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/statfs.h>
  34 #include <linux/compat.h>
  35 #include <linux/aio.h>
  36 #include <linux/bit_spinlock.h>
  37 #include <linux/xattr.h>
  38 #include <linux/posix_acl.h>
  39 #include <linux/falloc.h>
  40 #include <linux/slab.h>
  41 #include <linux/ratelimit.h>
  42 #include <linux/mount.h>
  43 #include <linux/btrfs.h>
  44 #include <linux/blkdev.h>
  45 #include <linux/posix_acl_xattr.h>
  46 #include "compat.h"
  47 #include "ctree.h"
  48 #include "disk-io.h"
  49 #include "transaction.h"
  50 #include "btrfs_inode.h"
  51 #include "print-tree.h"
  52 #include "ordered-data.h"
  53 #include "xattr.h"
  54 #include "tree-log.h"
  55 #include "volumes.h"
  56 #include "compression.h"
  57 #include "locking.h"
  58 #include "free-space-cache.h"
  59 #include "inode-map.h"
  60 #include "backref.h"
  61 #include "hash.h"
  62
  63 struct btrfs_iget_args {
  64         u64 ino;
  65         struct btrfs_root *root;
  66 };
  67
  68 static const struct inode_operations btrfs_dir_inode_operations;
  69 static const struct inode_operations btrfs_symlink_inode_operations;
  70 static const struct inode_operations btrfs_dir_ro_inode_operations;
  71 static const struct inode_operations btrfs_special_inode_operations;
  72 static const struct inode_operations btrfs_file_inode_operations;
  73 static const struct address_space_operations btrfs_aops;
  74 static const struct address_space_operations btrfs_symlink_aops;
  75 static const struct file_operations btrfs_dir_file_operations;
  76 static struct extent_io_ops btrfs_extent_io_ops;
  77
  78 static struct kmem_cache *btrfs_inode_cachep;
  79 static struct kmem_cache *btrfs_delalloc_work_cachep;
  80 struct kmem_cache *btrfs_trans_handle_cachep;
  81 struct kmem_cache *btrfs_transaction_cachep;
  82 struct kmem_cache *btrfs_path_cachep;
  83 struct kmem_cache *btrfs_free_space_cachep;
  84
  85 #define S_SHIFT 12
  86 static unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  87         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  88         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  89         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  90         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  91         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  92         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
  93         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
  94 };
  95
  96 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
  97 static int btrfs_truncate(struct inode *inode);
  98 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
  99 static noinline int cow_file_range(struct inode *inode,
 100                                    struct page *locked_page,
 101                                    u64 start, u64 end, int *page_started,
 102                                    unsigned long *nr_written, int unlock);
 103 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
 104                                            u64 len, u64 orig_start,
 105                                            u64 block_start, u64 block_len,
 106                                            u64 orig_block_len, u64 ram_bytes,
 107                                            int type);
 108
 109 static int btrfs_dirty_inode(struct inode *inode);
 110
 111 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 112                                      struct inode *inode,  struct inode *dir,
 113                                      const struct qstr *qstr)
 114 {
 115         int err;
 116
 117         err = btrfs_init_acl(trans, inode, dir);
 118         if (!err)
 119                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 120         return err;
 121 }
 122
 123 /*
 124  * this does all the hard work for inserting an inline extent into
 125  * the btree.  The caller should have done a btrfs_drop_extents so that
 126  * no overlapping inline items exist in the btree
 127  */
 128 static noinline int insert_inline_extent(struct btrfs_trans_handle *trans,
 129                                 struct btrfs_root *root, struct inode *inode,
 130                                 u64 start, size_t size, size_t compressed_size,
 131                                 int compress_type,
 132                                 struct page **compressed_pages)
 133 {
 134         struct btrfs_key key;
 135         struct btrfs_path *path;
 136         struct extent_buffer *leaf;
 137         struct page *page = NULL;
 138         char *kaddr;
 139         unsigned long ptr;
 140         struct btrfs_file_extent_item *ei;
 141         int err = 0;
 142         int ret;
 143         size_t cur_size = size;
 144         size_t datasize;
 145         unsigned long offset;
 146
 147         if (compressed_size && compressed_pages)
 148                 cur_size = compressed_size;
 149
 150         path = btrfs_alloc_path();
 151         if (!path)
 152                 return -ENOMEM;
 153
 154         path->leave_spinning = 1;
 155
 156         key.objectid = btrfs_ino(inode);
 157         key.offset = start;
 158         btrfs_set_key_type(&key, BTRFS_EXTENT_DATA_KEY);
 159         datasize = btrfs_file_extent_calc_inline_size(cur_size);
 160
 161         inode_add_bytes(inode, size);
 162         ret = btrfs_insert_empty_item(trans, root, path, &key,
 163                                       datasize);
 164         if (ret) {
 165                 err = ret;
 166                 goto fail;
 167         }
 168         leaf = path->nodes[0];
 169         ei = btrfs_item_ptr(leaf, path->slots[0],
 170                             struct btrfs_file_extent_item);
 171         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 172         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 173         btrfs_set_file_extent_encryption(leaf, ei, 0);
 174         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 175         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 176         ptr = btrfs_file_extent_inline_start(ei);
 177
 178         if (compress_type != BTRFS_COMPRESS_NONE) {
 179                 struct page *cpage;
 180                 int i = 0;
 181                 while (compressed_size > 0) {
 182                         cpage = compressed_pages[i];
 183                         cur_size = min_t(unsigned long, compressed_size,
 184                                        PAGE_CACHE_SIZE);
 185
 186                         kaddr = kmap_atomic(cpage);
 187                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 188                         kunmap_atomic(kaddr);
 189
 190                         i++;
 191                         ptr += cur_size;
 192                         compressed_size -= cur_size;
 193                 }
 194                 btrfs_set_file_extent_compression(leaf, ei,
 195                                                   compress_type);
 196         } else {
 197                 page = find_get_page(inode->i_mapping,
 198                                      start >> PAGE_CACHE_SHIFT);
 199                 btrfs_set_file_extent_compression(leaf, ei, 0);
 200                 kaddr = kmap_atomic(page);
 201                 offset = start & (PAGE_CACHE_SIZE - 1);
 202                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 203                 kunmap_atomic(kaddr);
 204                 page_cache_release(page);
 205         }
 206         btrfs_mark_buffer_dirty(leaf);
 207         btrfs_free_path(path);
 208
 209         /*
 210          * we're an inline extent, so nobody can
 211          * extend the file past i_size without locking
 212          * a page we already have locked.
 213          *
 214          * We must do any isize and inode updates
 215          * before we unlock the pages.  Otherwise we
 216          * could end up racing with unlink.
 217          */
 218         BTRFS_I(inode)->disk_i_size = inode->i_size;
 219         ret = btrfs_update_inode(trans, root, inode);
 220
 221         return ret;
 222 fail:
 223         btrfs_free_path(path);
 224         return err;
 225 }
 226
 227
 228 /*
 229  * conditionally insert an inline extent into the file.  This
 230  * does the checks required to make sure the data is small enough
 231  * to fit as an inline extent.
 232  */
 233 static noinline int cow_file_range_inline(struct btrfs_trans_handle *trans,
 234                                  struct btrfs_root *root,
 235                                  struct inode *inode, u64 start, u64 end,
 236                                  size_t compressed_size, int compress_type,
 237                                  struct page **compressed_pages)
 238 {
 239         u64 isize = i_size_read(inode);
 240         u64 actual_end = min(end + 1, isize);
 241         u64 inline_len = actual_end - start;
 242         u64 aligned_end = ALIGN(end, root->sectorsize);
 243         u64 data_len = inline_len;
 244         int ret;
 245
 246         if (compressed_size)
 247                 data_len = compressed_size;
 248
 249         if (start > 0 ||
 250             actual_end >= PAGE_CACHE_SIZE ||
 251             data_len >= BTRFS_MAX_INLINE_DATA_SIZE(root) ||
 252             (!compressed_size &&
 253             (actual_end & (root->sectorsize - 1)) == 0) ||
 254             end + 1 < isize ||
 255             data_len > root->fs_info->max_inline) {
 256                 return 1;
 257         }
 258
 259         ret = btrfs_drop_extents(trans, root, inode, start, aligned_end, 1);
 260         if (ret)
 261                 return ret;
 262
 263         if (isize > actual_end)
 264                 inline_len = min_t(u64, isize, actual_end);
 265         ret = insert_inline_extent(trans, root, inode, start,
 266                                    inline_len, compressed_size,
 267                                    compress_type, compressed_pages);
 268         if (ret && ret != -ENOSPC) {
 269                 btrfs_abort_transaction(trans, root, ret);
 270                 return ret;
 271         } else if (ret == -ENOSPC) {
 272                 return 1;
 273         }
 274
 275         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 276         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 277         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 278         return 0;
 279 }
 280
 281 struct async_extent {
 282         u64 start;
 283         u64 ram_size;
 284         u64 compressed_size;
 285         struct page **pages;
 286         unsigned long nr_pages;
 287         int compress_type;
 288         struct list_head list;
 289 };
 290
 291 struct async_cow {
 292         struct inode *inode;
 293         struct btrfs_root *root;
 294         struct page *locked_page;
 295         u64 start;
 296         u64 end;
 297         struct list_head extents;
 298         struct btrfs_work work;
 299 };
 300
 301 static noinline int add_async_extent(struct async_cow *cow,
 302                                      u64 start, u64 ram_size,
 303                                      u64 compressed_size,
 304                                      struct page **pages,
 305                                      unsigned long nr_pages,
 306                                      int compress_type)
 307 {
 308         struct async_extent *async_extent;
 309
 310         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 311         BUG_ON(!async_extent); /* -ENOMEM */
 312         async_extent->start = start;
 313         async_extent->ram_size = ram_size;
 314         async_extent->compressed_size = compressed_size;
 315         async_extent->pages = pages;
 316         async_extent->nr_pages = nr_pages;
 317         async_extent->compress_type = compress_type;
 318         list_add_tail(&async_extent->list, &cow->extents);
 319         return 0;
 320 }
 321
 322 /*
 323  * we create compressed extents in two phases.  The first
 324  * phase compresses a range of pages that have already been
 325  * locked (both pages and state bits are locked).
 326  *
 327  * This is done inside an ordered work queue, and the compression
 328  * is spread across many cpus.  The actual IO submission is step
 329  * two, and the ordered work queue takes care of making sure that
 330  * happens in the same order things were put onto the queue by
 331  * writepages and friends.
 332  *
 333  * If this code finds it can't get good compression, it puts an
 334  * entry onto the work queue to write the uncompressed bytes.  This
 335  * makes sure that both compressed inodes and uncompressed inodes
 336  * are written in the same order that the flusher thread sent them
 337  * down.
 338  */
 339 static noinline int compress_file_range(struct inode *inode,
 340                                         struct page *locked_page,
 341                                         u64 start, u64 end,
 342                                         struct async_cow *async_cow,
 343                                         int *num_added)
 344 {
 345         struct btrfs_root *root = BTRFS_I(inode)->root;
 346         struct btrfs_trans_handle *trans;
 347         u64 num_bytes;
 348         u64 blocksize = root->sectorsize;
 349         u64 actual_end;
 350         u64 isize = i_size_read(inode);
 351         int ret = 0;
 352         struct page **pages = NULL;
 353         unsigned long nr_pages;
 354         unsigned long nr_pages_ret = 0;
 355         unsigned long total_compressed = 0;
 356         unsigned long total_in = 0;
 357         unsigned long max_compressed = 128 * 1024;
 358         unsigned long max_uncompressed = 128 * 1024;
 359         int i;
 360         int will_compress;
 361         int compress_type = root->fs_info->compress_type;
 362         int redirty = 0;
 363
 364         /* if this is a small write inside eof, kick off a defrag */
 365         if ((end - start + 1) < 16 * 1024 &&
 366             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 367                 btrfs_add_inode_defrag(NULL, inode);
 368
 369         actual_end = min_t(u64, isize, end + 1);
 370 again:
 371         will_compress = 0;
 372         nr_pages = (end >> PAGE_CACHE_SHIFT) - (start >> PAGE_CACHE_SHIFT) + 1;
 373         nr_pages = min(nr_pages, (128 * 1024UL) / PAGE_CACHE_SIZE);
 374
 375         /*
 376          * we don't want to send crud past the end of i_size through
 377          * compression, that's just a waste of CPU time.  So, if the
 378          * end of the file is before the start of our current
 379          * requested range of bytes, we bail out to the uncompressed
 380          * cleanup code that can deal with all of this.
 381          *
 382          * It isn't really the fastest way to fix things, but this is a
 383          * very uncommon corner.
 384          */
 385         if (actual_end <= start)
 386                 goto cleanup_and_bail_uncompressed;
 387
 388         total_compressed = actual_end - start;
 389
 390         /* we want to make sure that amount of ram required to uncompress
 391          * an extent is reasonable, so we limit the total size in ram
 392          * of a compressed extent to 128k.  This is a crucial number
 393          * because it also controls how easily we can spread reads across
 394          * cpus for decompression.
 395          *
 396          * We also want to make sure the amount of IO required to do
 397          * a random read is reasonably small, so we limit the size of
 398          * a compressed extent to 128k.
 399          */
 400         total_compressed = min(total_compressed, max_uncompressed);
 401         num_bytes = ALIGN(end - start + 1, blocksize);
 402         num_bytes = max(blocksize,  num_bytes);
 403         total_in = 0;
 404         ret = 0;
 405
 406         /*
 407          * we do compression for mount -o compress and when the
 408          * inode has not been flagged as nocompress.  This flag can
 409          * change at any time if we discover bad compression ratios.
 410          */
 411         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS) &&
 412             (btrfs_test_opt(root, COMPRESS) ||
 413              (BTRFS_I(inode)->force_compress) ||
 414              (BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS))) {
 415                 WARN_ON(pages);
 416                 pages = kzalloc(sizeof(struct page *) * nr_pages, GFP_NOFS);
 417                 if (!pages) {
 418                         /* just bail out to the uncompressed code */
 419                         goto cont;
 420                 }
 421
 422                 if (BTRFS_I(inode)->force_compress)
 423                         compress_type = BTRFS_I(inode)->force_compress;
 424
 425                 /*
 426                  * we need to call clear_page_dirty_for_io on each
 427                  * page in the range.  Otherwise applications with the file
 428                  * mmap'd can wander in and change the page contents while
 429                  * we are compressing them.
 430                  *
 431                  * If the compression fails for any reason, we set the pages
 432                  * dirty again later on.
 433                  */
 434                 extent_range_clear_dirty_for_io(inode, start, end);
 435                 redirty = 1;
 436                 ret = btrfs_compress_pages(compress_type,
 437                                            inode->i_mapping, start,
 438                                            total_compressed, pages,
 439                                            nr_pages, &nr_pages_ret,
 440                                            &total_in,
 441                                            &total_compressed,
 442                                            max_compressed);
 443
 444                 if (!ret) {
 445                         unsigned long offset = total_compressed &
 446                                 (PAGE_CACHE_SIZE - 1);
 447                         struct page *page = pages[nr_pages_ret - 1];
 448                         char *kaddr;
 449
 450                         /* zero the tail end of the last page, we might be
 451                          * sending it down to disk
 452                          */
 453                         if (offset) {
 454                                 kaddr = kmap_atomic(page);
 455                                 memset(kaddr + offset, 0,
 456                                        PAGE_CACHE_SIZE - offset);
 457                                 kunmap_atomic(kaddr);
 458                         }
 459                         will_compress = 1;
 460                 }
 461         }
 462 cont:
 463         if (start == 0) {
 464                 trans = btrfs_join_transaction(root);
 465                 if (IS_ERR(trans)) {
 466                         ret = PTR_ERR(trans);
 467                         trans = NULL;
 468                         goto cleanup_and_out;
 469                 }
 470                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 471
 472                 /* lets try to make an inline extent */
 473                 if (ret || total_in < (actual_end - start)) {
 474                         /* we didn't compress the entire range, try
 475                          * to make an uncompressed inline extent.
 476                          */
 477                         ret = cow_file_range_inline(trans, root, inode,
 478                                                     start, end, 0, 0, NULL);
 479                 } else {
 480                         /* try making a compressed inline extent */
 481                         ret = cow_file_range_inline(trans, root, inode,
 482                                                     start, end,
 483                                                     total_compressed,
 484                                                     compress_type, pages);
 485                 }
 486                 if (ret <= 0) {
 487                         /*
 488                          * inline extent creation worked or returned error,
 489                          * we don't need to create any more async work items.
 490                          * Unlock and free up our temp pages.
 491                          */
 492                         extent_clear_unlock_delalloc(inode,
 493                              &BTRFS_I(inode)->io_tree,
 494                              start, end, NULL,
 495                              EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
 496                              EXTENT_CLEAR_DELALLOC |
 497                              EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK);
 498
 499                         btrfs_end_transaction(trans, root);
 500                         goto free_pages_out;
 501                 }
 502                 btrfs_end_transaction(trans, root);
 503         }
 504
 505         if (will_compress) {
 506                 /*
 507                  * we aren't doing an inline extent round the compressed size
 508                  * up to a block size boundary so the allocator does sane
 509                  * things
 510                  */
 511                 total_compressed = ALIGN(total_compressed, blocksize);
 512
 513                 /*
 514                  * one last check to make sure the compression is really a
 515                  * win, compare the page count read with the blocks on disk
 516                  */
 517                 total_in = ALIGN(total_in, PAGE_CACHE_SIZE);
 518                 if (total_compressed >= total_in) {
 519                         will_compress = 0;
 520                 } else {
 521                         num_bytes = total_in;
 522                 }
 523         }
 524         if (!will_compress && pages) {
 525                 /*
 526                  * the compression code ran but failed to make things smaller,
 527                  * free any pages it allocated and our page pointer array
 528                  */
 529                 for (i = 0; i < nr_pages_ret; i++) {
 530                         WARN_ON(pages[i]->mapping);
 531                         page_cache_release(pages[i]);
 532                 }
 533                 kfree(pages);
 534                 pages = NULL;
 535                 total_compressed = 0;
 536                 nr_pages_ret = 0;
 537
 538                 /* flag the file so we don't compress in the future */
 539                 if (!btrfs_test_opt(root, FORCE_COMPRESS) &&
 540                     !(BTRFS_I(inode)->force_compress)) {
 541                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 542                 }
 543         }
 544         if (will_compress) {
 545                 *num_added += 1;
 546
 547                 /* the async work queues will take care of doing actual
 548                  * allocation on disk for these compressed pages,
 549                  * and will submit them to the elevator.
 550                  */
 551                 add_async_extent(async_cow, start, num_bytes,
 552                                  total_compressed, pages, nr_pages_ret,
 553                                  compress_type);
 554
 555                 if (start + num_bytes < end) {
 556                         start += num_bytes;
 557                         pages = NULL;
 558                         cond_resched();
 559                         goto again;
 560                 }
 561         } else {
 562 cleanup_and_bail_uncompressed:
 563                 /*
 564                  * No compression, but we still need to write the pages in
 565                  * the file we've been given so far.  redirty the locked
 566                  * page if it corresponds to our extent and set things up
 567                  * for the async work queue to run cow_file_range to do
 568                  * the normal delalloc dance
 569                  */
 570                 if (page_offset(locked_page) >= start &&
 571                     page_offset(locked_page) <= end) {
 572                         __set_page_dirty_nobuffers(locked_page);
 573                         /* unlocked later on in the async handlers */
 574                 }
 575                 if (redirty)
 576                         extent_range_redirty_for_io(inode, start, end);
 577                 add_async_extent(async_cow, start, end - start + 1,
 578                                  0, NULL, 0, BTRFS_COMPRESS_NONE);
 579                 *num_added += 1;
 580         }
 581
 582 out:
 583         return ret;
 584
 585 free_pages_out:
 586         for (i = 0; i < nr_pages_ret; i++) {
 587                 WARN_ON(pages[i]->mapping);
 588                 page_cache_release(pages[i]);
 589         }
 590         kfree(pages);
 591
 592         goto out;
 593
 594 cleanup_and_out:
 595         extent_clear_unlock_delalloc(inode, &BTRFS_I(inode)->io_tree,
 596                                      start, end, NULL,
 597                                      EXTENT_CLEAR_UNLOCK_PAGE |
 598                                      EXTENT_CLEAR_DIRTY |
 599                                      EXTENT_CLEAR_DELALLOC |
 600                                      EXTENT_SET_WRITEBACK |
 601                                      EXTENT_END_WRITEBACK);
 602         if (!trans || IS_ERR(trans))
 603                 btrfs_error(root->fs_info, ret, "Failed to join transaction");
 604         else
 605                 btrfs_abort_transaction(trans, root, ret);
 606         goto free_pages_out;
 607 }
 608
 609 /*
 610  * phase two of compressed writeback.  This is the ordered portion
 611  * of the code, which only gets called in the order the work was
 612  * queued.  We walk all the async extents created by compress_file_range
 613  * and send them down to the disk.
 614  */
 615 static noinline int submit_compressed_extents(struct inode *inode,
 616                                               struct async_cow *async_cow)
 617 {
 618         struct async_extent *async_extent;
 619         u64 alloc_hint = 0;
 620         struct btrfs_trans_handle *trans;
 621         struct btrfs_key ins;
 622         struct extent_map *em;
 623         struct btrfs_root *root = BTRFS_I(inode)->root;
 624         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 625         struct extent_io_tree *io_tree;
 626         int ret = 0;
 627
 628         if (list_empty(&async_cow->extents))
 629                 return 0;
 630
 631 again:
 632         while (!list_empty(&async_cow->extents)) {
 633                 async_extent = list_entry(async_cow->extents.next,
 634                                           struct async_extent, list);
 635                 list_del(&async_extent->list);
 636
 637                 io_tree = &BTRFS_I(inode)->io_tree;
 638
 639 retry:
 640                 /* did the compression code fall back to uncompressed IO? */
 641                 if (!async_extent->pages) {
 642                         int page_started = 0;
 643                         unsigned long nr_written = 0;
 644
 645                         lock_extent(io_tree, async_extent->start,
 646                                          async_extent->start +
 647                                          async_extent->ram_size - 1);
 648
 649                         /* allocate blocks */
 650                         ret = cow_file_range(inode, async_cow->locked_page,
 651                                              async_extent->start,
 652                                              async_extent->start +
 653                                              async_extent->ram_size - 1,
 654                                              &page_started, &nr_written, 0);
 655
 656                         /* JDM XXX */
 657
 658                         /*
 659                          * if page_started, cow_file_range inserted an
 660                          * inline extent and took care of all the unlocking
 661                          * and IO for us.  Otherwise, we need to submit
 662                          * all those pages down to the drive.
 663                          */
 664                         if (!page_started && !ret)
 665                                 extent_write_locked_range(io_tree,
 666                                                   inode, async_extent->start,
 667                                                   async_extent->start +
 668                                                   async_extent->ram_size - 1,
 669                                                   btrfs_get_extent,
 670                                                   WB_SYNC_ALL);
 671                         else if (ret)
 672                                 unlock_page(async_cow->locked_page);
 673                         kfree(async_extent);
 674                         cond_resched();
 675                         continue;
 676                 }
 677
 678                 lock_extent(io_tree, async_extent->start,
 679                             async_extent->start + async_extent->ram_size - 1);
 680
 681                 trans = btrfs_join_transaction(root);
 682                 if (IS_ERR(trans)) {
 683                         ret = PTR_ERR(trans);
 684                 } else {
 685                         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 686                         ret = btrfs_reserve_extent(trans, root,
 687                                            async_extent->compressed_size,
 688                                            async_extent->compressed_size,
 689                                            0, alloc_hint, &ins, 1);
 690                         if (ret && ret != -ENOSPC)
 691                                 btrfs_abort_transaction(trans, root, ret);
 692                         btrfs_end_transaction(trans, root);
 693                 }
 694
 695                 if (ret) {
 696                         int i;
 697
 698                         for (i = 0; i < async_extent->nr_pages; i++) {
 699                                 WARN_ON(async_extent->pages[i]->mapping);
 700                                 page_cache_release(async_extent->pages[i]);
 701                         }
 702                         kfree(async_extent->pages);
 703                         async_extent->nr_pages = 0;
 704                         async_extent->pages = NULL;
 705
 706                         if (ret == -ENOSPC) {
 707                                 unlock_extent(io_tree, async_extent->start,
 708                                               async_extent->start +
 709                                               async_extent->ram_size - 1);
 710                                 goto retry;
 711                         }
 712                         goto out_free;
 713                 }
 714
 715                 /*
 716                  * here we're doing allocation and writeback of the
 717                  * compressed pages
 718                  */
 719                 btrfs_drop_extent_cache(inode, async_extent->start,
 720                                         async_extent->start +
 721                                         async_extent->ram_size - 1, 0);
 722
 723                 em = alloc_extent_map();
 724                 if (!em) {
 725                         ret = -ENOMEM;
 726                         goto out_free_reserve;
 727                 }
 728                 em->start = async_extent->start;
 729                 em->len = async_extent->ram_size;
 730                 em->orig_start = em->start;
 731                 em->mod_start = em->start;
 732                 em->mod_len = em->len;
 733
 734                 em->block_start = ins.objectid;
 735                 em->block_len = ins.offset;
 736                 em->orig_block_len = ins.offset;
 737                 em->ram_bytes = async_extent->ram_size;
 738                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 739                 em->compress_type = async_extent->compress_type;
 740                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 741                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 742                 em->generation = -1;
 743
 744                 while (1) {
 745                         write_lock(&em_tree->lock);
 746                         ret = add_extent_mapping(em_tree, em, 1);
 747                         write_unlock(&em_tree->lock);
 748                         if (ret != -EEXIST) {
 749                                 free_extent_map(em);
 750                                 break;
 751                         }
 752                         btrfs_drop_extent_cache(inode, async_extent->start,
 753                                                 async_extent->start +
 754                                                 async_extent->ram_size - 1, 0);
 755                 }
 756
 757                 if (ret)
 758                         goto out_free_reserve;
 759
 760                 ret = btrfs_add_ordered_extent_compress(inode,
 761                                                 async_extent->start,
 762                                                 ins.objectid,
 763                                                 async_extent->ram_size,
 764                                                 ins.offset,
 765                                                 BTRFS_ORDERED_COMPRESSED,
 766                                                 async_extent->compress_type);
 767                 if (ret)
 768                         goto out_free_reserve;
 769
 770                 /*
 771                  * clear dirty, set writeback and unlock the pages.
 772                  */
 773                 extent_clear_unlock_delalloc(inode,
 774                                 &BTRFS_I(inode)->io_tree,
 775                                 async_extent->start,
 776                                 async_extent->start +
 777                                 async_extent->ram_size - 1,
 778                                 NULL, EXTENT_CLEAR_UNLOCK_PAGE |
 779                                 EXTENT_CLEAR_UNLOCK |
 780                                 EXTENT_CLEAR_DELALLOC |
 781                                 EXTENT_CLEAR_DIRTY | EXTENT_SET_WRITEBACK);
 782
 783                 ret = btrfs_submit_compressed_write(inode,
 784                                     async_extent->start,
 785                                     async_extent->ram_size,
 786                                     ins.objectid,
 787                                     ins.offset, async_extent->pages,
 788                                     async_extent->nr_pages);
 789                 alloc_hint = ins.objectid + ins.offset;
 790                 kfree(async_extent);
 791                 if (ret)
 792                         goto out;
 793                 cond_resched();
 794         }
 795         ret = 0;
 796 out:
 797         return ret;
 798 out_free_reserve:
 799         btrfs_free_reserved_extent(root, ins.objectid, ins.offset);
 800 out_free:
 801         extent_clear_unlock_delalloc(inode, &BTRFS_I(inode)->io_tree,
 802                                      async_extent->start,
 803                                      async_extent->start +
 804                                      async_extent->ram_size - 1,
 805                                      NULL, EXTENT_CLEAR_UNLOCK_PAGE |
 806                                      EXTENT_CLEAR_UNLOCK |
 807                                      EXTENT_CLEAR_DELALLOC |
 808                                      EXTENT_CLEAR_DIRTY |
 809                                      EXTENT_SET_WRITEBACK |
 810                                      EXTENT_END_WRITEBACK);
 811         kfree(async_extent);
 812         goto again;
 813 }
 814
 815 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 816                                       u64 num_bytes)
 817 {
 818         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 819         struct extent_map *em;
 820         u64 alloc_hint = 0;
 821
 822         read_lock(&em_tree->lock);
 823         em = search_extent_mapping(em_tree, start, num_bytes);
 824         if (em) {
 825                 /*
 826                  * if block start isn't an actual block number then find the
 827                  * first block in this inode and use that as a hint.  If that
 828                  * block is also bogus then just don't worry about it.
