrfs_ino(inode);
6432         u64 parent_ino = btrfs_ino(parent_inode);
6433
6434         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6435                 memcpy(&key, &inode->root->root_key, sizeof(key));
6436         } else {
6437                 key.objectid = ino;
6438                 key.type = BTRFS_INODE_ITEM_KEY;
6439                 key.offset = 0;
6440         }
6441
6442         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6443                 ret = btrfs_add_root_ref(trans, fs_info, key.objectid,
6444                                          root->root_key.objectid, parent_ino,
6445                                          index, name, name_len);
6446         } else if (add_backref) {
6447                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6448                                              parent_ino, index);
6449         }
6450
6451         /* Nothing to clean up yet */
6452         if (ret)
6453                 return ret;
6454
6455         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6456                                     parent_inode, &key,
6457                                     btrfs_inode_type(&inode->vfs_inode), index);
6458         if (ret == -EEXIST || ret == -EOVERFLOW)
6459                 goto fail_dir_item;
6460         else if (ret) {
6461                 btrfs_abort_transaction(trans, ret);
6462                 return ret;
6463         }
6464
6465         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6466                            name_len * 2);
6467         inode_inc_iversion(&parent_inode->vfs_inode);
6468         parent_inode->vfs_inode.i_mtime = parent_inode->vfs_inode.i_ctime =
6469                 current_time(&parent_inode->vfs_inode);
6470         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6471         if (ret)
6472                 btrfs_abort_transaction(trans, ret);
6473         return ret;
6474
6475 fail_dir_item:
6476         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6477                 u64 local_index;
6478                 int err;
6479                 err = btrfs_del_root_ref(trans, fs_info, key.objectid,
6480                                          root->root_key.objectid, parent_ino,
6481                                          &local_index, name, name_len);
6482
6483         } else if (add_backref) {
6484                 u64 local_index;
6485                 int err;
6486
6487                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6488                                           ino, parent_ino, &local_index);
6489         }
6490         return ret;
6491 }
6492
6493 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6494                             struct btrfs_inode *dir, struct dentry *dentry,
6495                             struct btrfs_inode *inode, int backref, u64 index)
6496 {
6497         int err = btrfs_add_link(trans, dir, inode,
6498                                  dentry->d_name.name, dentry->d_name.len,
6499                                  backref, index);
6500         if (err > 0)
6501                 err = -EEXIST;
6502         return err;
6503 }
6504
6505 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6506                         umode_t mode, dev_t rdev)
6507 {
6508         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6509         struct btrfs_trans_handle *trans;
6510         struct btrfs_root *root = BTRFS_I(dir)->root;
6511         struct inode *inode = NULL;
6512         int err;
6513         int drop_inode = 0;
6514         u64 objectid;
6515         u64 index = 0;
6516
6517         /*
6518          * 2 for inode item and ref
6519          * 2 for dir items
6520          * 1 for xattr if selinux is on
6521          */
6522         trans = btrfs_start_transaction(root, 5);
6523         if (IS_ERR(trans))
6524                 return PTR_ERR(trans);
6525
6526         err = btrfs_find_free_ino(root, &objectid);
6527         if (err)
6528                 goto out_unlock;
6529
6530         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6531                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6532                         mode, &index);
6533         if (IS_ERR(inode)) {
6534                 err = PTR_ERR(inode);
6535                 goto out_unlock;
6536         }
6537
6538         /*
6539         * If the active LSM wants to access the inode during
6540         * d_instantiate it needs these. Smack checks to see
6541         * if the filesystem supports xattrs by looking at the
6542         * ops vector.
6543         */
6544         inode->i_op = &btrfs_special_inode_operations;
6545         init_special_inode(inode, inode->i_mode, rdev);
6546
6547         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6548         if (err)
6549                 goto out_unlock_inode;
6550
6551         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6552                         0, index);
6553         if (err) {
6554                 goto out_unlock_inode;
6555         } else {
6556                 btrfs_update_inode(trans, root, inode);
6557                 unlock_new_inode(inode);
6558                 d_instantiate(dentry, inode);
6559         }
6560
6561 out_unlock:
6562         btrfs_end_transaction(trans);
6563         btrfs_balance_delayed_items(fs_info);
6564         btrfs_btree_balance_dirty(fs_info);
6565         if (drop_inode) {
6566                 inode_dec_link_count(inode);
6567                 iput(inode);
6568         }
6569         return err;
6570
6571 out_unlock_inode:
6572         drop_inode = 1;
6573         unlock_new_inode(inode);
6574         goto out_unlock;
6575
6576 }
6577
6578 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6579                         umode_t mode, bool excl)
6580 {
6581         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6582         struct btrfs_trans_handle *trans;
6583         struct btrfs_root *root = BTRFS_I(dir)->root;
6584         struct inode *inode = NULL;
6585         int drop_inode_on_err = 0;
6586         int err;
6587         u64 objectid;
6588         u64 index = 0;
6589
6590         /*
6591          * 2 for inode item and ref
6592          * 2 for dir items
6593          * 1 for xattr if selinux is on
6594          */
6595         trans = btrfs_start_transaction(root, 5);
6596         if (IS_ERR(trans))
6597                 return PTR_ERR(trans);
6598
6599         err = btrfs_find_free_ino(root, &objectid);
6600         if (err)
6601                 goto out_unlock;
6602
6603         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6604                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6605                         mode, &index);
6606         if (IS_ERR(inode)) {
6607                 err = PTR_ERR(inode);
6608                 goto out_unlock;
6609         }
6610         drop_inode_on_err = 1;
6611         /*
6612         * If the active LSM wants to access the inode during
6613         * d_instantiate it needs these. Smack checks to see
6614         * if the filesystem supports xattrs by looking at the
6615         * ops vector.
6616         */
6617         inode->i_fop = &btrfs_file_operations;
6618         inode->i_op = &btrfs_file_inode_operations;
6619         inode->i_mapping->a_ops = &btrfs_aops;
6620
6621         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6622         if (err)
6623                 goto out_unlock_inode;
6624
6625         err = btrfs_update_inode(trans, root, inode);
6626         if (err)
6627                 goto out_unlock_inode;
6628
6629         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6630                         0, index);
6631         if (err)
6632                 goto out_unlock_inode;
6633
6634         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6635         unlock_new_inode(inode);
6636         d_instantiate(dentry, inode);
6637
6638 out_unlock:
6639         btrfs_end_transaction(trans);
6640         if (err && drop_inode_on_err) {
6641                 inode_dec_link_count(inode);
6642                 iput(inode);
6643         }
6644         btrfs_balance_delayed_items(fs_info);
6645         btrfs_btree_balance_dirty(fs_info);
6646         return err;
6647
6648 out_unlock_inode:
6649         unlock_new_inode(inode);
6650         goto out_unlock;
6651
6652 }
6653
6654 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6655                       struct dentry *dentry)
6656 {
6657         struct btrfs_trans_handle *trans = NULL;
6658         struct btrfs_root *root = BTRFS_I(dir)->root;
6659         struct inode *inode = d_inode(old_dentry);
6660         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6661         u64 index;
6662         int err;
6663         int drop_inode = 0;
6664
6665         /* do not allow sys_link's with other subvols of the same device */
6666         if (root->objectid != BTRFS_I(inode)->root->objectid)
6667                 return -EXDEV;
6668
6669         if (inode->i_nlink >= BTRFS_LINK_MAX)
6670                 return -EMLINK;
6671
6672         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6673         if (err)
6674                 goto fail;
6675
6676         /*
6677          * 2 items for inode and inode ref
6678          * 2 items for dir items
6679          * 1 item for parent inode
6680          */
6681         trans = btrfs_start_transaction(root, 5);
6682         if (IS_ERR(trans)) {
6683                 err = PTR_ERR(trans);
6684                 trans = NULL;
6685                 goto fail;
6686         }
6687
6688         /* There are several dir indexes for this inode, clear the cache. */
6689         BTRFS_I(inode)->dir_index = 0ULL;
6690         inc_nlink(inode);
6691         inode_inc_iversion(inode);
6692         inode->i_ctime = current_time(inode);
6693         ihold(inode);
6694         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6695
6696         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6697                         1, index);
6698
6699         if (err) {
6700                 drop_inode = 1;
6701         } else {
6702                 struct dentry *parent = dentry->d_parent;
6703                 err = btrfs_update_inode(trans, root, inode);
6704                 if (err)
6705                         goto fail;
6706                 if (inode->i_nlink == 1) {
6707                         /*
6708                          * If new hard link count is 1, it's a file created
6709                          * with open(2) O_TMPFILE flag.
6710                          */
6711                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6712                         if (err)
6713                                 goto fail;
6714                 }
6715                 d_instantiate(dentry, inode);
6716                 btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent);
6717         }
6718
6719         btrfs_balance_delayed_items(fs_info);
6720 fail:
6721         if (trans)
6722                 btrfs_end_transaction(trans);
6723         if (drop_inode) {
6724                 inode_dec_link_count(inode);
6725                 iput(inode);
6726         }
6727         btrfs_btree_balance_dirty(fs_info);
6728         return err;
6729 }
6730
6731 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6732 {
6733         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6734         struct inode *inode = NULL;
6735         struct btrfs_trans_handle *trans;
6736         struct btrfs_root *root = BTRFS_I(dir)->root;
6737         int err = 0;
6738         int drop_on_err = 0;
6739         u64 objectid = 0;
6740         u64 index = 0;
6741
6742         /*
6743          * 2 items for inode and ref
6744          * 2 items for dir items
6745          * 1 for xattr if selinux is on
6746          */
6747         trans = btrfs_start_transaction(root, 5);
6748         if (IS_ERR(trans))
6749                 return PTR_ERR(trans);
6750
6751         err = btrfs_find_free_ino(root, &objectid);
6752         if (err)
6753                 goto out_fail;
6754
6755         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6756                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6757                         S_IFDIR | mode, &index);
6758         if (IS_ERR(inode)) {
6759                 err = PTR_ERR(inode);
6760                 goto out_fail;
6761         }
6762
6763         drop_on_err = 1;
6764         /* these must be set before we unlock the inode */
6765         inode->i_op = &btrfs_dir_inode_operations;
6766         inode->i_fop = &btrfs_dir_file_operations;
6767
6768         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6769         if (err)
6770                 goto out_fail_inode;
6771
6772         btrfs_i_size_write(BTRFS_I(inode), 0);
6773         err = btrfs_update_inode(trans, root, inode);
6774         if (err)
6775                 goto out_fail_inode;
6776
6777         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6778                         dentry->d_name.name,
6779                         dentry->d_name.len, 0, index);
6780         if (err)
6781                 goto out_fail_inode;
6782
6783         d_instantiate(dentry, inode);
6784         /*
6785          * mkdir is special.  We're unlocking after we call d_instantiate
6786          * to avoid a race with nfsd calling d_instantiate.
6787          */
6788         unlock_new_inode(inode);
6789         drop_on_err = 0;
6790
6791 out_fail:
6792         btrfs_end_transaction(trans);
6793         if (drop_on_err) {
6794                 inode_dec_link_count(inode);
6795                 iput(inode);
6796         }
6797         btrfs_balance_delayed_items(fs_info);
6798         btrfs_btree_balance_dirty(fs_info);
6799         return err;
6800
6801 out_fail_inode:
6802         unlock_new_inode(inode);
6803         goto out_fail;
6804 }
6805
6806 /* Find next extent map of a given extent map, caller needs to ensure locks */
6807 static struct extent_map *next_extent_map(struct extent_map *em)
6808 {
6809         struct rb_node *next;
6810
6811         next = rb_next(&em->rb_node);
6812         if (!next)
6813                 return NULL;
6814         return container_of(next, struct extent_map, rb_node);
6815 }
6816
6817 static struct extent_map *prev_extent_map(struct extent_map *em)
6818 {
6819         struct rb_node *prev;
6820
6821         prev = rb_prev(&em->rb_node);
6822         if (!prev)
6823                 return NULL;
6824         return container_of(prev, struct extent_map, rb_node);
6825 }
6826
6827 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6828  * the existing extent is the nearest extent to map_start,
6829  * and an extent that you want to insert, deal with overlap and insert
6830  * the best fitted new extent into the tree.
6831  */
6832 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6833                                 struct extent_map *existing,
6834                                 struct extent_map *em,
6835                                 u64 map_start)
6836 {
6837         struct extent_map *prev;
6838         struct extent_map *next;
6839         u64 start;
6840         u64 end;
6841         u64 start_diff;
6842
6843         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6844
6845         if (existing->start > map_start) {
6846                 next = existing;
6847                 prev = prev_extent_map(next);
6848         } else {
6849                 prev = existing;
6850                 next = next_extent_map(prev);
6851         }
6852
6853         start = prev ? extent_map_end(prev) : em->start;
6854         start = max_t(u64, start, em->start);
6855         end = next ? next->start : extent_map_end(em);
6856         end = min_t(u64, end, extent_map_end(em));
6857         start_diff = start - em->start;
6858         em->start = start;
6859         em->len = end - start;
6860         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6861             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6862                 em->block_start += start_diff;
6863                 em->block_len -= start_diff;
6864         }
6865         return add_extent_mapping(em_tree, em, 0);
6866 }
6867
6868 static noinline int uncompress_inline(struct btrfs_path *path,
6869                                       struct page *page,
6870                                       size_t pg_offset, u64 extent_offset,
6871                                       struct btrfs_file_extent_item *item)
6872 {
6873         int ret;
6874         struct extent_buffer *leaf = path->nodes[0];
6875         char *tmp;
6876         size_t max_size;
6877         unsigned long inline_size;
6878         unsigned long ptr;
6879         int compress_type;
6880
6881         WARN_ON(pg_offset != 0);
6882         compress_type = btrfs_file_extent_compression(leaf, item);
6883         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6884         inline_size = btrfs_file_extent_inline_item_len(leaf,
6885                                         btrfs_item_nr(path->slots[0]));
6886         tmp = kmalloc(inline_size, GFP_NOFS);
6887         if (!tmp)
6888                 return -ENOMEM;
6889         ptr = btrfs_file_extent_inline_start(item);
6890
6891         read_extent_buffer(leaf, tmp, ptr, inline_size);
6892
6893         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6894         ret = btrfs_decompress(compress_type, tmp, page,
6895                                extent_offset, inline_size, max_size);
6896
6897         /*
6898          * decompression code contains a memset to fill in any space between the end
6899          * of the uncompressed data and the end of max_size in case the decompressed
6900          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6901          * the end of an inline extent and the beginning of the next block, so we
6902          * cover that region here.
6903          */
6904
6905         if (max_size + pg_offset < PAGE_SIZE) {
6906                 char *map = kmap(page);
6907                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6908                 kunmap(page);
6909         }
6910         kfree(tmp);
6911         return ret;
6912 }
6913
6914 /*
6915  * a bit scary, this does extent mapping from logical file offset to the disk.
6916  * the ugly parts come from merging extents from the disk with the in-ram
6917  * representation.  This gets more complex because of the data=ordered code,
6918  * where the in-ram extents might be locked pending data=ordered completion.
6919  *
6920  * This also copies inline extents directly into the page.
6921  */
6922 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6923                 struct page *page,
6924             size_t pg_offset, u64 start, u64 len,
6925                 int create)
6926 {
6927         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6928         int ret;
6929         int err = 0;
6930         u64 extent_start = 0;
6931         u64 extent_end = 0;
6932         u64 objectid = btrfs_ino(inode);
6933         u32 found_type;
6934         struct btrfs_path *path = NULL;
6935         struct btrfs_root *root = inode->root;
6936         struct btrfs_file_extent_item *item;
6937         struct extent_buffer *leaf;
6938         struct btrfs_key found_key;
6939         struct extent_map *em = NULL;
6940         struct extent_map_tree *em_tree = &inode->extent_tree;
6941         struct extent_io_tree *io_tree = &inode->io_tree;
6942         struct btrfs_trans_handle *trans = NULL;
6943         const bool new_inline = !page || create;
6944
6945 again:
6946         read_lock(&em_tree->lock);
6947         em = lookup_extent_mapping(em_tree, start, len);
6948         if (em)
6949                 em->bdev = fs_info->fs_devices->latest_bdev;
6950         read_unlock(&em_tree->lock);
6951
6952         if (em) {
6953                 if (em->start > start || em->start + em->len <= start)
6954                         free_extent_map(em);
6955                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6956                         free_extent_map(em);
6957                 else
6958                         goto out;
6959         }
6960         em = alloc_extent_map();
6961         if (!em) {
6962                 err = -ENOMEM;
6963                 goto out;
6964         }
6965         em->bdev = fs_info->fs_devices->latest_bdev;
6966         em->start = EXTENT_MAP_HOLE;
6967         em->orig_start = EXTENT_MAP_HOLE;
6968         em->len = (u64)-1;
6969         em->block_len = (u64)-1;
6970
6971         if (!path) {
6972                 path = btrfs_alloc_path();
6973                 if (!path) {
6974                         err = -ENOMEM;
6975                         goto out;
6976                 }
6977                 /*
6978                  * Chances are we'll be called again, so go ahead and do
6979                  * readahead
6980                  */
6981                 path->reada = READA_FORWARD;
6982         }
6983
6984         ret = btrfs_lookup_file_extent(trans, root, path,
6985                                        objectid, start, trans != NULL);
6986         if (ret < 0) {
6987                 err = ret;
6988                 goto out;
6989         }
6990
6991         if (ret != 0) {
6992                 if (path->slots[0] == 0)
6993                         goto not_found;
6994                 path->slots[0]--;
6995         }
6996
6997         leaf = path->nodes[0];
6998         item = btrfs_item_ptr(leaf, path->slots[0],
6999                               struct btrfs_file_extent_item);
7000         /* are we inside the extent that was found? */
7001         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7002         found_type = found_key.type;
7003         if (found_key.objectid != objectid ||
7004             found_type != BTRFS_EXTENT_DATA_KEY) {
7005                 /*
7006                  * If we backup past the first extent we want to move forward
7007                  * and see if there is an extent in front of us, otherwise we'll
7008                  * say there is a hole for our whole search range which can
7009                  * cause problems.
7010                  */
7011                 extent_end = start;
7012                 goto next;
7013         }
7014
7015         found_type = btrfs_file_extent_type(leaf, item);
7016         extent_start = found_key.offset;
7017         if (found_type == BTRFS_FILE_EXTENT_REG ||
7018             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7019                 extent_end = extent_start +
7020                        btrfs_file_extent_num_bytes(leaf, item);
7021
7022                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
7023                                                        extent_start);
7024         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
7025                 size_t size;
7026                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
7027                 extent_end = ALIGN(extent_start + size,
7028                                    fs_info->sectorsize);
7029
7030                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
7031                                                       path->slots[0],
7032                                                       extent_start);
7033         }
7034 next:
7035         if (start >= extent_end) {
7036                 path->slots[0]++;
7037                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
7038                         ret = btrfs_next_leaf(root, path);
7039                         if (ret < 0) {
7040                                 err = ret;
7041                                 goto out;
7042                         }
7043                         if (ret > 0)
7044                                 goto not_found;
7045                         leaf = path->nodes[0];
7046                 }
7047                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7048                 if (found_key.objectid != objectid ||
7049                     found_key.type != BTRFS_EXTENT_DATA_KEY)
7050                         goto not_found;
7051                 if (start + len <= found_key.offset)
7052                         goto not_found;
7053                 if (start > found_key.offset)
7054                         goto next;
7055                 em->start = start;
7056                 em->orig_start = start;
7057                 em->len = found_key.offset - start;
7058                 goto not_found_em;
7059         }
7060
7061         btrfs_extent_item_to_extent_map(inode, path, item,
7062                         new_inline, em);
7063
7064         if (found_type == BTRFS_FILE_EXTENT_REG ||
7065             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7066                 goto insert;
7067         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
7068                 unsigned long ptr;
7069                 char *map;
7070                 size_t size;
7071                 size_t extent_offset;
7072                 size_t copy_size;
7073
7074                 if (new_inline)
7075                         goto out;
7076
7077                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
7078                 extent_offset = page_offset(page) + pg_offset - extent_start;
7079                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
7080                                   size - extent_offset);
7081                 em->start = extent_start + extent_offset;
7082                 em->len = ALIGN(copy_size, fs_info->sectorsize);
7083                 em->orig_block_len = em->len;
7084                 em->orig_start = em->start;
7085                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
7086                 if (create == 0 && !PageUptodate(page)) {
7087                         if (btrfs_file_extent_compression(leaf, item) !=
7088                             BTRFS_COMPRESS_NONE) {
7089                                 ret = uncompress_inline(path, page, pg_offset,
7090                                                         extent_offset, item);
7091                                 if (ret) {
7092                                         err = ret;
7093                                         goto out;
7094                                 }
7095                         } else {
7096                                 map = kmap(page);
7097                                 read_extent_buffer(leaf, map + pg_offset, ptr,
7098                                                    copy_size);
7099                                 if (pg_offset + copy_size < PAGE_SIZE) {
7100                                         memset(map + pg_offset + copy_size, 0,
7101                                                PAGE_SIZE - pg_offset -
7102                                                copy_size);
7103                                 }
7104                                 kunmap(page);
7105                         }
7106                         flush_dcache_page(page);
7107                 } else if (create && PageUptodate(page)) {
7108                         BUG();
7109                         if (!trans) {
7110                                 kunmap(page);
7111                                 free_extent_map(em);
7112                                 em = NULL;
7113
7114                                 btrfs_release_path(path);
7115                                 trans = btrfs_join_transaction(root);
7116
7117                                 if (IS_ERR(trans))
7118                                         return ERR_CAST(trans);
7119                                 goto again;
7120                         }
7121                         map = kmap(page);
7122                         write_extent_buffer(leaf, map + pg_offset, ptr,
7123                                             copy_size);
7124                         kunmap(page);
7125                         btrfs_mark_buffer_dirty(leaf);
7126                 }
7127                 set_extent_uptodate(io_tree, em->start,
7128                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7129                 goto insert;
7130         }
7131 not_found:
7132         em->start = start;
7133         em->orig_start = start;
7134         em->len = len;
7135 not_found_em:
7136         em->block_start = EXTENT_MAP_HOLE;
7137         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
7138 insert:
7139         btrfs_release_path(path);
7140         if (em->start > start || extent_map_end(em) <= start) {
7141                 btrfs_err(fs_info,
7142                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
7143                           em->start, em->len, start, len);
7144                 err = -EIO;
7145                 goto out;
7146         }
7147
7148         err = 0;
7149         write_lock(&em_tree->lock);
7150         ret = add_extent_mapping(em_tree, em, 0);
7151         /* it is possible that someone inserted the extent into the tree
7152          * while we had the lock dropped.  It is also possible that
7153          * an overlapping map exists in the tree
7154          */
7155         if (ret == -EEXIST) {
7156                 struct extent_map *existing;
7157
7158                 ret = 0;
7159
7160                 existing = search_extent_mapping(em_tree, start, len);
7161                 /*
7162                  * existing will always be non-NULL, since there must be
7163                  * extent causing the -EEXIST.
7164                  */
7165                 if (existing->start == em->start &&
7166                     extent_map_end(existing) >= extent_map_end(em) &&
7167                     em->block_start == existing->block_start) {
7168                         /*
7169                          * The existing extent map already encompasses the
7170                          * entire extent map we tried to add.
7171                          */
7172                         free_extent_map(em);
7173                         em = existing;
7174                         err = 0;
7175
7176                 } else if (start >= extent_map_end(existing) ||
7177                     start <= existing->start) {
7178                         /*
7179                          * The existing extent map is the one nearest to
7180                          * the [start, start + len) range which overlaps
7181                          */
7182                         err = merge_extent_mapping(em_tree, existing,
7183                                                    em, start);
7184                         free_extent_map(existing);
7185                         if (err) {
7186                                 free_extent_map(em);
7187                                 em = NULL;
7188                         }
7189                 } else {
7190                         free_extent_map(em);
7191                         em = existing;
7192                         err = 0;
7193                 }
7194         }
7195         write_unlock(&em_tree->lock);
7196 out:
7197
7198         trace_btrfs_get_extent(root, inode, em);
7199
7200         btrfs_free_path(path);
7201         if (trans) {
7202                 ret = btrfs_end_transaction(trans);
7203                 if (!err)
7204                         err = ret;
7205         }
7206         if (err) {
7207                 free_extent_map(em);
7208                 return ERR_PTR(err);
7209         }
7210         BUG_ON(!em); /* Error is always set */
7211         return em;
7212 }
7213
7214 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7215                 struct page *page,
7216                 size_t pg_offset, u64 start, u64 len,
7217                 int create)
7218 {
7219         struct extent_map *em;
7220         struct extent_map *hole_em = NULL;
7221         u64 range_start = start;
7222         u64 end;
7223         u64 found;
7224         u64 found_end;
7225         int err = 0;
7226
7227         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7228         if (IS_ERR(em))
7229                 return em;
7230         /*
7231          * If our em maps to:
7232          * - a hole or
7233          * - a pre-alloc extent,
7234          * there might actually be delalloc bytes behind it.
7235          */
7236         if (em->block_start != EXTENT_MAP_HOLE &&
7237             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7238                 return em;
7239         else
7240                 hole_em = em;
7241
7242         /* check to see if we've wrapped (len == -1 or similar) */
7243         end = start + len;
7244         if (end < start)
7245                 end = (u64)-1;
7246         else
7247                 end -= 1;
7248
7249         em = NULL;
7250
7251         /* ok, we didn't find anything, lets look for delalloc */
7252         found = count_range_bits(&inode->io_tree, &range_start,
7253                                  end, len, EXTENT_DELALLOC, 1);
7254         found_end = range_start + found;
7255         if (found_end < range_start)
7256                 found_end = (u64)-1;
7257
7258         /*
7259          * we didn't find anything useful, return
7260          * the original results from get_extent()
7261          */
7262         if (range_start > end || found_end <= start) {
7263                 em = hole_em;
7264                 hole_em = NULL;
7265                 goto out;
7266         }
7267
7268         /* adjust the range_start to make sure it doesn't
7269          * go backwards from the start they passed in
7270          */
7271         range_start = max(start, range_start);
7272         found = found_end - range_start;
7273
7274         if (found > 0) {
7275                 u64 hole_start = start;
7276                 u64 hole_len = len;
7277
7278                 em = alloc_extent_map();
7279                 if (!em) {
7280                         err = -ENOMEM;
7281                         goto out;
7282                 }
7283                 /*
7284                  * when btrfs_get_extent can't find anything it
7285                  * returns one huge hole
7286                  *
7287                  * make sure what it found really fits our range, and
7288                  * adjust to make sure it is based on the start from
7289                  * the caller
7290                  */
7291                 if (hole_em) {
7292                         u64 calc_end = extent_map_end(hole_em);
7293
7294                         if (calc_end <= start || (hole_em->start > end)) {
7295                                 free_extent_map(hole_em);
7296                                 hole_em = NULL;
7297                         } else {
7298                                 hole_start = max(hole_em->start, start);
7299                                 hole_len = calc_end - hole_start;
7300                         }
7301                 }
7302                 em->bdev = NULL;
7303                 if (hole_em && range_start > hole_start) {
7304                         /* our hole starts before our delalloc, so we
7305                          * have to return just the parts of the hole
7306                          * that go until  the delalloc starts
7307                          */
7308                         em->len = min(hole_len,
7309                                       range_start - hole_start);
7310                         em->start = hole_start;
7311                         em->orig_start = hole_start;
7312                         /*
7313                          * don't adjust block start at all,
7314                          * it is fixed at EXTENT_MAP_HOLE
7315                          */
7316                         em->block_start = hole_em->block_start;
7317                         em->block_len = hole_len;
7318                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7319                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7320                 } else {
7321                         em->start = range_start;
7322                         em->len = found;
7323                         em->orig_start = range_start;
7324                         em->block_start = EXTENT_MAP_DELALLOC;
7325                         em->block_len = found;
7326                 }
7327         } else if (hole_em) {
7328                 return hole_em;
7329         }
7330 out:
7331
7332         free_extent_map(hole_em);
7333         if (err) {
7334                 free_extent_map(em);
7335                 return ERR_PTR(err);
7336         }
7337         return em;
7338 }
7339
7340 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7341                                                   const u64 start,
7342                                                   const u64 len,
7343                                                   const u64 orig_start,
7344                                                   const u64 block_start,
7345                                                   const u64 block_len,
7346                                                   const u64 orig_block_len,
7347                                                   const u64 ram_bytes,
7348                                                   const int type)
7349 {
7350         struct extent_map *em = NULL;
7351         int ret;
7352
7353         if (type != BTRFS_ORDERED_NOCOW) {
7354                 em = create_io_em(inode, start, len, orig_start,
7355                                   block_start, block_len, orig_block_len,
7356                                   ram_bytes,
7357                                   BTRFS_COMPRESS_NONE, /* compress_type */
7358                                   type);
7359                 if (IS_ERR(em))
7360                         goto out;
7361         }
7362         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7363                                            len, block_len, type);
7364         if (ret) {
7365                 if (em) {
7366                         free_extent_map(em);
7367                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7368                                                 start + len - 1, 0);
7369                 }
7370                 em = ERR_PTR(ret);
7371         }
7372  out:
7373
7374         return em;
7375 }
7376
7377 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7378                                                   u64 start, u64 len)
7379 {
7380         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7381         struct btrfs_root *root = BTRFS_I(inode)->root;
7382         struct extent_map *em;
7383         struct btrfs_key ins;
7384         u64 alloc_hint;
7385         int ret;
7386
7387         alloc_hint = get_extent_allocation_hint(inode, start, len);
7388         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7389                                    0, alloc_hint, &ins, 1, 1);
7390         if (ret)
7391                 return ERR_PTR(ret);
7392
7393         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7394                                      ins.objectid, ins.offset, ins.offset,
7395                                      ins.offset, BTRFS_ORDERED_REGULAR);
7396         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7397         if (IS_ERR(em))
7398                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7399                                            ins.offset, 1);
7400
7401         return em;
7402 }
7403
7404 /*
7405  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7406  * block must be cow'd
7407  */
7408 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7409                               u64 *orig_start, u64 *orig_block_len,
7410                               u64 *ram_bytes)
7411 {
7412         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7413         struct btrfs_path *path;
7414         int ret;
7415         struct extent_buffer *leaf;
7416         struct btrfs_root *root = BTRFS_I(inode)->root;
7417         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7418         struct btrfs_file_extent_item *fi;
7419         struct btrfs_key key;
7420         u64 disk_bytenr;
7421         u64 backref_offset;
7422         u64 extent_end;
7423         u64 num_bytes;
7424         int slot;
7425         int found_type;
7426         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7427
7428         path = btrfs_alloc_path();
7429         if (!path)
7430                 return -ENOMEM;
7431
7432         ret = btrfs_lookup_file_extent(NULL, root, path,
7433                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7434         if (ret < 0)
7435                 goto out;
7436
7437         slot = path->slots[0];
7438         if (ret == 1) {
7439                 if (slot == 0) {
7440                         /* can't find the item, must cow */
7441                         ret = 0;
7442                         goto out;
7443                 }
7444                 slot--;
7445         }
7446         ret = 0;
7447         leaf = path->nodes[0];
7448         btrfs_item_key_to_cpu(leaf, &key, slot);
7449         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7450             key.type != BTRFS_EXTENT_DATA_KEY) {
7451                 /* not our file or wrong item type, must cow */
7452                 goto out;
7453         }
7454
7455         if (key.offset > offset) {
7456                 /* Wrong offset, must cow */
7457                 goto out;
7458         }
7459
7460         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7461         found_type = btrfs_file_extent_type(leaf, fi);
7462         if (found_type != BTRFS_FILE_EXTENT_REG &&
7463             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7464                 /* not a regular extent, must cow */
7465                 goto out;
7466         }
7467
7468         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7469                 goto out;
7470
7471         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7472         if (extent_end <= offset)
7473                 goto out;
7474
7475         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7476         if (disk_bytenr == 0)
7477                 goto out;
7478
7479         if (btrfs_file_extent_compression(leaf, fi) ||
7480             btrfs_file_extent_encryption(leaf, fi) ||
7481             btrfs_file_extent_other_encoding(leaf, fi))
7482                 goto out;
7483
7484         backref_offset = btrfs_file_extent_offset(leaf, fi);
7485
7486         if (orig_start) {
7487                 *orig_start = key.offset - backref_offset;
7488                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7489                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7490         }
7491
7492         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7493                 goto out;
7494
7495         num_bytes = min(offset + *len, extent_end) - offset;
7496         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7497                 u64 range_end;
7498
7499                 range_end = round_up(offset + num_bytes,
7500                                      root->fs_info->sectorsize) - 1;
7501                 ret = test_range_bit(io_tree, offset, range_end,
7502                                      EXTENT_DELALLOC, 0, NULL);
7503                 if (ret) {
7504                         ret = -EAGAIN;
7505                         goto out;
7506                 }
7507         }
7508
7509         btrfs_release_path(path);
7510
7511         /*
7512          * look for other files referencing this extent, if we
7513          * find any we must cow
7514          */
7515
7516         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7517                                     key.offset - backref_offset, disk_bytenr);
7518         if (ret) {
7519                 ret = 0;
7520                 goto out;
7521         }
7522
7523         /*
7524          * adjust disk_bytenr and num_bytes to cover just the bytes
7525          * in this extent we are about to write.  If there
7526          * are any csums in that range we have to cow in order
7527          * to keep the csums correct
7528          */
7529         disk_bytenr += backref_offset;
7530         disk_bytenr += offset - key.offset;
7531         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7532                 goto out;
7533         /*
7534          * all of the above have passed, it is safe to overwrite this extent
7535          * without cow
7536          */
7537         *len = num_bytes;
7538         ret = 1;
7539 out:
7540         btrfs_free_path(path);
7541         return ret;
7542 }
7543
7544 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7545 {
7546         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7547         bool found = false;
7548         void **pagep = NULL;
7549         struct page *page = NULL;
7550         unsigned long start_idx;
7551         unsigned long end_idx;
7552
7553         start_idx = start >> PAGE_SHIFT;
7554
7555         /*
7556          * end is the last byte in the last page.  end == start is legal
7557          */
7558         end_idx = end >> PAGE_SHIFT;
7559
7560         rcu_read_lock();
7561
7562         /* Most of the code in this while loop is lifted from
7563          * find_get_page.  It's been modified to begin searching from a
7564          * page and return just the first page found in that range.  If the
7565          * found idx is less than or equal to the end idx then we know that
7566          * a page exists.  If no pages are found or if those pages are
7567          * outside of the range then we're fine (yay!) */
7568         while (page == NULL &&
7569                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7570                 page = radix_tree_deref_slot(pagep);
7571                 if (unlikely(!page))
7572                         break;
7573
7574                 if (radix_tree_exception(page)) {
7575                         if (radix_tree_deref_retry(page)) {
7576                                 page = NULL;
7577                                 continue;
7578                         }
7579                         /*
7580                          * Otherwise, shmem/tmpfs must be storing a swap entry
7581                          * here as an exceptional entry: so return it without
7582                          * attempting to raise page count.