 829                  */
 830                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 831                         free_extent_map(em);
 832                         em = search_extent_mapping(em_tree, 0, 0);
 833                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 834                                 alloc_hint = em->block_start;
 835                         if (em)
 836                                 free_extent_map(em);
 837                 } else {
 838                         alloc_hint = em->block_start;
 839                         free_extent_map(em);
 840                 }
 841         }
 842         read_unlock(&em_tree->lock);
 843
 844         return alloc_hint;
 845 }
 846
 847 /*
 848  * when extent_io.c finds a delayed allocation range in the file,
 849  * the call backs end up in this code.  The basic idea is to
 850  * allocate extents on disk for the range, and create ordered data structs
 851  * in ram to track those extents.
 852  *
 853  * locked_page is the page that writepage had locked already.  We use
 854  * it to make sure we don't do extra locks or unlocks.
 855  *
 856  * *page_started is set to one if we unlock locked_page and do everything
 857  * required to start IO on it.  It may be clean and already done with
 858  * IO when we return.
 859  */
 860 static noinline int __cow_file_range(struct btrfs_trans_handle *trans,
 861                                      struct inode *inode,
 862                                      struct btrfs_root *root,
 863                                      struct page *locked_page,
 864                                      u64 start, u64 end, int *page_started,
 865                                      unsigned long *nr_written,
 866                                      int unlock)
 867 {
 868         u64 alloc_hint = 0;
 869         u64 num_bytes;
 870         unsigned long ram_size;
 871         u64 disk_num_bytes;
 872         u64 cur_alloc_size;
 873         u64 blocksize = root->sectorsize;
 874         struct btrfs_key ins;
 875         struct extent_map *em;
 876         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 877         int ret = 0;
 878
 879         BUG_ON(btrfs_is_free_space_inode(inode));
 880
 881         num_bytes = ALIGN(end - start + 1, blocksize);
 882         num_bytes = max(blocksize,  num_bytes);
 883         disk_num_bytes = num_bytes;
 884
 885         /* if this is a small write inside eof, kick off defrag */
 886         if (num_bytes < 64 * 1024 &&
 887             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 888                 btrfs_add_inode_defrag(trans, inode);
 889
 890         if (start == 0) {
 891                 /* lets try to make an inline extent */
 892                 ret = cow_file_range_inline(trans, root, inode,
 893                                             start, end, 0, 0, NULL);
 894                 if (ret == 0) {
 895                         extent_clear_unlock_delalloc(inode,
 896                                      &BTRFS_I(inode)->io_tree,
 897                                      start, end, NULL,
 898                                      EXTENT_CLEAR_UNLOCK_PAGE |
 899                                      EXTENT_CLEAR_UNLOCK |
 900                                      EXTENT_CLEAR_DELALLOC |
 901                                      EXTENT_CLEAR_DIRTY |
 902                                      EXTENT_SET_WRITEBACK |
 903                                      EXTENT_END_WRITEBACK);
 904
 905                         *nr_written = *nr_written +
 906                              (end - start + PAGE_CACHE_SIZE) / PAGE_CACHE_SIZE;
 907                         *page_started = 1;
 908                         goto out;
 909                 } else if (ret < 0) {
 910                         btrfs_abort_transaction(trans, root, ret);
 911                         goto out_unlock;
 912                 }
 913         }
 914
 915         BUG_ON(disk_num_bytes >
 916                btrfs_super_total_bytes(root->fs_info->super_copy));
 917
 918         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 919         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 920
 921         while (disk_num_bytes > 0) {
 922                 unsigned long op;
 923
 924                 cur_alloc_size = disk_num_bytes;
 925                 ret = btrfs_reserve_extent(trans, root, cur_alloc_size,
 926                                            root->sectorsize, 0, alloc_hint,
 927                                            &ins, 1);
 928                 if (ret < 0) {
 929                         btrfs_abort_transaction(trans, root, ret);
 930                         goto out_unlock;
 931                 }
 932
 933                 em = alloc_extent_map();
 934                 if (!em) {
 935                         ret = -ENOMEM;
 936                         goto out_reserve;
 937                 }
 938                 em->start = start;
 939                 em->orig_start = em->start;
 940                 ram_size = ins.offset;
 941                 em->len = ins.offset;
 942                 em->mod_start = em->start;
 943                 em->mod_len = em->len;
 944
 945                 em->block_start = ins.objectid;
 946                 em->block_len = ins.offset;
 947                 em->orig_block_len = ins.offset;
 948                 em->ram_bytes = ram_size;
 949                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 950                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 951                 em->generation = -1;
 952
 953                 while (1) {
 954                         write_lock(&em_tree->lock);
 955                         ret = add_extent_mapping(em_tree, em, 1);
 956                         write_unlock(&em_tree->lock);
 957                         if (ret != -EEXIST) {
 958                                 free_extent_map(em);
 959                                 break;
 960                         }
 961                         btrfs_drop_extent_cache(inode, start,
 962                                                 start + ram_size - 1, 0);
 963                 }
 964                 if (ret)
 965                         goto out_reserve;
 966
 967                 cur_alloc_size = ins.offset;
 968                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
 969                                                ram_size, cur_alloc_size, 0);
 970                 if (ret)
 971                         goto out_reserve;
 972
 973                 if (root->root_key.objectid ==
 974                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
 975                         ret = btrfs_reloc_clone_csums(inode, start,
 976                                                       cur_alloc_size);
 977                         if (ret) {
 978                                 btrfs_abort_transaction(trans, root, ret);
 979                                 goto out_reserve;
 980                         }
 981                 }
 982
 983                 if (disk_num_bytes < cur_alloc_size)
 984                         break;
 985
 986                 /* we're not doing compressed IO, don't unlock the first
 987                  * page (which the caller expects to stay locked), don't
 988                  * clear any dirty bits and don't set any writeback bits
 989                  *
 990                  * Do set the Private2 bit so we know this page was properly
 991                  * setup for writepage
 992                  */
 993                 op = unlock ? EXTENT_CLEAR_UNLOCK_PAGE : 0;
 994                 op |= EXTENT_CLEAR_UNLOCK | EXTENT_CLEAR_DELALLOC |
 995                         EXTENT_SET_PRIVATE2;
 996
 997                 extent_clear_unlock_delalloc(inode, &BTRFS_I(inode)->io_tree,
 998                                              start, start + ram_size - 1,
 999                                              locked_page, op);
1000                 disk_num_bytes -= cur_alloc_size;
1001                 num_bytes -= cur_alloc_size;
1002                 alloc_hint = ins.objectid + ins.offset;
1003                 start += cur_alloc_size;
1004         }
1005 out:
1006         return ret;
1007
1008 out_reserve:
1009         btrfs_free_reserved_extent(root, ins.objectid, ins.offset);
1010 out_unlock:
1011         extent_clear_unlock_delalloc(inode,
1012                      &BTRFS_I(inode)->io_tree,
1013                      start, end, locked_page,
1014                      EXTENT_CLEAR_UNLOCK_PAGE |
1015                      EXTENT_CLEAR_UNLOCK |
1016                      EXTENT_CLEAR_DELALLOC |
1017                      EXTENT_CLEAR_DIRTY |
1018                      EXTENT_SET_WRITEBACK |
1019                      EXTENT_END_WRITEBACK);
1020
1021         goto out;
1022 }
1023
1024 static noinline int cow_file_range(struct inode *inode,
1025                                    struct page *locked_page,
1026                                    u64 start, u64 end, int *page_started,
1027                                    unsigned long *nr_written,
1028                                    int unlock)
1029 {
1030         struct btrfs_trans_handle *trans;
1031         struct btrfs_root *root = BTRFS_I(inode)->root;
1032         int ret;
1033
1034         trans = btrfs_join_transaction(root);
1035         if (IS_ERR(trans)) {
1036                 extent_clear_unlock_delalloc(inode,
1037                              &BTRFS_I(inode)->io_tree,
1038                              start, end, locked_page,
1039                              EXTENT_CLEAR_UNLOCK_PAGE |
1040                              EXTENT_CLEAR_UNLOCK |
1041                              EXTENT_CLEAR_DELALLOC |
1042                              EXTENT_CLEAR_DIRTY |
1043                              EXTENT_SET_WRITEBACK |
1044                              EXTENT_END_WRITEBACK);
1045                 return PTR_ERR(trans);
1046         }
1047         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1048
1049         ret = __cow_file_range(trans, inode, root, locked_page, start, end,
1050                                page_started, nr_written, unlock);
1051
1052         btrfs_end_transaction(trans, root);
1053
1054         return ret;
1055 }
1056
1057 /*
1058  * work queue call back to started compression on a file and pages
1059  */
1060 static noinline void async_cow_start(struct btrfs_work *work)
1061 {
1062         struct async_cow *async_cow;
1063         int num_added = 0;
1064         async_cow = container_of(work, struct async_cow, work);
1065
1066         compress_file_range(async_cow->inode, async_cow->locked_page,
1067                             async_cow->start, async_cow->end, async_cow,
1068                             &num_added);
1069         if (num_added == 0) {
1070                 btrfs_add_delayed_iput(async_cow->inode);
1071                 async_cow->inode = NULL;
1072         }
1073 }
1074
1075 /*
1076  * work queue call back to submit previously compressed pages
1077  */
1078 static noinline void async_cow_submit(struct btrfs_work *work)
1079 {
1080         struct async_cow *async_cow;
1081         struct btrfs_root *root;
1082         unsigned long nr_pages;
1083
1084         async_cow = container_of(work, struct async_cow, work);
1085
1086         root = async_cow->root;
1087         nr_pages = (async_cow->end - async_cow->start + PAGE_CACHE_SIZE) >>
1088                 PAGE_CACHE_SHIFT;
1089
1090         if (atomic_sub_return(nr_pages, &root->fs_info->async_delalloc_pages) <
1091             5 * 1024 * 1024 &&
1092             waitqueue_active(&root->fs_info->async_submit_wait))
1093                 wake_up(&root->fs_info->async_submit_wait);
1094
1095         if (async_cow->inode)
1096                 submit_compressed_extents(async_cow->inode, async_cow);
1097 }
1098
1099 static noinline void async_cow_free(struct btrfs_work *work)
1100 {
1101         struct async_cow *async_cow;
1102         async_cow = container_of(work, struct async_cow, work);
1103         if (async_cow->inode)
1104                 btrfs_add_delayed_iput(async_cow->inode);
1105         kfree(async_cow);
1106 }
1107
1108 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1109                                 u64 start, u64 end, int *page_started,
1110                                 unsigned long *nr_written)
1111 {
1112         struct async_cow *async_cow;
1113         struct btrfs_root *root = BTRFS_I(inode)->root;
1114         unsigned long nr_pages;
1115         u64 cur_end;
1116         int limit = 10 * 1024 * 1024;
1117
1118         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1119                          1, 0, NULL, GFP_NOFS);
1120         while (start < end) {
1121                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1122                 BUG_ON(!async_cow); /* -ENOMEM */
1123                 async_cow->inode = igrab(inode);
1124                 async_cow->root = root;
1125                 async_cow->locked_page = locked_page;
1126                 async_cow->start = start;
1127
1128                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
1129                         cur_end = end;
1130                 else
1131                         cur_end = min(end, start + 512 * 1024 - 1);
1132
1133                 async_cow->end = cur_end;
1134                 INIT_LIST_HEAD(&async_cow->extents);
1135
1136                 async_cow->work.func = async_cow_start;
1137                 async_cow->work.ordered_func = async_cow_submit;
1138                 async_cow->work.ordered_free = async_cow_free;
1139                 async_cow->work.flags = 0;
1140
1141                 nr_pages = (cur_end - start + PAGE_CACHE_SIZE) >>
1142                         PAGE_CACHE_SHIFT;
1143                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
1144
1145                 btrfs_queue_worker(&root->fs_info->delalloc_workers,
1146                                    &async_cow->work);
1147
1148                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
1149                         wait_event(root->fs_info->async_submit_wait,
1150                            (atomic_read(&root->fs_info->async_delalloc_pages) <
1151                             limit));
1152                 }
1153
1154                 while (atomic_read(&root->fs_info->async_submit_draining) &&
1155                       atomic_read(&root->fs_info->async_delalloc_pages)) {
1156                         wait_event(root->fs_info->async_submit_wait,
1157                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1158                            0));
1159                 }
1160
1161                 *nr_written += nr_pages;
1162                 start = cur_end + 1;
1163         }
1164         *page_started = 1;
1165         return 0;
1166 }
1167
1168 static noinline int csum_exist_in_range(struct btrfs_root *root,
1169                                         u64 bytenr, u64 num_bytes)
1170 {
1171         int ret;
1172         struct btrfs_ordered_sum *sums;
1173         LIST_HEAD(list);
1174
1175         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1176                                        bytenr + num_bytes - 1, &list, 0);
1177         if (ret == 0 && list_empty(&list))
1178                 return 0;
1179
1180         while (!list_empty(&list)) {
1181                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1182                 list_del(&sums->list);
1183                 kfree(sums);
1184         }
1185         return 1;
1186 }
1187
1188 /*
1189  * when nowcow writeback call back.  This checks for snapshots or COW copies
1190  * of the extents that exist in the file, and COWs the file as required.
1191  *
1192  * If no cow copies or snapshots exist, we write directly to the existing
1193  * blocks on disk
1194  */
1195 static noinline int run_delalloc_nocow(struct inode *inode,
1196                                        struct page *locked_page,
1197                               u64 start, u64 end, int *page_started, int force,
1198                               unsigned long *nr_written)
1199 {
1200         struct btrfs_root *root = BTRFS_I(inode)->root;
1201         struct btrfs_trans_handle *trans;
1202         struct extent_buffer *leaf;
1203         struct btrfs_path *path;
1204         struct btrfs_file_extent_item *fi;
1205         struct btrfs_key found_key;
1206         u64 cow_start;
1207         u64 cur_offset;
1208         u64 extent_end;
1209         u64 extent_offset;
1210         u64 disk_bytenr;
1211         u64 num_bytes;
1212         u64 disk_num_bytes;
1213         u64 ram_bytes;
1214         int extent_type;
1215         int ret, err;
1216         int type;
1217         int nocow;
1218         int check_prev = 1;
1219         bool nolock;
1220         u64 ino = btrfs_ino(inode);
1221
1222         path = btrfs_alloc_path();
1223         if (!path) {
1224                 extent_clear_unlock_delalloc(inode,
1225                              &BTRFS_I(inode)->io_tree,
1226                              start, end, locked_page,
1227                              EXTENT_CLEAR_UNLOCK_PAGE |
1228                              EXTENT_CLEAR_UNLOCK |
1229                              EXTENT_CLEAR_DELALLOC |
1230                              EXTENT_CLEAR_DIRTY |
1231                              EXTENT_SET_WRITEBACK |
1232                              EXTENT_END_WRITEBACK);
1233                 return -ENOMEM;
1234         }
1235
1236         nolock = btrfs_is_free_space_inode(inode);
1237
1238         if (nolock)
1239                 trans = btrfs_join_transaction_nolock(root);
1240         else
1241                 trans = btrfs_join_transaction(root);
1242
1243         if (IS_ERR(trans)) {
1244                 extent_clear_unlock_delalloc(inode,
1245                              &BTRFS_I(inode)->io_tree,
1246                              start, end, locked_page,
1247                              EXTENT_CLEAR_UNLOCK_PAGE |
1248                              EXTENT_CLEAR_UNLOCK |
1249                              EXTENT_CLEAR_DELALLOC |
1250                              EXTENT_CLEAR_DIRTY |
1251                              EXTENT_SET_WRITEBACK |
1252                              EXTENT_END_WRITEBACK);
1253                 btrfs_free_path(path);
1254                 return PTR_ERR(trans);
1255         }
1256
1257         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1258
1259         cow_start = (u64)-1;
1260         cur_offset = start;
1261         while (1) {
1262                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1263                                                cur_offset, 0);
1264                 if (ret < 0) {
1265                         btrfs_abort_transaction(trans, root, ret);
1266                         goto error;
1267                 }
1268                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1269                         leaf = path->nodes[0];
1270                         btrfs_item_key_to_cpu(leaf, &found_key,
1271                                               path->slots[0] - 1);
1272                         if (found_key.objectid == ino &&
1273                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1274                                 path->slots[0]--;
1275                 }
1276                 check_prev = 0;
1277 next_slot:
1278                 leaf = path->nodes[0];
1279                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1280                         ret = btrfs_next_leaf(root, path);
1281                         if (ret < 0) {
1282                                 btrfs_abort_transaction(trans, root, ret);
1283                                 goto error;
1284                         }
1285                         if (ret > 0)
1286                                 break;
1287                         leaf = path->nodes[0];
1288                 }
1289
1290                 nocow = 0;
1291                 disk_bytenr = 0;
1292                 num_bytes = 0;
1293                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1294
1295                 if (found_key.objectid > ino ||
1296                     found_key.type > BTRFS_EXTENT_DATA_KEY ||
1297                     found_key.offset > end)
1298                         break;
1299
1300                 if (found_key.offset > cur_offset) {
1301                         extent_end = found_key.offset;
1302                         extent_type = 0;
1303                         goto out_check;
1304                 }
1305
1306                 fi = btrfs_item_ptr(leaf, path->slots[0],
1307                                     struct btrfs_file_extent_item);
1308                 extent_type = btrfs_file_extent_type(leaf, fi);
1309
1310                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1311                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1312                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1313                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1314                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1315                         extent_end = found_key.offset +
1316                                 btrfs_file_extent_num_bytes(leaf, fi);
1317                         disk_num_bytes =
1318                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1319                         if (extent_end <= start) {
1320                                 path->slots[0]++;
1321                                 goto next_slot;
1322                         }
1323                         if (disk_bytenr == 0)
1324                                 goto out_check;
1325                         if (btrfs_file_extent_compression(leaf, fi) ||
1326                             btrfs_file_extent_encryption(leaf, fi) ||
1327                             btrfs_file_extent_other_encoding(leaf, fi))
1328                                 goto out_check;
1329                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1330                                 goto out_check;
1331                         if (btrfs_extent_readonly(root, disk_bytenr))
1332                                 goto out_check;
1333                         if (btrfs_cross_ref_exist(trans, root, ino,
1334                                                   found_key.offset -
1335                                                   extent_offset, disk_bytenr))
1336                                 goto out_check;
1337                         disk_bytenr += extent_offset;
1338                         disk_bytenr += cur_offset - found_key.offset;
1339                         num_bytes = min(end + 1, extent_end) - cur_offset;
1340                         /*
1341                          * force cow if csum exists in the range.
1342                          * this ensure that csum for a given extent are
1343                          * either valid or do not exist.
1344                          */
1345                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1346                                 goto out_check;
1347                         nocow = 1;
1348                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1349                         extent_end = found_key.offset +
1350                                 btrfs_file_extent_inline_len(leaf, fi);
1351                         extent_end = ALIGN(extent_end, root->sectorsize);
1352                 } else {
1353                         BUG_ON(1);
1354                 }
1355 out_check:
1356                 if (extent_end <= start) {
1357                         path->slots[0]++;
1358                         goto next_slot;
1359                 }
1360                 if (!nocow) {
1361                         if (cow_start == (u64)-1)
1362                                 cow_start = cur_offset;
1363                         cur_offset = extent_end;
1364                         if (cur_offset > end)
1365                                 break;
1366                         path->slots[0]++;
1367                         goto next_slot;
1368                 }
1369
1370                 btrfs_release_path(path);
1371                 if (cow_start != (u64)-1) {
1372                         ret = __cow_file_range(trans, inode, root, locked_page,
1373                                                cow_start, found_key.offset - 1,
1374                                                page_started, nr_written, 1);
1375                         if (ret) {
1376                                 btrfs_abort_transaction(trans, root, ret);
1377                                 goto error;
1378                         }
1379                         cow_start = (u64)-1;
1380                 }
1381
1382                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1383                         struct extent_map *em;
1384                         struct extent_map_tree *em_tree;
1385                         em_tree = &BTRFS_I(inode)->extent_tree;
1386                         em = alloc_extent_map();
1387                         BUG_ON(!em); /* -ENOMEM */
1388                         em->start = cur_offset;
1389                         em->orig_start = found_key.offset - extent_offset;
1390                         em->len = num_bytes;
1391                         em->block_len = num_bytes;
1392                         em->block_start = disk_bytenr;
1393                         em->orig_block_len = disk_num_bytes;
1394                         em->ram_bytes = ram_bytes;
1395                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1396                         em->mod_start = em->start;
1397                         em->mod_len = em->len;
1398                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1399                         set_bit(EXTENT_FLAG_FILLING, &em->flags);
1400                         em->generation = -1;
1401                         while (1) {
1402                                 write_lock(&em_tree->lock);
1403                                 ret = add_extent_mapping(em_tree, em, 1);
1404                                 write_unlock(&em_tree->lock);
1405                                 if (ret != -EEXIST) {
1406                                         free_extent_map(em);
1407                                         break;
1408                                 }
1409                                 btrfs_drop_extent_cache(inode, em->start,
1410                                                 em->start + em->len - 1, 0);
1411                         }
1412                         type = BTRFS_ORDERED_PREALLOC;
1413                 } else {
1414                         type = BTRFS_ORDERED_NOCOW;
1415                 }
1416
1417                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1418                                                num_bytes, num_bytes, type);
1419                 BUG_ON(ret); /* -ENOMEM */
1420
1421                 if (root->root_key.objectid ==
1422                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1423                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1424                                                       num_bytes);
1425                         if (ret) {
1426                                 btrfs_abort_transaction(trans, root, ret);
1427                                 goto error;
1428                         }
1429                 }
1430
1431                 extent_clear_unlock_delalloc(inode, &BTRFS_I(inode)->io_tree,
1432                                 cur_offset, cur_offset + num_bytes - 1,
1433                                 locked_page, EXTENT_CLEAR_UNLOCK_PAGE |
1434                                 EXTENT_CLEAR_UNLOCK | EXTENT_CLEAR_DELALLOC |
1435                                 EXTENT_SET_PRIVATE2);
1436                 cur_offset = extent_end;
1437                 if (cur_offset > end)
1438                         break;
1439         }
1440         btrfs_release_path(path);
1441
1442         if (cur_offset <= end && cow_start == (u64)-1) {
1443                 cow_start = cur_offset;
1444                 cur_offset = end;
1445         }
1446
1447         if (cow_start != (u64)-1) {
1448                 ret = __cow_file_range(trans, inode, root, locked_page,
1449                                        cow_start, end,
1450                                        page_started, nr_written, 1);
1451                 if (ret) {
1452                         btrfs_abort_transaction(trans, root, ret);
1453                         goto error;
1454                 }
1455         }
1456
1457 error:
1458         err = btrfs_end_transaction(trans, root);
1459         if (!ret)
1460                 ret = err;
1461
1462         if (ret && cur_offset < end)
1463                 extent_clear_unlock_delalloc(inode,
1464                              &BTRFS_I(inode)->io_tree,
1465                              cur_offset, end, locked_page,
1466                              EXTENT_CLEAR_UNLOCK_PAGE |
1467                              EXTENT_CLEAR_UNLOCK |
1468                              EXTENT_CLEAR_DELALLOC |
1469                              EXTENT_CLEAR_DIRTY |
1470                              EXTENT_SET_WRITEBACK |
1471                              EXTENT_END_WRITEBACK);
1472
1473         btrfs_free_path(path);
1474         return ret;
1475 }
1476
1477 /*
1478  * extent_io.c call back to do delayed allocation processing
1479  */
1480 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1481                               u64 start, u64 end, int *page_started,
1482                               unsigned long *nr_written)
1483 {
1484         int ret;
1485         struct btrfs_root *root = BTRFS_I(inode)->root;
1486
1487         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) {
1488                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1489                                          page_started, 1, nr_written);
1490         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC) {
1491                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1492                                          page_started, 0, nr_written);
1493         } else if (!btrfs_test_opt(root, COMPRESS) &&
1494                    !(BTRFS_I(inode)->force_compress) &&
1495                    !(BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS)) {
1496                 ret = cow_file_range(inode, locked_page, start, end,
1497                                       page_started, nr_written, 1);
1498         } else {
1499                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1500                         &BTRFS_I(inode)->runtime_flags);
1501                 ret = cow_file_range_async(inode, locked_page, start, end,
1502                                            page_started, nr_written);
1503         }
1504         return ret;
1505 }
1506
1507 static void btrfs_split_extent_hook(struct inode *inode,
1508                                     struct extent_state *orig, u64 split)
1509 {
1510         /* not delalloc, ignore it */
1511         if (!(orig->state & EXTENT_DELALLOC))
1512                 return;
1513
1514         spin_lock(&BTRFS_I(inode)->lock);
1515         BTRFS_I(inode)->outstanding_extents++;
1516         spin_unlock(&BTRFS_I(inode)->lock);
1517 }
1518
1519 /*
1520  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1521  * extents so we can keep track of new extents that are just merged onto old
1522  * extents, such as when we are doing sequential writes, so we can properly
1523  * account for the metadata space we'll need.
1524  */
1525 static void btrfs_merge_extent_hook(struct inode *inode,
1526                                     struct extent_state *new,
1527                                     struct extent_state *other)
1528 {
1529         /* not delalloc, ignore it */
1530         if (!(other->state & EXTENT_DELALLOC))
1531                 return;
1532
1533         spin_lock(&BTRFS_I(inode)->lock);
1534         BTRFS_I(inode)->outstanding_extents--;
1535         spin_unlock(&BTRFS_I(inode)->lock);
1536 }
1537
1538 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1539                                       struct inode *inode)
1540 {
1541         spin_lock(&root->delalloc_lock);
1542         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1543                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1544                               &root->delalloc_inodes);
1545                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1546                         &BTRFS_I(inode)->runtime_flags);
1547                 root->nr_delalloc_inodes++;
1548                 if (root->nr_delalloc_inodes == 1) {
1549                         spin_lock(&root->fs_info->delalloc_root_lock);
1550                         BUG_ON(!list_empty(&root->delalloc_root));
1551                         list_add_tail(&root->delalloc_root,
1552                                       &root->fs_info->delalloc_roots);
1553                         spin_unlock(&root->fs_info->delalloc_root_lock);
1554                 }
1555         }
1556         spin_unlock(&root->delalloc_lock);
1557 }
1558
1559 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1560                                      struct inode *inode)
1561 {
1562         spin_lock(&root->delalloc_lock);
1563         if (!list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1564                 list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1565                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1566                           &BTRFS_I(inode)->runtime_flags);
1567                 root->nr_delalloc_inodes--;
1568                 if (!root->nr_delalloc_inodes) {
1569                         spin_lock(&root->fs_info->delalloc_root_lock);
1570                         BUG_ON(list_empty(&root->delalloc_root));
1571                         list_del_init(&root->delalloc_root);
1572                         spin_unlock(&root->fs_info->delalloc_root_lock);
1573                 }
1574         }
1575         spin_unlock(&root->delalloc_lock);
1576 }
1577
1578 /*
1579  * extent_io.c set_bit_hook, used to track delayed allocation
1580  * bytes in this file, and to maintain the list of inodes that
1581  * have pending delalloc work to be done.