7583                          */
7584                         page = NULL;
7585                         break; /* TODO: Is this relevant for this use case? */
7586                 }
7587
7588                 if (!page_cache_get_speculative(page)) {
7589                         page = NULL;
7590                         continue;
7591                 }
7592
7593                 /*
7594                  * Has the page moved?
7595                  * This is part of the lockless pagecache protocol. See
7596                  * include/linux/pagemap.h for details.
7597                  */
7598                 if (unlikely(page != *pagep)) {
7599                         put_page(page);
7600                         page = NULL;
7601                 }
7602         }
7603
7604         if (page) {
7605                 if (page->index <= end_idx)
7606                         found = true;
7607                 put_page(page);
7608         }
7609
7610         rcu_read_unlock();
7611         return found;
7612 }
7613
7614 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7615                               struct extent_state **cached_state, int writing)
7616 {
7617         struct btrfs_ordered_extent *ordered;
7618         int ret = 0;
7619
7620         while (1) {
7621                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7622                                  cached_state);
7623                 /*
7624                  * We're concerned with the entire range that we're going to be
7625                  * doing DIO to, so we need to make sure there's no ordered
7626                  * extents in this range.
7627                  */
7628                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7629                                                      lockend - lockstart + 1);
7630
7631                 /*
7632                  * We need to make sure there are no buffered pages in this
7633                  * range either, we could have raced between the invalidate in
7634                  * generic_file_direct_write and locking the extent.  The
7635                  * invalidate needs to happen so that reads after a write do not
7636                  * get stale data.
7637                  */
7638                 if (!ordered &&
7639                     (!writing ||
7640                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7641                         break;
7642
7643                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7644                                      cached_state, GFP_NOFS);
7645
7646                 if (ordered) {
7647                         /*
7648                          * If we are doing a DIO read and the ordered extent we
7649                          * found is for a buffered write, we can not wait for it
7650                          * to complete and retry, because if we do so we can
7651                          * deadlock with concurrent buffered writes on page
7652                          * locks. This happens only if our DIO read covers more
7653                          * than one extent map, if at this point has already
7654                          * created an ordered extent for a previous extent map
7655                          * and locked its range in the inode's io tree, and a
7656                          * concurrent write against that previous extent map's
7657                          * range and this range started (we unlock the ranges
7658                          * in the io tree only when the bios complete and
7659                          * buffered writes always lock pages before attempting
7660                          * to lock range in the io tree).
7661                          */
7662                         if (writing ||
7663                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7664                                 btrfs_start_ordered_extent(inode, ordered, 1);
7665                         else
7666                                 ret = -ENOTBLK;
7667                         btrfs_put_ordered_extent(ordered);
7668                 } else {
7669                         /*
7670                          * We could trigger writeback for this range (and wait
7671                          * for it to complete) and then invalidate the pages for
7672                          * this range (through invalidate_inode_pages2_range()),
7673                          * but that can lead us to a deadlock with a concurrent
7674                          * call to readpages() (a buffered read or a defrag call
7675                          * triggered a readahead) on a page lock due to an
7676                          * ordered dio extent we created before but did not have
7677                          * yet a corresponding bio submitted (whence it can not
7678                          * complete), which makes readpages() wait for that
7679                          * ordered extent to complete while holding a lock on
7680                          * that page.
7681                          */
7682                         ret = -ENOTBLK;
7683                 }
7684
7685                 if (ret)
7686                         break;
7687
7688                 cond_resched();
7689         }
7690
7691         return ret;
7692 }
7693
7694 /* The callers of this must take lock_extent() */
7695 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7696                                        u64 orig_start, u64 block_start,
7697                                        u64 block_len, u64 orig_block_len,
7698                                        u64 ram_bytes, int compress_type,
7699                                        int type)
7700 {
7701         struct extent_map_tree *em_tree;
7702         struct extent_map *em;
7703         struct btrfs_root *root = BTRFS_I(inode)->root;
7704         int ret;
7705
7706         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7707                type == BTRFS_ORDERED_COMPRESSED ||
7708                type == BTRFS_ORDERED_NOCOW ||
7709                type == BTRFS_ORDERED_REGULAR);
7710
7711         em_tree = &BTRFS_I(inode)->extent_tree;
7712         em = alloc_extent_map();
7713         if (!em)
7714                 return ERR_PTR(-ENOMEM);
7715
7716         em->start = start;
7717         em->orig_start = orig_start;
7718         em->len = len;
7719         em->block_len = block_len;
7720         em->block_start = block_start;
7721         em->bdev = root->fs_info->fs_devices->latest_bdev;
7722         em->orig_block_len = orig_block_len;
7723         em->ram_bytes = ram_bytes;
7724         em->generation = -1;
7725         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7726         if (type == BTRFS_ORDERED_PREALLOC) {
7727                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7728         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7729                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7730                 em->compress_type = compress_type;
7731         }
7732
7733         do {
7734                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7735                                 em->start + em->len - 1, 0);
7736                 write_lock(&em_tree->lock);
7737                 ret = add_extent_mapping(em_tree, em, 1);
7738                 write_unlock(&em_tree->lock);
7739                 /*
7740                  * The caller has taken lock_extent(), who could race with us
7741                  * to add em?
7742                  */
7743         } while (ret == -EEXIST);
7744
7745         if (ret) {
7746                 free_extent_map(em);
7747                 return ERR_PTR(ret);
7748         }
7749
7750         /* em got 2 refs now, callers needs to do free_extent_map once. */
7751         return em;
7752 }
7753
7754 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7755                                    struct buffer_head *bh_result, int create)
7756 {
7757         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7758         struct extent_map *em;
7759         struct extent_state *cached_state = NULL;
7760         struct btrfs_dio_data *dio_data = NULL;
7761         u64 start = iblock << inode->i_blkbits;
7762         u64 lockstart, lockend;
7763         u64 len = bh_result->b_size;
7764         int unlock_bits = EXTENT_LOCKED;
7765         int ret = 0;
7766
7767         if (create)
7768                 unlock_bits |= EXTENT_DIRTY;
7769         else
7770                 len = min_t(u64, len, fs_info->sectorsize);
7771
7772         lockstart = start;
7773         lockend = start + len - 1;
7774
7775         if (current->journal_info) {
7776                 /*
7777                  * Need to pull our outstanding extents and set journal_info to NULL so
7778                  * that anything that needs to check if there's a transaction doesn't get
7779                  * confused.
7780                  */
7781                 dio_data = current->journal_info;
7782                 current->journal_info = NULL;
7783         }
7784
7785         /*
7786          * If this errors out it's because we couldn't invalidate pagecache for
7787          * this range and we need to fallback to buffered.
7788          */
7789         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7790                                create)) {
7791                 ret = -ENOTBLK;
7792                 goto err;
7793         }
7794
7795         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7796         if (IS_ERR(em)) {
7797                 ret = PTR_ERR(em);
7798                 goto unlock_err;
7799         }
7800
7801         /*
7802          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7803          * io.  INLINE is special, and we could probably kludge it in here, but
7804          * it's still buffered so for safety lets just fall back to the generic
7805          * buffered path.
7806          *
7807          * For COMPRESSED we _have_ to read the entire extent in so we can
7808          * decompress it, so there will be buffering required no matter what we
7809          * do, so go ahead and fallback to buffered.
7810          *
7811          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7812          * to buffered IO.  Don't blame me, this is the price we pay for using
7813          * the generic code.
7814          */
7815         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7816             em->block_start == EXTENT_MAP_INLINE) {
7817                 free_extent_map(em);
7818                 ret = -ENOTBLK;
7819                 goto unlock_err;
7820         }
7821
7822         /* Just a good old fashioned hole, return */
7823         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7824                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7825                 free_extent_map(em);
7826                 goto unlock_err;
7827         }
7828
7829         /*
7830          * We don't allocate a new extent in the following cases
7831          *
7832          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7833          * existing extent.
7834          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7835          * just use the extent.
7836          *
7837          */
7838         if (!create) {
7839                 len = min(len, em->len - (start - em->start));
7840                 lockstart = start + len;
7841                 goto unlock;
7842         }
7843
7844         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7845             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7846              em->block_start != EXTENT_MAP_HOLE)) {
7847                 int type;
7848                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7849
7850                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7851                         type = BTRFS_ORDERED_PREALLOC;
7852                 else
7853                         type = BTRFS_ORDERED_NOCOW;
7854                 len = min(len, em->len - (start - em->start));
7855                 block_start = em->block_start + (start - em->start);
7856
7857                 if (can_nocow_extent(inode, start, &len, &orig_start,
7858                                      &orig_block_len, &ram_bytes) == 1 &&
7859                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7860                         struct extent_map *em2;
7861
7862                         em2 = btrfs_create_dio_extent(inode, start, len,
7863                                                       orig_start, block_start,
7864                                                       len, orig_block_len,
7865                                                       ram_bytes, type);
7866                         btrfs_dec_nocow_writers(fs_info, block_start);
7867                         if (type == BTRFS_ORDERED_PREALLOC) {
7868                                 free_extent_map(em);
7869                                 em = em2;
7870                         }
7871                         if (em2 && IS_ERR(em2)) {
7872                                 ret = PTR_ERR(em2);
7873                                 goto unlock_err;
7874                         }
7875                         /*
7876                          * For inode marked NODATACOW or extent marked PREALLOC,
7877                          * use the existing or preallocated extent, so does not
7878                          * need to adjust btrfs_space_info's bytes_may_use.
7879                          */
7880                         btrfs_free_reserved_data_space_noquota(inode,
7881                                         start, len);
7882                         goto unlock;
7883                 }
7884         }
7885
7886         /*
7887          * this will cow the extent, reset the len in case we changed
7888          * it above
7889          */
7890         len = bh_result->b_size;
7891         free_extent_map(em);
7892         em = btrfs_new_extent_direct(inode, start, len);
7893         if (IS_ERR(em)) {
7894                 ret = PTR_ERR(em);
7895                 goto unlock_err;
7896         }
7897         len = min(len, em->len - (start - em->start));
7898 unlock:
7899         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7900                 inode->i_blkbits;
7901         bh_result->b_size = len;
7902         bh_result->b_bdev = em->bdev;
7903         set_buffer_mapped(bh_result);
7904         if (create) {
7905                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7906                         set_buffer_new(bh_result);
7907
7908                 /*
7909                  * Need to update the i_size under the extent lock so buffered
7910                  * readers will get the updated i_size when we unlock.
7911                  */
7912                 if (!dio_data->overwrite && start + len > i_size_read(inode))
7913                         i_size_write(inode, start + len);
7914
7915                 WARN_ON(dio_data->reserve < len);
7916                 dio_data->reserve -= len;
7917                 dio_data->unsubmitted_oe_range_end = start + len;
7918                 current->journal_info = dio_data;
7919         }
7920
7921         /*
7922          * In the case of write we need to clear and unlock the entire range,
7923          * in the case of read we need to unlock only the end area that we
7924          * aren't using if there is any left over space.
7925          */
7926         if (lockstart < lockend) {
7927                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7928                                  lockend, unlock_bits, 1, 0,
7929                                  &cached_state, GFP_NOFS);
7930         } else {
7931                 free_extent_state(cached_state);
7932         }
7933
7934         free_extent_map(em);
7935
7936         return 0;
7937
7938 unlock_err:
7939         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7940                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7941 err:
7942         if (dio_data)
7943                 current->journal_info = dio_data;
7944         return ret;
7945 }
7946
7947 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7948                                                  struct bio *bio,
7949                                                  int mirror_num)
7950 {
7951         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7952         blk_status_t ret;
7953
7954         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7955
7956         bio_get(bio);
7957
7958         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7959         if (ret)
7960                 goto err;
7961
7962         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
7963 err:
7964         bio_put(bio);
7965         return ret;
7966 }
7967
7968 static int btrfs_check_dio_repairable(struct inode *inode,
7969                                       struct bio *failed_bio,
7970                                       struct io_failure_record *failrec,
7971                                       int failed_mirror)
7972 {
7973         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7974         int num_copies;
7975
7976         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7977         if (num_copies == 1) {
7978                 /*
7979                  * we only have a single copy of the data, so don't bother with
7980                  * all the retry and error correction code that follows. no
7981                  * matter what the error is, it is very likely to persist.
7982                  */
7983                 btrfs_debug(fs_info,
7984                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7985                         num_copies, failrec->this_mirror, failed_mirror);
7986                 return 0;
7987         }
7988
7989         failrec->failed_mirror = failed_mirror;
7990         failrec->this_mirror++;
7991         if (failrec->this_mirror == failed_mirror)
7992                 failrec->this_mirror++;
7993
7994         if (failrec->this_mirror > num_copies) {
7995                 btrfs_debug(fs_info,
7996                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7997                         num_copies, failrec->this_mirror, failed_mirror);
7998                 return 0;
7999         }
8000
8001         return 1;
8002 }
8003
8004 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
8005                                    struct page *page, unsigned int pgoff,
8006                                    u64 start, u64 end, int failed_mirror,
8007                                    bio_end_io_t *repair_endio, void *repair_arg)
8008 {
8009         struct io_failure_record *failrec;
8010         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8011         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
8012         struct bio *bio;
8013         int isector;
8014         unsigned int read_mode = 0;
8015         int segs;
8016         int ret;
8017         blk_status_t status;
8018
8019         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
8020
8021         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
8022         if (ret)
8023                 return errno_to_blk_status(ret);
8024
8025         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
8026                                          failed_mirror);
8027         if (!ret) {
8028                 free_io_failure(failure_tree, io_tree, failrec);
8029                 return BLK_STS_IOERR;
8030         }
8031
8032         segs = bio_segments(failed_bio);
8033         if (segs > 1 ||
8034             (failed_bio->bi_io_vec->bv_len > btrfs_inode_sectorsize(inode)))
8035                 read_mode |= REQ_FAILFAST_DEV;
8036
8037         isector = start - btrfs_io_bio(failed_bio)->logical;
8038         isector >>= inode->i_sb->s_blocksize_bits;
8039         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
8040                                 pgoff, isector, repair_endio, repair_arg);
8041         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
8042
8043         btrfs_debug(BTRFS_I(inode)->root->fs_info,
8044                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
8045                     read_mode, failrec->this_mirror, failrec->in_validation);
8046
8047         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
8048         if (status) {
8049                 free_io_failure(failure_tree, io_tree, failrec);
8050                 bio_put(bio);
8051         }
8052
8053         return status;
8054 }
8055
8056 struct btrfs_retry_complete {
8057         struct completion done;
8058         struct inode *inode;
8059         u64 start;
8060         int uptodate;
8061 };
8062
8063 static void btrfs_retry_endio_nocsum(struct bio *bio)
8064 {
8065         struct btrfs_retry_complete *done = bio->bi_private;
8066         struct inode *inode = done->inode;
8067         struct bio_vec *bvec;
8068         struct extent_io_tree *io_tree, *failure_tree;
8069         int i;
8070
8071         if (bio->bi_status)
8072                 goto end;
8073
8074         ASSERT(bio->bi_vcnt == 1);
8075         io_tree = &BTRFS_I(inode)->io_tree;
8076         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8077         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(inode));
8078
8079         done->uptodate = 1;
8080         ASSERT(!bio_flagged(bio, BIO_CLONED));
8081         bio_for_each_segment_all(bvec, bio, i)
8082                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
8083                                  io_tree, done->start, bvec->bv_page,
8084                                  btrfs_ino(BTRFS_I(inode)), 0);
8085 end:
8086         complete(&done->done);
8087         bio_put(bio);
8088 }
8089
8090 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
8091                                                 struct btrfs_io_bio *io_bio)
8092 {
8093         struct btrfs_fs_info *fs_info;
8094         struct bio_vec bvec;
8095         struct bvec_iter iter;
8096         struct btrfs_retry_complete done;
8097         u64 start;
8098         unsigned int pgoff;
8099         u32 sectorsize;
8100         int nr_sectors;
8101         blk_status_t ret;
8102         blk_status_t err = BLK_STS_OK;
8103
8104         fs_info = BTRFS_I(inode)->root->fs_info;
8105         sectorsize = fs_info->sectorsize;
8106
8107         start = io_bio->logical;
8108         done.inode = inode;
8109         io_bio->bio.bi_iter = io_bio->iter;
8110
8111         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8112                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8113                 pgoff = bvec.bv_offset;
8114
8115 next_block_or_try_again:
8116                 done.uptodate = 0;
8117                 done.start = start;
8118                 init_completion(&done.done);
8119
8120                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8121                                 pgoff, start, start + sectorsize - 1,
8122                                 io_bio->mirror_num,
8123                                 btrfs_retry_endio_nocsum, &done);
8124                 if (ret) {
8125                         err = ret;
8126                         goto next;
8127                 }
8128
8129                 wait_for_completion_io(&done.done);
8130
8131                 if (!done.uptodate) {
8132                         /* We might have another mirror, so try again */
8133                         goto next_block_or_try_again;
8134                 }
8135
8136 next:
8137                 start += sectorsize;
8138
8139                 nr_sectors--;
8140                 if (nr_sectors) {
8141                         pgoff += sectorsize;
8142                         ASSERT(pgoff < PAGE_SIZE);
8143                         goto next_block_or_try_again;
8144                 }
8145         }
8146
8147         return err;
8148 }
8149
8150 static void btrfs_retry_endio(struct bio *bio)
8151 {
8152         struct btrfs_retry_complete *done = bio->bi_private;
8153         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8154         struct extent_io_tree *io_tree, *failure_tree;
8155         struct inode *inode = done->inode;
8156         struct bio_vec *bvec;
8157         int uptodate;
8158         int ret;
8159         int i;
8160
8161         if (bio->bi_status)
8162                 goto end;
8163
8164         uptodate = 1;
8165
8166         ASSERT(bio->bi_vcnt == 1);
8167         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(done->inode));
8168
8169         io_tree = &BTRFS_I(inode)->io_tree;
8170         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8171
8172         ASSERT(!bio_flagged(bio, BIO_CLONED));
8173         bio_for_each_segment_all(bvec, bio, i) {
8174                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
8175                                              bvec->bv_offset, done->start,
8176                                              bvec->bv_len);
8177                 if (!ret)
8178                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
8179                                          failure_tree, io_tree, done->start,
8180                                          bvec->bv_page,
8181                                          btrfs_ino(BTRFS_I(inode)),
8182                                          bvec->bv_offset);
8183                 else
8184                         uptodate = 0;
8185         }
8186
8187         done->uptodate = uptodate;
8188 end:
8189         complete(&done->done);
8190         bio_put(bio);
8191 }
8192
8193 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
8194                 struct btrfs_io_bio *io_bio, blk_status_t err)
8195 {
8196         struct btrfs_fs_info *fs_info;
8197         struct bio_vec bvec;
8198         struct bvec_iter iter;
8199         struct btrfs_retry_complete done;
8200         u64 start;
8201         u64 offset = 0;
8202         u32 sectorsize;
8203         int nr_sectors;
8204         unsigned int pgoff;
8205         int csum_pos;
8206         bool uptodate = (err == 0);
8207         int ret;
8208         blk_status_t status;
8209
8210         fs_info = BTRFS_I(inode)->root->fs_info;
8211         sectorsize = fs_info->sectorsize;
8212
8213         err = BLK_STS_OK;
8214         start = io_bio->logical;
8215         done.inode = inode;
8216         io_bio->bio.bi_iter = io_bio->iter;
8217
8218         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8219                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8220
8221                 pgoff = bvec.bv_offset;
8222 next_block:
8223                 if (uptodate) {
8224                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8225                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
8226                                         bvec.bv_page, pgoff, start, sectorsize);
8227                         if (likely(!ret))
8228                                 goto next;
8229                 }
8230 try_again:
8231                 done.uptodate = 0;
8232                 done.start = start;
8233                 init_completion(&done.done);
8234
8235                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8236                                         pgoff, start, start + sectorsize - 1,
8237                                         io_bio->mirror_num, btrfs_retry_endio,
8238                                         &done);
8239                 if (status) {
8240                         err = status;
8241                         goto next;
8242                 }
8243
8244                 wait_for_completion_io(&done.done);
8245
8246                 if (!done.uptodate) {
8247                         /* We might have another mirror, so try again */
8248                         goto try_again;
8249                 }
8250 next:
8251                 offset += sectorsize;
8252                 start += sectorsize;
8253
8254                 ASSERT(nr_sectors);
8255
8256                 nr_sectors--;
8257                 if (nr_sectors) {
8258                         pgoff += sectorsize;
8259                         ASSERT(pgoff < PAGE_SIZE);
8260                         goto next_block;
8261                 }
8262         }
8263
8264         return err;
8265 }
8266
8267 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8268                 struct btrfs_io_bio *io_bio, blk_status_t err)
8269 {
8270         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8271
8272         if (skip_csum) {
8273                 if (unlikely(err))
8274                         return __btrfs_correct_data_nocsum(inode, io_bio);
8275                 else
8276                         return BLK_STS_OK;
8277         } else {
8278                 return __btrfs_subio_endio_read(inode, io_bio, err);
8279         }
8280 }
8281
8282 static void btrfs_endio_direct_read(struct bio *bio)
8283 {
8284         struct btrfs_dio_private *dip = bio->bi_private;
8285         struct inode *inode = dip->inode;
8286         struct bio *dio_bio;
8287         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8288         blk_status_t err = bio->bi_status;
8289
8290         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8291                 err = btrfs_subio_endio_read(inode, io_bio, err);
8292
8293         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8294                       dip->logical_offset + dip->bytes - 1);
8295         dio_bio = dip->dio_bio;
8296
8297         kfree(dip);
8298
8299         dio_bio->bi_status = err;
8300         dio_end_io(dio_bio);
8301
8302         if (io_bio->end_io)
8303                 io_bio->end_io(io_bio, blk_status_to_errno(err));
8304         bio_put(bio);
8305 }
8306
8307 static void __endio_write_update_ordered(struct inode *inode,
8308                                          const u64 offset, const u64 bytes,
8309                                          const bool uptodate)
8310 {
8311         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8312         struct btrfs_ordered_extent *ordered = NULL;
8313         struct btrfs_workqueue *wq;
8314         btrfs_work_func_t func;
8315         u64 ordered_offset = offset;
8316         u64 ordered_bytes = bytes;
8317         u64 last_offset;
8318         int ret;
8319
8320         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
8321                 wq = fs_info->endio_freespace_worker;
8322                 func = btrfs_freespace_write_helper;
8323         } else {
8324                 wq = fs_info->endio_write_workers;
8325                 func = btrfs_endio_write_helper;
8326         }
8327
8328 again:
8329         last_offset = ordered_offset;
8330         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8331                                                    &ordered_offset,
8332                                                    ordered_bytes,
8333                                                    uptodate);
8334         if (!ret)
8335                 goto out_test;
8336
8337         btrfs_init_work(&ordered->work, func, finish_ordered_fn, NULL, NULL);
8338         btrfs_queue_work(wq, &ordered->work);
8339 out_test:
8340         /*
8341          * If btrfs_dec_test_ordered_pending does not find any ordered extent
8342          * in the range, we can exit.