1582  */
1583 static void btrfs_set_bit_hook(struct inode *inode,
1584                                struct extent_state *state, unsigned long *bits)
1585 {
1586
1587         /*
1588          * set_bit and clear bit hooks normally require _irqsave/restore
1589          * but in this case, we are only testing for the DELALLOC
1590          * bit, which is only set or cleared with irqs on
1591          */
1592         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1593                 struct btrfs_root *root = BTRFS_I(inode)->root;
1594                 u64 len = state->end + 1 - state->start;
1595                 bool do_list = !btrfs_is_free_space_inode(inode);
1596
1597                 if (*bits & EXTENT_FIRST_DELALLOC) {
1598                         *bits &= ~EXTENT_FIRST_DELALLOC;
1599                 } else {
1600                         spin_lock(&BTRFS_I(inode)->lock);
1601                         BTRFS_I(inode)->outstanding_extents++;
1602                         spin_unlock(&BTRFS_I(inode)->lock);
1603                 }
1604
1605                 __percpu_counter_add(&root->fs_info->delalloc_bytes, len,
1606                                      root->fs_info->delalloc_batch);
1607                 spin_lock(&BTRFS_I(inode)->lock);
1608                 BTRFS_I(inode)->delalloc_bytes += len;
1609                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1610                                          &BTRFS_I(inode)->runtime_flags))
1611                         btrfs_add_delalloc_inodes(root, inode);
1612                 spin_unlock(&BTRFS_I(inode)->lock);
1613         }
1614 }
1615
1616 /*
1617  * extent_io.c clear_bit_hook, see set_bit_hook for why
1618  */
1619 static void btrfs_clear_bit_hook(struct inode *inode,
1620                                  struct extent_state *state,
1621                                  unsigned long *bits)
1622 {
1623         /*
1624          * set_bit and clear bit hooks normally require _irqsave/restore
1625          * but in this case, we are only testing for the DELALLOC
1626          * bit, which is only set or cleared with irqs on
1627          */
1628         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1629                 struct btrfs_root *root = BTRFS_I(inode)->root;
1630                 u64 len = state->end + 1 - state->start;
1631                 bool do_list = !btrfs_is_free_space_inode(inode);
1632
1633                 if (*bits & EXTENT_FIRST_DELALLOC) {
1634                         *bits &= ~EXTENT_FIRST_DELALLOC;
1635                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1636                         spin_lock(&BTRFS_I(inode)->lock);
1637                         BTRFS_I(inode)->outstanding_extents--;
1638                         spin_unlock(&BTRFS_I(inode)->lock);
1639                 }
1640
1641                 if (*bits & EXTENT_DO_ACCOUNTING)
1642                         btrfs_delalloc_release_metadata(inode, len);
1643
1644                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1645                     && do_list && !(state->state & EXTENT_NORESERVE))
1646                         btrfs_free_reserved_data_space(inode, len);
1647
1648                 __percpu_counter_add(&root->fs_info->delalloc_bytes, -len,
1649                                      root->fs_info->delalloc_batch);
1650                 spin_lock(&BTRFS_I(inode)->lock);
1651                 BTRFS_I(inode)->delalloc_bytes -= len;
1652                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1653                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1654                              &BTRFS_I(inode)->runtime_flags))
1655                         btrfs_del_delalloc_inode(root, inode);
1656                 spin_unlock(&BTRFS_I(inode)->lock);
1657         }
1658 }
1659
1660 /*
1661  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1662  * we don't create bios that span stripes or chunks
1663  */
1664 int btrfs_merge_bio_hook(int rw, struct page *page, unsigned long offset,
1665                          size_t size, struct bio *bio,
1666                          unsigned long bio_flags)
1667 {
1668         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1669         u64 logical = (u64)bio->bi_sector << 9;
1670         u64 length = 0;
1671         u64 map_length;
1672         int ret;
1673
1674         if (bio_flags & EXTENT_BIO_COMPRESSED)
1675                 return 0;
1676
1677         length = bio->bi_size;
1678         map_length = length;
1679         ret = btrfs_map_block(root->fs_info, rw, logical,
1680                               &map_length, NULL, 0);
1681         /* Will always return 0 with map_multi == NULL */
1682         BUG_ON(ret < 0);
1683         if (map_length < length + size)
1684                 return 1;
1685         return 0;
1686 }
1687
1688 /*
1689  * in order to insert checksums into the metadata in large chunks,
1690  * we wait until bio submission time.   All the pages in the bio are
1691  * checksummed and sums are attached onto the ordered extent record.
1692  *
1693  * At IO completion time the cums attached on the ordered extent record
1694  * are inserted into the btree
1695  */
1696 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1697                                     struct bio *bio, int mirror_num,
1698                                     unsigned long bio_flags,
1699                                     u64 bio_offset)
1700 {
1701         struct btrfs_root *root = BTRFS_I(inode)->root;
1702         int ret = 0;
1703
1704         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1705         BUG_ON(ret); /* -ENOMEM */
1706         return 0;
1707 }
1708
1709 /*
1710  * in order to insert checksums into the metadata in large chunks,
1711  * we wait until bio submission time.   All the pages in the bio are
1712  * checksummed and sums are attached onto the ordered extent record.
1713  *
1714  * At IO completion time the cums attached on the ordered extent record
1715  * are inserted into the btree
1716  */
1717 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1718                           int mirror_num, unsigned long bio_flags,
1719                           u64 bio_offset)
1720 {
1721         struct btrfs_root *root = BTRFS_I(inode)->root;
1722         int ret;
1723
1724         ret = btrfs_map_bio(root, rw, bio, mirror_num, 1);
1725         if (ret)
1726                 bio_endio(bio, ret);
1727         return ret;
1728 }
1729
1730 /*
1731  * extent_io.c submission hook. This does the right thing for csum calculation
1732  * on write, or reading the csums from the tree before a read
1733  */
1734 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1735                           int mirror_num, unsigned long bio_flags,
1736                           u64 bio_offset)
1737 {
1738         struct btrfs_root *root = BTRFS_I(inode)->root;
1739         int ret = 0;
1740         int skip_sum;
1741         int metadata = 0;
1742         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1743
1744         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1745
1746         if (btrfs_is_free_space_inode(inode))
1747                 metadata = 2;
1748
1749         if (!(rw & REQ_WRITE)) {
1750                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, metadata);
1751                 if (ret)
1752                         goto out;
1753
1754                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1755                         ret = btrfs_submit_compressed_read(inode, bio,
1756                                                            mirror_num,
1757                                                            bio_flags);
1758                         goto out;
1759                 } else if (!skip_sum) {
1760                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1761                         if (ret)
1762                                 goto out;
1763                 }
1764                 goto mapit;
1765         } else if (async && !skip_sum) {
1766                 /* csum items have already been cloned */
1767                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1768                         goto mapit;
1769                 /* we're doing a write, do the async checksumming */
1770                 ret = btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1771                                    inode, rw, bio, mirror_num,
1772                                    bio_flags, bio_offset,
1773                                    __btrfs_submit_bio_start,
1774                                    __btrfs_submit_bio_done);
1775                 goto out;
1776         } else if (!skip_sum) {
1777                 ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1778                 if (ret)
1779                         goto out;
1780         }
1781
1782 mapit:
1783         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
1784
1785 out:
1786         if (ret < 0)
1787                 bio_endio(bio, ret);
1788         return ret;
1789 }
1790
1791 /*
1792  * given a list of ordered sums record them in the inode.  This happens
1793  * at IO completion time based on sums calculated at bio submission time.
1794  */
1795 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1796                              struct inode *inode, u64 file_offset,
1797                              struct list_head *list)
1798 {
1799         struct btrfs_ordered_sum *sum;
1800
1801         list_for_each_entry(sum, list, list) {
1802                 trans->adding_csums = 1;
1803                 btrfs_csum_file_blocks(trans,
1804                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1805                 trans->adding_csums = 0;
1806         }
1807         return 0;
1808 }
1809
1810 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1811                               struct extent_state **cached_state)
1812 {
1813         WARN_ON((end & (PAGE_CACHE_SIZE - 1)) == 0);
1814         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1815                                    cached_state, GFP_NOFS);
1816 }
1817
1818 /* see btrfs_writepage_start_hook for details on why this is required */
1819 struct btrfs_writepage_fixup {
1820         struct page *page;
1821         struct btrfs_work work;
1822 };
1823
1824 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1825 {
1826         struct btrfs_writepage_fixup *fixup;
1827         struct btrfs_ordered_extent *ordered;
1828         struct extent_state *cached_state = NULL;
1829         struct page *page;
1830         struct inode *inode;
1831         u64 page_start;
1832         u64 page_end;
1833         int ret;
1834
1835         fixup = container_of(work, struct btrfs_writepage_fixup, work);
1836         page = fixup->page;
1837 again:
1838         lock_page(page);
1839         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
1840                 ClearPageChecked(page);
1841                 goto out_page;
1842         }
1843
1844         inode = page->mapping->host;
1845         page_start = page_offset(page);
1846         page_end = page_offset(page) + PAGE_CACHE_SIZE - 1;
1847
1848         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end, 0,
1849                          &cached_state);
1850
1851         /* already ordered? We're done */
1852         if (PagePrivate2(page))
1853                 goto out;
1854
1855         ordered = btrfs_lookup_ordered_extent(inode, page_start);
1856         if (ordered) {
1857                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
1858                                      page_end, &cached_state, GFP_NOFS);
1859                 unlock_page(page);
1860                 btrfs_start_ordered_extent(inode, ordered, 1);
1861                 btrfs_put_ordered_extent(ordered);
1862                 goto again;
1863         }
1864
1865         ret = btrfs_delalloc_reserve_space(inode, PAGE_CACHE_SIZE);
1866         if (ret) {
1867                 mapping_set_error(page->mapping, ret);
1868                 end_extent_writepage(page, ret, page_start, page_end);
1869                 ClearPageChecked(page);
1870                 goto out;
1871          }
1872
1873         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
1874         ClearPageChecked(page);
1875         set_page_dirty(page);
1876 out:
1877         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
1878                              &cached_state, GFP_NOFS);
1879 out_page:
1880         unlock_page(page);
1881         page_cache_release(page);
1882         kfree(fixup);
1883 }
1884
1885 /*
1886  * There are a few paths in the higher layers of the kernel that directly
1887  * set the page dirty bit without asking the filesystem if it is a
1888  * good idea.  This causes problems because we want to make sure COW
1889  * properly happens and the data=ordered rules are followed.
1890  *
1891  * In our case any range that doesn't have the ORDERED bit set
1892  * hasn't been properly setup for IO.  We kick off an async process
1893  * to fix it up.  The async helper will wait for ordered extents, set
1894  * the delalloc bit and make it safe to write the page.
1895  */
1896 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
1897 {
1898         struct inode *inode = page->mapping->host;
1899         struct btrfs_writepage_fixup *fixup;
1900         struct btrfs_root *root = BTRFS_I(inode)->root;
1901
1902         /* this page is properly in the ordered list */
1903         if (TestClearPagePrivate2(page))
1904                 return 0;
1905
1906         if (PageChecked(page))
1907                 return -EAGAIN;
1908
1909         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
1910         if (!fixup)
1911                 return -EAGAIN;
1912
1913         SetPageChecked(page);
1914         page_cache_get(page);
1915         fixup->work.func = btrfs_writepage_fixup_worker;
1916         fixup->page = page;
1917         btrfs_queue_worker(&root->fs_info->fixup_workers, &fixup->work);
1918         return -EBUSY;
1919 }
1920
1921 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
1922                                        struct inode *inode, u64 file_pos,
1923                                        u64 disk_bytenr, u64 disk_num_bytes,
1924                                        u64 num_bytes, u64 ram_bytes,
1925                                        u8 compression, u8 encryption,
1926                                        u16 other_encoding, int extent_type)
1927 {
1928         struct btrfs_root *root = BTRFS_I(inode)->root;
1929         struct btrfs_file_extent_item *fi;
1930         struct btrfs_path *path;
1931         struct extent_buffer *leaf;
1932         struct btrfs_key ins;
1933         int ret;
1934
1935         path = btrfs_alloc_path();
1936         if (!path)
1937                 return -ENOMEM;
1938
1939         path->leave_spinning = 1;
1940
1941         /*
1942          * we may be replacing one extent in the tree with another.
1943          * The new extent is pinned in the extent map, and we don't want
1944          * to drop it from the cache until it is completely in the btree.
1945          *
1946          * So, tell btrfs_drop_extents to leave this extent in the cache.
1947          * the caller is expected to unpin it and allow it to be merged
1948          * with the others.
1949          */
1950         ret = btrfs_drop_extents(trans, root, inode, file_pos,
1951                                  file_pos + num_bytes, 0);
1952         if (ret)
1953                 goto out;
1954
1955         ins.objectid = btrfs_ino(inode);
1956         ins.offset = file_pos;
1957         ins.type = BTRFS_EXTENT_DATA_KEY;
1958         ret = btrfs_insert_empty_item(trans, root, path, &ins, sizeof(*fi));
1959         if (ret)
1960                 goto out;
1961         leaf = path->nodes[0];
1962         fi = btrfs_item_ptr(leaf, path->slots[0],
1963                             struct btrfs_file_extent_item);
1964         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1965         btrfs_set_file_extent_type(leaf, fi, extent_type);
1966         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
1967         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
1968         btrfs_set_file_extent_offset(leaf, fi, 0);
1969         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
1970         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
1971         btrfs_set_file_extent_compression(leaf, fi, compression);
1972         btrfs_set_file_extent_encryption(leaf, fi, encryption);
1973         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
1974
1975         btrfs_mark_buffer_dirty(leaf);
1976         btrfs_release_path(path);
1977
1978         inode_add_bytes(inode, num_bytes);
1979
1980         ins.objectid = disk_bytenr;
1981         ins.offset = disk_num_bytes;
1982         ins.type = BTRFS_EXTENT_ITEM_KEY;
1983         ret = btrfs_alloc_reserved_file_extent(trans, root,
1984                                         root->root_key.objectid,
1985                                         btrfs_ino(inode), file_pos, &ins);
1986 out:
1987         btrfs_free_path(path);
1988
1989         return ret;
1990 }
1991
1992 /* snapshot-aware defrag */
1993 struct sa_defrag_extent_backref {
1994         struct rb_node node;
1995         struct old_sa_defrag_extent *old;
1996         u64 root_id;
1997         u64 inum;
1998         u64 file_pos;
1999         u64 extent_offset;
2000         u64 num_bytes;
2001         u64 generation;
2002 };
2003
2004 struct old_sa_defrag_extent {
2005         struct list_head list;
2006         struct new_sa_defrag_extent *new;
2007
2008         u64 extent_offset;
2009         u64 bytenr;
2010         u64 offset;
2011         u64 len;
2012         int count;
2013 };
2014
2015 struct new_sa_defrag_extent {
2016         struct rb_root root;
2017         struct list_head head;
2018         struct btrfs_path *path;
2019         struct inode *inode;
2020         u64 file_pos;
2021         u64 len;
2022         u64 bytenr;
2023         u64 disk_len;
2024         u8 compress_type;
2025 };
2026
2027 static int backref_comp(struct sa_defrag_extent_backref *b1,
2028                         struct sa_defrag_extent_backref *b2)
2029 {
2030         if (b1->root_id < b2->root_id)
2031                 return -1;
2032         else if (b1->root_id > b2->root_id)
2033                 return 1;
2034
2035         if (b1->inum < b2->inum)
2036                 return -1;
2037         else if (b1->inum > b2->inum)
2038                 return 1;
2039
2040         if (b1->file_pos < b2->file_pos)
2041                 return -1;
2042         else if (b1->file_pos > b2->file_pos)
2043                 return 1;
2044
2045         /*
2046          * [------------------------------] ===> (a range of space)
2047          *     |<--->|   |<---->| =============> (fs/file tree A)
2048          * |<---------------------------->| ===> (fs/file tree B)
2049          *
2050          * A range of space can refer to two file extents in one tree while
2051          * refer to only one file extent in another tree.
2052          *
2053          * So we may process a disk offset more than one time(two extents in A)
2054          * and locate at the same extent(one extent in B), then insert two same
2055          * backrefs(both refer to the extent in B).
2056          */
2057         return 0;
2058 }
2059
2060 static void backref_insert(struct rb_root *root,
2061                            struct sa_defrag_extent_backref *backref)
2062 {
2063         struct rb_node **p = &root->rb_node;
2064         struct rb_node *parent = NULL;
2065         struct sa_defrag_extent_backref *entry;
2066         int ret;
2067
2068         while (*p) {
2069                 parent = *p;
2070                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2071
2072                 ret = backref_comp(backref, entry);
2073                 if (ret < 0)
2074                         p = &(*p)->rb_left;
2075                 else
2076                         p = &(*p)->rb_right;
2077         }
2078
2079         rb_link_node(&backref->node, parent, p);
2080         rb_insert_color(&backref->node, root);
2081 }
2082
2083 /*
2084  * Note the backref might has changed, and in this case we just return 0.
2085  */
2086 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2087                                        void *ctx)
2088 {
2089         struct btrfs_file_extent_item *extent;
2090         struct btrfs_fs_info *fs_info;
2091         struct old_sa_defrag_extent *old = ctx;
2092         struct new_sa_defrag_extent *new = old->new;
2093         struct btrfs_path *path = new->path;
2094         struct btrfs_key key;
2095         struct btrfs_root *root;
2096         struct sa_defrag_extent_backref *backref;
2097         struct extent_buffer *leaf;
2098         struct inode *inode = new->inode;
2099         int slot;
2100         int ret;
2101         u64 extent_offset;
2102         u64 num_bytes;
2103
2104         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2105             inum == btrfs_ino(inode))
2106                 return 0;
2107
2108         key.objectid = root_id;
2109         key.type = BTRFS_ROOT_ITEM_KEY;
2110         key.offset = (u64)-1;
2111
2112         fs_info = BTRFS_I(inode)->root->fs_info;
2113         root = btrfs_read_fs_root_no_name(fs_info, &key);
2114         if (IS_ERR(root)) {
2115                 if (PTR_ERR(root) == -ENOENT)
2116                         return 0;
2117                 WARN_ON(1);
2118                 pr_debug("inum=%llu, offset=%llu, root_id=%llu\n",
2119                          inum, offset, root_id);
2120                 return PTR_ERR(root);
2121         }
2122
2123         key.objectid = inum;
2124         key.type = BTRFS_EXTENT_DATA_KEY;
2125         if (offset > (u64)-1 << 32)
2126                 key.offset = 0;
2127         else
2128                 key.offset = offset;
2129
2130         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2131         if (ret < 0) {
2132                 WARN_ON(1);
2133                 return ret;
2134         }
2135
2136         while (1) {
2137                 cond_resched();
2138
2139                 leaf = path->nodes[0];
2140                 slot = path->slots[0];
2141
2142                 if (slot >= btrfs_header_nritems(leaf)) {
2143                         ret = btrfs_next_leaf(root, path);
2144                         if (ret < 0) {
2145                                 goto out;
2146                         } else if (ret > 0) {
2147                                 ret = 0;
2148                                 goto out;
2149                         }
2150                         continue;
2151                 }
2152
2153                 path->slots[0]++;
2154
2155                 btrfs_item_key_to_cpu(leaf, &key, slot);
2156
2157                 if (key.objectid > inum)
2158                         goto out;
2159
2160                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2161                         continue;
2162
2163                 extent = btrfs_item_ptr(leaf, slot,
2164                                         struct btrfs_file_extent_item);
2165
2166                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2167                         continue;
2168
2169                 /*
2170                  * 'offset' refers to the exact key.offset,
2171                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2172                  * (key.offset - extent_offset).
2173                  */
2174                 if (key.offset != offset)
2175                         continue;
2176
2177                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2178                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2179
2180                 if (extent_offset >= old->extent_offset + old->offset +
2181                     old->len || extent_offset + num_bytes <=
2182                     old->extent_offset + old->offset)
2183                         continue;
2184
2185                 ret = 0;
2186                 break;
2187         }
2188
2189         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2190         if (!backref) {
2191                 ret = -ENOENT;
2192                 goto out;
2193         }
2194
2195         backref->root_id = root_id;
2196         backref->inum = inum;
2197         backref->file_pos = offset;
2198         backref->num_bytes = num_bytes;
2199         backref->extent_offset = extent_offset;
2200         backref->generation = btrfs_file_extent_generation(leaf, extent);
2201         backref->old = old;
2202         backref_insert(&new->root, backref);
2203         old->count++;
2204 out:
2205         btrfs_release_path(path);
2206         WARN_ON(ret);
2207         return ret;
2208 }
2209
2210 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2211                                    struct new_sa_defrag_extent *new)
2212 {
2213         struct btrfs_fs_info *fs_info = BTRFS_I(new->inode)->root->fs_info;
2214         struct old_sa_defrag_extent *old, *tmp;
2215         int ret;
2216
2217         new->path = path;
2218
2219         list_for_each_entry_safe(old, tmp, &new->head, list) {
2220                 ret = iterate_inodes_from_logical(old->bytenr +
2221                                                   old->extent_offset, fs_info,
2222                                                   path, record_one_backref,
2223                                                   old);
2224                 BUG_ON(ret < 0 && ret != -ENOENT);
2225
2226                 /* no backref to be processed for this extent */
2227                 if (!old->count) {
2228                         list_del(&old->list);
2229                         kfree(old);
2230                 }
2231         }
2232
2233         if (list_empty(&new->head))
2234                 return false;
2235
2236         return true;
2237 }
2238
2239 static int relink_is_mergable(struct extent_buffer *leaf,
2240                               struct btrfs_file_extent_item *fi,
2241                               u64 disk_bytenr)
2242 {
2243         if (btrfs_file_extent_disk_bytenr(leaf, fi) != disk_bytenr)
2244                 return 0;
2245
2246         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2247                 return 0;
2248
2249         if (btrfs_file_extent_compression(leaf, fi) ||
2250             btrfs_file_extent_encryption(leaf, fi) ||
2251             btrfs_file_extent_other_encoding(leaf, fi))
2252                 return 0;
2253
2254         return 1;
2255 }
2256
2257 /*
2258  * Note the backref might has changed, and in this case we just return 0.
2259  */
2260 static noinline int relink_extent_backref(struct btrfs_path *path,
2261                                  struct sa_defrag_extent_backref *prev,
2262                                  struct sa_defrag_extent_backref *backref)
2263 {
2264         struct btrfs_file_extent_item *extent;
2265         struct btrfs_file_extent_item *item;
2266         struct btrfs_ordered_extent *ordered;
2267         struct btrfs_trans_handle *trans;
2268         struct btrfs_fs_info *fs_info;
2269         struct btrfs_root *root;
2270         struct btrfs_key key;
2271         struct extent_buffer *leaf;
2272         struct old_sa_defrag_extent *old = backref->old;
2273         struct new_sa_defrag_extent *new = old->new;
2274         struct inode *src_inode = new->inode;
2275         struct inode *inode;
2276         struct extent_state *cached = NULL;
2277         int ret = 0;
2278         u64 start;
2279         u64 len;
2280         u64 lock_start;
2281         u64 lock_end;
2282         bool merge = false;
2283         int index;
2284
2285         if (prev && prev->root_id == backref->root_id &&
2286             prev->inum == backref->inum &&
2287             prev->file_pos + prev->num_bytes == backref->file_pos)
2288                 merge = true;
2289
2290         /* step 1: get root */
2291         key.objectid = backref->root_id;
2292         key.type = BTRFS_ROOT_ITEM_KEY;
2293         key.offset = (u64)-1;
2294
2295         fs_info = BTRFS_I(src_inode)->root->fs_info;
2296         index = srcu_read_lock(&fs_info->subvol_srcu);
2297
2298         root = btrfs_read_fs_root_no_name(fs_info, &key);
2299         if (IS_ERR(root)) {
2300                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2301                 if (PTR_ERR(root) == -ENOENT)
2302                         return 0;
2303                 return PTR_ERR(root);
2304         }
2305
2306         /* step 2: get inode */
2307         key.objectid = backref->inum;
2308         key.type = BTRFS_INODE_ITEM_KEY;
2309         key.offset = 0;
2310
2311         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2312         if (IS_ERR(inode)) {
2313                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2314                 return 0;
2315         }
2316
2317         srcu_read_unlock(&fs_info->subvol_srcu, index);
2318
2319         /* step 3: relink backref */
2320         lock_start = backref->file_pos;
2321         lock_end = backref->file_pos + backref->num_bytes - 1;
2322         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2323                          0, &cached);
2324
2325         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2326         if (ordered) {
2327                 btrfs_put_ordered_extent(ordered);
2328                 goto out_unlock;
2329         }
2330
2331         trans = btrfs_join_transaction(root);
2332         if (IS_ERR(trans)) {
2333                 ret = PTR_ERR(trans);
2334                 goto out_unlock;
2335         }
2336
2337         key.objectid = backref->inum;
2338         key.type = BTRFS_EXTENT_DATA_KEY;
2339         key.offset = backref->file_pos;
2340
2341         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2342         if (ret < 0) {
2343                 goto out_free_path;
2344         } else if (ret > 0) {
2345                 ret = 0;
2346                 goto out_free_path;
2347         }
2348
2349         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2350                                 struct btrfs_file_extent_item);
2351
2352         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2353             backref->generation)
2354                 goto out_free_path;
2355
2356         btrfs_release_path(path);
2357
2358         start = backref->file_pos;
2359         if (backref->extent_offset < old->extent_offset + old->offset)
2360                 start += old->extent_offset + old->offset -
2361                          backref->extent_offset;
2362
2363         len = min(backref->extent_offset + backref->num_bytes,
2364                   old->extent_offset + old->offset + old->len);
2365         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2366
2367         ret = btrfs_drop_extents(trans, root, inode, start,
2368                                  start + len, 1);
2369         if (ret)
2370                 goto out_free_path;
2371 again:
2372         key.objectid = btrfs_ino(inode);
2373         key.type = BTRFS_EXTENT_DATA_KEY;
2374         key.offset = start;
2375
2376         path->leave_spinning = 1;
2377         if (merge) {
2378                 struct btrfs_file_extent_item *fi;
2379                 u64 extent_len;
2380                 struct btrfs_key found_key;
2381
2382                 ret = btrfs_search_slot(trans, root, &key, path, 1, 1);
2383                 if (ret < 0)
2384                         goto out_free_path;
2385
2386                 path->slots[0]--;
2387                 leaf = path->nodes[0];
2388                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2389
2390                 fi = btrfs_item_ptr(leaf, path->slots[0],
2391                                     struct btrfs_file_extent_item);
2392                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2393
2394                 if (relink_is_mergable(leaf, fi, new->bytenr) &&
2395                     extent_len + found_key.offset == start) {
2396                         btrfs_set_file_extent_num_bytes(leaf, fi,
2397                                                         extent_len + len);
2398                         btrfs_mark_buffer_dirty(leaf);
2399                         inode_add_bytes(inode, len);
2400
2401                         ret = 1;
2402                         goto out_free_path;
2403                 } else {
2404                         merge = false;
2405                         btrfs_release_path(path);
2406                         goto again;
2407                 }
2408         }
2409
2410         ret = btrfs_insert_empty_item(trans, root, path, &key,
2411                                         sizeof(*extent));
2412         if (ret) {
2413                 btrfs_abort_transaction(trans, root, ret);
2414                 goto out_free_path;
2415         }
2416
2417         leaf = path->nodes[0];
2418         item = btrfs_item_ptr(leaf, path->slots[0],
2419                                 struct btrfs_file_extent_item);
2420         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2421         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2422         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2423         btrfs_set_file_extent_num_bytes(leaf, item, len);
2424         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2425         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2426         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2427         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2428         btrfs_set_file_extent_encryption(leaf, item, 0);
2429         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2430
2431         btrfs_mark_buffer_dirty(leaf);
2432         inode_add_bytes(inode, len);
2433         btrfs_release_path(path);
2434
2435         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2436                         new->disk_len, 0,
2437                         backref->root_id, backref->inum,
2438                         new->file_pos, 0);      /* start - extent_offset */
2439         if (ret) {
2440                 btrfs_abort_transaction(trans, root, ret);
2441                 goto out_free_path;
2442         }
2443
2444         ret = 1;
2445 out_free_path:
2446         btrfs_release_path(path);
2447         path->leave_spinning = 0;
2448         btrfs_end_transaction(trans, root);
2449 out_unlock:
2450         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2451                              &cached, GFP_NOFS);
2452         iput(inode);
2453         return ret;
2454 }
2455
2456 static void relink_file_extents(struct new_sa_defrag_extent *new)
2457 {
2458         struct btrfs_path *path;
2459         struct old_sa_defrag_extent *old, *tmp;
2460         struct sa_defrag_extent_backref *backref;
2461         struct sa_defrag_extent_backref *prev = NULL;
2462         struct inode *inode;
2463         struct btrfs_root *root;
2464         struct rb_node *node;
2465         int ret;
2466
2467         inode = new->inode;
2468         root = BTRFS_I(inode)->root;
2469
2470         path = btrfs_alloc_path();
2471         if (!path)
2472                 return;
2473
2474         if (!record_extent_backrefs(path, new)) {
2475                 btrfs_free_path(path);
2476                 goto out;
2477         }
2478         btrfs_release_path(path);
2479
2480         while (1) {
2481                 node = rb_first(&new->root);
2482                 if (!node)
2483                         break;
2484                 rb_erase(node, &new->root);
2485
2486                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2487
2488                 ret = relink_extent_backref(path, prev, backref);
2489                 WARN_ON(ret < 0);
2490
2491                 kfree(prev);
2492
2493                 if (ret == 1)
2494                         prev = backref;
2495                 else
2496                         prev = NULL;
2497                 cond_resched();
2498         }
2499         kfree(prev);
2500
2501         btrfs_free_path(path);
2502
2503         list_for_each_entry_safe(old, tmp, &new->head, list) {
2504                 list_del(&old->list);
2505                 kfree(old);
2506         }
2507 out:
2508         atomic_dec(&root->fs_info->defrag_running);
2509         wake_up(&root->fs_info->transaction_wait);
2510
2511         kfree(new);
2512 }
2513
2514 static struct new_sa_defrag_extent *
2515 record_old_file_extents(struct inode *inode,
2516                         struct btrfs_ordered_extent *ordered)
2517 {
2518         struct btrfs_root *root = BTRFS_I(inode)->root;
2519         struct btrfs_path *path;
2520         struct btrfs_key key;
2521         struct old_sa_defrag_extent *old, *tmp;
2522         struct new_sa_defrag_extent *new;
2523         int ret;
2524
2525         new = kmalloc(sizeof(*new), GFP_NOFS);
2526         if (!new)
2527                 return NULL;
2528
2529         new->inode = inode;
2530         new->file_pos = ordered->file_offset;
2531         new->len = ordered->len;
2532         new->bytenr = ordered->start;
2533         new->disk_len = ordered->disk_len;
2534         new->compress_type = ordered->compress_type;
2535         new->root = RB_ROOT;
2536         INIT_LIST_HEAD(&new->head);
2537
2538         path = btrfs_alloc_path();
2539         if (!path)
2540                 goto out_kfree;
2541
2542         key.objectid = btrfs_ino(inode);
2543         key.type = BTRFS_EXTENT_DATA_KEY;
2544         key.offset = new->file_pos;
2545
2546         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2547         if (ret < 0)
2548                 goto out_free_path;
2549         if (ret > 0 && path->slots[0] > 0)
2550                 path->slots[0]--;
2551
2552         /* find out all the old extents for the file range */
2553         while (1) {
2554                 struct btrfs_file_extent_item *extent;
2555                 struct extent_buffer *l;
2556                 int slot;
2557                 u64 num_bytes;
2558                 u64 offset;
2559                 u64 end;
2560                 u64 disk_bytenr;
2561                 u64 extent_offset;
2562
2563                 l = path->nodes[0];
2564                 slot = path->slots[0];
2565
2566                 if (slot >= btrfs_header_nritems(l)) {
2567                         ret = btrfs_next_leaf(root, path);
2568                         if (ret < 0)
2569                                 goto out_free_list;
2570                         else if (ret > 0)
2571                                 break;
2572                         continue;
2573                 }
2574
2575                 btrfs_item_key_to_cpu(l, &key, slot);
2576
2577                 if (key.objectid != btrfs_ino(inode))
2578                         break;
2579                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2580                         break;
2581                 if (key.offset >= new->file_pos + new->len)
2582                         break;
2583
2584                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2585
2586                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2587                 if (key.offset + num_bytes < new->file_pos)
2588                         goto next;
2589
2590                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2591                 if (!disk_bytenr)
2592                         goto next;
2593
2594                 extent_offset = btrfs_file_extent_offset(l, extent);
2595
2596                 old = kmalloc(sizeof(*old), GFP_NOFS);
2597                 if (!old)
2598                         goto out_free_list;
2599
2600                 offset = max(new->file_pos, key.offset);
2601                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2602
2603                 old->bytenr = disk_bytenr;
2604                 old->extent_offset = extent_offset;
2605                 old->offset = offset - key.offset;
2606                 old->len = end - offset;
2607                 old->new = new;
2608                 old->count = 0;
2609                 list_add_tail(&old->list, &new->head);
2610 next:
2611                 path->slots[0]++;
2612                 cond_resched();
2613         }
2614
2615         btrfs_free_path(path);
2616         atomic_inc(&root->fs_info->defrag_running);
2617
2618         return new;
2619
2620 out_free_list:
2621         list_for_each_entry_safe(old, tmp, &new->head, list) {
2622                 list_del(&old->list);
2623                 kfree(old);
2624         }
2625 out_free_path:
2626         btrfs_free_path(path);
2627 out_kfree:
2628         kfree(new);
2629         return NULL;
2630 }
2631
2632 /*
2633  * helper function for btrfs_finish_ordered_io, this
2634  * just reads in some of the csum leaves to prime them into ram
2635  * before we start the transaction.  It limits the amount of btree
2636  * reads required while inside the transaction.
2637  */
2638 /* as ordered data IO finishes, this gets called so we can finish
2639  * an ordered extent if the range of bytes in the file it covers are
2640  * fully written.
2641  */
2642 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2643 {
2644         struct inode *inode = ordered_extent->inode;
2645         struct btrfs_root *root = BTRFS_I(inode)->root;
2646         struct btrfs_trans_handle *trans = NULL;
2647         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2648         struct extent_state *cached_state = NULL;
2649         struct new_sa_defrag_extent *new = NULL;
2650         int compress_type = 0;
2651         int ret;
2652         bool nolock;
2653
2654         nolock = btrfs_is_free_space_inode(inode);
2655
2656         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2657                 ret = -EIO;
2658                 goto out;
2659         }
2660
2661         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2662                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2663                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2664                 if (nolock)
2665                         trans = btrfs_join_transaction_nolock(root);
2666                 else
2667                         trans = btrfs_join_transaction(root);
2668                 if (IS_ERR(trans)) {
2669                         ret = PTR_ERR(trans);
2670                         trans = NULL;
2671                         goto out;
2672                 }
2673                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2674                 ret = btrfs_update_inode_fallback(trans, root, inode);
2675                 if (ret) /* -ENOMEM or corruption */
2676                         btrfs_abort_transaction(trans, root, ret);
2677                 goto out;
2678         }
2679
2680         lock_extent_bits(io_tree, ordered_extent->file_offset,
2681                          ordered_extent->file_offset + ordered_extent->len - 1,
2682                          0, &cached_state);
2683
2684         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2685                         ordered_extent->file_offset + ordered_extent->len - 1,
2686                         EXTENT_DEFRAG, 1, cached_state);
2687         if (ret) {
2688                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2689                 if (last_snapshot >= BTRFS_I(inode)->generation)
2690                         /* the inode is shared */
2691                         new = record_old_file_extents(inode, ordered_extent);
2692
2693                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2694                         ordered_extent->file_offset + ordered_extent->len - 1,
2695                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2696         }
2697
2698         if (nolock)
2699                 trans = btrfs_join_transaction_nolock(root);
2700         else
2701                 trans = btrfs_join_transaction(root);
2702         if (IS_ERR(trans)) {
2703                 ret = PTR_ERR(trans);
2704                 trans = NULL;
2705                 goto out_unlock;
2706         }
2707         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2708
2709         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2710                 compress_type = ordered_extent->compress_type;
2711         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2712                 BUG_ON(compress_type);
2713                 ret = btrfs_mark_extent_written(trans, inode,
2714                                                 ordered_extent->file_offset,
2715                                                 ordered_extent->file_offset +
2716                                                 ordered_extent->len);
2717         } else {
2718                 BUG_ON(root == root->fs_info->tree_root);
2719                 ret = insert_reserved_file_extent(trans, inode,
2720                                                 ordered_extent->file_offset,
2721                                                 ordered_extent->start,
2722                                                 ordered_extent->disk_len,
2723                                                 ordered_extent->len,
2724                                                 ordered_extent->len,
2725                                                 compress_type, 0, 0,
2726                                                 BTRFS_FILE_EXTENT_REG);
2727         }
2728         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2729                            ordered_extent->file_offset, ordered_extent->len,
2730                            trans->transid);
2731         if (ret < 0) {
2732                 btrfs_abort_transaction(trans, root, ret);
2733                 goto out_unlock;
2734         }
2735
2736         add_pending_csums(trans, inode, ordered_extent->file_offset,
2737                           &ordered_extent->list);
2738
2739         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2740         ret = btrfs_update_inode_fallback(trans, root, inode);
2741         if (ret) { /* -ENOMEM or corruption */
2742                 btrfs_abort_transaction(trans, root, ret);
2743                 goto out_unlock;
2744         }
2745         ret = 0;
2746 out_unlock:
2747         unlock_extent_cached(io_tree, ordered_extent->file_offset,
2748                              ordered_extent->file_offset +
2749                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
2750 out:
2751         if (root != root->fs_info->tree_root)
2752                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
2753         if (trans)
2754                 btrfs_end_transaction(trans, root);
2755
2756         if (ret) {
2757                 clear_extent_uptodate(io_tree, ordered_extent->file_offset,
2758                                       ordered_extent->file_offset +
2759                                       ordered_extent->len - 1, NULL, GFP_NOFS);
2760
2761                 /*
2762                  * If the ordered extent had an IOERR or something else went
2763                  * wrong we need to return the space for this ordered extent
2764                  * back to the allocator.
2765                  */
2766                 if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2767                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
2768                         btrfs_free_reserved_extent(root, ordered_extent->start,
2769                                                    ordered_extent->disk_len);
2770         }
2771
2772
2773         /*
2774          * This needs to be done to make sure anybody waiting knows we are done
2775          * updating everything for this ordered extent.
2776          */
2777         btrfs_remove_ordered_extent(inode, ordered_extent);
2778
2779         /* for snapshot-aware defrag */
2780         if (new)
2781                 relink_file_extents(new);
2782
2783         /* once for us */
2784         btrfs_put_ordered_extent(ordered_extent);
2785         /* once for the tree */
2786         btrfs_put_ordered_extent(ordered_extent);
2787
2788         return ret;
2789 }
2790
2791 static void finish_ordered_fn(struct btrfs_work *work)
2792 {
2793         struct btrfs_ordered_extent *ordered_extent;
2794         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
2795         btrfs_finish_ordered_io(ordered_extent);
2796 }
2797
2798 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
2799                                 struct extent_state *state, int uptodate)
2800 {
2801         struct inode *inode = page->mapping->host;
2802         struct btrfs_root *root = BTRFS_I(inode)->root;
2803         struct btrfs_ordered_extent *ordered_extent = NULL;
2804         struct btrfs_workers *workers;
2805
2806         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
2807
2808         ClearPagePrivate2(page);
2809         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
2810                                             end - start + 1, uptodate))
2811                 return 0;
2812
2813         ordered_extent->work.func = finish_ordered_fn;
2814         ordered_extent->work.flags = 0;
2815
2816         if (btrfs_is_free_space_inode(inode))
2817                 workers = &root->fs_info->endio_freespace_worker;
2818         else
2819                 workers = &root->fs_info->endio_write_workers;
2820         btrfs_queue_worker(workers, &ordered_extent->work);
2821
2822         return 0;
2823 }
2824
2825 /*
2826  * when reads are done, we need to check csums to verify the data is correct
2827  * if there's a match, we allow the bio to finish.  If not, the code in
2828  * extent_io.c will try to find good copies for us.
2829  */
2830 static int btrfs_readpage_end_io_hook(struct page *page, u64 start, u64 end,
2831                                struct extent_state *state, int mirror)
2832 {
2833         size_t offset = start - page_offset(page);
2834         struct inode *inode = page->mapping->host;
2835         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2836         char *kaddr;
2837         u64 private = ~(u32)0;
2838         int ret;
2839         struct btrfs_root *root = BTRFS_I(inode)->root;
2840         u32 csum = ~(u32)0;
2841         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
2842                                       DEFAULT_RATELIMIT_BURST);
2843
2844         if (PageChecked(page)) {
2845                 ClearPageChecked(page);
2846                 goto good;
2847         }
2848
2849         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
2850                 goto good;
2851
2852         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
2853             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
2854                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM,
2855                                   GFP_NOFS);
2856                 return 0;
2857         }
2858
2859         if (state && state->start == start) {
2860                 private = state->private;
2861                 ret = 0;
2862         } else {
2863                 ret = get_state_private(io_tree, start, &private);
2864         }
2865         kaddr = kmap_atomic(page);
2866         if (ret)
2867                 goto zeroit;
2868
2869         csum = btrfs_csum_data(kaddr + offset, csum,  end - start + 1);
2870         btrfs_csum_final(csum, (char *)&csum);
2871         if (csum != private)
2872                 goto zeroit;
2873
2874         kunmap_atomic(kaddr);
2875 good:
2876         return 0;
2877
2878 zeroit:
2879         if (__ratelimit(&_rs))
2880                 btrfs_info(root->fs_info, "csum failed ino %llu off %llu csum %u private %llu",
2881                         (unsigned long long)btrfs_ino(page->mapping->host),
2882                         (unsigned long long)start, csum,
2883                         (unsigned long long)private);
2884         memset(kaddr + offset, 1, end - start + 1);
2885         flush_dcache_page(page);
2886         kunmap_atomic(kaddr);
2887         if (private == 0)
2888                 return 0;
2889         return -EIO;
2890 }
2891
2892 struct delayed_iput {
2893         struct list_head list;
2894         struct inode *inode;
2895 };
2896
2897 /* JDM: If this is fs-wide, why can't we add a pointer to
2898  * btrfs_inode instead and avoid the allocation? */
2899 void btrfs_add_delayed_iput(struct inode *inode)
2900 {
2901         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2902         struct delayed_iput *delayed;
2903
2904         if (atomic_add_unless(&inode->i_count, -1, 1))
2905                 return;
2906
2907         delayed = kmalloc(sizeof(*delayed), GFP_NOFS | __GFP_NOFAIL);
2908         delayed->inode = inode;
2909
2910         spin_lock(&fs_info->delayed_iput_lock);
2911         list_add_tail(&delayed->list, &fs_info->delayed_iputs);
2912         spin_unlock(&fs_info->delayed_iput_lock);
2913 }
2914
2915 void btrfs_run_delayed_iputs(struct btrfs_root *root)
2916 {
2917         LIST_HEAD(list);
2918         struct btrfs_fs_info *fs_info = root->fs_info;
2919         struct delayed_iput *delayed;
2920         int empty;
2921
2922         spin_lock(&fs_info->delayed_iput_lock);
2923         empty = list_empty(&fs_info->delayed_iputs);
2924         spin_unlock(&fs_info->delayed_iput_lock);
2925         if (empty)
2926                 return;
2927
2928         spin_lock(&fs_info->delayed_iput_lock);
2929         list_splice_init(&fs_info->delayed_iputs, &list);
2930         spin_unlock(&fs_info->delayed_iput_lock);
2931
2932         while (!list_empty(&list)) {
2933                 delayed = list_entry(list.next, struct delayed_iput, list);
2934                 list_del(&delayed->list);
2935                 iput(delayed->inode);
2936                 kfree(delayed);
2937         }
2938 }
2939
2940 /*
2941  * This is called in transaction commit time. If there are no orphan
2942  * files in the subvolume, it removes orphan item and frees block_rsv
2943  * structure.
2944  */
2945 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
2946                               struct btrfs_root *root)
2947 {
2948         struct btrfs_block_rsv *block_rsv;
2949         int ret;
2950
2951         if (atomic_read(&root->orphan_inodes) ||
2952             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
2953                 return;
2954
2955         spin_lock(&root->orphan_lock);
2956         if (atomic_read(&root->orphan_inodes)) {
2957                 spin_unlock(&root->orphan_lock);
2958                 return;
2959         }
2960
2961         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
2962                 spin_unlock(&root->orphan_lock);
2963                 return;
2964         }
2965
2966         block_rsv = root->orphan_block_rsv;
2967         root->orphan_block_rsv = NULL;
2968         spin_unlock(&root->orphan_lock);
2969
2970         if (root->orphan_item_inserted &&
2971             btrfs_root_refs(&root->root_item) > 0) {
2972                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
2973                                             root->root_key.objectid);
2974                 BUG_ON(ret);
2975                 root->orphan_item_inserted = 0;
2976         }
2977
2978         if (block_rsv) {
2979                 WARN_ON(block_rsv->size > 0);
2980                 btrfs_free_block_rsv(root, block_rsv);
2981         }
2982 }
2983
2984 /*
2985  * This creates an orphan entry for the given inode in case something goes
2986  * wrong in the middle of an unlink/truncate.
2987  *
2988  * NOTE: caller of this function should reserve 5 units of metadata for
2989  *       this function.
2990  */
2991 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
2992 {
2993         struct btrfs_root *root = BTRFS_I(inode)->root;
2994         struct btrfs_block_rsv *block_rsv = NULL;
2995         int reserve = 0;
2996         int insert = 0;
2997         int ret;
2998
2999         if (!root->orphan_block_rsv) {
3000                 block_rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
3001                 if (!block_rsv)
3002                         return -ENOMEM;
3003         }
3004
3005         spin_lock(&root->orphan_lock);
3006         if (!root->orphan_block_rsv) {
3007                 root->orphan_block_rsv = block_rsv;
3008         } else if (block_rsv) {
3009                 btrfs_free_block_rsv(root, block_rsv);
3010                 block_rsv = NULL;
3011         }
3012
3013         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3014                               &BTRFS_I(inode)->runtime_flags)) {
3015 #if 0
3016                 /*
3017                  * For proper ENOSPC handling, we should do orphan
3018                  * cleanup when mounting. But this introduces backward
3019                  * compatibility issue.
3020                  */
3021                 if (!xchg(&root->orphan_item_inserted, 1))
3022                         insert = 2;
3023                 else
3024                         insert = 1;
3025 #endif
3026                 insert = 1;
3027                 atomic_inc(&root->orphan_inodes);
3028         }
3029
3030         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3031                               &BTRFS_I(inode)->runtime_flags))
3032                 reserve = 1;
3033         spin_unlock(&root->orphan_lock);
3034
3035         /* grab metadata reservation from transaction handle */
3036         if (reserve) {
3037                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3038                 BUG_ON(ret); /* -ENOSPC in reservation; Logic error? JDM */
3039         }
3040
3041         /* insert an orphan item to track this unlinked/truncated file */
3042         if (insert >= 1) {
3043                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3044                 if (ret && ret != -EEXIST) {
3045                         clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3046                                   &BTRFS_I(inode)->runtime_flags);
3047                         btrfs_abort_transaction(trans, root, ret);
3048                         return ret;
3049                 }
3050                 ret = 0;
3051         }
3052
3053         /* insert an orphan item to track subvolume contains orphan files */
3054         if (insert >= 2) {
3055                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
3056                                                root->root_key.objectid);
3057                 if (ret && ret != -EEXIST) {
3058                         btrfs_abort_transaction(trans, root, ret);
3059                         return ret;
3060                 }
3061         }
3062         return 0;
3063 }
3064
3065 /*
3066  * We have done the truncate/delete so we can go ahead and remove the orphan
3067  * item for this particular inode.
3068  */
3069 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3070                             struct inode *inode)
3071 {
3072         struct btrfs_root *root = BTRFS_I(inode)->root;
3073         int delete_item = 0;
3074         int release_rsv = 0;
3075         int ret = 0;
3076
3077         spin_lock(&root->orphan_lock);
3078         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3079                                &BTRFS_I(inode)->runtime_flags))
3080                 delete_item = 1;
3081
3082         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3083                                &BTRFS_I(inode)->runtime_flags))
3084                 release_rsv = 1;
3085         spin_unlock(&root->orphan_lock);
3086
3087         if (trans && delete_item) {
3088                 ret = btrfs_del_orphan_item(trans, root, btrfs_ino(inode));
3089                 BUG_ON(ret); /* -ENOMEM or corruption (JDM: Recheck) */
3090         }
3091
3092         if (release_rsv) {
3093                 btrfs_orphan_release_metadata(inode);
3094                 atomic_dec(&root->orphan_inodes);
3095         }
3096
3097         return 0;
3098 }
3099
3100 /*
3101  * this cleans up any orphans that may be left on the list from the last use
3102  * of this root.
3103  */
3104 int btrfs_orphan_cleanup(struct btrfs_root *root)
3105 {
3106         struct btrfs_path *path;
3107         struct extent_buffer *leaf;
3108         struct btrfs_key key, found_key;
3109         struct btrfs_trans_handle *trans;
3110         struct inode *inode;
3111         u64 last_objectid = 0;
3112         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3113
3114         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3115                 return 0;
3116
3117         path = btrfs_alloc_path();
3118         if (!path) {
3119                 ret = -ENOMEM;
3120                 goto out;
3121         }
3122         path->reada = -1;
3123
3124         key.objectid = BTRFS_ORPHAN_OBJECTID;
3125         btrfs_set_key_type(&key, BTRFS_ORPHAN_ITEM_KEY);
3126         key.offset = (u64)-1;
3127
3128         while (1) {
3129                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3130                 if (ret < 0)
3131                         goto out;
3132
3133                 /*
3134                  * if ret == 0 means we found what we were searching for, which
3135                  * is weird, but possible, so only screw with path if we didn't
3136                  * find the key and see if we have stuff that matches
3137                  */
3138                 if (ret > 0) {
3139                         ret = 0;
3140                         if (path->slots[0] == 0)
3141                                 break;
3142                         path->slots[0]--;
3143                 }
3144
3145                 /* pull out the item */
3146                 leaf = path->nodes[0];
3147                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3148
3149                 /* make sure the item matches what we want */
3150                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3151                         break;
3152                 if (btrfs_key_type(&found_key) != BTRFS_ORPHAN_ITEM_KEY)
3153                         break;
3154
3155                 /* release the path since we're done with it */
3156                 btrfs_release_path(path);
3157
3158                 /*
3159                  * this is where we are basically btrfs_lookup, without the
3160                  * crossing root thing.  we store the inode number in the
3161                  * offset of the orphan item.
3162                  */
3163
3164                 if (found_key.offset == last_objectid) {
3165                         btrfs_err(root->fs_info,
3166                                 "Error removing orphan entry, stopping orphan cleanup");
3167                         ret = -EINVAL;
3168                         goto out;
3169                 }
3170
3171                 last_objectid = found_key.offset;
3172
3173                 found_key.objectid = found_key.offset;
3174                 found_key.type = BTRFS_INODE_ITEM_KEY;
3175                 found_key.offset = 0;
3176                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
3177                 ret = PTR_RET(inode);
3178                 if (ret && ret != -ESTALE)
3179                         goto out;
3180
3181                 if (ret == -ESTALE && root == root->fs_info->tree_root) {
3182                         struct btrfs_root *dead_root;
3183                         struct btrfs_fs_info *fs_info = root->fs_info;
3184                         int is_dead_root = 0;
3185
3186                         /*
3187                          * this is an orphan in the tree root. Currently these
3188                          * could come from 2 sources:
3189                          *  a) a snapshot deletion in progress
3190                          *  b) a free space cache inode
3191                          * We need to distinguish those two, as the snapshot
3192                          * orphan must not get deleted.
3193                          * find_dead_roots already ran before us, so if this
3194                          * is a snapshot deletion, we should find the root
3195                          * in the dead_roots list
3196                          */
3197                         spin_lock(&fs_info->trans_lock);
3198                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3199                                             root_list) {
3200                                 if (dead_root->root_key.objectid ==
3201                                     found_key.objectid) {
3202                                         is_dead_root = 1;
3203                                         break;
3204                                 }
3205                         }
3206                         spin_unlock(&fs_info->trans_lock);
3207                         if (is_dead_root) {
3208                                 /* prevent this orphan from being found again */
3209                                 key.offset = found_key.objectid - 1;
3210                                 continue;
3211                         }
3212                 }
3213                 /*
3214                  * Inode is already gone but the orphan item is still there,
3215                  * kill the orphan item.
3216                  */
3217                 if (ret == -ESTALE) {
3218                         trans = btrfs_start_transaction(root, 1);
3219                         if (IS_ERR(trans)) {
3220                                 ret = PTR_ERR(trans);
3221                                 goto out;
3222                         }
3223                         btrfs_debug(root->fs_info, "auto deleting %Lu",
3224                                 found_key.objectid);
3225                         ret = btrfs_del_orphan_item(trans, root,
3226                                                     found_key.objectid);
3227                         BUG_ON(ret); /* -ENOMEM or corruption (JDM: Recheck) */
3228                         btrfs_end_transaction(trans, root);
3229                         continue;
3230                 }
3231
3232                 /*
3233                  * add this inode to the orphan list so btrfs_orphan_del does
3234                  * the proper thing when we hit it
3235                  */
3236                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3237                         &BTRFS_I(inode)->runtime_flags);
3238                 atomic_inc(&root->orphan_inodes);
3239
3240                 /* if we have links, this was a truncate, lets do that */
3241                 if (inode->i_nlink) {
3242                         if (!S_ISREG(inode->i_mode)) {
3243                                 WARN_ON(1);
3244                                 iput(inode);
3245                                 continue;
3246                         }
3247                         nr_truncate++;
3248
3249                         /* 1 for the orphan item deletion. */
3250                         trans = btrfs_start_transaction(root, 1);
3251                         if (IS_ERR(trans)) {
3252                                 iput(inode);
3253                                 ret = PTR_ERR(trans);
3254                                 goto out;
3255                         }
3256                         ret = btrfs_orphan_add(trans, inode);
3257                         btrfs_end_transaction(trans, root);
3258                         if (ret) {
3259                                 iput(inode);
3260                                 goto out;
3261                         }
3262
3263                         ret = btrfs_truncate(inode);
3264                         if (ret)
3265                                 btrfs_orphan_del(NULL, inode);
3266                 } else {
3267                         nr_unlink++;
3268                 }
3269
3270                 /* this will do delete_inode and everything for us */
3271                 iput(inode);
3272                 if (ret)
3273                         goto out;
3274         }
3275         /* release the path since we're done with it */
3276         btrfs_release_path(path);
3277
3278         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3279
3280         if (root->orphan_block_rsv)
3281                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
3282                                         (u64)-1);
3283
3284         if (root->orphan_block_rsv || root->orphan_item_inserted) {
3285                 trans = btrfs_join_transaction(root);
3286                 if (!IS_ERR(trans))
3287                         btrfs_end_transaction(trans, root);
3288         }
3289
3290         if (nr_unlink)
3291                 btrfs_debug(root->fs_info, "unlinked %d orphans", nr_unlink);
3292         if (nr_truncate)
3293                 btrfs_debug(root->fs_info, "truncated %d orphans", nr_truncate);
3294
3295 out:
3296         if (ret)
3297                 btrfs_crit(root->fs_info,
3298                         "could not do orphan cleanup %d", ret);
3299         btrfs_free_path(path);
3300         return ret;
3301 }
3302
3303 /*
3304  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3305  * don't find any xattrs, we know there can't be any acls.