8343          */
8344         if (ordered_offset == last_offset)
8345                 return;
8346         /*
8347          * our bio might span multiple ordered extents.  If we haven't
8348          * completed the accounting for the whole dio, go back and try again
8349          */
8350         if (ordered_offset < offset + bytes) {
8351                 ordered_bytes = offset + bytes - ordered_offset;
8352                 ordered = NULL;
8353                 goto again;
8354         }
8355 }
8356
8357 static void btrfs_endio_direct_write(struct bio *bio)
8358 {
8359         struct btrfs_dio_private *dip = bio->bi_private;
8360         struct bio *dio_bio = dip->dio_bio;
8361
8362         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8363                                      dip->bytes, !bio->bi_status);
8364
8365         kfree(dip);
8366
8367         dio_bio->bi_status = bio->bi_status;
8368         dio_end_io(dio_bio);
8369         bio_put(bio);
8370 }
8371
8372 static blk_status_t __btrfs_submit_bio_start_direct_io(void *private_data,
8373                                     struct bio *bio, int mirror_num,
8374                                     unsigned long bio_flags, u64 offset)
8375 {
8376         struct inode *inode = private_data;
8377         blk_status_t ret;
8378         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8379         BUG_ON(ret); /* -ENOMEM */
8380         return 0;
8381 }
8382
8383 static void btrfs_end_dio_bio(struct bio *bio)
8384 {
8385         struct btrfs_dio_private *dip = bio->bi_private;
8386         blk_status_t err = bio->bi_status;
8387
8388         if (err)
8389                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8390                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8391                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8392                            bio->bi_opf,
8393                            (unsigned long long)bio->bi_iter.bi_sector,
8394                            bio->bi_iter.bi_size, err);
8395
8396         if (dip->subio_endio)
8397                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8398
8399         if (err) {
8400                 dip->errors = 1;
8401
8402                 /*
8403                  * before atomic variable goto zero, we must make sure
8404                  * dip->errors is perceived to be set.
8405                  */
8406                 smp_mb__before_atomic();
8407         }
8408
8409         /* if there are more bios still pending for this dio, just exit */
8410         if (!atomic_dec_and_test(&dip->pending_bios))
8411                 goto out;
8412
8413         if (dip->errors) {
8414                 bio_io_error(dip->orig_bio);
8415         } else {
8416                 dip->dio_bio->bi_status = BLK_STS_OK;
8417                 bio_endio(dip->orig_bio);
8418         }
8419 out:
8420         bio_put(bio);
8421 }
8422
8423 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8424                                                  struct btrfs_dio_private *dip,
8425                                                  struct bio *bio,
8426                                                  u64 file_offset)
8427 {
8428         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8429         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8430         blk_status_t ret;
8431
8432         /*
8433          * We load all the csum data we need when we submit
8434          * the first bio to reduce the csum tree search and
8435          * contention.
8436          */
8437         if (dip->logical_offset == file_offset) {
8438                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8439                                                 file_offset);
8440                 if (ret)
8441                         return ret;
8442         }
8443
8444         if (bio == dip->orig_bio)
8445                 return 0;
8446
8447         file_offset -= dip->logical_offset;
8448         file_offset >>= inode->i_sb->s_blocksize_bits;
8449         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8450
8451         return 0;
8452 }
8453
8454 static inline blk_status_t
8455 __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode, u64 file_offset,
8456                        int async_submit)
8457 {
8458         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8459         struct btrfs_dio_private *dip = bio->bi_private;
8460         bool write = bio_op(bio) == REQ_OP_WRITE;
8461         blk_status_t ret;
8462
8463         if (async_submit)
8464                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8465
8466         bio_get(bio);
8467
8468         if (!write) {
8469                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8470                 if (ret)
8471                         goto err;
8472         }
8473
8474         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8475                 goto map;
8476
8477         if (write && async_submit) {
8478                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8479                                           file_offset, inode,
8480                                           __btrfs_submit_bio_start_direct_io,
8481                                           __btrfs_submit_bio_done);
8482                 goto err;
8483         } else if (write) {
8484                 /*
8485                  * If we aren't doing async submit, calculate the csum of the
8486                  * bio now.
8487                  */
8488                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8489                 if (ret)
8490                         goto err;
8491         } else {
8492                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8493                                                      file_offset);
8494                 if (ret)
8495                         goto err;
8496         }
8497 map:
8498         ret = btrfs_map_bio(fs_info, bio, 0, 0);
8499 err:
8500         bio_put(bio);
8501         return ret;
8502 }
8503
8504 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8505 {
8506         struct inode *inode = dip->inode;
8507         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8508         struct bio *bio;
8509         struct bio *orig_bio = dip->orig_bio;
8510         u64 start_sector = orig_bio->bi_iter.bi_sector;
8511         u64 file_offset = dip->logical_offset;
8512         u64 map_length;
8513         int async_submit = 0;
8514         u64 submit_len;
8515         int clone_offset = 0;
8516         int clone_len;
8517         int ret;
8518         blk_status_t status;
8519
8520         map_length = orig_bio->bi_iter.bi_size;
8521         submit_len = map_length;
8522         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio), start_sector << 9,
8523                               &map_length, NULL, 0);
8524         if (ret)
8525                 return -EIO;
8526
8527         if (map_length >= submit_len) {
8528                 bio = orig_bio;
8529                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8530                 goto submit;
8531         }
8532
8533         /* async crcs make it difficult to collect full stripe writes. */
8534         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8535                 async_submit = 0;
8536         else
8537                 async_submit = 1;
8538
8539         /* bio split */
8540         ASSERT(map_length <= INT_MAX);
8541         atomic_inc(&dip->pending_bios);
8542         do {
8543                 clone_len = min_t(int, submit_len, map_length);
8544
8545                 /*
8546                  * This will never fail as it's passing GPF_NOFS and
8547                  * the allocation is backed by btrfs_bioset.
8548                  */
8549                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8550                                               clone_len);
8551                 bio->bi_private = dip;
8552                 bio->bi_end_io = btrfs_end_dio_bio;
8553                 btrfs_io_bio(bio)->logical = file_offset;
8554
8555                 ASSERT(submit_len >= clone_len);
8556                 submit_len -= clone_len;
8557                 if (submit_len == 0)
8558                         break;
8559
8560                 /*
8561                  * Increase the count before we submit the bio so we know
8562                  * the end IO handler won't happen before we increase the
8563                  * count. Otherwise, the dip might get freed before we're
8564                  * done setting it up.
8565                  */
8566                 atomic_inc(&dip->pending_bios);
8567
8568                 status = __btrfs_submit_dio_bio(bio, inode, file_offset,
8569                                                 async_submit);
8570                 if (status) {
8571                         bio_put(bio);
8572                         atomic_dec(&dip->pending_bios);
8573                         goto out_err;
8574                 }
8575
8576                 clone_offset += clone_len;
8577                 start_sector += clone_len >> 9;
8578                 file_offset += clone_len;
8579
8580                 map_length = submit_len;
8581                 ret = btrfs_map_block(fs_info, btrfs_op(orig_bio),
8582                                       start_sector << 9, &map_length, NULL, 0);
8583                 if (ret)
8584                         goto out_err;
8585         } while (submit_len > 0);
8586
8587 submit:
8588         status = __btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8589         if (!status)
8590                 return 0;
8591
8592         bio_put(bio);
8593 out_err:
8594         dip->errors = 1;
8595         /*
8596          * before atomic variable goto zero, we must
8597          * make sure dip->errors is perceived to be set.
8598          */
8599         smp_mb__before_atomic();
8600         if (atomic_dec_and_test(&dip->pending_bios))
8601                 bio_io_error(dip->orig_bio);
8602
8603         /* bio_end_io() will handle error, so we needn't return it */
8604         return 0;
8605 }
8606
8607 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8608                                 loff_t file_offset)
8609 {
8610         struct btrfs_dio_private *dip = NULL;
8611         struct bio *bio = NULL;
8612         struct btrfs_io_bio *io_bio;
8613         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8614         int ret = 0;
8615
8616         bio = btrfs_bio_clone(dio_bio);
8617
8618         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8619         if (!dip) {
8620                 ret = -ENOMEM;
8621                 goto free_ordered;
8622         }
8623
8624         dip->private = dio_bio->bi_private;
8625         dip->inode = inode;
8626         dip->logical_offset = file_offset;
8627         dip->bytes = dio_bio->bi_iter.bi_size;
8628         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8629         bio->bi_private = dip;
8630         dip->orig_bio = bio;
8631         dip->dio_bio = dio_bio;
8632         atomic_set(&dip->pending_bios, 0);
8633         io_bio = btrfs_io_bio(bio);
8634         io_bio->logical = file_offset;
8635
8636         if (write) {
8637                 bio->bi_end_io = btrfs_endio_direct_write;
8638         } else {
8639                 bio->bi_end_io = btrfs_endio_direct_read;
8640                 dip->subio_endio = btrfs_subio_endio_read;
8641         }
8642
8643         /*
8644          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8645          * even if we fail to submit a bio, because in such case we do the
8646          * corresponding error handling below and it must not be done a second
8647          * time by btrfs_direct_IO().
8648          */
8649         if (write) {
8650                 struct btrfs_dio_data *dio_data = current->journal_info;
8651
8652                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8653                         dip->bytes;
8654                 dio_data->unsubmitted_oe_range_start =
8655                         dio_data->unsubmitted_oe_range_end;
8656         }
8657
8658         ret = btrfs_submit_direct_hook(dip);
8659         if (!ret)
8660                 return;
8661
8662         if (io_bio->end_io)
8663                 io_bio->end_io(io_bio, ret);
8664
8665 free_ordered:
8666         /*
8667          * If we arrived here it means either we failed to submit the dip
8668          * or we either failed to clone the dio_bio or failed to allocate the
8669          * dip. If we cloned the dio_bio and allocated the dip, we can just
8670          * call bio_endio against our io_bio so that we get proper resource
8671          * cleanup if we fail to submit the dip, otherwise, we must do the
8672          * same as btrfs_endio_direct_[write|read] because we can't call these
8673          * callbacks - they require an allocated dip and a clone of dio_bio.
8674          */
8675         if (bio && dip) {
8676                 bio_io_error(bio);
8677                 /*
8678                  * The end io callbacks free our dip, do the final put on bio
8679                  * and all the cleanup and final put for dio_bio (through
8680                  * dio_end_io()).
8681                  */
8682                 dip = NULL;
8683                 bio = NULL;
8684         } else {
8685                 if (write)
8686                         __endio_write_update_ordered(inode,
8687                                                 file_offset,
8688                                                 dio_bio->bi_iter.bi_size,
8689                                                 false);
8690                 else
8691                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8692                               file_offset + dio_bio->bi_iter.bi_size - 1);
8693
8694                 dio_bio->bi_status = BLK_STS_IOERR;
8695                 /*
8696                  * Releases and cleans up our dio_bio, no need to bio_put()
8697                  * nor bio_endio()/bio_io_error() against dio_bio.
8698                  */
8699                 dio_end_io(dio_bio);
8700         }
8701         if (bio)
8702                 bio_put(bio);
8703         kfree(dip);
8704 }
8705
8706 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8707                                const struct iov_iter *iter, loff_t offset)
8708 {
8709         int seg;
8710         int i;
8711         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8712         ssize_t retval = -EINVAL;
8713
8714         if (offset & blocksize_mask)
8715                 goto out;
8716
8717         if (iov_iter_alignment(iter) & blocksize_mask)
8718                 goto out;
8719
8720         /* If this is a write we don't need to check anymore */
8721         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8722                 return 0;
8723         /*
8724          * Check to make sure we don't have duplicate iov_base's in this
8725          * iovec, if so return EINVAL, otherwise we'll get csum errors
8726          * when reading back.
8727          */
8728         for (seg = 0; seg < iter->nr_segs; seg++) {
8729                 for (i = seg + 1; i < iter->nr_segs; i++) {
8730                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8731                                 goto out;
8732                 }
8733         }
8734         retval = 0;
8735 out:
8736         return retval;
8737 }
8738
8739 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8740 {
8741         struct file *file = iocb->ki_filp;
8742         struct inode *inode = file->f_mapping->host;
8743         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8744         struct btrfs_dio_data dio_data = { 0 };
8745         struct extent_changeset *data_reserved = NULL;
8746         loff_t offset = iocb->ki_pos;
8747         size_t count = 0;
8748         int flags = 0;
8749         bool wakeup = true;
8750         bool relock = false;
8751         ssize_t ret;
8752
8753         if (check_direct_IO(fs_info, iter, offset))
8754                 return 0;
8755
8756         inode_dio_begin(inode);
8757
8758         /*
8759          * The generic stuff only does filemap_write_and_wait_range, which
8760          * isn't enough if we've written compressed pages to this area, so
8761          * we need to flush the dirty pages again to make absolutely sure
8762          * that any outstanding dirty pages are on disk.