3306  *
3307  * slot is the slot the inode is in, objectid is the objectid of the inode
3308  */
3309 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3310                                           int slot, u64 objectid)
3311 {
3312         u32 nritems = btrfs_header_nritems(leaf);
3313         struct btrfs_key found_key;
3314         static u64 xattr_access = 0;
3315         static u64 xattr_default = 0;
3316         int scanned = 0;
3317
3318         if (!xattr_access) {
3319                 xattr_access = btrfs_name_hash(POSIX_ACL_XATTR_ACCESS,
3320                                         strlen(POSIX_ACL_XATTR_ACCESS));
3321                 xattr_default = btrfs_name_hash(POSIX_ACL_XATTR_DEFAULT,
3322                                         strlen(POSIX_ACL_XATTR_DEFAULT));
3323         }
3324
3325         slot++;
3326         while (slot < nritems) {
3327                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3328
3329                 /* we found a different objectid, there must not be acls */
3330                 if (found_key.objectid != objectid)
3331                         return 0;
3332
3333                 /* we found an xattr, assume we've got an acl */
3334                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3335                         if (found_key.offset == xattr_access ||
3336                             found_key.offset == xattr_default)
3337                                 return 1;
3338                 }
3339
3340                 /*
3341                  * we found a key greater than an xattr key, there can't
3342                  * be any acls later on
3343                  */
3344                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3345                         return 0;
3346
3347                 slot++;
3348                 scanned++;
3349
3350                 /*
3351                  * it goes inode, inode backrefs, xattrs, extents,
3352                  * so if there are a ton of hard links to an inode there can
3353                  * be a lot of backrefs.  Don't waste time searching too hard,
3354                  * this is just an optimization
3355                  */
3356                 if (scanned >= 8)
3357                         break;
3358         }
3359         /* we hit the end of the leaf before we found an xattr or
3360          * something larger than an xattr.  We have to assume the inode
3361          * has acls
3362          */
3363         return 1;
3364 }
3365
3366 /*
3367  * read an inode from the btree into the in-memory inode
3368  */
3369 static void btrfs_read_locked_inode(struct inode *inode)
3370 {
3371         struct btrfs_path *path;
3372         struct extent_buffer *leaf;
3373         struct btrfs_inode_item *inode_item;
3374         struct btrfs_timespec *tspec;
3375         struct btrfs_root *root = BTRFS_I(inode)->root;
3376         struct btrfs_key location;
3377         int maybe_acls;
3378         u32 rdev;
3379         int ret;
3380         bool filled = false;
3381
3382         ret = btrfs_fill_inode(inode, &rdev);
3383         if (!ret)
3384                 filled = true;
3385
3386         path = btrfs_alloc_path();
3387         if (!path)
3388                 goto make_bad;
3389
3390         path->leave_spinning = 1;
3391         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3392
3393         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3394         if (ret)
3395                 goto make_bad;
3396
3397         leaf = path->nodes[0];
3398
3399         if (filled)
3400                 goto cache_acl;
3401
3402         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3403                                     struct btrfs_inode_item);
3404         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3405         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3406         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3407         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3408         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3409
3410         tspec = btrfs_inode_atime(inode_item);
3411         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, tspec);
3412         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
3413
3414         tspec = btrfs_inode_mtime(inode_item);
3415         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, tspec);
3416         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
3417
3418         tspec = btrfs_inode_ctime(inode_item);
3419         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, tspec);
3420         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
3421
3422         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3423         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3424         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3425
3426         /*
3427          * If we were modified in the current generation and evicted from memory
3428          * and then re-read we need to do a full sync since we don't have any
3429          * idea about which extents were modified before we were evicted from
3430          * cache.
3431          */
3432         if (BTRFS_I(inode)->last_trans == root->fs_info->generation)
3433                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3434                         &BTRFS_I(inode)->runtime_flags);
3435
3436         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3437         inode->i_generation = BTRFS_I(inode)->generation;
3438         inode->i_rdev = 0;
3439         rdev = btrfs_inode_rdev(leaf, inode_item);
3440
3441         BTRFS_I(inode)->index_cnt = (u64)-1;
3442         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3443 cache_acl:
3444         /*
3445          * try to precache a NULL acl entry for files that don't have
3446          * any xattrs or acls
3447          */
3448         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3449                                            btrfs_ino(inode));
3450         if (!maybe_acls)
3451                 cache_no_acl(inode);
3452
3453         btrfs_free_path(path);
3454
3455         switch (inode->i_mode & S_IFMT) {
3456         case S_IFREG:
3457                 inode->i_mapping->a_ops = &btrfs_aops;
3458                 inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
3459                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3460                 inode->i_fop = &btrfs_file_operations;
3461                 inode->i_op = &btrfs_file_inode_operations;
3462                 break;
3463         case S_IFDIR:
3464                 inode->i_fop = &btrfs_dir_file_operations;
3465                 if (root == root->fs_info->tree_root)
3466                         inode->i_op = &btrfs_dir_ro_inode_operations;
3467                 else
3468                         inode->i_op = &btrfs_dir_inode_operations;
3469                 break;
3470         case S_IFLNK:
3471                 inode->i_op = &btrfs_symlink_inode_operations;
3472                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3473                 inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
3474                 break;
3475         default:
3476                 inode->i_op = &btrfs_special_inode_operations;
3477                 init_special_inode(inode, inode->i_mode, rdev);
3478                 break;
3479         }
3480
3481         btrfs_update_iflags(inode);
3482         return;
3483
3484 make_bad:
3485         btrfs_free_path(path);
3486         make_bad_inode(inode);
3487 }
3488
3489 /*
3490  * given a leaf and an inode, copy the inode fields into the leaf
3491  */
3492 static void fill_inode_item(struct btrfs_trans_handle *trans,
3493                             struct extent_buffer *leaf,
3494                             struct btrfs_inode_item *item,
3495                             struct inode *inode)
3496 {
3497         struct btrfs_map_token token;
3498
3499         btrfs_init_map_token(&token);
3500
3501         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3502         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3503         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3504                                    &token);
3505         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3506         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3507
3508         btrfs_set_token_timespec_sec(leaf, btrfs_inode_atime(item),
3509                                      inode->i_atime.tv_sec, &token);
3510         btrfs_set_token_timespec_nsec(leaf, btrfs_inode_atime(item),
3511                                       inode->i_atime.tv_nsec, &token);
3512
3513         btrfs_set_token_timespec_sec(leaf, btrfs_inode_mtime(item),
3514                                      inode->i_mtime.tv_sec, &token);
3515         btrfs_set_token_timespec_nsec(leaf, btrfs_inode_mtime(item),
3516                                       inode->i_mtime.tv_nsec, &token);
3517
3518         btrfs_set_token_timespec_sec(leaf, btrfs_inode_ctime(item),
3519                                      inode->i_ctime.tv_sec, &token);
3520         btrfs_set_token_timespec_nsec(leaf, btrfs_inode_ctime(item),
3521                                       inode->i_ctime.tv_nsec, &token);
3522
3523         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3524                                      &token);
3525         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3526                                          &token);
3527         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3528         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3529         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3530         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3531         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3532 }
3533
3534 /*
3535  * copy everything in the in-memory inode into the btree.
3536  */
3537 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3538                                 struct btrfs_root *root, struct inode *inode)
3539 {
3540         struct btrfs_inode_item *inode_item;
3541         struct btrfs_path *path;
3542         struct extent_buffer *leaf;
3543         int ret;
3544
3545         path = btrfs_alloc_path();
3546         if (!path)
3547                 return -ENOMEM;
3548
3549         path->leave_spinning = 1;
3550         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3551                                  1);
3552         if (ret) {
3553                 if (ret > 0)
3554                         ret = -ENOENT;
3555                 goto failed;
3556         }
3557
3558         btrfs_unlock_up_safe(path, 1);
3559         leaf = path->nodes[0];
3560         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3561                                     struct btrfs_inode_item);
3562
3563         fill_inode_item(trans, leaf, inode_item, inode);
3564         btrfs_mark_buffer_dirty(leaf);
3565         btrfs_set_inode_last_trans(trans, inode);
3566         ret = 0;
3567 failed:
3568         btrfs_free_path(path);
3569         return ret;
3570 }
3571
3572 /*
3573  * copy everything in the in-memory inode into the btree.
3574  */
3575 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3576                                 struct btrfs_root *root, struct inode *inode)
3577 {
3578         int ret;
3579
3580         /*
3581          * If the inode is a free space inode, we can deadlock during commit
3582          * if we put it into the delayed code.
3583          *
3584          * The data relocation inode should also be directly updated
3585          * without delay
3586          */
3587         if (!btrfs_is_free_space_inode(inode)
3588             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID) {
3589                 btrfs_update_root_times(trans, root);
3590
3591                 ret = btrfs_delayed_update_inode(trans, root, inode);
3592                 if (!ret)
3593                         btrfs_set_inode_last_trans(trans, inode);
3594                 return ret;
3595         }
3596
3597         return btrfs_update_inode_item(trans, root, inode);
3598 }
3599
3600 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3601                                          struct btrfs_root *root,
3602                                          struct inode *inode)
3603 {
3604         int ret;
3605
3606         ret = btrfs_update_inode(trans, root, inode);
3607         if (ret == -ENOSPC)
3608                 return btrfs_update_inode_item(trans, root, inode);
3609         return ret;
3610 }
3611
3612 /*
3613  * unlink helper that gets used here in inode.c and in the tree logging
3614  * recovery code.  It remove a link in a directory with a given name, and
3615  * also drops the back refs in the inode to the directory
3616  */
3617 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3618                                 struct btrfs_root *root,
3619                                 struct inode *dir, struct inode *inode,
3620                                 const char *name, int name_len)
3621 {
3622         struct btrfs_path *path;
3623         int ret = 0;
3624         struct extent_buffer *leaf;
3625         struct btrfs_dir_item *di;
3626         struct btrfs_key key;
3627         u64 index;
3628         u64 ino = btrfs_ino(inode);
3629         u64 dir_ino = btrfs_ino(dir);
3630
3631         path = btrfs_alloc_path();
3632         if (!path) {
3633                 ret = -ENOMEM;
3634                 goto out;
3635         }
3636
3637         path->leave_spinning = 1;
3638         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3639                                     name, name_len, -1);
3640         if (IS_ERR(di)) {
3641                 ret = PTR_ERR(di);
3642                 goto err;
3643         }
3644         if (!di) {
3645                 ret = -ENOENT;
3646                 goto err;
3647         }
3648         leaf = path->nodes[0];
3649         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3650         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3651         if (ret)
3652                 goto err;
3653         btrfs_release_path(path);
3654
3655         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3656                                   dir_ino, &index);
3657         if (ret) {
3658                 btrfs_info(root->fs_info,
3659                         "failed to delete reference to %.*s, inode %llu parent %llu",
3660                         name_len, name,
3661                         (unsigned long long)ino, (unsigned long long)dir_ino);
3662                 btrfs_abort_transaction(trans, root, ret);
3663                 goto err;
3664         }
3665
3666         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
3667         if (ret) {
3668                 btrfs_abort_transaction(trans, root, ret);
3669                 goto err;
3670         }
3671
3672         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
3673                                          inode, dir_ino);
3674         if (ret != 0 && ret != -ENOENT) {
3675                 btrfs_abort_transaction(trans, root, ret);
3676                 goto err;
3677         }
3678
3679         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
3680                                            dir, index);
3681         if (ret == -ENOENT)
3682                 ret = 0;
3683         else if (ret)
3684                 btrfs_abort_transaction(trans, root, ret);
3685 err:
3686         btrfs_free_path(path);
3687         if (ret)
3688                 goto out;
3689
3690         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
3691         inode_inc_iversion(inode);
3692         inode_inc_iversion(dir);
3693         inode->i_ctime = dir->i_mtime = dir->i_ctime = CURRENT_TIME;
3694         ret = btrfs_update_inode(trans, root, dir);
3695 out:
3696         return ret;
3697 }
3698
3699 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3700                        struct btrfs_root *root,
3701                        struct inode *dir, struct inode *inode,
3702                        const char *name, int name_len)
3703 {
3704         int ret;
3705         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
3706         if (!ret) {
3707                 btrfs_drop_nlink(inode);
3708                 ret = btrfs_update_inode(trans, root, inode);
3709         }
3710         return ret;
3711 }
3712
3713 /*
3714  * helper to start transaction for unlink and rmdir.
3715  *
3716  * unlink and rmdir are special in btrfs, they do not always free space, so
3717  * if we cannot make our reservations the normal way try and see if there is
3718  * plenty of slack room in the global reserve to migrate, otherwise we cannot
3719  * allow the unlink to occur.
3720  */
3721 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
3722 {
3723         struct btrfs_trans_handle *trans;
3724         struct btrfs_root *root = BTRFS_I(dir)->root;
3725         int ret;
3726
3727         /*
3728          * 1 for the possible orphan item
3729          * 1 for the dir item
3730          * 1 for the dir index
3731          * 1 for the inode ref
3732          * 1 for the inode
3733          */
3734         trans = btrfs_start_transaction(root, 5);
3735         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
3736                 return trans;
3737
3738         if (PTR_ERR(trans) == -ENOSPC) {
3739                 u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
3740
3741                 trans = btrfs_start_transaction(root, 0);
3742                 if (IS_ERR(trans))
3743                         return trans;
3744                 ret = btrfs_cond_migrate_bytes(root->fs_info,
3745                                                &root->fs_info->trans_block_rsv,
3746                                                num_bytes, 5);
3747                 if (ret) {
3748                         btrfs_end_transaction(trans, root);
3749                         return ERR_PTR(ret);
3750                 }
3751                 trans->block_rsv = &root->fs_info->trans_block_rsv;
3752                 trans->bytes_reserved = num_bytes;
3753         }
3754         return trans;
3755 }
3756
3757 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
3758 {
3759         struct btrfs_root *root = BTRFS_I(dir)->root;
3760         struct btrfs_trans_handle *trans;
3761         struct inode *inode = dentry->d_inode;
3762         int ret;
3763
3764         trans = __unlink_start_trans(dir);
3765         if (IS_ERR(trans))
3766                 return PTR_ERR(trans);
3767
3768         btrfs_record_unlink_dir(trans, dir, dentry->d_inode, 0);
3769
3770         ret = btrfs_unlink_inode(trans, root, dir, dentry->d_inode,
3771                                  dentry->d_name.name, dentry->d_name.len);
3772         if (ret)
3773                 goto out;
3774
3775         if (inode->i_nlink == 0) {
3776                 ret = btrfs_orphan_add(trans, inode);
3777                 if (ret)
3778                         goto out;
3779         }
3780
3781 out:
3782         btrfs_end_transaction(trans, root);
3783         btrfs_btree_balance_dirty(root);
3784         return ret;
3785 }
3786
3787 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
3788                         struct btrfs_root *root,
3789                         struct inode *dir, u64 objectid,
3790                         const char *name, int name_len)
3791 {
3792         struct btrfs_path *path;
3793         struct extent_buffer *leaf;
3794         struct btrfs_dir_item *di;
3795         struct btrfs_key key;
3796         u64 index;
3797         int ret;
3798         u64 dir_ino = btrfs_ino(dir);
3799
3800         path = btrfs_alloc_path();
3801         if (!path)
3802                 return -ENOMEM;
3803
3804         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3805                                    name, name_len, -1);
3806         if (IS_ERR_OR_NULL(di)) {
3807                 if (!di)
3808                         ret = -ENOENT;
3809                 else
3810                         ret = PTR_ERR(di);
3811                 goto out;
3812         }
3813
3814         leaf = path->nodes[0];
3815         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3816         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
3817         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3818         if (ret) {
3819                 btrfs_abort_transaction(trans, root, ret);
3820                 goto out;
3821         }
3822         btrfs_release_path(path);
3823
3824         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
3825                                  objectid, root->root_key.objectid,
3826                                  dir_ino, &index, name, name_len);
3827         if (ret < 0) {
3828                 if (ret != -ENOENT) {
3829                         btrfs_abort_transaction(trans, root, ret);
3830                         goto out;
3831                 }
3832                 di = btrfs_search_dir_index_item(root, path, dir_ino,
3833                                                  name, name_len);
3834                 if (IS_ERR_OR_NULL(di)) {
3835                         if (!di)
3836                                 ret = -ENOENT;
3837                         else
3838                                 ret = PTR_ERR(di);
3839                         btrfs_abort_transaction(trans, root, ret);
3840                         goto out;
3841                 }
3842
3843                 leaf = path->nodes[0];
3844                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3845                 btrfs_release_path(path);
3846                 index = key.offset;
3847         }
3848         btrfs_release_path(path);
3849
3850         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
3851         if (ret) {
3852                 btrfs_abort_transaction(trans, root, ret);
3853                 goto out;
3854         }
3855
3856         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
3857         inode_inc_iversion(dir);
3858         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
3859         ret = btrfs_update_inode_fallback(trans, root, dir);
3860         if (ret)
3861                 btrfs_abort_transaction(trans, root, ret);
3862 out:
3863         btrfs_free_path(path);
3864         return ret;
3865 }
3866
3867 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
3868 {
3869         struct inode *inode = dentry->d_inode;
3870         int err = 0;
3871         struct btrfs_root *root = BTRFS_I(dir)->root;
3872         struct btrfs_trans_handle *trans;
3873
3874         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
3875                 return -ENOTEMPTY;
3876         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
3877                 return -EPERM;
3878
3879         trans = __unlink_start_trans(dir);
3880         if (IS_ERR(trans))
3881                 return PTR_ERR(trans);
3882
3883         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
3884                 err = btrfs_unlink_subvol(trans, root, dir,
3885                                           BTRFS_I(inode)->location.objectid,
3886                                           dentry->d_name.name,
3887                                           dentry->d_name.len);
3888                 goto out;
3889         }
3890
3891         err = btrfs_orphan_add(trans, inode);
3892         if (err)
3893                 goto out;
3894
3895         /* now the directory is empty */
3896         err = btrfs_unlink_inode(trans, root, dir, dentry->d_inode,
3897                                  dentry->d_name.name, dentry->d_name.len);
3898         if (!err)
3899                 btrfs_i_size_write(inode, 0);
3900 out:
3901         btrfs_end_transaction(trans, root);
3902         btrfs_btree_balance_dirty(root);
3903
3904         return err;
3905 }
3906
3907 /*
3908  * this can truncate away extent items, csum items and directory items.
3909  * It starts at a high offset and removes keys until it can't find
3910  * any higher than new_size
3911  *
3912  * csum items that cross the new i_size are truncated to the new size
3913  * as well.
3914  *
3915  * min_type is the minimum key type to truncate down to.  If set to 0, this
3916  * will kill all the items on this inode, including the INODE_ITEM_KEY.
3917  */
3918 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
3919                                struct btrfs_root *root,
3920                                struct inode *inode,
3921                                u64 new_size, u32 min_type)
3922 {
3923         struct btrfs_path *path;
3924         struct extent_buffer *leaf;
3925         struct btrfs_file_extent_item *fi;
3926         struct btrfs_key key;
3927         struct btrfs_key found_key;
3928         u64 extent_start = 0;
3929         u64 extent_num_bytes = 0;
3930         u64 extent_offset = 0;
3931         u64 item_end = 0;
3932         u32 found_type = (u8)-1;
3933         int found_extent;
3934         int del_item;
3935         int pending_del_nr = 0;
3936         int pending_del_slot = 0;
3937         int extent_type = -1;
3938         int ret;
3939         int err = 0;
3940         u64 ino = btrfs_ino(inode);
3941
3942         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
3943
3944         path = btrfs_alloc_path();
3945         if (!path)
3946                 return -ENOMEM;
3947         path->reada = -1;
3948
3949         /*
3950          * We want to drop from the next block forward in case this new size is
3951          * not block aligned since we will be keeping the last block of the
3952          * extent just the way it is.
3953          */
3954         if (root->ref_cows || root == root->fs_info->tree_root)
3955                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
3956                                         root->sectorsize), (u64)-1, 0);
3957
3958         /*
3959          * This function is also used to drop the items in the log tree before
3960          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
3961          * it is used to drop the loged items. So we shouldn't kill the delayed
3962          * items.
3963          */
3964         if (min_type == 0 && root == BTRFS_I(inode)->root)
3965                 btrfs_kill_delayed_inode_items(inode);
3966
3967         key.objectid = ino;
3968         key.offset = (u64)-1;
3969         key.type = (u8)-1;
3970
3971 search_again:
3972         path->leave_spinning = 1;
3973         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
3974         if (ret < 0) {
3975                 err = ret;
3976                 goto out;
3977         }
3978
3979         if (ret > 0) {
3980                 /* there are no items in the tree for us to truncate, we're
3981                  * done
3982                  */
3983                 if (path->slots[0] == 0)
3984                         goto out;
3985                 path->slots[0]--;
3986         }
3987
3988         while (1) {
3989                 fi = NULL;
3990                 leaf = path->nodes[0];
3991                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3992                 found_type = btrfs_key_type(&found_key);
3993
3994                 if (found_key.objectid != ino)
3995                         break;
3996
3997                 if (found_type < min_type)
3998                         break;
3999
4000                 item_end = found_key.offset;
4001                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4002                         fi = btrfs_item_ptr(leaf, path->slots[0],
4003                                             struct btrfs_file_extent_item);
4004                         extent_type = btrfs_file_extent_type(leaf, fi);
4005                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4006                                 item_end +=
4007                                     btrfs_file_extent_num_bytes(leaf, fi);
4008                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4009                                 item_end += btrfs_file_extent_inline_len(leaf,
4010                                                                          fi);
4011                         }
4012                         item_end--;
4013                 }
4014                 if (found_type > min_type) {
4015                         del_item = 1;
4016                 } else {
4017                         if (item_end < new_size)
4018                                 break;
4019                         if (found_key.offset >= new_size)
4020                                 del_item = 1;
4021                         else
4022                                 del_item = 0;
4023                 }
4024                 found_extent = 0;
4025                 /* FIXME, shrink the extent if the ref count is only 1 */
4026                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4027                         goto delete;
4028
4029                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4030                         u64 num_dec;
4031                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4032                         if (!del_item) {
4033                                 u64 orig_num_bytes =
4034                                         btrfs_file_extent_num_bytes(leaf, fi);
4035                                 extent_num_bytes = ALIGN(new_size -
4036                                                 found_key.offset,
4037                                                 root->sectorsize);
4038                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4039                                                          extent_num_bytes);
4040                                 num_dec = (orig_num_bytes -
4041                                            extent_num_bytes);
4042                                 if (root->ref_cows && extent_start != 0)
4043                                         inode_sub_bytes(inode, num_dec);
4044                                 btrfs_mark_buffer_dirty(leaf);
4045                         } else {
4046                                 extent_num_bytes =
4047                                         btrfs_file_extent_disk_num_bytes(leaf,
4048                                                                          fi);
4049                                 extent_offset = found_key.offset -
4050                                         btrfs_file_extent_offset(leaf, fi);
4051
4052                                 /* FIXME blocksize != 4096 */
4053                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4054                                 if (extent_start != 0) {
4055                                         found_extent = 1;
4056                                         if (root->ref_cows)
4057                                                 inode_sub_bytes(inode, num_dec);
4058                                 }
4059                         }
4060                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4061                         /*
4062                          * we can't truncate inline items that have had
4063                          * special encodings
4064                          */
4065                         if (!del_item &&
4066                             btrfs_file_extent_compression(leaf, fi) == 0 &&
4067                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4068                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4069                                 u32 size = new_size - found_key.offset;
4070
4071                                 if (root->ref_cows) {
4072                                         inode_sub_bytes(inode, item_end + 1 -
4073                                                         new_size);
4074                                 }
4075                                 size =
4076                                     btrfs_file_extent_calc_inline_size(size);
4077                                 btrfs_truncate_item(root, path, size, 1);
4078                         } else if (root->ref_cows) {
4079                                 inode_sub_bytes(inode, item_end + 1 -
4080                                                 found_key.offset);
4081                         }
4082                 }
4083 delete:
4084                 if (del_item) {
4085                         if (!pending_del_nr) {
4086                                 /* no pending yet, add ourselves */
4087                                 pending_del_slot = path->slots[0];
4088                                 pending_del_nr = 1;
4089                         } else if (pending_del_nr &&
4090                                    path->slots[0] + 1 == pending_del_slot) {
4091                                 /* hop on the pending chunk */
4092                                 pending_del_nr++;
4093                                 pending_del_slot = path->slots[0];
4094                         } else {
4095                                 BUG();
4096                         }
4097                 } else {
4098                         break;
4099                 }
4100                 if (found_extent && (root->ref_cows ||
4101                                      root == root->fs_info->tree_root)) {
4102                         btrfs_set_path_blocking(path);
4103                         ret = btrfs_free_extent(trans, root, extent_start,
4104                                                 extent_num_bytes, 0,
4105                                                 btrfs_header_owner(leaf),
4106                                                 ino, extent_offset, 0);
4107                         BUG_ON(ret);
4108                 }
4109
4110                 if (found_type == BTRFS_INODE_ITEM_KEY)
4111                         break;
4112
4113                 if (path->slots[0] == 0 ||
4114                     path->slots[0] != pending_del_slot) {
4115                         if (pending_del_nr) {
4116                                 ret = btrfs_del_items(trans, root, path,
4117                                                 pending_del_slot,
4118                                                 pending_del_nr);
4119                                 if (ret) {
4120                                         btrfs_abort_transaction(trans,
4121                                                                 root, ret);
4122                                         goto error;
4123                                 }
4124                                 pending_del_nr = 0;
4125                         }
4126                         btrfs_release_path(path);
4127                         goto search_again;
4128                 } else {
4129                         path->slots[0]--;
4130                 }
4131         }
4132 out:
4133         if (pending_del_nr) {
4134                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4135                                       pending_del_nr);
4136                 if (ret)
4137                         btrfs_abort_transaction(trans, root, ret);
4138         }
4139 error:
4140         btrfs_free_path(path);
4141         return err;
4142 }
4143
4144 /*
4145  * btrfs_truncate_page - read, zero a chunk and write a page
4146  * @inode - inode that we're zeroing
4147  * @from - the offset to start zeroing
4148  * @len - the length to zero, 0 to zero the entire range respective to the
4149  *      offset
4150  * @front - zero up to the offset instead of from the offset on
4151  *
4152  * This will find the page for the "from" offset and cow the page and zero the
4153  * part we want to zero.  This is used with truncate and hole punching.
4154  */
4155 int btrfs_truncate_page(struct inode *inode, loff_t from, loff_t len,
4156                         int front)
4157 {
4158         struct address_space *mapping = inode->i_mapping;
4159         struct btrfs_root *root = BTRFS_I(inode)->root;
4160         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4161         struct btrfs_ordered_extent *ordered;
4162         struct extent_state *cached_state = NULL;
4163         char *kaddr;
4164         u32 blocksize = root->sectorsize;
4165         pgoff_t index = from >> PAGE_CACHE_SHIFT;
4166         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4167         struct page *page;
4168         gfp_t mask = btrfs_alloc_write_mask(mapping);
4169         int ret = 0;
4170         u64 page_start;
4171         u64 page_end;
4172
4173         if ((offset & (blocksize - 1)) == 0 &&
4174             (!len || ((len & (blocksize - 1)) == 0)))
4175                 goto out;
4176         ret = btrfs_delalloc_reserve_space(inode, PAGE_CACHE_SIZE);
4177         if (ret)
4178                 goto out;
4179
4180 again:
4181         page = find_or_create_page(mapping, index, mask);
4182         if (!page) {
4183                 btrfs_delalloc_release_space(inode, PAGE_CACHE_SIZE);
4184                 ret = -ENOMEM;
4185                 goto out;
4186         }
4187
4188         page_start = page_offset(page);
4189         page_end = page_start + PAGE_CACHE_SIZE - 1;
4190
4191         if (!PageUptodate(page)) {
4192                 ret = btrfs_readpage(NULL, page);
4193                 lock_page(page);
4194                 if (page->mapping != mapping) {
4195                         unlock_page(page);
4196                         page_cache_release(page);
4197                         goto again;
4198                 }
4199                 if (!PageUptodate(page)) {
4200                         ret = -EIO;
4201                         goto out_unlock;
4202                 }
4203         }
4204         wait_on_page_writeback(page);
4205
4206         lock_extent_bits(io_tree, page_start, page_end, 0, &cached_state);
4207         set_page_extent_mapped(page);
4208
4209         ordered = btrfs_lookup_ordered_extent(inode, page_start);
4210         if (ordered) {
4211                 unlock_extent_cached(io_tree, page_start, page_end,
4212                                      &cached_state, GFP_NOFS);
4213                 unlock_page(page);
4214                 page_cache_release(page);
4215                 btrfs_start_ordered_extent(inode, ordered, 1);
4216                 btrfs_put_ordered_extent(ordered);
4217                 goto again;
4218         }
4219
4220         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
4221                           EXTENT_DIRTY | EXTENT_DELALLOC |
4222                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4223                           0, 0, &cached_state, GFP_NOFS);
4224
4225         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
4226                                         &cached_state);
4227         if (ret) {
4228                 unlock_extent_cached(io_tree, page_start, page_end,
4229                                      &cached_state, GFP_NOFS);
4230                 goto out_unlock;
4231         }
4232
4233         if (offset != PAGE_CACHE_SIZE) {
4234                 if (!len)
4235                         len = PAGE_CACHE_SIZE - offset;
4236                 kaddr = kmap(page);
4237                 if (front)
4238                         memset(kaddr, 0, offset);
4239                 else
4240                         memset(kaddr + offset, 0, len);
4241                 flush_dcache_page(page);
4242                 kunmap(page);
4243         }
4244         ClearPageChecked(page);
4245         set_page_dirty(page);
4246         unlock_extent_cached(io_tree, page_start, page_end, &cached_state,
4247                              GFP_NOFS);
4248
4249 out_unlock:
4250         if (ret)
4251                 btrfs_delalloc_release_space(inode, PAGE_CACHE_SIZE);
4252         unlock_page(page);
4253         page_cache_release(page);
4254 out:
4255         return ret;
4256 }
4257
4258 /*
4259  * This function puts in dummy file extents for the area we're creating a hole
4260  * for.  So if we are truncating this file to a larger size we need to insert
4261  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4262  * the range between oldsize and size
4263  */
4264 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4265 {
4266         struct btrfs_trans_handle *trans;
4267         struct btrfs_root *root = BTRFS_I(inode)->root;
4268         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4269         struct extent_map *em = NULL;
4270         struct extent_state *cached_state = NULL;
4271         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4272         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4273         u64 block_end = ALIGN(size, root->sectorsize);
4274         u64 last_byte;
4275         u64 cur_offset;
4276         u64 hole_size;
4277         int err = 0;
4278
4279         /*
4280          * If our size started in the middle of a page we need to zero out the
4281          * rest of the page before we expand the i_size, otherwise we could
4282          * expose stale data.
4283          */
4284         err = btrfs_truncate_page(inode, oldsize, 0, 0);
4285         if (err)
4286                 return err;
4287
4288         if (size <= hole_start)
4289                 return 0;
4290
4291         while (1) {
4292                 struct btrfs_ordered_extent *ordered;
4293                 btrfs_wait_ordered_range(inode, hole_start,
4294                                          block_end - hole_start);
4295                 lock_extent_bits(io_tree, hole_start, block_end - 1, 0,
4296                                  &cached_state);
4297                 ordered = btrfs_lookup_ordered_extent(inode, hole_start);
4298                 if (!ordered)
4299                         break;
4300                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4301                                      &cached_state, GFP_NOFS);
4302                 btrfs_put_ordered_extent(ordered);
4303         }
4304
4305         cur_offset = hole_start;
4306         while (1) {
4307                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4308                                 block_end - cur_offset, 0);
4309                 if (IS_ERR(em)) {
4310                         err = PTR_ERR(em);
4311                         em = NULL;
4312                         break;
4313                 }
4314                 last_byte = min(extent_map_end(em), block_end);
4315                 last_byte = ALIGN(last_byte , root->sectorsize);
4316                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4317                         struct extent_map *hole_em;
4318                         hole_size = last_byte - cur_offset;
4319
4320                         trans = btrfs_start_transaction(root, 3);
4321                         if (IS_ERR(trans)) {
4322                                 err = PTR_ERR(trans);
4323                                 break;
4324                         }
4325
4326                         err = btrfs_drop_extents(trans, root, inode,
4327                                                  cur_offset,
4328                                                  cur_offset + hole_size, 1);
4329                         if (err) {
4330                                 btrfs_abort_transaction(trans, root, err);
4331                                 btrfs_end_transaction(trans, root);
4332                                 break;
4333                         }
4334
4335                         err = btrfs_insert_file_extent(trans, root,
4336                                         btrfs_ino(inode), cur_offset, 0,
4337                                         0, hole_size, 0, hole_size,
4338                                         0, 0, 0);
4339                         if (err) {
4340                                 btrfs_abort_transaction(trans, root, err);
4341                                 btrfs_end_transaction(trans, root);
4342                                 break;
4343                         }
4344
4345                         btrfs_drop_extent_cache(inode, cur_offset,
4346                                                 cur_offset + hole_size - 1, 0);
4347                         hole_em = alloc_extent_map();
4348                         if (!hole_em) {
4349                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4350                                         &BTRFS_I(inode)->runtime_flags);
4351                                 goto next;
4352                         }
4353                         hole_em->start = cur_offset;
4354                         hole_em->len = hole_size;
4355                         hole_em->orig_start = cur_offset;
4356
4357                         hole_em->block_start = EXTENT_MAP_HOLE;
4358                         hole_em->block_len = 0;
4359                         hole_em->orig_block_len = 0;
4360                         hole_em->ram_bytes = hole_size;
4361                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4362                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4363                         hole_em->generation = trans->transid;
4364
4365                         while (1) {
4366                                 write_lock(&em_tree->lock);
4367                                 err = add_extent_mapping(em_tree, hole_em, 1);
4368                                 write_unlock(&em_tree->lock);
4369                                 if (err != -EEXIST)
4370                                         break;
4371                                 btrfs_drop_extent_cache(inode, cur_offset,
4372                                                         cur_offset +
4373                                                         hole_size - 1, 0);
4374                         }
4375                         free_extent_map(hole_em);
4376 next:
4377                         btrfs_update_inode(trans, root, inode);
4378                         btrfs_end_transaction(trans, root);
4379                 }
4380                 free_extent_map(em);
4381                 em = NULL;
4382                 cur_offset = last_byte;
4383                 if (cur_offset >= block_end)
4384                         break;
4385         }
4386
4387         free_extent_map(em);
4388         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4389                              GFP_NOFS);
4390         return err;
4391 }
4392
4393 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4394 {
4395         struct btrfs_root *root = BTRFS_I(inode)->root;
4396         struct btrfs_trans_handle *trans;
4397         loff_t oldsize = i_size_read(inode);
4398         loff_t newsize = attr->ia_size;
4399         int mask = attr->ia_valid;
4400         int ret;
4401
4402         /*
4403          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4404          * special case where we need to update the times despite not having
4405          * these flags set.  For all other operations the VFS set these flags
4406          * explicitly if it wants a timestamp update.
4407          */
4408         if (newsize != oldsize && (!(mask & (ATTR_CTIME | ATTR_MTIME))))
4409                 inode->i_ctime = inode->i_mtime = current_fs_time(inode->i_sb);
4410
4411         if (newsize > oldsize) {
4412                 truncate_pagecache(inode, oldsize, newsize);
4413                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4414                 if (ret)
4415                         return ret;
4416
4417                 trans = btrfs_start_transaction(root, 1);
4418                 if (IS_ERR(trans))
4419                         return PTR_ERR(trans);
4420
4421                 i_size_write(inode, newsize);
4422                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4423                 ret = btrfs_update_inode(trans, root, inode);
4424                 btrfs_end_transaction(trans, root);
4425         } else {
4426
4427                 /*
4428                  * We're truncating a file that used to have good data down to
4429                  * zero. Make sure it gets into the ordered flush list so that
4430                  * any new writes get down to disk quickly.
4431                  */
4432                 if (newsize == 0)
4433                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4434                                 &BTRFS_I(inode)->runtime_flags);
4435
4436                 /*
4437                  * 1 for the orphan item we're going to add
4438                  * 1 for the orphan item deletion.
4439                  */
4440                 trans = btrfs_start_transaction(root, 2);
4441                 if (IS_ERR(trans))
4442                         return PTR_ERR(trans);
4443
4444                 /*
4445                  * We need to do this in case we fail at _any_ point during the
4446                  * actual truncate.  Once we do the truncate_setsize we could
4447                  * invalidate pages which forces any outstanding ordered io to
4448                  * be instantly completed which will give us extents that need
4449                  * to be truncated.  If we fail to get an orphan inode down we
4450                  * could have left over extents that were never meant to live,
4451                  * so we need to garuntee from this point on that everything
4452                  * will be consistent.
4453                  */
4454                 ret = btrfs_orphan_add(trans, inode);
4455                 btrfs_end_transaction(trans, root);
4456                 if (ret)
4457                         return ret;
4458
4459                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
4460                 truncate_setsize(inode, newsize);
4461
4462                 /* Disable nonlocked read DIO to avoid the end less truncate */
4463                 btrfs_inode_block_unlocked_dio(inode);
4464                 inode_dio_wait(inode);
4465                 btrfs_inode_resume_unlocked_dio(inode);
4466
4467                 ret = btrfs_truncate(inode);
4468                 if (ret && inode->i_nlink)
4469                         btrfs_orphan_del(NULL, inode);
4470         }
4471
4472         return ret;
4473 }
4474
4475 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
4476 {
4477         struct inode *inode = dentry->d_inode;
4478         struct btrfs_root *root = BTRFS_I(inode)->root;
4479         int err;
4480
4481         if (btrfs_root_readonly(root))
4482                 return -EROFS;
4483
4484         err = inode_change_ok(inode, attr);
4485         if (err)
4486                 return err;
4487
4488         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
4489                 err = btrfs_setsize(inode, attr);
4490                 if (err)
4491                         return err;
4492         }
4493
4494         if (attr->ia_valid) {
4495                 setattr_copy(inode, attr);
4496                 inode_inc_iversion(inode);
4497                 err = btrfs_dirty_inode(inode);
4498
4499                 if (!err && attr->ia_valid & ATTR_MODE)
4500                         err = btrfs_acl_chmod(inode);
4501         }
4502
4503         return err;
4504 }
4505
4506 void btrfs_evict_inode(struct inode *inode)
4507 {
4508         struct btrfs_trans_handle *trans;
4509         struct btrfs_root *root = BTRFS_I(inode)->root;
4510         struct btrfs_block_rsv *rsv, *global_rsv;
4511         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
4512         int ret;
4513
4514         trace_btrfs_inode_evict(inode);
4515
4516         truncate_inode_pages(&inode->i_data, 0);
4517         if (inode->i_nlink && (btrfs_root_refs(&root->root_item) != 0 ||
4518                                btrfs_is_free_space_inode(inode)))
4519                 goto no_delete;
4520
4521         if (is_bad_inode(inode)) {
4522                 btrfs_orphan_del(NULL, inode);
4523                 goto no_delete;
4524         }
4525         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
4526         btrfs_wait_ordered_range(inode, 0, (u64)-1);
4527
4528         if (root->fs_info->log_root_recovering) {
4529                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
4530                                  &BTRFS_I(inode)->runtime_flags));
4531                 goto no_delete;
4532         }
4533
4534         if (inode->i_nlink > 0) {
4535                 BUG_ON(btrfs_root_refs(&root->root_item) != 0);
4536                 goto no_delete;
4537         }
4538
4539         ret = btrfs_commit_inode_delayed_inode(inode);
4540         if (ret) {
4541                 btrfs_orphan_del(NULL, inode);
4542                 goto no_delete;
4543         }
4544
4545         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
4546         if (!rsv) {
4547                 btrfs_orphan_del(NULL, inode);
4548                 goto no_delete;
4549         }
4550         rsv->size = min_size;
4551         rsv->failfast = 1;
4552         global_rsv = &root->fs_info->global_block_rsv;
4553
4554         btrfs_i_size_write(inode, 0);
4555
4556         /*
4557          * This is a bit simpler than btrfs_truncate since we've already
4558          * reserved our space for our orphan item in the unlink, so we just
4559          * need to reserve some slack space in case we add bytes and update
4560          * inode item when doing the truncate.
4561          */
4562         while (1) {
4563                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
4564                                              BTRFS_RESERVE_FLUSH_LIMIT);
4565
4566                 /*
4567                  * Try and steal from the global reserve since we will
4568                  * likely not use this space anyway, we want to try as
4569                  * hard as possible to get this to work.
4570                  */
4571                 if (ret)
4572                         ret = btrfs_block_rsv_migrate(global_rsv, rsv, min_size);
4573
4574                 if (ret) {
4575                         btrfs_warn(root->fs_info,
4576                                 "Could not get space for a delete, will truncate on mount %d",
4577                                 ret);
4578                         btrfs_orphan_del(NULL, inode);
4579                         btrfs_free_block_rsv(root, rsv);
4580                         goto no_delete;
4581                 }
4582
4583                 trans = btrfs_join_transaction(root);
4584                 if (IS_ERR(trans)) {
4585                         btrfs_orphan_del(NULL, inode);
4586                         btrfs_free_block_rsv(root, rsv);
4587                         goto no_delete;
4588                 }
4589
4590                 trans->block_rsv = rsv;
4591
4592                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
4593                 if (ret != -ENOSPC)
4594                         break;
4595
4596                 trans->block_rsv = &root->fs_info->trans_block_rsv;
4597                 btrfs_end_transaction(trans, root);
4598                 trans = NULL;
4599                 btrfs_btree_balance_dirty(root);
4600         }
4601
4602         btrfs_free_block_rsv(root, rsv);
4603
4604         if (ret == 0) {
4605                 trans->block_rsv = root->orphan_block_rsv;
4606                 ret = btrfs_orphan_del(trans, inode);
4607                 BUG_ON(ret);
4608         }
4609
4610         trans->block_rsv = &root->fs_info->trans_block_rsv;
4611         if (!(root == root->fs_info->tree_root ||
4612               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
4613                 btrfs_return_ino(root, btrfs_ino(inode));
4614
4615         btrfs_end_transaction(trans, root);
4616         btrfs_btree_balance_dirty(root);
4617 no_delete:
4618         btrfs_remove_delayed_node(inode);
4619         clear_inode(inode);
4620         return;
4621 }
4622
4623 /*
4624  * this returns the key found in the dir entry in the location pointer.
4625  * If no dir entries were found, location->objectid is 0.
4626  */
4627 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
4628                                struct btrfs_key *location)
4629 {
4630         const char *name = dentry->d_name.name;
4631         int namelen = dentry->d_name.len;
4632         struct btrfs_dir_item *di;
4633         struct btrfs_path *path;
4634         struct btrfs_root *root = BTRFS_I(dir)->root;
4635         int ret = 0;
4636
4637         path = btrfs_alloc_path();
4638         if (!path)
4639                 return -ENOMEM;
4640
4641         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
4642                                     namelen, 0);
4643         if (IS_ERR(di))
4644                 ret = PTR_ERR(di);
4645
4646         if (IS_ERR_OR_NULL(di))
4647                 goto out_err;
4648
4649         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
4650 out:
4651         btrfs_free_path(path);
4652         return ret;
4653 out_err:
4654         location->objectid = 0;
4655         goto out;
4656 }
4657
4658 /*
4659  * when we hit a tree root in a directory, the btrfs part of the inode
4660  * needs to be changed to reflect the root directory of the tree root.  This
4661  * is kind of like crossing a mount point.
4662  */
4663 static int fixup_tree_root_location(struct btrfs_root *root,
4664                                     struct inode *dir,
4665                                     struct dentry *dentry,
4666                                     struct btrfs_key *location,
4667                                     struct btrfs_root **sub_root)
4668 {
4669         struct btrfs_path *path;
4670         struct btrfs_root *new_root;
4671         struct btrfs_root_ref *ref;
4672         struct extent_buffer *leaf;
4673         int ret;
4674         int err = 0;
4675
4676         path = btrfs_alloc_path();
4677         if (!path) {
4678                 err = -ENOMEM;
4679                 goto out;
4680         }
4681
4682         err = -ENOENT;
4683         ret = btrfs_find_root_ref(root->fs_info->tree_root, path,
4684                                   BTRFS_I(dir)->root->root_key.objectid,
4685                                   location->objectid);
4686         if (ret) {
4687                 if (ret < 0)
4688                         err = ret;
4689                 goto out;
4690         }
4691
4692         leaf = path->nodes[0];
4693         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
4694         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
4695             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
4696                 goto out;
4697
4698         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
4699                                    (unsigned long)(ref + 1),
4700                                    dentry->d_name.len);
4701         if (ret)
4702                 goto out;
4703
4704         btrfs_release_path(path);
4705
4706         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
4707         if (IS_ERR(new_root)) {
4708                 err = PTR_ERR(new_root);
4709                 goto out;
4710         }
4711
4712         *sub_root = new_root;
4713         location->objectid = btrfs_root_dirid(&new_root->root_item);
4714         location->type = BTRFS_INODE_ITEM_KEY;
4715         location->offset = 0;
4716         err = 0;
4717 out:
4718         btrfs_free_path(path);
4719         return err;
4720 }
4721
4722 static void inode_tree_add(struct inode *inode)
4723 {
4724         struct btrfs_root *root = BTRFS_I(inode)->root;
4725         struct btrfs_inode *entry;
4726         struct rb_node **p;
4727         struct rb_node *parent;
4728         u64 ino = btrfs_ino(inode);
4729
4730         if (inode_unhashed(inode))
4731                 return;
4732 again:
4733         parent = NULL;
4734         spin_lock(&root->inode_lock);
4735         p = &root->inode_tree.rb_node;
4736         while (*p) {
4737                 parent = *p;
4738                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
4739
4740                 if (ino < btrfs_ino(&entry->vfs_inode))
4741                         p = &parent->rb_left;
4742                 else if (ino > btrfs_ino(&entry->vfs_inode))
4743                         p = &parent->rb_right;
4744                 else {
4745                         WARN_ON(!(entry->vfs_inode.i_state &
4746                                   (I_WILL_FREE | I_FREEING)));
4747                         rb_erase(parent, &root->inode_tree);
4748                         RB_CLEAR_NODE(parent);
4749                         spin_unlock(&root->inode_lock);
4750                         goto again;
4751                 }
4752         }
4753         rb_link_node(&BTRFS_I(inode)->rb_node, parent, p);
4754         rb_insert_color(&BTRFS_I(inode)->rb_node, &root->inode_tree);
4755         spin_unlock(&root->inode_lock);
4756 }
4757
4758 static void inode_tree_del(struct inode *inode)
4759 {
4760         struct btrfs_root *root = BTRFS_I(inode)->root;
4761         int empty = 0;
4762
4763         spin_lock(&root->inode_lock);
4764         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
4765                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
4766                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
4767                 empty = RB_EMPTY_ROOT(&root->inode_tree);
4768         }
4769         spin_unlock(&root->inode_lock);
4770
4771         /*
4772          * Free space cache has inodes in the tree root, but the tree root has a
4773          * root_refs of 0, so this could end up dropping the tree root as a
4774          * snapshot, so we need the extra !root->fs_info->tree_root check to
4775          * make sure we don't drop it.
4776          */
4777         if (empty && btrfs_root_refs(&root->root_item) == 0 &&
4778             root != root->fs_info->tree_root) {
4779                 synchronize_srcu(&root->fs_info->subvol_srcu);
4780                 spin_lock(&root->inode_lock);
4781                 empty = RB_EMPTY_ROOT(&root->inode_tree);
4782                 spin_unlock(&root->inode_lock);
4783                 if (empty)
4784                         btrfs_add_dead_root(root);
4785         }
4786 }
4787
4788 void btrfs_invalidate_inodes(struct btrfs_root *root)
4789 {
4790         struct rb_node *node;
4791         struct rb_node *prev;
4792         struct btrfs_inode *entry;
4793         struct inode *inode;
4794         u64 objectid = 0;
4795
4796         WARN_ON(btrfs_root_refs(&root->root_item) != 0);
4797
4798         spin_lock(&root->inode_lock);
4799 again:
4800         node = root->inode_tree.rb_node;
4801         prev = NULL;
4802         while (node) {
4803                 prev = node;
4804                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4805
4806                 if (objectid < btrfs_ino(&entry->vfs_inode))
4807                         node = node->rb_left;
4808                 else if (objectid > btrfs_ino(&entry->vfs_inode))
4809                         node = node->rb_right;
4810                 else
4811                         break;
4812         }
4813         if (!node) {
4814                 while (prev) {
4815                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
4816                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
4817                                 node = prev;
4818                                 break;
4819                         }
4820                         prev = rb_next(prev);
4821                 }
4822         }
4823         while (node) {
4824                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4825                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
4826                 inode = igrab(&entry->vfs_inode);
4827                 if (inode) {
4828                         spin_unlock(&root->inode_lock);
4829                         if (atomic_read(&inode->i_count) > 1)
4830                                 d_prune_aliases(inode);
4831                         /*
4832                          * btrfs_drop_inode will have it removed from
4833                          * the inode cache when its usage count
4834                          * hits zero.
4835                          */
4836                         iput(inode);
4837                         cond_resched();
4838                         spin_lock(&root->inode_lock);
4839                         goto again;
4840                 }
4841
4842                 if (cond_resched_lock(&root->inode_lock))
4843                         goto again;
4844
4845                 node = rb_next(node);
4846         }
4847         spin_unlock(&root->inode_lock);
4848 }
4849
4850 static int btrfs_init_locked_inode(struct inode *inode, void *p)
4851 {
4852         struct btrfs_iget_args *args = p;
4853         inode->i_ino = args->ino;
4854         BTRFS_I(inode)->root = args->root;
4855         return 0;
4856 }
4857
4858 static int btrfs_find_actor(struct inode *inode, void *opaque)
4859 {
4860         struct btrfs_iget_args *args = opaque;
4861         return args->ino == btrfs_ino(inode) &&
4862                 args->root == BTRFS_I(inode)->root;
4863 }
4864
4865 static struct inode *btrfs_iget_locked(struct super_block *s,
4866                                        u64 objectid,
4867                                        struct btrfs_root *root)
4868 {
4869         struct inode *inode;
4870         struct btrfs_iget_args args;
4871         args.ino = objectid;
4872         args.root = root;
4873
4874         inode = iget5_locked(s, objectid, btrfs_find_actor,
4875                              btrfs_init_locked_inode,
4876                              (void *)&args);
4877         return inode;
4878 }
4879
4880 /* Get an inode object given its location and corresponding root.
4881  * Returns in *is_new if the inode was read from disk
4882  */
4883 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
4884                          struct btrfs_root *root, int *new)
4885 {
4886         struct inode *inode;
4887
4888         inode = btrfs_iget_locked(s, location->objectid, root);
4889         if (!inode)
4890                 return ERR_PTR(-ENOMEM);
4891
4892         if (inode->i_state & I_NEW) {
4893                 BTRFS_I(inode)->root = root;
4894                 memcpy(&BTRFS_I(inode)->location, location, sizeof(*location));
4895                 btrfs_read_locked_inode(inode);
4896                 if (!is_bad_inode(inode)) {
4897                         inode_tree_add(inode);
4898                         unlock_new_inode(inode);
4899                         if (new)
4900                                 *new = 1;
4901                 } else {
4902                         unlock_new_inode(inode);
4903                         iput(inode);
4904                         inode = ERR_PTR(-ESTALE);
4905                 }
4906         }
4907
4908         return inode;
4909 }
4910
4911 static struct inode *new_simple_dir(struct super_block *s,
4912                                     struct btrfs_key *key,
4913                                     struct btrfs_root *root)
4914 {
4915         struct inode *inode = new_inode(s);
4916
4917         if (!inode)
4918                 return ERR_PTR(-ENOMEM);
4919
4920         BTRFS_I(inode)->root = root;
4921         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
4922         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
4923
4924         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
4925         inode->i_op = &btrfs_dir_ro_inode_operations;
4926         inode->i_fop = &simple_dir_operations;
4927         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
4928         inode->i_mtime = inode->i_atime = inode->i_ctime = CURRENT_TIME;
4929
4930         return inode;
4931 }
4932
4933 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
4934 {
4935         struct inode *inode;
4936         struct btrfs_root *root = BTRFS_I(dir)->root;
4937         struct btrfs_root *sub_root = root;
4938         struct btrfs_key location;
4939         int index;
4940         int ret = 0;
4941
4942         if (dentry->d_name.len > BTRFS_NAME_LEN)
4943                 return ERR_PTR(-ENAMETOOLONG);
4944
4945         ret = btrfs_inode_by_name(dir, dentry, &location);
4946         if (ret < 0)
4947                 return ERR_PTR(ret);
4948
4949         if (location.objectid == 0)
4950                 return NULL;
4951
4952         if (location.type == BTRFS_INODE_ITEM_KEY) {
4953                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
4954                 return inode;
4955         }
4956
4957         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
4958
4959         index = srcu_read_lock(&root->fs_info->subvol_srcu);
4960         ret = fixup_tree_root_location(root, dir, dentry,
4961                                        &location, &sub_root);
4962         if (ret < 0) {
4963                 if (ret != -ENOENT)
4964                         inode = ERR_PTR(ret);
4965                 else
4966                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
4967         } else {
4968                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
4969         }
4970         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
4971
4972         if (!IS_ERR(inode) && root != sub_root) {
4973                 down_read(&root->fs_info->cleanup_work_sem);
4974                 if (!(inode->i_sb->s_flags & MS_RDONLY))
4975                         ret = btrfs_orphan_cleanup(sub_root);
4976                 up_read(&root->fs_info->cleanup_work_sem);
4977                 if (ret) {
4978                         iput(inode);
4979                         inode = ERR_PTR(ret);
4980                 }
4981         }
4982
4983         return inode;
4984 }
4985
4986 static int btrfs_dentry_delete(const struct dentry *dentry)
4987 {
4988         struct btrfs_root *root;
4989         struct inode *inode = dentry->d_inode;
4990
4991         if (!inode && !IS_ROOT(dentry))
4992                 inode = dentry->d_parent->d_inode;
4993
4994         if (inode) {
4995                 root = BTRFS_I(inode)->root;
4996                 if (btrfs_root_refs(&root->root_item) == 0)
4997                         return 1;
4998
4999                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5000                         return 1;
5001         }
5002         return 0;
5003 }
5004
5005 static void btrfs_dentry_release(struct dentry *dentry)
5006 {
5007         if (dentry->d_fsdata)
5008                 kfree(dentry->d_fsdata);
5009 }
5010
5011 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5012                                    unsigned int flags)
5013 {
5014         struct dentry *ret;
5015
5016         ret = d_splice_alias(btrfs_lookup_dentry(dir, dentry), dentry);
5017         return ret;
5018 }
5019
5020 unsigned char btrfs_filetype_table[] = {
5021         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5022 };
5023
5024 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5025 {
5026         struct inode *inode = file_inode(file);
5027         struct btrfs_root *root = BTRFS_I(inode)->root;
5028         struct btrfs_item *item;
5029         struct btrfs_dir_item *di;
5030         struct btrfs_key key;
5031         struct btrfs_key found_key;
5032         struct btrfs_path *path;
5033         struct list_head ins_list;
5034         struct list_head del_list;
5035         int ret;
5036         struct extent_buffer *leaf;
5037         int slot;
5038         unsigned char d_type;
5039         int over = 0;
5040         u32 di_cur;
5041         u32 di_total;
5042         u32 di_len;
5043         int key_type = BTRFS_DIR_INDEX_KEY;
5044         char tmp_name[32];
5045         char *name_ptr;
5046         int name_len;
5047         int is_curr = 0;        /* ctx->pos points to the current index? */
5048
5049         /* FIXME, use a real flag for deciding about the key type */
5050         if (root->fs_info->tree_root == root)
5051                 key_type = BTRFS_DIR_ITEM_KEY;
5052
5053         if (!dir_emit_dots(file, ctx))
5054                 return 0;
5055
5056         path = btrfs_alloc_path();
5057         if (!path)
5058                 return -ENOMEM;
5059
5060         path->reada = 1;
5061
5062         if (key_type == BTRFS_DIR_INDEX_KEY) {
5063                 INIT_LIST_HEAD(&ins_list);
5064                 INIT_LIST_HEAD(&del_list);
5065                 btrfs_get_delayed_items(inode, &ins_list, &del_list);
5066         }
5067
5068         btrfs_set_key_type(&key, key_type);
5069         key.offset = ctx->pos;
5070         key.objectid = btrfs_ino(inode);
5071
5072         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5073         if (ret < 0)
5074                 goto err;
5075
5076         while (1) {
5077                 leaf = path->nodes[0];
5078                 slot = path->slots[0];
5079                 if (slot >= btrfs_header_nritems(leaf)) {
5080                         ret = btrfs_next_leaf(root, path);
5081                         if (ret < 0)
5082                                 goto err;
5083                         else if (ret > 0)
5084                                 break;
5085                         continue;
5086                 }
5087
5088                 item = btrfs_item_nr(leaf, slot);
5089                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5090
5091                 if (found_key.objectid != key.objectid)
5092                         break;
5093                 if (btrfs_key_type(&found_key) != key_type)
5094                         break;
5095                 if (found_key.offset < ctx->pos)
5096                         goto next;
5097                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5098                     btrfs_should_delete_dir_index(&del_list,
5099                                                   found_key.offset))
5100                         goto next;
5101
5102                 ctx->pos = found_key.offset;
5103                 is_curr = 1;
5104
5105                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5106                 di_cur = 0;
5107                 di_total = btrfs_item_size(leaf, item);
5108
5109                 while (di_cur < di_total) {
5110                         struct btrfs_key location;
5111
5112                         if (verify_dir_item(root, leaf, di))
5113                                 break;
5114
5115                         name_len = btrfs_dir_name_len(leaf, di);
5116                         if (name_len <= sizeof(tmp_name)) {
5117                                 name_ptr = tmp_name;
5118                         } else {
5119                                 name_ptr = kmalloc(name_len, GFP_NOFS);
5120                                 if (!name_ptr) {
5121                                         ret = -ENOMEM;
5122                                         goto err;
5123                                 }
5124                         }
5125                         read_extent_buffer(leaf, name_ptr,
5126                                            (unsigned long)(di + 1), name_len);
5127
5128                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5129                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5130
5131
5132                         /* is this a reference to our own snapshot? If so
5133                          * skip it.