8763          */
8764         count = iov_iter_count(iter);
8765         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8766                      &BTRFS_I(inode)->runtime_flags))
8767                 filemap_fdatawrite_range(inode->i_mapping, offset,
8768                                          offset + count - 1);
8769
8770         if (iov_iter_rw(iter) == WRITE) {
8771                 /*
8772                  * If the write DIO is beyond the EOF, we need update
8773                  * the isize, but it is protected by i_mutex. So we can
8774                  * not unlock the i_mutex at this case.
8775                  */
8776                 if (offset + count <= inode->i_size) {
8777                         dio_data.overwrite = 1;
8778                         inode_unlock(inode);
8779                         relock = true;
8780                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8781                         ret = -EAGAIN;
8782                         goto out;
8783                 }
8784                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8785                                                    offset, count);
8786                 if (ret)
8787                         goto out;
8788
8789                 /*
8790                  * We need to know how many extents we reserved so that we can
8791                  * do the accounting properly if we go over the number we
8792                  * originally calculated.  Abuse current->journal_info for this.
8793                  */
8794                 dio_data.reserve = round_up(count,
8795                                             fs_info->sectorsize);
8796                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8797                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8798                 current->journal_info = &dio_data;
8799                 down_read(&BTRFS_I(inode)->dio_sem);
8800         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8801                                      &BTRFS_I(inode)->runtime_flags)) {
8802                 inode_dio_end(inode);
8803                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8804                 wakeup = false;
8805         }
8806
8807         ret = __blockdev_direct_IO(iocb, inode,
8808                                    fs_info->fs_devices->latest_bdev,
8809                                    iter, btrfs_get_blocks_direct, NULL,
8810                                    btrfs_submit_direct, flags);
8811         if (iov_iter_rw(iter) == WRITE) {
8812                 up_read(&BTRFS_I(inode)->dio_sem);
8813                 current->journal_info = NULL;
8814                 if (ret < 0 && ret != -EIOCBQUEUED) {
8815                         if (dio_data.reserve)
8816                                 btrfs_delalloc_release_space(inode, data_reserved,
8817                                         offset, dio_data.reserve);
8818                         /*
8819                          * On error we might have left some ordered extents
8820                          * without submitting corresponding bios for them, so
8821                          * cleanup them up to avoid other tasks getting them
8822                          * and waiting for them to complete forever.
8823                          */
8824                         if (dio_data.unsubmitted_oe_range_start <
8825                             dio_data.unsubmitted_oe_range_end)
8826                                 __endio_write_update_ordered(inode,
8827                                         dio_data.unsubmitted_oe_range_start,
8828                                         dio_data.unsubmitted_oe_range_end -
8829                                         dio_data.unsubmitted_oe_range_start,
8830                                         false);
8831                 } else if (ret >= 0 && (size_t)ret < count)
8832                         btrfs_delalloc_release_space(inode, data_reserved,
8833                                         offset, count - (size_t)ret);
8834                 btrfs_delalloc_release_extents(BTRFS_I(inode), count);
8835         }
8836 out:
8837         if (wakeup)
8838                 inode_dio_end(inode);
8839         if (relock)
8840                 inode_lock(inode);
8841
8842         extent_changeset_free(data_reserved);
8843         return ret;
8844 }
8845
8846 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8847
8848 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8849                 __u64 start, __u64 len)
8850 {
8851         int     ret;
8852
8853         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8854         if (ret)
8855                 return ret;
8856
8857         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8858 }
8859
8860 int btrfs_readpage(struct file *file, struct page *page)
8861 {
8862         struct extent_io_tree *tree;
8863         tree = &BTRFS_I(page->mapping->host)->io_tree;
8864         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8865 }
8866
8867 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8868 {
8869         struct extent_io_tree *tree;
8870         struct inode *inode = page->mapping->host;
8871         int ret;
8872
8873         if (current->flags & PF_MEMALLOC) {
8874                 redirty_page_for_writepage(wbc, page);
8875                 unlock_page(page);
8876                 return 0;
8877         }
8878
8879         /*
8880          * If we are under memory pressure we will call this directly from the
8881          * VM, we need to make sure we have the inode referenced for the ordered
8882          * extent.  If not just return like we didn't do anything.
8883          */
8884         if (!igrab(inode)) {
8885                 redirty_page_for_writepage(wbc, page);
8886                 return AOP_WRITEPAGE_ACTIVATE;
8887         }
8888         tree = &BTRFS_I(page->mapping->host)->io_tree;
8889         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8890         btrfs_add_delayed_iput(inode);
8891         return ret;
8892 }
8893
8894 static int btrfs_writepages(struct address_space *mapping,
8895                             struct writeback_control *wbc)
8896 {
8897         struct extent_io_tree *tree;
8898
8899         tree = &BTRFS_I(mapping->host)->io_tree;
8900         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8901 }
8902
8903 static int
8904 btrfs_readpages(struct file *file, struct address_space *mapping,
8905                 struct list_head *pages, unsigned nr_pages)
8906 {
8907         struct extent_io_tree *tree;
8908         tree = &BTRFS_I(mapping->host)->io_tree;
8909         return extent_readpages(tree, mapping, pages, nr_pages,
8910                                 btrfs_get_extent);
8911 }
8912 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8913 {
8914         struct extent_io_tree *tree;
8915         struct extent_map_tree *map;
8916         int ret;
8917
8918         tree = &BTRFS_I(page->mapping->host)->io_tree;
8919         map = &BTRFS_I(page->mapping->host)->extent_tree;
8920         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8921         if (ret == 1) {
8922                 ClearPagePrivate(page);
8923                 set_page_private(page, 0);
8924                 put_page(page);
8925         }
8926         return ret;
8927 }
8928
8929 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8930 {
8931         if (PageWriteback(page) || PageDirty(page))
8932                 return 0;
8933         return __btrfs_releasepage(page, gfp_flags);
8934 }
8935
8936 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8937                                  unsigned int length)
8938 {
8939         struct inode *inode = page->mapping->host;
8940         struct extent_io_tree *tree;
8941         struct btrfs_ordered_extent *ordered;
8942         struct extent_state *cached_state = NULL;
8943         u64 page_start = page_offset(page);
8944         u64 page_end = page_start + PAGE_SIZE - 1;
8945         u64 start;
8946         u64 end;
8947         int inode_evicting = inode->i_state & I_FREEING;
8948
8949         /*
8950          * we have the page locked, so new writeback can't start,
8951          * and the dirty bit won't be cleared while we are here.
8952          *
8953          * Wait for IO on this page so that we can safely clear
8954          * the PagePrivate2 bit and do ordered accounting
8955          */
8956         wait_on_page_writeback(page);
8957
8958         tree = &BTRFS_I(inode)->io_tree;
8959         if (offset) {
8960                 btrfs_releasepage(page, GFP_NOFS);
8961                 return;
8962         }
8963
8964         if (!inode_evicting)
8965                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8966 again:
8967         start = page_start;
8968         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8969                                         page_end - start + 1);
8970         if (ordered) {
8971                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8972                 /*
8973                  * IO on this page will never be started, so we need
8974                  * to account for any ordered extents now
8975                  */
8976                 if (!inode_evicting)
8977                         clear_extent_bit(tree, start, end,
8978                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8979                                          EXTENT_DELALLOC_NEW |
8980                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8981                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8982                                          GFP_NOFS);
8983                 /*
8984                  * whoever cleared the private bit is responsible
8985                  * for the finish_ordered_io
8986                  */
8987                 if (TestClearPagePrivate2(page)) {
8988                         struct btrfs_ordered_inode_tree *tree;
8989                         u64 new_len;
8990
8991                         tree = &BTRFS_I(inode)->ordered_tree;
8992
8993                         spin_lock_irq(&tree->lock);
8994                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8995                         new_len = start - ordered->file_offset;
8996                         if (new_len < ordered->truncated_len)
8997                                 ordered->truncated_len = new_len;
8998                         spin_unlock_irq(&tree->lock);
8999
9000                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
9001                                                            start,
9002                                                            end - start + 1, 1))
9003                                 btrfs_finish_ordered_io(ordered);
9004                 }
9005                 btrfs_put_ordered_extent(ordered);
9006                 if (!inode_evicting) {
9007                         cached_state = NULL;
9008                         lock_extent_bits(tree, start, end,
9009                                          &cached_state);
9010                 }
9011
9012                 start = end + 1;
9013                 if (start < page_end)
9014                         goto again;
9015         }
9016
9017         /*
9018          * Qgroup reserved space handler
9019          * Page here will be either
9020          * 1) Already written to disk
9021          *    In this case, its reserved space is released from data rsv map
9022          *    and will be freed by delayed_ref handler finally.
9023          *    So even we call qgroup_free_data(), it won't decrease reserved
9024          *    space.
9025          * 2) Not written to disk
9026          *    This means the reserved space should be freed here. However,
9027          *    if a truncate invalidates the page (by clearing PageDirty)
9028          *    and the page is accounted for while allocating extent
9029          *    in btrfs_check_data_free_space() we let delayed_ref to
9030          *    free the entire extent.
9031          */
9032         if (PageDirty(page))
9033                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
9034         if (!inode_evicting) {
9035                 clear_extent_bit(tree, page_start, page_end,
9036                                  EXTENT_LOCKED | EXTENT_DIRTY |
9037                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
9038                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
9039                                  &cached_state, GFP_NOFS);
9040
9041                 __btrfs_releasepage(page, GFP_NOFS);
9042         }
9043
9044         ClearPageChecked(page);
9045         if (PagePrivate(page)) {
9046                 ClearPagePrivate(page);
9047                 set_page_private(page, 0);
9048                 put_page(page);
9049         }
9050 }
9051
9052 /*
9053  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
9054  * called from a page fault handler when a page is first dirtied. Hence we must
9055  * be careful to check for EOF conditions here. We set the page up correctly
9056  * for a written page which means we get ENOSPC checking when writing into
9057  * holes and correct delalloc and unwritten extent mapping on filesystems that
9058  * support these features.
9059  *
9060  * We are not allowed to take the i_mutex here so we have to play games to
9061  * protect against truncate races as the page could now be beyond EOF.  Because
9062  * vmtruncate() writes the inode size before removing pages, once we have the
9063  * page lock we can determine safely if the page is beyond EOF. If it is not
9064  * beyond EOF, then the page is guaranteed safe against truncation until we
9065  * unlock the page.
9066  */
9067 int btrfs_page_mkwrite(struct vm_fault *vmf)
9068 {
9069         struct page *page = vmf->page;
9070         struct inode *inode = file_inode(vmf->vma->vm_file);
9071         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9072         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
9073         struct btrfs_ordered_extent *ordered;
9074         struct extent_state *cached_state = NULL;
9075         struct extent_changeset *data_reserved = NULL;
9076         char *kaddr;
9077         unsigned long zero_start;
9078         loff_t size;
9079         int ret;
9080         int reserved = 0;
9081         u64 reserved_space;
9082         u64 page_start;
9083         u64 page_end;
9084         u64 end;
9085
9086         reserved_space = PAGE_SIZE;
9087
9088         sb_start_pagefault(inode->i_sb);
9089         page_start = page_offset(page);
9090         page_end = page_start + PAGE_SIZE - 1;
9091         end = page_end;
9092
9093         /*
9094          * Reserving delalloc space after obtaining the page lock can lead to
9095          * deadlock. For example, if a dirty page is locked by this function
9096          * and the call to btrfs_delalloc_reserve_space() ends up triggering
9097          * dirty page write out, then the btrfs_writepage() function could
9098          * end up waiting indefinitely to get a lock on the page currently
9099          * being processed by btrfs_page_mkwrite() function.
9100          */
9101         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
9102                                            reserved_space);
9103         if (!ret) {
9104                 ret = file_update_time(vmf->vma->vm_file);
9105                 reserved = 1;
9106         }
9107         if (ret) {
9108                 if (ret == -ENOMEM)
9109                         ret = VM_FAULT_OOM;
9110                 else /* -ENOSPC, -EIO, etc */
9111                         ret = VM_FAULT_SIGBUS;
9112                 if (reserved)
9113                         goto out;
9114                 goto out_noreserve;
9115         }
9116
9117         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
9118 again:
9119         lock_page(page);
9120         size = i_size_read(inode);
9121
9122         if ((page->mapping != inode->i_mapping) ||
9123             (page_start >= size)) {
9124                 /* page got truncated out from underneath us */
9125                 goto out_unlock;
9126         }
9127         wait_on_page_writeback(page);
9128
9129         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
9130         set_page_extent_mapped(page);
9131
9132         /*
9133          * we can't set the delalloc bits if there are pending ordered
9134          * extents.  Drop our locks and wait for them to finish
9135          */
9136         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
9137                         PAGE_SIZE);
9138         if (ordered) {
9139                 unlock_extent_cached(io_tree, page_start, page_end,
9140                                      &cached_state, GFP_NOFS);
9141                 unlock_page(page);
9142                 btrfs_start_ordered_extent(inode, ordered, 1);
9143                 btrfs_put_ordered_extent(ordered);
9144                 goto again;
9145         }
9146
9147         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9148                 reserved_space = round_up(size - page_start,
9149                                           fs_info->sectorsize);
9150                 if (reserved_space < PAGE_SIZE) {
9151                         end = page_start + reserved_space - 1;
9152                         btrfs_delalloc_release_space(inode, data_reserved,
9153                                         page_start, PAGE_SIZE - reserved_space);
9154                 }
9155         }
9156
9157         /*
9158          * page_mkwrite gets called when the page is firstly dirtied after it's
9159          * faulted in, but write(2) could also dirty a page and set delalloc
9160          * bits, thus in this case for space account reason, we still need to
9161          * clear any delalloc bits within this page range since we have to
9162          * reserve data&meta space before lock_page() (see above comments).
9163          */
9164         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9165                           EXTENT_DIRTY | EXTENT_DELALLOC |
9166                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9167                           0, 0, &cached_state, GFP_NOFS);
9168
9169         ret = btrfs_set_extent_delalloc(inode, page_start, end, 0,
9170                                         &cached_state, 0);
9171         if (ret) {
9172                 unlock_extent_cached(io_tree, page_start, page_end,
9173                                      &cached_state, GFP_NOFS);
9174                 ret = VM_FAULT_SIGBUS;
9175                 goto out_unlock;
9176         }
9177         ret = 0;
9178
9179         /* page is wholly or partially inside EOF */
9180         if (page_start + PAGE_SIZE > size)
9181                 zero_start = size & ~PAGE_MASK;
9182         else
9183                 zero_start = PAGE_SIZE;
9184
9185         if (zero_start != PAGE_SIZE) {
9186                 kaddr = kmap(page);
9187                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9188                 flush_dcache_page(page);
9189                 kunmap(page);
9190         }
9191         ClearPageChecked(page);
9192         set_page_dirty(page);
9193         SetPageUptodate(page);
9194
9195         BTRFS_I(inode)->last_trans = fs_info->generation;
9196         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9197         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9198
9199         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
9200
9201 out_unlock:
9202         if (!ret) {
9203                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9204                 sb_end_pagefault(inode->i_sb);
9205                 extent_changeset_free(data_reserved);
9206                 return VM_FAULT_LOCKED;
9207         }
9208         unlock_page(page);
9209 out:
9210         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9211         btrfs_delalloc_release_space(inode, data_reserved, page_start,
9212                                      reserved_space);
9213 out_noreserve:
9214         sb_end_pagefault(inode->i_sb);
9215         extent_changeset_free(data_reserved);
9216         return ret;
9217 }
9218
9219 static int btrfs_truncate(struct inode *inode)
9220 {
9221         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9222         struct btrfs_root *root = BTRFS_I(inode)->root;
9223         struct btrfs_block_rsv *rsv;
9224         int ret = 0;
9225         int err = 0;
9226         struct btrfs_trans_handle *trans;
9227         u64 mask = fs_info->sectorsize - 1;
9228         u64 min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
9229
9230         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9231                                        (u64)-1);
9232         if (ret)
9233                 return ret;
9234
9235         /*
9236          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9237          * 3 things going on here
9238          *
9239          * 1) We need to reserve space for our orphan item and the space to
9240          * delete our orphan item.  Lord knows we don't want to have a dangling
9241          * orphan item because we didn't reserve space to remove it.
9242          *
9243          * 2) We need to reserve space to update our inode.
9244          *
9245          * 3) We need to have something to cache all the space that is going to
9246          * be free'd up by the truncate operation, but also have some slack
9247          * space reserved in case it uses space during the truncate (thank you
9248          * very much snapshotting).
9249          *
9250          * And we need these to all be separate.  The fact is we can use a lot of
9251          * space doing the truncate, and we have no earthly idea how much space
9252          * we will use, so we need the truncate reservation to be separate so it
9253          * doesn't end up using space reserved for updating the inode or
9254          * removing the orphan item.  We also need to be able to stop the
9255          * transaction and start a new one, which means we need to be able to
9256          * update the inode several times, and we have no idea of knowing how
9257          * many times that will be, so we can't just reserve 1 item for the
9258          * entirety of the operation, so that has to be done separately as well.
9259          * Then there is the orphan item, which does indeed need to be held on
9260          * to for the whole operation, and we need nobody to touch this reserved
9261          * space except the orphan code.
9262          *
9263          * So that leaves us with
9264          *
9265          * 1) root->orphan_block_rsv - for the orphan deletion.
9266          * 2) rsv - for the truncate reservation, which we will steal from the
9267          * transaction reservation.
9268          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9269          * updating the inode.
9270          */
9271         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9272         if (!rsv)
9273                 return -ENOMEM;
9274         rsv->size = min_size;
9275         rsv->failfast = 1;
9276
9277         /*
9278          * 1 for the truncate slack space
9279          * 1 for updating the inode.
9280          */
9281         trans = btrfs_start_transaction(root, 2);
9282         if (IS_ERR(trans)) {
9283                 err = PTR_ERR(trans);
9284                 goto out;
9285         }
9286
9287         /* Migrate the slack space for the truncate to our reserve */
9288         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9289                                       min_size, 0);
9290         BUG_ON(ret);
9291
9292         /*
9293          * So if we truncate and then write and fsync we normally would just
9294          * write the extents that changed, which is a problem if we need to
9295          * first truncate that entire inode.  So set this flag so we write out
9296          * all of the extents in the inode to the sync log so we're completely
9297          * safe.
9298          */
9299         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9300         trans->block_rsv = rsv;
9301
9302         while (1) {
9303                 ret = btrfs_truncate_inode_items(trans, root, inode,
9304                                                  inode->i_size,
9305                                                  BTRFS_EXTENT_DATA_KEY);
9306                 trans->block_rsv = &fs_info->trans_block_rsv;
9307                 if (ret != -ENOSPC && ret != -EAGAIN) {
9308                         err = ret;
9309                         break;
9310                 }
9311
9312                 ret = btrfs_update_inode(trans, root, inode);
9313                 if (ret) {
9314                         err = ret;
9315                         break;
9316                 }
9317
9318                 btrfs_end_transaction(trans);
9319                 btrfs_btree_balance_dirty(fs_info);
9320
9321                 trans = btrfs_start_transaction(root, 2);
9322                 if (IS_ERR(trans)) {
9323                         ret = err = PTR_ERR(trans);
9324                         trans = NULL;
9325                         break;
9326                 }
9327
9328                 btrfs_block_rsv_release(fs_info, rsv, -1);
9329                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9330                                               rsv, min_size, 0);
9331                 BUG_ON(ret);    /* shouldn't happen */
9332                 trans->block_rsv = rsv;
9333         }
9334
9335         /*
9336          * We can't call btrfs_truncate_block inside a trans handle as we could
9337          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
9338          * we've truncated everything except the last little bit, and can do
9339          * btrfs_truncate_block and then update the disk_i_size.
9340          */
9341         if (ret == NEED_TRUNCATE_BLOCK) {
9342                 btrfs_end_transaction(trans);
9343                 btrfs_btree_balance_dirty(fs_info);
9344
9345                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
9346                 if (ret)
9347                         goto out;
9348                 trans = btrfs_start_transaction(root, 1);
9349                 if (IS_ERR(trans)) {
9350                         ret = PTR_ERR(trans);
9351                         goto out;
9352                 }
9353                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
9354         }
9355
9356         if (ret == 0 && inode->i_nlink > 0) {
9357                 trans->block_rsv = root->orphan_block_rsv;
9358                 ret = btrfs_orphan_del(trans, BTRFS_I(inode));
9359                 if (ret)
9360                         err = ret;
9361         }
9362
9363         if (trans) {
9364                 trans->block_rsv = &fs_info->trans_block_rsv;
9365                 ret = btrfs_update_inode(trans, root, inode);
9366                 if (ret && !err)
9367                         err = ret;
9368
9369                 ret = btrfs_end_transaction(trans);
9370                 btrfs_btree_balance_dirty(fs_info);
9371         }
9372 out:
9373         btrfs_free_block_rsv(fs_info, rsv);
9374
9375         if (ret && !err)
9376                 err = ret;
9377
9378         return err;
9379 }
9380
9381 /*
9382  * create a new subvolume directory/inode (helper for the ioctl).
9383  */
9384 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9385                              struct btrfs_root *new_root,
9386                              struct btrfs_root *parent_root,
9387                              u64 new_dirid)
9388 {
9389         struct inode *inode;
9390         int err;
9391         u64 index = 0;
9392
9393         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9394                                 new_dirid, new_dirid,
9395                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9396                                 &index);
9397         if (IS_ERR(inode))
9398                 return PTR_ERR(inode);
9399         inode->i_op = &btrfs_dir_inode_operations;
9400         inode->i_fop = &btrfs_dir_file_operations;
9401
9402         set_nlink(inode, 1);
9403         btrfs_i_size_write(BTRFS_I(inode), 0);
9404         unlock_new_inode(inode);
9405
9406         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9407         if (err)
9408                 btrfs_err(new_root->fs_info,
9409                           "error inheriting subvolume %llu properties: %d",
9410                           new_root->root_key.objectid, err);
9411
9412         err = btrfs_update_inode(trans, new_root, inode);
9413
9414         iput(inode);
9415         return err;
9416 }
9417
9418 struct inode *btrfs_alloc_inode(struct super_block *sb)
9419 {
9420         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
9421         struct btrfs_inode *ei;
9422         struct inode *inode;
9423
9424         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9425         if (!ei)
9426                 return NULL;
9427
9428         ei->root = NULL;
9429         ei->generation = 0;
9430         ei->last_trans = 0;
9431         ei->last_sub_trans = 0;
9432         ei->logged_trans = 0;
9433         ei->delalloc_bytes = 0;
9434         ei->new_delalloc_bytes = 0;
9435         ei->defrag_bytes = 0;
9436         ei->disk_i_size = 0;
9437         ei->flags = 0;
9438         ei->csum_bytes = 0;
9439         ei->index_cnt = (u64)-1;
9440         ei->dir_index = 0;
9441         ei->last_unlink_trans = 0;
9442         ei->last_log_commit = 0;
9443         ei->delayed_iput_count = 0;
9444
9445         spin_lock_init(&ei->lock);
9446         ei->outstanding_extents = 0;
9447         if (sb->s_magic != BTRFS_TEST_MAGIC)
9448                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
9449                                               BTRFS_BLOCK_RSV_DELALLOC);
9450         ei->runtime_flags = 0;
9451         ei->prop_compress = BTRFS_COMPRESS_NONE;
9452         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9453
9454         ei->delayed_node = NULL;
9455
9456         ei->i_otime.tv_sec = 0;
9457         ei->i_otime.tv_nsec = 0;
9458
9459         inode = &ei->vfs_inode;
9460         extent_map_tree_init(&ei->extent_tree);
9461         extent_io_tree_init(&ei->io_tree, inode);
9462         extent_io_tree_init(&ei->io_failure_tree, inode);
9463         ei->io_tree.track_uptodate = 1;
9464         ei->io_failure_tree.track_uptodate = 1;
9465         atomic_set(&ei->sync_writers, 0);
9466         mutex_init(&ei->log_mutex);
9467         mutex_init(&ei->delalloc_mutex);
9468         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9469         INIT_LIST_HEAD(&ei->delalloc_inodes);
9470         INIT_LIST_HEAD(&ei->delayed_iput);
9471         RB_CLEAR_NODE(&ei->rb_node);
9472         init_rwsem(&ei->dio_sem);
9473
9474         return inode;
9475 }
9476
9477 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9478 void btrfs_test_destroy_inode(struct inode *inode)
9479 {
9480         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9481         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9482 }
9483 #endif
9484
9485 static void btrfs_i_callback(struct rcu_head *head)
9486 {
9487         struct inode *inode = container_of(head, struct inode, i_rcu);
9488         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9489 }
9490
9491 void btrfs_destroy_inode(struct inode *inode)
9492 {
9493         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9494         struct btrfs_ordered_extent *ordered;
9495         struct btrfs_root *root = BTRFS_I(inode)->root;
9496
9497         WARN_ON(!hlist_empty(&inode->i_dentry));
9498         WARN_ON(inode->i_data.nrpages);
9499         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
9500         WARN_ON(BTRFS_I(inode)->block_rsv.size);
9501         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9502         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9503         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9504         WARN_ON(BTRFS_I(inode)->csum_bytes);
9505         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9506
9507         /*
9508          * This can happen where we create an inode, but somebody else also
9509          * created the same inode and we need to destroy the one we already
9510          * created.
9511          */
9512         if (!root)
9513                 goto free;
9514
9515         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9516                      &BTRFS_I(inode)->runtime_flags)) {
9517                 btrfs_info(fs_info, "inode %llu still on the orphan list",
9518                            btrfs_ino(BTRFS_I(inode)));
9519                 atomic_dec(&root->orphan_inodes);
9520         }
9521
9522         while (1) {
9523                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9524                 if (!ordered)
9525                         break;
9526                 else {
9527                         btrfs_err(fs_info,
9528                                   "found ordered extent %llu %llu on inode cleanup",
9529                                   ordered->file_offset, ordered->len);
9530                         btrfs_remove_ordered_extent(inode, ordered);
9531                         btrfs_put_ordered_extent(ordered);
9532                         btrfs_put_ordered_extent(ordered);
9533                 }
9534         }
9535         btrfs_qgroup_check_reserved_leak(inode);
9536         inode_tree_del(inode);
9537         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9538 free:
9539         call_rcu(&inode->i_rcu, btrfs_i_callback);
9540 }
9541
9542 int btrfs_drop_inode(struct inode *inode)
9543 {
9544         struct btrfs_root *root = BTRFS_I(inode)->root;
9545
9546         if (root == NULL)
9547                 return 1;
9548
9549         /* the snap/subvol tree is on deleting */
9550         if (btrfs_root_refs(&root->root_item) == 0)
9551                 return 1;
9552         else
9553                 return generic_drop_inode(inode);
9554 }
9555
9556 static void init_once(void *foo)
9557 {
9558         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9559
9560         inode_init_once(&ei->vfs_inode);
9561 }
9562
9563 void btrfs_destroy_cachep(void)
9564 {
9565         /*
9566          * Make sure all delayed rcu free inodes are flushed before we
9567          * destroy cache.
9568          */
9569         rcu_barrier();
9570         kmem_cache_destroy(btrfs_inode_cachep);
9571         kmem_cache_destroy(btrfs_trans_handle_cachep);
9572         kmem_cache_destroy(btrfs_path_cachep);
9573         kmem_cache_destroy(btrfs_free_space_cachep);
9574 }
9575
9576 int btrfs_init_cachep(void)
9577 {
9578         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9579                         sizeof(struct btrfs_inode), 0,
9580                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9581                         init_once);
9582         if (!btrfs_inode_cachep)
9583                 goto fail;
9584
9585         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9586                         sizeof(struct btrfs_trans_handle), 0,
9587                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9588         if (!btrfs_trans_handle_cachep)
9589                 goto fail;
9590
9591         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9592                         sizeof(struct btrfs_path), 0,
9593                         SLAB_MEM_SPREAD, NULL);
9594         if (!btrfs_path_cachep)
9595                 goto fail;
9596
9597         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9598                         sizeof(struct btrfs_free_space), 0,
9599                         SLAB_MEM_SPREAD, NULL);
9600         if (!btrfs_free_space_cachep)
9601                 goto fail;
9602
9603         return 0;
9604 fail:
9605         btrfs_destroy_cachep();
9606         return -ENOMEM;
9607 }
9608
9609 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9610                          u32 request_mask, unsigned int flags)
9611 {
9612         u64 delalloc_bytes;
9613         struct inode *inode = d_inode(path->dentry);
9614         u32 blocksize = inode->i_sb->s_blocksize;
9615         u32 bi_flags = BTRFS_I(inode)->flags;
9616
9617         stat->result_mask |= STATX_BTIME;
9618         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9619         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9620         if (bi_flags & BTRFS_INODE_APPEND)
9621                 stat->attributes |= STATX_ATTR_APPEND;
9622         if (bi_flags & BTRFS_INODE_COMPRESS)
9623                 stat->attributes |= STATX_ATTR_COMPRESSED;
9624         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9625                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9626         if (bi_flags & BTRFS_INODE_NODUMP)
9627                 stat->attributes |= STATX_ATTR_NODUMP;
9628
9629         stat->attributes_mask |= (STATX_ATTR_APPEND |
9630                                   STATX_ATTR_COMPRESSED |
9631                                   STATX_ATTR_IMMUTABLE |
9632                                   STATX_ATTR_NODUMP);
9633
9634         generic_fillattr(inode, stat);
9635         stat->dev = BTRFS_I(inode)->root->anon_dev;
9636
9637         spin_lock(&BTRFS_I(inode)->lock);
9638         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9639         spin_unlock(&BTRFS_I(inode)->lock);
9640         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9641                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9642         return 0;
9643 }
9644
9645 static int btrfs_rename_exchange(struct inode *old_dir,
9646                               struct dentry *old_dentry,
9647                               struct inode *new_dir,
9648                               struct dentry *new_dentry)
9649 {
9650         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9651         struct btrfs_trans_handle *trans;
9652         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9653         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9654         struct inode *new_inode = new_dentry->d_inode;
9655         struct inode *old_inode = old_dentry->d_inode;
9656         struct timespec ctime = current_time(old_inode);
9657         struct dentry *parent;
9658         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9659         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9660         u64 old_idx = 0;
9661         u64 new_idx = 0;
9662         u64 root_objectid;
9663         int ret;
9664         bool root_log_pinned = false;
9665         bool dest_log_pinned = false;
9666
9667         /* we only allow rename subvolume link between subvolumes */
9668         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9669                 return -EXDEV;
9670
9671         /* close the race window with snapshot create/destroy ioctl */
9672         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9673                 down_read(&fs_info->subvol_sem);
9674         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9675                 down_read(&fs_info->subvol_sem);
9676
9677         /*
9678          * We want to reserve the absolute worst case amount of items.  So if
9679          * both inodes are subvols and we need to unlink them then that would
9680          * require 4 item modifications, but if they are both normal inodes it
9681          * would require 5 item modifications, so we'll assume their normal
9682          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9683          * should cover the worst case number of items we'll modify.
9684          */
9685         trans = btrfs_start_transaction(root, 12);
9686         if (IS_ERR(trans)) {
9687                 ret = PTR_ERR(trans);
9688                 goto out_notrans;
9689         }
9690
9691         /*
9692          * We need to find a free sequence number both in the source and
9693          * in the destination directory for the exchange.
9694          */
9695         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9696         if (ret)
9697                 goto out_fail;
9698         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9699         if (ret)
9700                 goto out_fail;
9701
9702         BTRFS_I(old_inode)->dir_index = 0ULL;
9703         BTRFS_I(new_inode)->dir_index = 0ULL;
9704
9705         /* Reference for the source. */
9706         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9707                 /* force full log commit if subvolume involved. */
9708                 btrfs_set_log_full_commit(fs_info, trans);
9709         } else {
9710                 btrfs_pin_log_trans(root);
9711                 root_log_pinned = true;
9712                 ret = btrfs_insert_inode_ref(trans, dest,
9713                                              new_dentry->d_name.name,
9714                                              new_dentry->d_name.len,
9715                                              old_ino,
9716                                              btrfs_ino(BTRFS_I(new_dir)),
9717                                              old_idx);
9718                 if (ret)
9719                         goto out_fail;
9720         }
9721
9722         /* And now for the dest. */
9723         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9724                 /* force full log commit if subvolume involved. */
9725                 btrfs_set_log_full_commit(fs_info, trans);
9726         } else {
9727                 btrfs_pin_log_trans(dest);
9728                 dest_log_pinned = true;
9729                 ret = btrfs_insert_inode_ref(trans, root,
9730                                              old_dentry->d_name.name,
9731                                              old_dentry->d_name.len,
9732                                              new_ino,
9733                                              btrfs_ino(BTRFS_I(old_dir)),
9734                                              new_idx);
9735                 if (ret)
9736                         goto out_fail;
9737         }
9738
9739         /* Update inode version and ctime/mtime. */
9740         inode_inc_iversion(old_dir);
9741         inode_inc_iversion(new_dir);
9742         inode_inc_iversion(old_inode);
9743         inode_inc_iversion(new_inode);
9744         old_dir->i_ctime = old_dir->i_mtime = ctime;
9745         new_dir->i_ctime = new_dir->i_mtime = ctime;
9746         old_inode->i_ctime = ctime;
9747         new_inode->i_ctime = ctime;
9748
9749         if (old_dentry->d_parent != new_dentry->d_parent) {
9750                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9751                                 BTRFS_I(old_inode), 1);
9752                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9753                                 BTRFS_I(new_inode), 1);
9754         }
9755
9756         /* src is a subvolume */
9757         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9758                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9759                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9760                                           root_objectid,
9761                                           old_dentry->d_name.name,
9762                                           old_dentry->d_name.len);
9763         } else { /* src is an inode */
9764                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9765                                            BTRFS_I(old_dentry->d_inode),
9766                                            old_dentry->d_name.name,
9767                                            old_dentry->d_name.len);
9768                 if (!ret)
9769                         ret = btrfs_update_inode(trans, root, old_inode);
9770         }
9771         if (ret) {
9772                 btrfs_abort_transaction(trans, ret);
9773                 goto out_fail;
9774         }
9775
9776         /* dest is a subvolume */
9777         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9778                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9779                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9780                                           root_objectid,
9781                                           new_dentry->d_name.name,
9782                                           new_dentry->d_name.len);
9783         } else { /* dest is an inode */
9784                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9785                                            BTRFS_I(new_dentry->d_inode),
9786                                            new_dentry->d_name.name,
9787                                            new_dentry->d_name.len);
9788                 if (!ret)
9789                         ret = btrfs_update_inode(trans, dest, new_inode);
9790         }
9791         if (ret) {
9792                 btrfs_abort_transaction(trans, ret);
9793                 goto out_fail;
9794         }
9795
9796         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9797                              new_dentry->d_name.name,
9798                              new_dentry->d_name.len, 0, old_idx);
9799         if (ret) {
9800                 btrfs_abort_transaction(trans, ret);
9801                 goto out_fail;
9802         }
9803
9804         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9805                              old_dentry->d_name.name,
9806                              old_dentry->d_name.len, 0, new_idx);
9807         if (ret) {
9808                 btrfs_abort_transaction(trans, ret);
9809                 goto out_fail;
9810         }
9811
9812         if (old_inode->i_nlink == 1)
9813                 BTRFS_I(old_inode)->dir_index = old_idx;
9814         if (new_inode->i_nlink == 1)
9815                 BTRFS_I(new_inode)->dir_index = new_idx;
9816
9817         if (root_log_pinned) {
9818                 parent = new_dentry->d_parent;
9819                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9820                                 parent);
9821                 btrfs_end_log_trans(root);
9822                 root_log_pinned = false;
9823         }
9824         if (dest_log_pinned) {
9825                 parent = old_dentry->d_parent;
9826                 btrfs_log_new_name(trans, BTRFS_I(new_inode), BTRFS_I(new_dir),
9827                                 parent);
9828                 btrfs_end_log_trans(dest);
9829                 dest_log_pinned = false;
9830         }
9831 out_fail:
9832         /*
9833          * If we have pinned a log and an error happened, we unpin tasks
9834          * trying to sync the log and force them to fallback to a transaction
9835          * commit if the log currently contains any of the inodes involved in
9836          * this rename operation (to ensure we do not persist a log with an
9837          * inconsistent state for any of these inodes or leading to any
9838          * inconsistencies when replayed). If the transaction was aborted, the
9839          * abortion reason is propagated to userspace when attempting to commit
9840          * the transaction. If the log does not contain any of these inodes, we
9841          * allow the tasks to sync it.