5134                          *
5135                          * In contrast to old kernels, we insert the snapshot's
5136                          * dir item and dir index after it has been created, so
5137                          * we won't find a reference to our own snapshot. We
5138                          * still keep the following code for backward
5139                          * compatibility.
5140                          */
5141                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5142                             location.objectid == root->root_key.objectid) {
5143                                 over = 0;
5144                                 goto skip;
5145                         }
5146                         over = !dir_emit(ctx, name_ptr, name_len,
5147                                        location.objectid, d_type);
5148
5149 skip:
5150                         if (name_ptr != tmp_name)
5151                                 kfree(name_ptr);
5152
5153                         if (over)
5154                                 goto nopos;
5155                         di_len = btrfs_dir_name_len(leaf, di) +
5156                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5157                         di_cur += di_len;
5158                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5159                 }
5160 next:
5161                 path->slots[0]++;
5162         }
5163
5164         if (key_type == BTRFS_DIR_INDEX_KEY) {
5165                 if (is_curr)
5166                         ctx->pos++;
5167                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5168                 if (ret)
5169                         goto nopos;
5170         }
5171
5172         /* Reached end of directory/root. Bump pos past the last item. */
5173         if (key_type == BTRFS_DIR_INDEX_KEY)
5174                 /*
5175                  * 32-bit glibc will use getdents64, but then strtol -
5176                  * so the last number we can serve is this.
5177                  */
5178                 ctx->pos = 0x7fffffff;
5179         else
5180                 ctx->pos++;
5181 nopos:
5182         ret = 0;
5183 err:
5184         if (key_type == BTRFS_DIR_INDEX_KEY)
5185                 btrfs_put_delayed_items(&ins_list, &del_list);
5186         btrfs_free_path(path);
5187         return ret;
5188 }
5189
5190 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5191 {
5192         struct btrfs_root *root = BTRFS_I(inode)->root;
5193         struct btrfs_trans_handle *trans;
5194         int ret = 0;
5195         bool nolock = false;
5196
5197         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5198                 return 0;
5199
5200         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5201                 nolock = true;
5202
5203         if (wbc->sync_mode == WB_SYNC_ALL) {
5204                 if (nolock)
5205                         trans = btrfs_join_transaction_nolock(root);
5206                 else
5207                         trans = btrfs_join_transaction(root);
5208                 if (IS_ERR(trans))
5209                         return PTR_ERR(trans);
5210                 ret = btrfs_commit_transaction(trans, root);
5211         }
5212         return ret;
5213 }
5214
5215 /*
5216  * This is somewhat expensive, updating the tree every time the
5217  * inode changes.  But, it is most likely to find the inode in cache.
5218  * FIXME, needs more benchmarking...there are no reasons other than performance
5219  * to keep or drop this code.
5220  */
5221 static int btrfs_dirty_inode(struct inode *inode)
5222 {
5223         struct btrfs_root *root = BTRFS_I(inode)->root;
5224         struct btrfs_trans_handle *trans;
5225         int ret;
5226
5227         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5228                 return 0;
5229
5230         trans = btrfs_join_transaction(root);
5231         if (IS_ERR(trans))
5232                 return PTR_ERR(trans);
5233
5234         ret = btrfs_update_inode(trans, root, inode);
5235         if (ret && ret == -ENOSPC) {
5236                 /* whoops, lets try again with the full transaction */
5237                 btrfs_end_transaction(trans, root);
5238                 trans = btrfs_start_transaction(root, 1);
5239                 if (IS_ERR(trans))
5240                         return PTR_ERR(trans);
5241
5242                 ret = btrfs_update_inode(trans, root, inode);
5243         }
5244         btrfs_end_transaction(trans, root);
5245         if (BTRFS_I(inode)->delayed_node)
5246                 btrfs_balance_delayed_items(root);
5247
5248         return ret;
5249 }
5250
5251 /*
5252  * This is a copy of file_update_time.  We need this so we can return error on
5253  * ENOSPC for updating the inode in the case of file write and mmap writes.
5254  */
5255 static int btrfs_update_time(struct inode *inode, struct timespec *now,
5256                              int flags)
5257 {
5258         struct btrfs_root *root = BTRFS_I(inode)->root;
5259
5260         if (btrfs_root_readonly(root))
5261                 return -EROFS;
5262
5263         if (flags & S_VERSION)
5264                 inode_inc_iversion(inode);
5265         if (flags & S_CTIME)
5266                 inode->i_ctime = *now;
5267         if (flags & S_MTIME)
5268                 inode->i_mtime = *now;
5269         if (flags & S_ATIME)
5270                 inode->i_atime = *now;
5271         return btrfs_dirty_inode(inode);
5272 }
5273
5274 /*
5275  * find the highest existing sequence number in a directory
5276  * and then set the in-memory index_cnt variable to reflect
5277  * free sequence numbers
5278  */
5279 static int btrfs_set_inode_index_count(struct inode *inode)
5280 {
5281         struct btrfs_root *root = BTRFS_I(inode)->root;
5282         struct btrfs_key key, found_key;
5283         struct btrfs_path *path;
5284         struct extent_buffer *leaf;
5285         int ret;
5286
5287         key.objectid = btrfs_ino(inode);
5288         btrfs_set_key_type(&key, BTRFS_DIR_INDEX_KEY);
5289         key.offset = (u64)-1;
5290
5291         path = btrfs_alloc_path();
5292         if (!path)
5293                 return -ENOMEM;
5294
5295         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5296         if (ret < 0)
5297                 goto out;
5298         /* FIXME: we should be able to handle this */
5299         if (ret == 0)
5300                 goto out;
5301         ret = 0;
5302
5303         /*
5304          * MAGIC NUMBER EXPLANATION:
5305          * since we search a directory based on f_pos we have to start at 2
5306          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
5307          * else has to start at 2
5308          */
5309         if (path->slots[0] == 0) {
5310                 BTRFS_I(inode)->index_cnt = 2;
5311                 goto out;
5312         }
5313
5314         path->slots[0]--;
5315
5316         leaf = path->nodes[0];
5317         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5318
5319         if (found_key.objectid != btrfs_ino(inode) ||
5320             btrfs_key_type(&found_key) != BTRFS_DIR_INDEX_KEY) {
5321                 BTRFS_I(inode)->index_cnt = 2;
5322                 goto out;
5323         }
5324
5325         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
5326 out:
5327         btrfs_free_path(path);
5328         return ret;
5329 }
5330
5331 /*
5332  * helper to find a free sequence number in a given directory.  This current
5333  * code is very simple, later versions will do smarter things in the btree
5334  */
5335 int btrfs_set_inode_index(struct inode *dir, u64 *index)
5336 {
5337         int ret = 0;
5338
5339         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
5340                 ret = btrfs_inode_delayed_dir_index_count(dir);
5341                 if (ret) {
5342                         ret = btrfs_set_inode_index_count(dir);
5343                         if (ret)
5344                                 return ret;
5345                 }
5346         }
5347
5348         *index = BTRFS_I(dir)->index_cnt;
5349         BTRFS_I(dir)->index_cnt++;
5350
5351         return ret;
5352 }
5353
5354 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
5355                                      struct btrfs_root *root,
5356                                      struct inode *dir,
5357                                      const char *name, int name_len,
5358                                      u64 ref_objectid, u64 objectid,
5359                                      umode_t mode, u64 *index)
5360 {
5361         struct inode *inode;
5362         struct btrfs_inode_item *inode_item;
5363         struct btrfs_key *location;
5364         struct btrfs_path *path;
5365         struct btrfs_inode_ref *ref;
5366         struct btrfs_key key[2];
5367         u32 sizes[2];
5368         unsigned long ptr;
5369         int ret;
5370         int owner;
5371
5372         path = btrfs_alloc_path();
5373         if (!path)
5374                 return ERR_PTR(-ENOMEM);
5375
5376         inode = new_inode(root->fs_info->sb);
5377         if (!inode) {
5378                 btrfs_free_path(path);
5379                 return ERR_PTR(-ENOMEM);
5380         }
5381
5382         /*
5383          * we have to initialize this early, so we can reclaim the inode
5384          * number if we fail afterwards in this function.
5385          */
5386         inode->i_ino = objectid;
5387
5388         if (dir) {
5389                 trace_btrfs_inode_request(dir);
5390
5391                 ret = btrfs_set_inode_index(dir, index);
5392                 if (ret) {
5393                         btrfs_free_path(path);
5394                         iput(inode);
5395                         return ERR_PTR(ret);
5396                 }
5397         }
5398         /*
5399          * index_cnt is ignored for everything but a dir,
5400          * btrfs_get_inode_index_count has an explanation for the magic
5401          * number
5402          */
5403         BTRFS_I(inode)->index_cnt = 2;
5404         BTRFS_I(inode)->root = root;
5405         BTRFS_I(inode)->generation = trans->transid;
5406         inode->i_generation = BTRFS_I(inode)->generation;
5407
5408         /*
5409          * We could have gotten an inode number from somebody who was fsynced
5410          * and then removed in this same transaction, so let's just set full
5411          * sync since it will be a full sync anyway and this will blow away the
5412          * old info in the log.
5413          */
5414         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
5415
5416         if (S_ISDIR(mode))
5417                 owner = 0;
5418         else
5419                 owner = 1;
5420
5421         key[0].objectid = objectid;
5422         btrfs_set_key_type(&key[0], BTRFS_INODE_ITEM_KEY);
5423         key[0].offset = 0;
5424
5425         /*
5426          * Start new inodes with an inode_ref. This is slightly more
5427          * efficient for small numbers of hard links since they will
5428          * be packed into one item. Extended refs will kick in if we
5429          * add more hard links than can fit in the ref item.
5430          */
5431         key[1].objectid = objectid;
5432         btrfs_set_key_type(&key[1], BTRFS_INODE_REF_KEY);
5433         key[1].offset = ref_objectid;
5434
5435         sizes[0] = sizeof(struct btrfs_inode_item);
5436         sizes[1] = name_len + sizeof(*ref);
5437
5438         path->leave_spinning = 1;
5439         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, 2);
5440         if (ret != 0)
5441                 goto fail;
5442
5443         inode_init_owner(inode, dir, mode);
5444         inode_set_bytes(inode, 0);
5445         inode->i_mtime = inode->i_atime = inode->i_ctime = CURRENT_TIME;
5446         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
5447                                   struct btrfs_inode_item);
5448         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
5449                              sizeof(*inode_item));
5450         fill_inode_item(trans, path->nodes[0], inode_item, inode);
5451
5452         ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
5453                              struct btrfs_inode_ref);
5454         btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
5455         btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
5456         ptr = (unsigned long)(ref + 1);
5457         write_extent_buffer(path->nodes[0], name, ptr, name_len);
5458
5459         btrfs_mark_buffer_dirty(path->nodes[0]);
5460         btrfs_free_path(path);
5461
5462         location = &BTRFS_I(inode)->location;
5463         location->objectid = objectid;
5464         location->offset = 0;
5465         btrfs_set_key_type(location, BTRFS_INODE_ITEM_KEY);
5466
5467         btrfs_inherit_iflags(inode, dir);
5468
5469         if (S_ISREG(mode)) {
5470                 if (btrfs_test_opt(root, NODATASUM))
5471                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
5472                 if (btrfs_test_opt(root, NODATACOW))
5473                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
5474                                 BTRFS_INODE_NODATASUM;
5475         }
5476
5477         insert_inode_hash(inode);
5478         inode_tree_add(inode);
5479
5480         trace_btrfs_inode_new(inode);
5481         btrfs_set_inode_last_trans(trans, inode);
5482
5483         btrfs_update_root_times(trans, root);
5484
5485         return inode;
5486 fail:
5487         if (dir)
5488                 BTRFS_I(dir)->index_cnt--;
5489         btrfs_free_path(path);
5490         iput(inode);
5491         return ERR_PTR(ret);
5492 }
5493
5494 static inline u8 btrfs_inode_type(struct inode *inode)
5495 {
5496         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
5497 }
5498
5499 /*
5500  * utility function to add 'inode' into 'parent_inode' with
5501  * a give name and a given sequence number.
5502  * if 'add_backref' is true, also insert a backref from the
5503  * inode to the parent directory.
5504  */
5505 int btrfs_add_link(struct btrfs_trans_handle *trans,
5506                    struct inode *parent_inode, struct inode *inode,
5507                    const char *name, int name_len, int add_backref, u64 index)
5508 {
5509         int ret = 0;
5510         struct btrfs_key key;
5511         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
5512         u64 ino = btrfs_ino(inode);
5513         u64 parent_ino = btrfs_ino(parent_inode);
5514
5515         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
5516                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
5517         } else {
5518                 key.objectid = ino;
5519                 btrfs_set_key_type(&key, BTRFS_INODE_ITEM_KEY);
5520                 key.offset = 0;
5521         }
5522
5523         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
5524                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
5525                                          key.objectid, root->root_key.objectid,
5526                                          parent_ino, index, name, name_len);
5527         } else if (add_backref) {
5528                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
5529                                              parent_ino, index);
5530         }
5531
5532         /* Nothing to clean up yet */
5533         if (ret)
5534                 return ret;
5535
5536         ret = btrfs_insert_dir_item(trans, root, name, name_len,
5537                                     parent_inode, &key,
5538                                     btrfs_inode_type(inode), index);
5539         if (ret == -EEXIST || ret == -EOVERFLOW)
5540                 goto fail_dir_item;
5541         else if (ret) {
5542                 btrfs_abort_transaction(trans, root, ret);
5543                 return ret;
5544         }
5545
5546         btrfs_i_size_write(parent_inode, parent_inode->i_size +
5547                            name_len * 2);
5548         inode_inc_iversion(parent_inode);
5549         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
5550         ret = btrfs_update_inode(trans, root, parent_inode);
5551         if (ret)
5552                 btrfs_abort_transaction(trans, root, ret);
5553         return ret;
5554
5555 fail_dir_item:
5556         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
5557                 u64 local_index;
5558                 int err;
5559                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
5560                                  key.objectid, root->root_key.objectid,
5561                                  parent_ino, &local_index, name, name_len);
5562
5563         } else if (add_backref) {
5564                 u64 local_index;
5565                 int err;
5566
5567                 err = btrfs_del_inode_ref(trans, root, name, name_len,
5568                                           ino, parent_ino, &local_index);
5569         }
5570         return ret;
5571 }
5572
5573 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
5574                             struct inode *dir, struct dentry *dentry,
5575                             struct inode *inode, int backref, u64 index)
5576 {
5577         int err = btrfs_add_link(trans, dir, inode,
5578                                  dentry->d_name.name, dentry->d_name.len,
5579                                  backref, index);
5580         if (err > 0)
5581                 err = -EEXIST;
5582         return err;
5583 }
5584
5585 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
5586                         umode_t mode, dev_t rdev)
5587 {
5588         struct btrfs_trans_handle *trans;
5589         struct btrfs_root *root = BTRFS_I(dir)->root;
5590         struct inode *inode = NULL;
5591         int err;
5592         int drop_inode = 0;
5593         u64 objectid;
5594         u64 index = 0;
5595
5596         if (!new_valid_dev(rdev))
5597                 return -EINVAL;
5598
5599         /*
5600          * 2 for inode item and ref
5601          * 2 for dir items
5602          * 1 for xattr if selinux is on
5603          */
5604         trans = btrfs_start_transaction(root, 5);
5605         if (IS_ERR(trans))
5606                 return PTR_ERR(trans);
5607
5608         err = btrfs_find_free_ino(root, &objectid);
5609         if (err)
5610                 goto out_unlock;
5611
5612         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
5613                                 dentry->d_name.len, btrfs_ino(dir), objectid,
5614                                 mode, &index);
5615         if (IS_ERR(inode)) {
5616                 err = PTR_ERR(inode);
5617                 goto out_unlock;
5618         }
5619
5620         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
5621         if (err) {
5622                 drop_inode = 1;
5623                 goto out_unlock;
5624         }
5625
5626         /*
5627         * If the active LSM wants to access the inode during
5628         * d_instantiate it needs these. Smack checks to see
5629         * if the filesystem supports xattrs by looking at the
5630         * ops vector.
5631         */
5632
5633         inode->i_op = &btrfs_special_inode_operations;
5634         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
5635         if (err)
5636                 drop_inode = 1;
5637         else {
5638                 init_special_inode(inode, inode->i_mode, rdev);
5639                 btrfs_update_inode(trans, root, inode);
5640                 d_instantiate(dentry, inode);
5641         }
5642 out_unlock:
5643         btrfs_end_transaction(trans, root);
5644         btrfs_btree_balance_dirty(root);
5645         if (drop_inode) {
5646                 inode_dec_link_count(inode);
5647                 iput(inode);
5648         }
5649         return err;
5650 }
5651
5652 static int btrfs_create(struct inode *dir, struct dentry *dentry,
5653                         umode_t mode, bool excl)
5654 {
5655         struct btrfs_trans_handle *trans;
5656         struct btrfs_root *root = BTRFS_I(dir)->root;
5657         struct inode *inode = NULL;
5658         int drop_inode_on_err = 0;
5659         int err;
5660         u64 objectid;
5661         u64 index = 0;
5662
5663         /*
5664          * 2 for inode item and ref
5665          * 2 for dir items
5666          * 1 for xattr if selinux is on
5667          */
5668         trans = btrfs_start_transaction(root, 5);
5669         if (IS_ERR(trans))
5670                 return PTR_ERR(trans);
5671
5672         err = btrfs_find_free_ino(root, &objectid);
5673         if (err)
5674                 goto out_unlock;
5675
5676         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
5677                                 dentry->d_name.len, btrfs_ino(dir), objectid,
5678                                 mode, &index);
5679         if (IS_ERR(inode)) {
5680                 err = PTR_ERR(inode);
5681                 goto out_unlock;
5682         }
5683         drop_inode_on_err = 1;
5684
5685         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
5686         if (err)
5687                 goto out_unlock;
5688
5689         err = btrfs_update_inode(trans, root, inode);
5690         if (err)
5691                 goto out_unlock;
5692
5693         /*
5694         * If the active LSM wants to access the inode during
5695         * d_instantiate it needs these. Smack checks to see
5696         * if the filesystem supports xattrs by looking at the
5697         * ops vector.
5698         */
5699         inode->i_fop = &btrfs_file_operations;
5700         inode->i_op = &btrfs_file_inode_operations;
5701
5702         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
5703         if (err)
5704                 goto out_unlock;
5705
5706         inode->i_mapping->a_ops = &btrfs_aops;
5707         inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
5708         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
5709         d_instantiate(dentry, inode);
5710
5711 out_unlock:
5712         btrfs_end_transaction(trans, root);
5713         if (err && drop_inode_on_err) {
5714                 inode_dec_link_count(inode);
5715                 iput(inode);
5716         }
5717         btrfs_btree_balance_dirty(root);
5718         return err;
5719 }
5720
5721 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
5722                       struct dentry *dentry)
5723 {
5724         struct btrfs_trans_handle *trans;
5725         struct btrfs_root *root = BTRFS_I(dir)->root;
5726         struct inode *inode = old_dentry->d_inode;
5727         u64 index;
5728         int err;
5729         int drop_inode = 0;
5730
5731         /* do not allow sys_link's with other subvols of the same device */
5732         if (root->objectid != BTRFS_I(inode)->root->objectid)
5733                 return -EXDEV;
5734
5735         if (inode->i_nlink >= BTRFS_LINK_MAX)
5736                 return -EMLINK;
5737
5738         err = btrfs_set_inode_index(dir, &index);
5739         if (err)
5740                 goto fail;
5741
5742         /*
5743          * 2 items for inode and inode ref
5744          * 2 items for dir items
5745          * 1 item for parent inode
5746          */
5747         trans = btrfs_start_transaction(root, 5);
5748         if (IS_ERR(trans)) {
5749                 err = PTR_ERR(trans);
5750                 goto fail;
5751         }
5752
5753         btrfs_inc_nlink(inode);
5754         inode_inc_iversion(inode);
5755         inode->i_ctime = CURRENT_TIME;
5756         ihold(inode);
5757         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
5758
5759         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
5760
5761         if (err) {
5762                 drop_inode = 1;
5763         } else {
5764                 struct dentry *parent = dentry->d_parent;
5765                 err = btrfs_update_inode(trans, root, inode);
5766                 if (err)
5767                         goto fail;
5768                 d_instantiate(dentry, inode);
5769                 btrfs_log_new_name(trans, inode, NULL, parent);
5770         }
5771
5772         btrfs_end_transaction(trans, root);
5773 fail:
5774         if (drop_inode) {
5775                 inode_dec_link_count(inode);
5776                 iput(inode);
5777         }
5778         btrfs_btree_balance_dirty(root);
5779         return err;
5780 }
5781
5782 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
5783 {
5784         struct inode *inode = NULL;
5785         struct btrfs_trans_handle *trans;
5786         struct btrfs_root *root = BTRFS_I(dir)->root;
5787         int err = 0;
5788         int drop_on_err = 0;
5789         u64 objectid = 0;
5790         u64 index = 0;
5791
5792         /*
5793          * 2 items for inode and ref
5794          * 2 items for dir items
5795          * 1 for xattr if selinux is on
5796          */
5797         trans = btrfs_start_transaction(root, 5);
5798         if (IS_ERR(trans))
5799                 return PTR_ERR(trans);
5800
5801         err = btrfs_find_free_ino(root, &objectid);
5802         if (err)
5803                 goto out_fail;
5804
5805         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
5806                                 dentry->d_name.len, btrfs_ino(dir), objectid,
5807                                 S_IFDIR | mode, &index);
5808         if (IS_ERR(inode)) {
5809                 err = PTR_ERR(inode);
5810                 goto out_fail;
5811         }
5812
5813         drop_on_err = 1;
5814
5815         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
5816         if (err)
5817                 goto out_fail;
5818
5819         inode->i_op = &btrfs_dir_inode_operations;
5820         inode->i_fop = &btrfs_dir_file_operations;
5821
5822         btrfs_i_size_write(inode, 0);
5823         err = btrfs_update_inode(trans, root, inode);
5824         if (err)
5825                 goto out_fail;
5826
5827         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
5828                              dentry->d_name.len, 0, index);
5829         if (err)
5830                 goto out_fail;
5831
5832         d_instantiate(dentry, inode);
5833         drop_on_err = 0;
5834
5835 out_fail:
5836         btrfs_end_transaction(trans, root);
5837         if (drop_on_err)
5838                 iput(inode);
5839         btrfs_btree_balance_dirty(root);
5840         return err;
5841 }
5842
5843 /* helper for btfs_get_extent.  Given an existing extent in the tree,
5844  * and an extent that you want to insert, deal with overlap and insert
5845  * the new extent into the tree.
5846  */
5847 static int merge_extent_mapping(struct extent_map_tree *em_tree,
5848                                 struct extent_map *existing,
5849                                 struct extent_map *em,
5850                                 u64 map_start, u64 map_len)
5851 {
5852         u64 start_diff;
5853
5854         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
5855         start_diff = map_start - em->start;
5856         em->start = map_start;
5857         em->len = map_len;
5858         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
5859             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
5860                 em->block_start += start_diff;
5861                 em->block_len -= start_diff;
5862         }
5863         return add_extent_mapping(em_tree, em, 0);
5864 }
5865
5866 static noinline int uncompress_inline(struct btrfs_path *path,
5867                                       struct inode *inode, struct page *page,
5868                                       size_t pg_offset, u64 extent_offset,
5869                                       struct btrfs_file_extent_item *item)
5870 {
5871         int ret;
5872         struct extent_buffer *leaf = path->nodes[0];
5873         char *tmp;
5874         size_t max_size;
5875         unsigned long inline_size;
5876         unsigned long ptr;
5877         int compress_type;
5878
5879         WARN_ON(pg_offset != 0);
5880         compress_type = btrfs_file_extent_compression(leaf, item);
5881         max_size = btrfs_file_extent_ram_bytes(leaf, item);
5882         inline_size = btrfs_file_extent_inline_item_len(leaf,
5883                                         btrfs_item_nr(leaf, path->slots[0]));
5884         tmp = kmalloc(inline_size, GFP_NOFS);
5885         if (!tmp)
5886                 return -ENOMEM;
5887         ptr = btrfs_file_extent_inline_start(item);
5888
5889         read_extent_buffer(leaf, tmp, ptr, inline_size);
5890
5891         max_size = min_t(unsigned long, PAGE_CACHE_SIZE, max_size);
5892         ret = btrfs_decompress(compress_type, tmp, page,
5893                                extent_offset, inline_size, max_size);
5894         if (ret) {
5895                 char *kaddr = kmap_atomic(page);
5896                 unsigned long copy_size = min_t(u64,
5897                                   PAGE_CACHE_SIZE - pg_offset,
5898                                   max_size - extent_offset);
5899                 memset(kaddr + pg_offset, 0, copy_size);
5900                 kunmap_atomic(kaddr);
5901         }
5902         kfree(tmp);
5903         return 0;
5904 }
5905
5906 /*
5907  * a bit scary, this does extent mapping from logical file offset to the disk.
5908  * the ugly parts come from merging extents from the disk with the in-ram
5909  * representation.  This gets more complex because of the data=ordered code,
5910  * where the in-ram extents might be locked pending data=ordered completion.
5911  *
5912  * This also copies inline extents directly into the page.