9842          */
9843         if (ret && (root_log_pinned || dest_log_pinned)) {
9844                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9845                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9846                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9847                     (new_inode &&
9848                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9849                         btrfs_set_log_full_commit(fs_info, trans);
9850
9851                 if (root_log_pinned) {
9852                         btrfs_end_log_trans(root);
9853                         root_log_pinned = false;
9854                 }
9855                 if (dest_log_pinned) {
9856                         btrfs_end_log_trans(dest);
9857                         dest_log_pinned = false;
9858                 }
9859         }
9860         ret = btrfs_end_transaction(trans);
9861 out_notrans:
9862         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9863                 up_read(&fs_info->subvol_sem);
9864         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9865                 up_read(&fs_info->subvol_sem);
9866
9867         return ret;
9868 }
9869
9870 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9871                                      struct btrfs_root *root,
9872                                      struct inode *dir,
9873                                      struct dentry *dentry)
9874 {
9875         int ret;
9876         struct inode *inode;
9877         u64 objectid;
9878         u64 index;
9879
9880         ret = btrfs_find_free_ino(root, &objectid);
9881         if (ret)
9882                 return ret;
9883
9884         inode = btrfs_new_inode(trans, root, dir,
9885                                 dentry->d_name.name,
9886                                 dentry->d_name.len,
9887                                 btrfs_ino(BTRFS_I(dir)),
9888                                 objectid,
9889                                 S_IFCHR | WHITEOUT_MODE,
9890                                 &index);
9891
9892         if (IS_ERR(inode)) {
9893                 ret = PTR_ERR(inode);
9894                 return ret;
9895         }
9896
9897         inode->i_op = &btrfs_special_inode_operations;
9898         init_special_inode(inode, inode->i_mode,
9899                 WHITEOUT_DEV);
9900
9901         ret = btrfs_init_inode_security(trans, inode, dir,
9902                                 &dentry->d_name);
9903         if (ret)
9904                 goto out;
9905
9906         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9907                                 BTRFS_I(inode), 0, index);
9908         if (ret)
9909                 goto out;
9910
9911         ret = btrfs_update_inode(trans, root, inode);
9912 out:
9913         unlock_new_inode(inode);
9914         if (ret)
9915                 inode_dec_link_count(inode);
9916         iput(inode);
9917
9918         return ret;
9919 }
9920
9921 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9922                            struct inode *new_dir, struct dentry *new_dentry,
9923                            unsigned int flags)
9924 {
9925         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9926         struct btrfs_trans_handle *trans;
9927         unsigned int trans_num_items;
9928         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9929         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9930         struct inode *new_inode = d_inode(new_dentry);
9931         struct inode *old_inode = d_inode(old_dentry);
9932         u64 index = 0;
9933         u64 root_objectid;
9934         int ret;
9935         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9936         bool log_pinned = false;
9937
9938         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9939                 return -EPERM;
9940
9941         /* we only allow rename subvolume link between subvolumes */
9942         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9943                 return -EXDEV;
9944
9945         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9946             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9947                 return -ENOTEMPTY;
9948
9949         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9950             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9951                 return -ENOTEMPTY;
9952
9953
9954         /* check for collisions, even if the  name isn't there */
9955         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9956                              new_dentry->d_name.name,
9957                              new_dentry->d_name.len);
9958
9959         if (ret) {
9960                 if (ret == -EEXIST) {
9961                         /* we shouldn't get
9962                          * eexist without a new_inode */
9963                         if (WARN_ON(!new_inode)) {
9964                                 return ret;
9965                         }
9966                 } else {
9967                         /* maybe -EOVERFLOW */
9968                         return ret;
9969                 }
9970         }
9971         ret = 0;
9972
9973         /*
9974          * we're using rename to replace one file with another.  Start IO on it
9975          * now so  we don't add too much work to the end of the transaction
9976          */
9977         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9978                 filemap_flush(old_inode->i_mapping);
9979
9980         /* close the racy window with snapshot create/destroy ioctl */
9981         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9982                 down_read(&fs_info->subvol_sem);
9983         /*
9984          * We want to reserve the absolute worst case amount of items.  So if
9985          * both inodes are subvols and we need to unlink them then that would
9986          * require 4 item modifications, but if they are both normal inodes it
9987          * would require 5 item modifications, so we'll assume they are normal
9988          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9989          * should cover the worst case number of items we'll modify.
9990          * If our rename has the whiteout flag, we need more 5 units for the
9991          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9992          * when selinux is enabled).
9993          */
9994         trans_num_items = 11;
9995         if (flags & RENAME_WHITEOUT)
9996                 trans_num_items += 5;
9997         trans = btrfs_start_transaction(root, trans_num_items);
9998         if (IS_ERR(trans)) {
9999                 ret = PTR_ERR(trans);
10000                 goto out_notrans;
10001         }
10002
10003         if (dest != root)
10004                 btrfs_record_root_in_trans(trans, dest);
10005
10006         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
10007         if (ret)
10008                 goto out_fail;
10009
10010         BTRFS_I(old_inode)->dir_index = 0ULL;
10011         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
10012                 /* force full log commit if subvolume involved. */
10013                 btrfs_set_log_full_commit(fs_info, trans);
10014         } else {
10015                 btrfs_pin_log_trans(root);
10016                 log_pinned = true;
10017                 ret = btrfs_insert_inode_ref(trans, dest,
10018                                              new_dentry->d_name.name,
10019                                              new_dentry->d_name.len,
10020                                              old_ino,
10021                                              btrfs_ino(BTRFS_I(new_dir)), index);
10022                 if (ret)
10023                         goto out_fail;
10024         }
10025
10026         inode_inc_iversion(old_dir);
10027         inode_inc_iversion(new_dir);
10028         inode_inc_iversion(old_inode);
10029         old_dir->i_ctime = old_dir->i_mtime =
10030         new_dir->i_ctime = new_dir->i_mtime =
10031         old_inode->i_ctime = current_time(old_dir);
10032
10033         if (old_dentry->d_parent != new_dentry->d_parent)
10034                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
10035                                 BTRFS_I(old_inode), 1);
10036
10037         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
10038                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
10039                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
10040                                         old_dentry->d_name.name,
10041                                         old_dentry->d_name.len);
10042         } else {
10043                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
10044                                         BTRFS_I(d_inode(old_dentry)),
10045                                         old_dentry->d_name.name,
10046                                         old_dentry->d_name.len);
10047                 if (!ret)
10048                         ret = btrfs_update_inode(trans, root, old_inode);
10049         }
10050         if (ret) {
10051                 btrfs_abort_transaction(trans, ret);
10052                 goto out_fail;
10053         }
10054
10055         if (new_inode) {
10056                 inode_inc_iversion(new_inode);
10057                 new_inode->i_ctime = current_time(new_inode);
10058                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
10059                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
10060                         root_objectid = BTRFS_I(new_inode)->location.objectid;
10061                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
10062                                                 root_objectid,
10063                                                 new_dentry->d_name.name,
10064                                                 new_dentry->d_name.len);
10065                         BUG_ON(new_inode->i_nlink == 0);
10066                 } else {
10067                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
10068                                                  BTRFS_I(d_inode(new_dentry)),
10069                                                  new_dentry->d_name.name,
10070                                                  new_dentry->d_name.len);
10071                 }
10072                 if (!ret && new_inode->i_nlink == 0)
10073                         ret = btrfs_orphan_add(trans,
10074                                         BTRFS_I(d_inode(new_dentry)));
10075                 if (ret) {
10076                         btrfs_abort_transaction(trans, ret);
10077                         goto out_fail;
10078                 }
10079         }
10080
10081         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
10082                              new_dentry->d_name.name,
10083                              new_dentry->d_name.len, 0, index);
10084         if (ret) {
10085                 btrfs_abort_transaction(trans, ret);
10086                 goto out_fail;
10087         }
10088
10089         if (old_inode->i_nlink == 1)
10090                 BTRFS_I(old_inode)->dir_index = index;
10091
10092         if (log_pinned) {
10093                 struct dentry *parent = new_dentry->d_parent;
10094
10095                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
10096                                 parent);
10097                 btrfs_end_log_trans(root);
10098                 log_pinned = false;
10099         }
10100
10101         if (flags & RENAME_WHITEOUT) {
10102                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
10103                                                 old_dentry);
10104
10105                 if (ret) {
10106                         btrfs_abort_transaction(trans, ret);
10107                         goto out_fail;
10108                 }
10109         }
10110 out_fail:
10111         /*
10112          * If we have pinned the log and an error happened, we unpin tasks
10113          * trying to sync the log and force them to fallback to a transaction
10114          * commit if the log currently contains any of the inodes involved in
10115          * this rename operation (to ensure we do not persist a log with an
10116          * inconsistent state for any of these inodes or leading to any
10117          * inconsistencies when replayed). If the transaction was aborted, the
10118          * abortion reason is propagated to userspace when attempting to commit
10119          * the transaction. If the log does not contain any of these inodes, we
10120          * allow the tasks to sync it.
10121          */
10122         if (ret && log_pinned) {
10123                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
10124                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
10125                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
10126                     (new_inode &&
10127                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
10128                         btrfs_set_log_full_commit(fs_info, trans);
10129
10130                 btrfs_end_log_trans(root);
10131                 log_pinned = false;
10132         }
10133         btrfs_end_transaction(trans);
10134 out_notrans:
10135         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
10136                 up_read(&fs_info->subvol_sem);
10137
10138         return ret;
10139 }
10140
10141 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
10142                          struct inode *new_dir, struct dentry *new_dentry,
10143                          unsigned int flags)
10144 {
10145         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
10146                 return -EINVAL;
10147
10148         if (flags & RENAME_EXCHANGE)
10149                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
10150                                           new_dentry);
10151
10152         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
10153 }
10154
10155 static void btrfs_run_delalloc_work(struct btrfs_work *work)
10156 {
10157         struct btrfs_delalloc_work *delalloc_work;
10158         struct inode *inode;
10159
10160         delalloc_work = container_of(work, struct btrfs_delalloc_work,
10161                                      work);
10162         inode = delalloc_work->inode;
10163         filemap_flush(inode->i_mapping);
10164         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
10165                                 &BTRFS_I(inode)->runtime_flags))
10166                 filemap_flush(inode->i_mapping);
10167
10168         if (delalloc_work->delay_iput)
10169                 btrfs_add_delayed_iput(inode);
10170         else
10171                 iput(inode);
10172         complete(&delalloc_work->completion);
10173 }
10174
10175 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
10176                                                     int delay_iput)
10177 {
10178         struct btrfs_delalloc_work *work;
10179
10180         work = kmalloc(sizeof(*work), GFP_NOFS);
10181         if (!work)
10182                 return NULL;
10183
10184         init_completion(&work->completion);
10185         INIT_LIST_HEAD(&work->list);
10186         work->inode = inode;
10187         work->delay_iput = delay_iput;
10188         WARN_ON_ONCE(!inode);
10189         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
10190                         btrfs_run_delalloc_work, NULL, NULL);
10191
10192         return work;
10193 }
10194
10195 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
10196 {
10197         wait_for_completion(&work->completion);
10198         kfree(work);
10199 }
10200
10201 /*
10202  * some fairly slow code that needs optimization. This walks the list
10203  * of all the inodes with pending delalloc and forces them to disk.
10204  */
10205 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
10206                                    int nr)
10207 {
10208         struct btrfs_inode *binode;
10209         struct inode *inode;
10210         struct btrfs_delalloc_work *work, *next;
10211         struct list_head works;
10212         struct list_head splice;
10213         int ret = 0;
10214
10215         INIT_LIST_HEAD(&works);
10216         INIT_LIST_HEAD(&splice);
10217
10218         mutex_lock(&root->delalloc_mutex);
10219         spin_lock(&root->delalloc_lock);
10220         list_splice_init(&root->delalloc_inodes, &splice);
10221         while (!list_empty(&splice)) {
10222                 binode = list_entry(splice.next, struct btrfs_inode,
10223                                     delalloc_inodes);
10224
10225                 list_move_tail(&binode->delalloc_inodes,
10226                                &root->delalloc_inodes);
10227                 inode = igrab(&binode->vfs_inode);
10228                 if (!inode) {
10229                         cond_resched_lock(&root->delalloc_lock);
10230                         continue;
10231                 }
10232                 spin_unlock(&root->delalloc_lock);
10233
10234                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10235                 if (!work) {
10236                         if (delay_iput)
10237                                 btrfs_add_delayed_iput(inode);
10238                         else
10239                                 iput(inode);
10240                         ret = -ENOMEM;
10241                         goto out;
10242                 }
10243                 list_add_tail(&work->list, &works);
10244                 btrfs_queue_work(root->fs_info->flush_workers,
10245                                  &work->work);
10246                 ret++;
10247                 if (nr != -1 && ret >= nr)
10248                         goto out;
10249                 cond_resched();
10250                 spin_lock(&root->delalloc_lock);
10251         }
10252         spin_unlock(&root->delalloc_lock);
10253
10254 out:
10255         list_for_each_entry_safe(work, next, &works, list) {
10256                 list_del_init(&work->list);
10257                 btrfs_wait_and_free_delalloc_work(work);
10258         }
10259
10260         if (!list_empty_careful(&splice)) {
10261                 spin_lock(&root->delalloc_lock);
10262                 list_splice_tail(&splice, &root->delalloc_inodes);
10263                 spin_unlock(&root->delalloc_lock);
10264         }
10265         mutex_unlock(&root->delalloc_mutex);
10266         return ret;
10267 }
10268
10269 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10270 {
10271         struct btrfs_fs_info *fs_info = root->fs_info;
10272         int ret;
10273
10274         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10275                 return -EROFS;
10276
10277         ret = __start_delalloc_inodes(root, delay_iput, -1);
10278         if (ret > 0)
10279                 ret = 0;
10280         return ret;
10281 }
10282
10283 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10284                                int nr)
10285 {
10286         struct btrfs_root *root;
10287         struct list_head splice;
10288         int ret;
10289
10290         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10291                 return -EROFS;
10292
10293         INIT_LIST_HEAD(&splice);
10294
10295         mutex_lock(&fs_info->delalloc_root_mutex);
10296         spin_lock(&fs_info->delalloc_root_lock);
10297         list_splice_init(&fs_info->delalloc_roots, &splice);
10298         while (!list_empty(&splice) && nr) {
10299                 root = list_first_entry(&splice, struct btrfs_root,
10300                                         delalloc_root);
10301                 root = btrfs_grab_fs_root(root);
10302                 BUG_ON(!root);
10303                 list_move_tail(&root->delalloc_root,
10304                                &fs_info->delalloc_roots);
10305                 spin_unlock(&fs_info->delalloc_root_lock);
10306
10307                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10308                 btrfs_put_fs_root(root);
10309                 if (ret < 0)
10310                         goto out;
10311
10312                 if (nr != -1) {
10313                         nr -= ret;
10314                         WARN_ON(nr < 0);
10315                 }
10316                 spin_lock(&fs_info->delalloc_root_lock);
10317         }
10318         spin_unlock(&fs_info->delalloc_root_lock);
10319
10320         ret = 0;
10321 out:
10322         if (!list_empty_careful(&splice)) {
10323                 spin_lock(&fs_info->delalloc_root_lock);
10324                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10325                 spin_unlock(&fs_info->delalloc_root_lock);
10326         }
10327         mutex_unlock(&fs_info->delalloc_root_mutex);
10328         return ret;
10329 }
10330
10331 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10332                          const char *symname)
10333 {
10334         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10335         struct btrfs_trans_handle *trans;
10336         struct btrfs_root *root = BTRFS_I(dir)->root;
10337         struct btrfs_path *path;
10338         struct btrfs_key key;
10339         struct inode *inode = NULL;
10340         int err;
10341         int drop_inode = 0;
10342         u64 objectid;
10343         u64 index = 0;
10344         int name_len;
10345         int datasize;
10346         unsigned long ptr;
10347         struct btrfs_file_extent_item *ei;
10348         struct extent_buffer *leaf;
10349
10350         name_len = strlen(symname);
10351         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10352                 return -ENAMETOOLONG;
10353
10354         /*
10355          * 2 items for inode item and ref
10356          * 2 items for dir items
10357          * 1 item for updating parent inode item
10358          * 1 item for the inline extent item
10359          * 1 item for xattr if selinux is on
10360          */
10361         trans = btrfs_start_transaction(root, 7);
10362         if (IS_ERR(trans))
10363                 return PTR_ERR(trans);
10364
10365         err = btrfs_find_free_ino(root, &objectid);
10366         if (err)
10367                 goto out_unlock;
10368
10369         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10370                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10371                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10372         if (IS_ERR(inode)) {
10373                 err = PTR_ERR(inode);
10374                 goto out_unlock;
10375         }
10376
10377         /*
10378         * If the active LSM wants to access the inode during
10379         * d_instantiate it needs these. Smack checks to see
10380         * if the filesystem supports xattrs by looking at the
10381         * ops vector.
10382         */
10383         inode->i_fop = &btrfs_file_operations;
10384         inode->i_op = &btrfs_file_inode_operations;
10385         inode->i_mapping->a_ops = &btrfs_aops;
10386         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10387
10388         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10389         if (err)
10390                 goto out_unlock_inode;
10391
10392         path = btrfs_alloc_path();
10393         if (!path) {
10394                 err = -ENOMEM;
10395                 goto out_unlock_inode;
10396         }
10397         key.objectid = btrfs_ino(BTRFS_I(inode));
10398         key.offset = 0;
10399         key.type = BTRFS_EXTENT_DATA_KEY;
10400         datasize = btrfs_file_extent_calc_inline_size(name_len);
10401         err = btrfs_insert_empty_item(trans, root, path, &key,
10402                                       datasize);
10403         if (err) {
10404                 btrfs_free_path(path);
10405                 goto out_unlock_inode;
10406         }
10407         leaf = path->nodes[0];
10408         ei = btrfs_item_ptr(leaf, path->slots[0],
10409                             struct btrfs_file_extent_item);
10410         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10411         btrfs_set_file_extent_type(leaf, ei,
10412                                    BTRFS_FILE_EXTENT_INLINE);
10413         btrfs_set_file_extent_encryption(leaf, ei, 0);
10414         btrfs_set_file_extent_compression(leaf, ei, 0);
10415         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10416         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10417
10418         ptr = btrfs_file_extent_inline_start(ei);
10419         write_extent_buffer(leaf, symname, ptr, name_len);
10420         btrfs_mark_buffer_dirty(leaf);
10421         btrfs_free_path(path);
10422
10423         inode->i_op = &btrfs_symlink_inode_operations;
10424         inode_nohighmem(inode);
10425         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10426         inode_set_bytes(inode, name_len);
10427         btrfs_i_size_write(BTRFS_I(inode), name_len);
10428         err = btrfs_update_inode(trans, root, inode);
10429         /*
10430          * Last step, add directory indexes for our symlink inode. This is the
10431          * last step to avoid extra cleanup of these indexes if an error happens
10432          * elsewhere above.
10433          */
10434         if (!err)
10435                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10436                                 BTRFS_I(inode), 0, index);
10437         if (err) {
10438                 drop_inode = 1;
10439                 goto out_unlock_inode;
10440         }
10441
10442         unlock_new_inode(inode);
10443         d_instantiate(dentry, inode);
10444
10445 out_unlock:
10446         btrfs_end_transaction(trans);
10447         if (drop_inode) {
10448                 inode_dec_link_count(inode);
10449                 iput(inode);
10450         }
10451         btrfs_btree_balance_dirty(fs_info);
10452         return err;
10453
10454 out_unlock_inode:
10455         drop_inode = 1;
10456         unlock_new_inode(inode);
10457         goto out_unlock;
10458 }
10459
10460 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10461                                        u64 start, u64 num_bytes, u64 min_size,
10462                                        loff_t actual_len, u64 *alloc_hint,
10463                                        struct btrfs_trans_handle *trans)
10464 {
10465         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10466         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10467         struct extent_map *em;
10468         struct btrfs_root *root = BTRFS_I(inode)->root;
10469         struct btrfs_key ins;
10470         u64 cur_offset = start;
10471         u64 i_size;
10472         u64 cur_bytes;
10473         u64 last_alloc = (u64)-1;
10474         int ret = 0;
10475         bool own_trans = true;
10476         u64 end = start + num_bytes - 1;
10477
10478         if (trans)
10479                 own_trans = false;
10480         while (num_bytes > 0) {
10481                 if (own_trans) {
10482                         trans = btrfs_start_transaction(root, 3);
10483                         if (IS_ERR(trans)) {
10484                                 ret = PTR_ERR(trans);
10485                                 break;
10486                         }
10487                 }
10488
10489                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10490                 cur_bytes = max(cur_bytes, min_size);
10491                 /*
10492                  * If we are severely fragmented we could end up with really
10493                  * small allocations, so if the allocator is returning small
10494                  * chunks lets make its job easier by only searching for those
10495                  * sized chunks.
10496                  */
10497                 cur_bytes = min(cur_bytes, last_alloc);
10498                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10499                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10500                 if (ret) {
10501                         if (own_trans)
10502                                 btrfs_end_transaction(trans);
10503                         break;
10504                 }
10505                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10506
10507                 last_alloc = ins.offset;
10508                 ret = insert_reserved_file_extent(trans, inode,
10509                                                   cur_offset, ins.objectid,
10510                                                   ins.offset, ins.offset,
10511                                                   ins.offset, 0, 0, 0,
10512                                                   BTRFS_FILE_EXTENT_PREALLOC);
10513                 if (ret) {
10514                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10515                                                    ins.offset, 0);
10516                         btrfs_abort_transaction(trans, ret);
10517                         if (own_trans)
10518                                 btrfs_end_transaction(trans);
10519                         break;
10520                 }
10521
10522                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10523                                         cur_offset + ins.offset -1, 0);
10524
10525                 em = alloc_extent_map();
10526                 if (!em) {
10527                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10528                                 &BTRFS_I(inode)->runtime_flags);
10529                         goto next;
10530                 }
10531
10532                 em->start = cur_offset;
10533                 em->orig_start = cur_offset;
10534                 em->len = ins.offset;
10535                 em->block_start = ins.objectid;
10536                 em->block_len = ins.offset;
10537                 em->orig_block_len = ins.offset;
10538                 em->ram_bytes = ins.offset;
10539                 em->bdev = fs_info->fs_devices->latest_bdev;
10540                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10541                 em->generation = trans->transid;
10542
10543                 while (1) {
10544                         write_lock(&em_tree->lock);
10545                         ret = add_extent_mapping(em_tree, em, 1);
10546                         write_unlock(&em_tree->lock);
10547                         if (ret != -EEXIST)
10548                                 break;
10549                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10550                                                 cur_offset + ins.offset - 1,
10551                                                 0);
10552                 }
10553                 free_extent_map(em);
10554 next:
10555                 num_bytes -= ins.offset;
10556                 cur_offset += ins.offset;
10557                 *alloc_hint = ins.objectid + ins.offset;
10558
10559                 inode_inc_iversion(inode);
10560                 inode->i_ctime = current_time(inode);
10561                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10562                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10563                     (actual_len > inode->i_size) &&
10564                     (cur_offset > inode->i_size)) {
10565                         if (cur_offset > actual_len)
10566                                 i_size = actual_len;
10567                         else
10568                                 i_size = cur_offset;
10569                         i_size_write(inode, i_size);
10570                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10571                 }
10572
10573                 ret = btrfs_update_inode(trans, root, inode);
10574
10575                 if (ret) {
10576                         btrfs_abort_transaction(trans, ret);
10577                         if (own_trans)
10578                                 btrfs_end_transaction(trans);
10579                         break;
10580                 }
10581
10582                 if (own_trans)
10583                         btrfs_end_transaction(trans);
10584         }
10585         if (cur_offset < end)
10586                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10587                         end - cur_offset + 1);
10588         return ret;
10589 }
10590
10591 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10592                               u64 start, u64 num_bytes, u64 min_size,
10593                               loff_t actual_len, u64 *alloc_hint)
10594 {
10595         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10596                                            min_size, actual_len, alloc_hint,
10597                                            NULL);
10598 }
10599
10600 int btrfs_prealloc_file_range_trans(struct inode *inode,
10601                                     struct btrfs_trans_handle *trans, int mode,
10602                                     u64 start, u64 num_bytes, u64 min_size,
10603                                     loff_t actual_len, u64 *alloc_hint)
10604 {
10605         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10606                                            min_size, actual_len, alloc_hint, trans);
10607 }
10608
10609 static int btrfs_set_page_dirty(struct page *page)
10610 {
10611         return __set_page_dirty_nobuffers(page);
10612 }
10613
10614 static int btrfs_permission(struct inode *inode, int mask)
10615 {
10616         struct btrfs_root *root = BTRFS_I(inode)->root;
10617         umode_t mode = inode->i_mode;
10618
10619         if (mask & MAY_WRITE &&
10620             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10621                 if (btrfs_root_readonly(root))
10622                         return -EROFS;
10623                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10624                         return -EACCES;
10625         }
10626         return generic_permission(inode, mask);
10627 }
10628
10629 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10630 {
10631         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10632         struct btrfs_trans_handle *trans;
10633         struct btrfs_root *root = BTRFS_I(dir)->root;
10634         struct inode *inode = NULL;
10635         u64 objectid;
10636         u64 index;
10637         int ret = 0;
10638
10639         /*
10640          * 5 units required for adding orphan entry
10641          */
10642         trans = btrfs_start_transaction(root, 5);
10643         if (IS_ERR(trans))
10644                 return PTR_ERR(trans);
10645
10646         ret = btrfs_find_free_ino(root, &objectid);
10647         if (ret)
10648                 goto out;
10649
10650         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10651                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10652         if (IS_ERR(inode)) {
10653                 ret = PTR_ERR(inode);
10654                 inode = NULL;
10655                 goto out;
10656         }
10657
10658         inode->i_fop = &btrfs_file_operations;
10659         inode->i_op = &btrfs_file_inode_operations;
10660
10661         inode->i_mapping->a_ops = &btrfs_aops;
10662         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10663
10664         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10665         if (ret)
10666                 goto out_inode;
10667
10668         ret = btrfs_update_inode(trans, root, inode);
10669         if (ret)
10670                 goto out_inode;
10671         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10672         if (ret)
10673                 goto out_inode;
10674
10675         /*
10676          * We set number of links to 0 in btrfs_new_inode(), and here we set
10677          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10678          * through:
10679          *
10680          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10681          */
10682         set_nlink(inode, 1);
10683         unlock_new_inode(inode);
10684         d_tmpfile(dentry, inode);
10685         mark_inode_dirty(inode);
10686
10687 out:
10688         btrfs_end_transaction(trans);
10689         if (ret)
10690                 iput(inode);
10691         btrfs_balance_delayed_items(fs_info);
10692         btrfs_btree_balance_dirty(fs_info);
10693         return ret;
10694
10695 out_inode:
10696         unlock_new_inode(inode);
10697         goto out;
10698
10699 }
10700
10701 __attribute__((const))
10702 static int btrfs_readpage_io_failed_hook(struct page *page, int failed_mirror)
10703 {
10704         return -EAGAIN;
10705 }
10706
10707 static struct btrfs_fs_info *iotree_fs_info(void *private_data)
10708 {
10709         struct inode *inode = private_data;
10710         return btrfs_sb(inode->i_sb);
10711 }
10712
10713 static void btrfs_check_extent_io_range(void *private_data, const char *caller,
10714                                         u64 start, u64 end)
10715 {
10716         struct inode *inode = private_data;
10717         u64 isize;
10718
10719         isize = i_size_read(inode);
10720         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
10721                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
10722                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
10723                         caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
10724         }
10725 }
10726
10727 void btrfs_set_range_writeback(void *private_data, u64 start, u64 end)
10728 {
10729         struct inode *inode = private_data;
10730         unsigned long index = start >> PAGE_SHIFT;
10731         unsigned long end_index = end >> PAGE_SHIFT;
10732         struct page *page;
10733
10734         while (index <= end_index) {
10735                 page = find_get_page(inode->i_mapping, index);
10736                 ASSERT(page); /* Pages should be in the extent_io_tree */
10737                 set_page_writeback(page);
10738                 put_page(page);
10739                 index++;
10740         }
10741 }
10742
10743 static const struct inode_operations btrfs_dir_inode_operations = {
10744         .getattr        = btrfs_getattr,
10745         .lookup         = btrfs_lookup,
10746         .create         = btrfs_create,
10747         .unlink         = btrfs_unlink,
10748         .link           = btrfs_link,
10749         .mkdir          = btrfs_mkdir,
10750         .rmdir          = btrfs_rmdir,
10751         .rename         = btrfs_rename2,
10752         .symlink        = btrfs_symlink,
10753         .setattr        = btrfs_setattr,
10754         .mknod          = btrfs_mknod,
10755         .listxattr      = btrfs_listxattr,
10756         .permission     = btrfs_permission,
10757         .get_acl        = btrfs_get_acl,
10758         .set_acl        = btrfs_set_acl,
10759         .update_time    = btrfs_update_time,
10760         .tmpfile        = btrfs_tmpfile,
10761 };
10762 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10763         .lookup         = btrfs_lookup,
10764         .permission     = btrfs_permission,
10765         .update_time    = btrfs_update_time,
10766 };
10767
10768 static const struct file_operations btrfs_dir_file_operations = {
10769         .llseek         = generic_file_llseek,
10770         .read           = generic_read_dir,
10771         .iterate_shared = btrfs_real_readdir,
10772         .open           = btrfs_opendir,
10773         .unlocked_ioctl = btrfs_ioctl,
10774 #ifdef CONFIG_COMPAT
10775         .compat_ioctl   = btrfs_compat_ioctl,
10776 #endif
10777         .release        = btrfs_release_file,
10778         .fsync          = btrfs_sync_file,
10779 };
10780
10781 static const struct extent_io_ops btrfs_extent_io_ops = {
10782         /* mandatory callbacks */
10783         .submit_bio_hook = btrfs_submit_bio_hook,
10784         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10785         .merge_bio_hook = btrfs_merge_bio_hook,
10786         .readpage_io_failed_hook = btrfs_readpage_io_failed_hook,
10787         .tree_fs_info = iotree_fs_info,
10788         .set_range_writeback = btrfs_set_range_writeback,
10789
10790         /* optional callbacks */
10791         .fill_delalloc = run_delalloc_range,
10792         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10793         .writepage_start_hook = btrfs_writepage_start_hook,
10794         .set_bit_hook = btrfs_set_bit_hook,
10795         .clear_bit_hook = btrfs_clear_bit_hook,
10796         .merge_extent_hook = btrfs_merge_extent_hook,
10797         .split_extent_hook = btrfs_split_extent_hook,
10798         .check_extent_io_range = btrfs_check_extent_io_range,
10799 };
10800
10801 /*
10802  * btrfs doesn't support the bmap operation because swapfiles
10803  * use bmap to make a mapping of extents in the file.  They assume
10804  * these extents won't change over the life of the file and they
10805  * use the bmap result to do IO directly to the drive.
10806  *
10807  * the btrfs bmap call would return logical addresses that aren't
10808  * suitable for IO and they also will change frequently as COW
10809  * operations happen.  So, swapfile + btrfs == corruption.
10810  *
10811  * For now we're avoiding this by dropping bmap.
10812  */
10813 static const struct address_space_operations btrfs_aops = {
10814         .readpage       = btrfs_readpage,
10815         .writepage      = btrfs_writepage,
10816         .writepages     = btrfs_writepages,
10817         .readpages      = btrfs_readpages,
10818         .direct_IO      = btrfs_direct_IO,
10819         .invalidatepage = btrfs_invalidatepage,
10820         .releasepage    = btrfs_releasepage,
10821         .set_page_dirty = btrfs_set_page_dirty,
10822         .error_remove_page = generic_error_remove_page,
10823 };
10824
10825 static const struct address_space_operations btrfs_symlink_aops = {
10826         .readpage       = btrfs_readpage,
10827         .writepage      = btrfs_writepage,
10828         .invalidatepage = btrfs_invalidatepage,
10829         .releasepage    = btrfs_releasepage,
10830 };
10831
10832 static const struct inode_operations btrfs_file_inode_operations = {
10833         .getattr        = btrfs_getattr,
10834         .setattr        = btrfs_setattr,
10835         .listxattr      = btrfs_listxattr,
10836         .permission     = btrfs_permission,
10837         .fiemap         = btrfs_fiemap,
10838         .get_acl        = btrfs_get_acl,
10839         .set_acl        = btrfs_set_acl,
10840         .update_time    = btrfs_update_time,
10841 };
10842 static const struct inode_operations btrfs_special_inode_operations = {
10843         .getattr        = btrfs_getattr,
10844         .setattr        = btrfs_setattr,
10845         .permission     = btrfs_permission,
10846         .listxattr      = btrfs_listxattr,
10847         .get_acl        = btrfs_get_acl,
10848         .set_acl        = btrfs_set_acl,
10849         .update_time    = btrfs_update_time,
10850 };
10851 static const struct inode_operations btrfs_symlink_inode_operations = {
10852         .get_link       = page_get_link,
10853         .getattr        = btrfs_getattr,
10854         .setattr        = btrfs_setattr,
10855         .permission     = btrfs_permission,
10856         .listxattr      = btrfs_listxattr,
10857         .update_time    = btrfs_update_time,
10858 };
10859
10860 const struct dentry_operations btrfs_dentry_operations = {
10861         .d_delete       = btrfs_dentry_delete,
10862         .d_release      = btrfs_dentry_release,
10863 };