5913  */
5914
5915 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
5916                                     size_t pg_offset, u64 start, u64 len,
5917                                     int create)
5918 {
5919         int ret;
5920         int err = 0;
5921         u64 bytenr;
5922         u64 extent_start = 0;
5923         u64 extent_end = 0;
5924         u64 objectid = btrfs_ino(inode);
5925         u32 found_type;
5926         struct btrfs_path *path = NULL;
5927         struct btrfs_root *root = BTRFS_I(inode)->root;
5928         struct btrfs_file_extent_item *item;
5929         struct extent_buffer *leaf;
5930         struct btrfs_key found_key;
5931         struct extent_map *em = NULL;
5932         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
5933         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5934         struct btrfs_trans_handle *trans = NULL;
5935         int compress_type;
5936
5937 again:
5938         read_lock(&em_tree->lock);
5939         em = lookup_extent_mapping(em_tree, start, len);
5940         if (em)
5941                 em->bdev = root->fs_info->fs_devices->latest_bdev;
5942         read_unlock(&em_tree->lock);
5943
5944         if (em) {
5945                 if (em->start > start || em->start + em->len <= start)
5946                         free_extent_map(em);
5947                 else if (em->block_start == EXTENT_MAP_INLINE && page)
5948                         free_extent_map(em);
5949                 else
5950                         goto out;
5951         }
5952         em = alloc_extent_map();
5953         if (!em) {
5954                 err = -ENOMEM;
5955                 goto out;
5956         }
5957         em->bdev = root->fs_info->fs_devices->latest_bdev;
5958         em->start = EXTENT_MAP_HOLE;
5959         em->orig_start = EXTENT_MAP_HOLE;
5960         em->len = (u64)-1;
5961         em->block_len = (u64)-1;
5962
5963         if (!path) {
5964                 path = btrfs_alloc_path();
5965                 if (!path) {
5966                         err = -ENOMEM;
5967                         goto out;
5968                 }
5969                 /*
5970                  * Chances are we'll be called again, so go ahead and do
5971                  * readahead
5972                  */
5973                 path->reada = 1;
5974         }
5975
5976         ret = btrfs_lookup_file_extent(trans, root, path,
5977                                        objectid, start, trans != NULL);
5978         if (ret < 0) {
5979                 err = ret;
5980                 goto out;
5981         }
5982
5983         if (ret != 0) {
5984                 if (path->slots[0] == 0)
5985                         goto not_found;
5986                 path->slots[0]--;
5987         }
5988
5989         leaf = path->nodes[0];
5990         item = btrfs_item_ptr(leaf, path->slots[0],
5991                               struct btrfs_file_extent_item);
5992         /* are we inside the extent that was found? */
5993         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5994         found_type = btrfs_key_type(&found_key);
5995         if (found_key.objectid != objectid ||
5996             found_type != BTRFS_EXTENT_DATA_KEY) {
5997                 goto not_found;
5998         }
5999
6000         found_type = btrfs_file_extent_type(leaf, item);
6001         extent_start = found_key.offset;
6002         compress_type = btrfs_file_extent_compression(leaf, item);
6003         if (found_type == BTRFS_FILE_EXTENT_REG ||
6004             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6005                 extent_end = extent_start +
6006                        btrfs_file_extent_num_bytes(leaf, item);
6007         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6008                 size_t size;
6009                 size = btrfs_file_extent_inline_len(leaf, item);
6010                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6011         }
6012
6013         if (start >= extent_end) {
6014                 path->slots[0]++;
6015                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6016                         ret = btrfs_next_leaf(root, path);
6017                         if (ret < 0) {
6018                                 err = ret;
6019                                 goto out;
6020                         }
6021                         if (ret > 0)
6022                                 goto not_found;
6023                         leaf = path->nodes[0];
6024                 }
6025                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6026                 if (found_key.objectid != objectid ||
6027                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6028                         goto not_found;
6029                 if (start + len <= found_key.offset)
6030                         goto not_found;
6031                 em->start = start;
6032                 em->orig_start = start;
6033                 em->len = found_key.offset - start;
6034                 goto not_found_em;
6035         }
6036
6037         em->ram_bytes = btrfs_file_extent_ram_bytes(leaf, item);
6038         if (found_type == BTRFS_FILE_EXTENT_REG ||
6039             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6040                 em->start = extent_start;
6041                 em->len = extent_end - extent_start;
6042                 em->orig_start = extent_start -
6043                                  btrfs_file_extent_offset(leaf, item);
6044                 em->orig_block_len = btrfs_file_extent_disk_num_bytes(leaf,
6045                                                                       item);
6046                 bytenr = btrfs_file_extent_disk_bytenr(leaf, item);
6047                 if (bytenr == 0) {
6048                         em->block_start = EXTENT_MAP_HOLE;
6049                         goto insert;
6050                 }
6051                 if (compress_type != BTRFS_COMPRESS_NONE) {
6052                         set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
6053                         em->compress_type = compress_type;
6054                         em->block_start = bytenr;
6055                         em->block_len = em->orig_block_len;
6056                 } else {
6057                         bytenr += btrfs_file_extent_offset(leaf, item);
6058                         em->block_start = bytenr;
6059                         em->block_len = em->len;
6060                         if (found_type == BTRFS_FILE_EXTENT_PREALLOC)
6061                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
6062                 }
6063                 goto insert;
6064         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6065                 unsigned long ptr;
6066                 char *map;
6067                 size_t size;
6068                 size_t extent_offset;
6069                 size_t copy_size;
6070
6071                 em->block_start = EXTENT_MAP_INLINE;
6072                 if (!page || create) {
6073                         em->start = extent_start;
6074                         em->len = extent_end - extent_start;
6075                         goto out;
6076                 }
6077
6078                 size = btrfs_file_extent_inline_len(leaf, item);
6079                 extent_offset = page_offset(page) + pg_offset - extent_start;
6080                 copy_size = min_t(u64, PAGE_CACHE_SIZE - pg_offset,
6081                                 size - extent_offset);
6082                 em->start = extent_start + extent_offset;
6083                 em->len = ALIGN(copy_size, root->sectorsize);
6084                 em->orig_block_len = em->len;
6085                 em->orig_start = em->start;
6086                 if (compress_type) {
6087                         set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
6088                         em->compress_type = compress_type;
6089                 }
6090                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6091                 if (create == 0 && !PageUptodate(page)) {
6092                         if (btrfs_file_extent_compression(leaf, item) !=
6093                             BTRFS_COMPRESS_NONE) {
6094                                 ret = uncompress_inline(path, inode, page,
6095                                                         pg_offset,
6096                                                         extent_offset, item);
6097                                 BUG_ON(ret); /* -ENOMEM */
6098                         } else {
6099                                 map = kmap(page);
6100                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6101                                                    copy_size);
6102                                 if (pg_offset + copy_size < PAGE_CACHE_SIZE) {
6103                                         memset(map + pg_offset + copy_size, 0,
6104                                                PAGE_CACHE_SIZE - pg_offset -
6105                                                copy_size);
6106                                 }
6107                                 kunmap(page);
6108                         }
6109                         flush_dcache_page(page);
6110                 } else if (create && PageUptodate(page)) {
6111                         BUG();
6112                         if (!trans) {
6113                                 kunmap(page);
6114                                 free_extent_map(em);
6115                                 em = NULL;
6116
6117                                 btrfs_release_path(path);
6118                                 trans = btrfs_join_transaction(root);
6119
6120                                 if (IS_ERR(trans))
6121                                         return ERR_CAST(trans);
6122                                 goto again;
6123                         }
6124                         map = kmap(page);
6125                         write_extent_buffer(leaf, map + pg_offset, ptr,
6126                                             copy_size);
6127                         kunmap(page);
6128                         btrfs_mark_buffer_dirty(leaf);
6129                 }
6130                 set_extent_uptodate(io_tree, em->start,
6131                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6132                 goto insert;
6133         } else {
6134                 WARN(1, KERN_ERR "btrfs unknown found_type %d\n", found_type);
6135         }
6136 not_found:
6137         em->start = start;
6138         em->orig_start = start;
6139         em->len = len;
6140 not_found_em:
6141         em->block_start = EXTENT_MAP_HOLE;
6142         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
6143 insert:
6144         btrfs_release_path(path);
6145         if (em->start > start || extent_map_end(em) <= start) {
6146                 btrfs_err(root->fs_info, "bad extent! em: [%llu %llu] passed [%llu %llu]",
6147                         (unsigned long long)em->start,
6148                         (unsigned long long)em->len,
6149                         (unsigned long long)start,
6150                         (unsigned long long)len);
6151                 err = -EIO;
6152                 goto out;
6153         }
6154
6155         err = 0;
6156         write_lock(&em_tree->lock);
6157         ret = add_extent_mapping(em_tree, em, 0);
6158         /* it is possible that someone inserted the extent into the tree
6159          * while we had the lock dropped.  It is also possible that
6160          * an overlapping map exists in the tree
6161          */
6162         if (ret == -EEXIST) {
6163                 struct extent_map *existing;
6164
6165                 ret = 0;
6166
6167                 existing = lookup_extent_mapping(em_tree, start, len);
6168                 if (existing && (existing->start > start ||
6169                     existing->start + existing->len <= start)) {
6170                         free_extent_map(existing);
6171                         existing = NULL;
6172                 }
6173                 if (!existing) {
6174                         existing = lookup_extent_mapping(em_tree, em->start,
6175                                                          em->len);
6176                         if (existing) {
6177                                 err = merge_extent_mapping(em_tree, existing,
6178                                                            em, start,
6179                                                            root->sectorsize);
6180                                 free_extent_map(existing);
6181                                 if (err) {
6182                                         free_extent_map(em);
6183                                         em = NULL;
6184                                 }
6185                         } else {
6186                                 err = -EIO;
6187                                 free_extent_map(em);
6188                                 em = NULL;
6189                         }
6190                 } else {
6191                         free_extent_map(em);
6192                         em = existing;
6193                         err = 0;
6194                 }
6195         }
6196         write_unlock(&em_tree->lock);
6197 out:
6198
6199         if (em)
6200                 trace_btrfs_get_extent(root, em);
6201
6202         if (path)
6203                 btrfs_free_path(path);
6204         if (trans) {
6205                 ret = btrfs_end_transaction(trans, root);
6206                 if (!err)
6207                         err = ret;
6208         }
6209         if (err) {
6210                 free_extent_map(em);
6211                 return ERR_PTR(err);
6212         }
6213         BUG_ON(!em); /* Error is always set */
6214         return em;
6215 }
6216
6217 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
6218                                            size_t pg_offset, u64 start, u64 len,
6219                                            int create)
6220 {
6221         struct extent_map *em;
6222         struct extent_map *hole_em = NULL;
6223         u64 range_start = start;
6224         u64 end;
6225         u64 found;
6226         u64 found_end;
6227         int err = 0;
6228
6229         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
6230         if (IS_ERR(em))
6231                 return em;
6232         if (em) {
6233                 /*
6234                  * if our em maps to
6235                  * -  a hole or
6236                  * -  a pre-alloc extent,
6237                  * there might actually be delalloc bytes behind it.
6238                  */
6239                 if (em->block_start != EXTENT_MAP_HOLE &&
6240                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
6241                         return em;
6242                 else
6243                         hole_em = em;
6244         }
6245
6246         /* check to see if we've wrapped (len == -1 or similar) */
6247         end = start + len;
6248         if (end < start)
6249                 end = (u64)-1;
6250         else
6251                 end -= 1;
6252
6253         em = NULL;
6254
6255         /* ok, we didn't find anything, lets look for delalloc */
6256         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
6257                                  end, len, EXTENT_DELALLOC, 1);
6258         found_end = range_start + found;
6259         if (found_end < range_start)
6260                 found_end = (u64)-1;
6261
6262         /*
6263          * we didn't find anything useful, return
6264          * the original results from get_extent()
6265          */
6266         if (range_start > end || found_end <= start) {
6267                 em = hole_em;
6268                 hole_em = NULL;
6269                 goto out;
6270         }
6271
6272         /* adjust the range_start to make sure it doesn't
6273          * go backwards from the start they passed in
6274          */
6275         range_start = max(start,range_start);
6276         found = found_end - range_start;
6277
6278         if (found > 0) {
6279                 u64 hole_start = start;
6280                 u64 hole_len = len;
6281
6282                 em = alloc_extent_map();
6283                 if (!em) {
6284                         err = -ENOMEM;
6285                         goto out;
6286                 }
6287                 /*
6288                  * when btrfs_get_extent can't find anything it
6289                  * returns one huge hole
6290                  *
6291                  * make sure what it found really fits our range, and
6292                  * adjust to make sure it is based on the start from
6293                  * the caller
6294                  */
6295                 if (hole_em) {
6296                         u64 calc_end = extent_map_end(hole_em);
6297
6298                         if (calc_end <= start || (hole_em->start > end)) {
6299                                 free_extent_map(hole_em);
6300                                 hole_em = NULL;
6301                         } else {
6302                                 hole_start = max(hole_em->start, start);
6303                                 hole_len = calc_end - hole_start;
6304                         }
6305                 }
6306                 em->bdev = NULL;
6307                 if (hole_em && range_start > hole_start) {
6308                         /* our hole starts before our delalloc, so we
6309                          * have to return just the parts of the hole
6310                          * that go until  the delalloc starts
6311                          */
6312                         em->len = min(hole_len,
6313                                       range_start - hole_start);
6314                         em->start = hole_start;
6315                         em->orig_start = hole_start;
6316                         /*
6317                          * don't adjust block start at all,
6318                          * it is fixed at EXTENT_MAP_HOLE
6319                          */
6320                         em->block_start = hole_em->block_start;
6321                         em->block_len = hole_len;
6322                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
6323                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
6324                 } else {
6325                         em->start = range_start;
6326                         em->len = found;
6327                         em->orig_start = range_start;
6328                         em->block_start = EXTENT_MAP_DELALLOC;
6329                         em->block_len = found;
6330                 }
6331         } else if (hole_em) {
6332                 return hole_em;
6333         }
6334 out:
6335
6336         free_extent_map(hole_em);
6337         if (err) {
6338                 free_extent_map(em);
6339                 return ERR_PTR(err);
6340         }
6341         return em;
6342 }
6343
6344 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
6345                                                   u64 start, u64 len)
6346 {
6347         struct btrfs_root *root = BTRFS_I(inode)->root;
6348         struct btrfs_trans_handle *trans;
6349         struct extent_map *em;
6350         struct btrfs_key ins;
6351         u64 alloc_hint;
6352         int ret;
6353
6354         trans = btrfs_join_transaction(root);
6355         if (IS_ERR(trans))
6356                 return ERR_CAST(trans);
6357
6358         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
6359
6360         alloc_hint = get_extent_allocation_hint(inode, start, len);
6361         ret = btrfs_reserve_extent(trans, root, len, root->sectorsize, 0,
6362                                    alloc_hint, &ins, 1);
6363         if (ret) {
6364                 em = ERR_PTR(ret);
6365                 goto out;
6366         }
6367
6368         em = create_pinned_em(inode, start, ins.offset, start, ins.objectid,
6369                               ins.offset, ins.offset, ins.offset, 0);
6370         if (IS_ERR(em))
6371                 goto out;
6372
6373         ret = btrfs_add_ordered_extent_dio(inode, start, ins.objectid,
6374                                            ins.offset, ins.offset, 0);
6375         if (ret) {
6376                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset);
6377                 em = ERR_PTR(ret);
6378         }
6379 out:
6380         btrfs_end_transaction(trans, root);
6381         return em;
6382 }
6383
6384 /*
6385  * returns 1 when the nocow is safe, < 1 on error, 0 if the
6386  * block must be cow'd
6387  */
6388 noinline int can_nocow_extent(struct btrfs_trans_handle *trans,
6389                               struct inode *inode, u64 offset, u64 *len,
6390                               u64 *orig_start, u64 *orig_block_len,
6391                               u64 *ram_bytes)
6392 {
6393         struct btrfs_path *path;
6394         int ret;
6395         struct extent_buffer *leaf;
6396         struct btrfs_root *root = BTRFS_I(inode)->root;
6397         struct btrfs_file_extent_item *fi;
6398         struct btrfs_key key;
6399         u64 disk_bytenr;
6400         u64 backref_offset;
6401         u64 extent_end;
6402         u64 num_bytes;
6403         int slot;
6404         int found_type;
6405         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
6406         path = btrfs_alloc_path();
6407         if (!path)
6408                 return -ENOMEM;
6409
6410         ret = btrfs_lookup_file_extent(trans, root, path, btrfs_ino(inode),
6411                                        offset, 0);
6412         if (ret < 0)
6413                 goto out;
6414
6415         slot = path->slots[0];
6416         if (ret == 1) {
6417                 if (slot == 0) {
6418                         /* can't find the item, must cow */
6419                         ret = 0;
6420                         goto out;
6421                 }
6422                 slot--;
6423         }
6424         ret = 0;
6425         leaf = path->nodes[0];
6426         btrfs_item_key_to_cpu(leaf, &key, slot);
6427         if (key.objectid != btrfs_ino(inode) ||
6428             key.type != BTRFS_EXTENT_DATA_KEY) {
6429                 /* not our file or wrong item type, must cow */
6430                 goto out;
6431         }
6432
6433         if (key.offset > offset) {
6434                 /* Wrong offset, must cow */
6435                 goto out;
6436         }
6437
6438         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
6439         found_type = btrfs_file_extent_type(leaf, fi);
6440         if (found_type != BTRFS_FILE_EXTENT_REG &&
6441             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
6442                 /* not a regular extent, must cow */
6443                 goto out;
6444         }
6445
6446         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
6447                 goto out;
6448
6449         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
6450         if (disk_bytenr == 0)
6451                 goto out;
6452
6453         if (btrfs_file_extent_compression(leaf, fi) ||
6454             btrfs_file_extent_encryption(leaf, fi) ||
6455             btrfs_file_extent_other_encoding(leaf, fi))
6456                 goto out;
6457
6458         backref_offset = btrfs_file_extent_offset(leaf, fi);
6459
6460         if (orig_start) {
6461                 *orig_start = key.offset - backref_offset;
6462                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
6463                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
6464         }
6465
6466         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
6467
6468         if (btrfs_extent_readonly(root, disk_bytenr))
6469                 goto out;
6470
6471         /*
6472          * look for other files referencing this extent, if we
6473          * find any we must cow
6474          */
6475         if (btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
6476                                   key.offset - backref_offset, disk_bytenr))
6477                 goto out;
6478
6479         /*
6480          * adjust disk_bytenr and num_bytes to cover just the bytes
6481          * in this extent we are about to write.  If there
6482          * are any csums in that range we have to cow in order
6483          * to keep the csums correct
6484          */
6485         disk_bytenr += backref_offset;
6486         disk_bytenr += offset - key.offset;
6487         num_bytes = min(offset + *len, extent_end) - offset;
6488         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
6489                                 goto out;
6490         /*
6491          * all of the above have passed, it is safe to overwrite this extent
6492          * without cow
6493          */
6494         *len = num_bytes;
6495         ret = 1;
6496 out:
6497         btrfs_free_path(path);
6498         return ret;
6499 }
6500
6501 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
6502                               struct extent_state **cached_state, int writing)
6503 {
6504         struct btrfs_ordered_extent *ordered;
6505         int ret = 0;
6506
6507         while (1) {
6508                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
6509                                  0, cached_state);
6510                 /*
6511                  * We're concerned with the entire range that we're going to be
6512                  * doing DIO to, so we need to make sure theres no ordered
6513                  * extents in this range.
6514                  */
6515                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
6516                                                      lockend - lockstart + 1);
6517
6518                 /*
6519                  * We need to make sure there are no buffered pages in this
6520                  * range either, we could have raced between the invalidate in
6521                  * generic_file_direct_write and locking the extent.  The
6522                  * invalidate needs to happen so that reads after a write do not
6523                  * get stale data.
6524                  */
6525                 if (!ordered && (!writing ||
6526                     !test_range_bit(&BTRFS_I(inode)->io_tree,
6527                                     lockstart, lockend, EXTENT_UPTODATE, 0,
6528                                     *cached_state)))
6529                         break;
6530
6531                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
6532                                      cached_state, GFP_NOFS);
6533
6534                 if (ordered) {
6535                         btrfs_start_ordered_extent(inode, ordered, 1);
6536                         btrfs_put_ordered_extent(ordered);
6537                 } else {
6538                         /* Screw you mmap */
6539                         ret = filemap_write_and_wait_range(inode->i_mapping,
6540                                                            lockstart,
6541                                                            lockend);
6542                         if (ret)
6543                                 break;
6544
6545                         /*
6546                          * If we found a page that couldn't be invalidated just
6547                          * fall back to buffered.
6548                          */
6549                         ret = invalidate_inode_pages2_range(inode->i_mapping,
6550                                         lockstart >> PAGE_CACHE_SHIFT,
6551                                         lockend >> PAGE_CACHE_SHIFT);
6552                         if (ret)
6553                                 break;
6554                 }
6555
6556                 cond_resched();
6557         }
6558
6559         return ret;
6560 }
6561
6562 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
6563                                            u64 len, u64 orig_start,
6564                                            u64 block_start, u64 block_len,
6565                                            u64 orig_block_len, u64 ram_bytes,
6566                                            int type)
6567 {
6568         struct extent_map_tree *em_tree;
6569         struct extent_map *em;
6570         struct btrfs_root *root = BTRFS_I(inode)->root;
6571         int ret;
6572
6573         em_tree = &BTRFS_I(inode)->extent_tree;
6574         em = alloc_extent_map();
6575         if (!em)
6576                 return ERR_PTR(-ENOMEM);
6577
6578         em->start = start;
6579         em->orig_start = orig_start;
6580         em->mod_start = start;
6581         em->mod_len = len;
6582         em->len = len;
6583         em->block_len = block_len;
6584         em->block_start = block_start;
6585         em->bdev = root->fs_info->fs_devices->latest_bdev;
6586         em->orig_block_len = orig_block_len;
6587         em->ram_bytes = ram_bytes;
6588         em->generation = -1;
6589         set_bit(EXTENT_FLAG_PINNED, &em->flags);
6590         if (type == BTRFS_ORDERED_PREALLOC)
6591                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
6592
6593         do {
6594                 btrfs_drop_extent_cache(inode, em->start,
6595                                 em->start + em->len - 1, 0);
6596                 write_lock(&em_tree->lock);
6597                 ret = add_extent_mapping(em_tree, em, 1);
6598                 write_unlock(&em_tree->lock);
6599         } while (ret == -EEXIST);
6600
6601         if (ret) {
6602                 free_extent_map(em);
6603                 return ERR_PTR(ret);
6604         }
6605
6606         return em;
6607 }
6608
6609
6610 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
6611                                    struct buffer_head *bh_result, int create)
6612 {
6613         struct extent_map *em;
6614         struct btrfs_root *root = BTRFS_I(inode)->root;
6615         struct extent_state *cached_state = NULL;
6616         u64 start = iblock << inode->i_blkbits;
6617         u64 lockstart, lockend;
6618         u64 len = bh_result->b_size;
6619         struct btrfs_trans_handle *trans;
6620         int unlock_bits = EXTENT_LOCKED;
6621         int ret = 0;
6622
6623         if (create)
6624                 unlock_bits |= EXTENT_DELALLOC | EXTENT_DIRTY;
6625         else
6626                 len = min_t(u64, len, root->sectorsize);
6627
6628         lockstart = start;
6629         lockend = start + len - 1;
6630
6631         /*
6632          * If this errors out it's because we couldn't invalidate pagecache for
6633          * this range and we need to fallback to buffered.
6634          */
6635         if (lock_extent_direct(inode, lockstart, lockend, &cached_state, create))
6636                 return -ENOTBLK;
6637
6638         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
6639         if (IS_ERR(em)) {
6640                 ret = PTR_ERR(em);
6641                 goto unlock_err;
6642         }
6643
6644         /*
6645          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
6646          * io.  INLINE is special, and we could probably kludge it in here, but
6647          * it's still buffered so for safety lets just fall back to the generic
6648          * buffered path.
6649          *
6650          * For COMPRESSED we _have_ to read the entire extent in so we can
6651          * decompress it, so there will be buffering required no matter what we
6652          * do, so go ahead and fallback to buffered.
6653          *
6654          * We return -ENOTBLK because thats what makes DIO go ahead and go back
6655          * to buffered IO.  Don't blame me, this is the price we pay for using
6656          * the generic code.
6657          */
6658         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
6659             em->block_start == EXTENT_MAP_INLINE) {
6660                 free_extent_map(em);
6661                 ret = -ENOTBLK;
6662                 goto unlock_err;
6663         }
6664
6665         /* Just a good old fashioned hole, return */
6666         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
6667                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
6668                 free_extent_map(em);
6669                 goto unlock_err;
6670         }
6671
6672         /*
6673          * We don't allocate a new extent in the following cases
6674          *
6675          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
6676          * existing extent.
6677          * 2) The extent is marked as PREALLOC.  We're good to go here and can
6678          * just use the extent.
6679          *
6680          */
6681         if (!create) {
6682                 len = min(len, em->len - (start - em->start));
6683                 lockstart = start + len;
6684                 goto unlock;
6685         }
6686
6687         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
6688             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
6689              em->block_start != EXTENT_MAP_HOLE)) {
6690                 int type;
6691                 int ret;
6692                 u64 block_start, orig_start, orig_block_len, ram_bytes;
6693
6694                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
6695                         type = BTRFS_ORDERED_PREALLOC;
6696                 else
6697                         type = BTRFS_ORDERED_NOCOW;
6698                 len = min(len, em->len - (start - em->start));
6699                 block_start = em->block_start + (start - em->start);
6700
6701                 /*
6702                  * we're not going to log anything, but we do need
6703                  * to make sure the current transaction stays open
6704                  * while we look for nocow cross refs
6705                  */
6706                 trans = btrfs_join_transaction(root);
6707                 if (IS_ERR(trans))
6708                         goto must_cow;
6709
6710                 if (can_nocow_extent(trans, inode, start, &len, &orig_start,
6711                                      &orig_block_len, &ram_bytes) == 1) {
6712                         if (type == BTRFS_ORDERED_PREALLOC) {
6713                                 free_extent_map(em);
6714                                 em = create_pinned_em(inode, start, len,
6715                                                        orig_start,
6716                                                        block_start, len,
6717                                                        orig_block_len,
6718                                                        ram_bytes, type);
6719                                 if (IS_ERR(em)) {
6720                                         btrfs_end_transaction(trans, root);
6721                                         goto unlock_err;
6722                                 }
6723                         }
6724
6725                         ret = btrfs_add_ordered_extent_dio(inode, start,
6726                                            block_start, len, len, type);
6727                         btrfs_end_transaction(trans, root);
6728                         if (ret) {
6729                                 free_extent_map(em);
6730                                 goto unlock_err;
6731                         }
6732                         goto unlock;
6733                 }
6734                 btrfs_end_transaction(trans, root);
6735         }
6736 must_cow:
6737         /*
6738          * this will cow the extent, reset the len in case we changed
6739          * it above
6740          */
6741         len = bh_result->b_size;
6742         free_extent_map(em);
6743         em = btrfs_new_extent_direct(inode, start, len);
6744         if (IS_ERR(em)) {
6745                 ret = PTR_ERR(em);
6746                 goto unlock_err;
6747         }
6748         len = min(len, em->len - (start - em->start));
6749 unlock:
6750         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
6751                 inode->i_blkbits;
6752         bh_result->b_size = len;
6753         bh_result->b_bdev = em->bdev;
6754         set_buffer_mapped(bh_result);
6755         if (create) {
6756                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
6757                         set_buffer_new(bh_result);
6758
6759                 /*
6760                  * Need to update the i_size under the extent lock so buffered
6761                  * readers will get the updated i_size when we unlock.
6762                  */
6763                 if (start + len > i_size_read(inode))
6764                         i_size_write(inode, start + len);
6765
6766                 spin_lock(&BTRFS_I(inode)->lock);
6767                 BTRFS_I(inode)->outstanding_extents++;
6768                 spin_unlock(&BTRFS_I(inode)->lock);
6769
6770                 ret = set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
6771                                      lockstart + len - 1, EXTENT_DELALLOC, NULL,
6772                                      &cached_state, GFP_NOFS);
6773                 BUG_ON(ret);
6774         }
6775
6776         /*
6777          * In the case of write we need to clear and unlock the entire range,
6778          * in the case of read we need to unlock only the end area that we
6779          * aren't using if there is any left over space.
6780          */
6781         if (lockstart < lockend) {
6782                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
6783                                  lockend, unlock_bits, 1, 0,
6784                                  &cached_state, GFP_NOFS);
6785         } else {
6786                 free_extent_state(cached_state);
6787         }
6788
6789         free_extent_map(em);
6790
6791         return 0;
6792
6793 unlock_err:
6794         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
6795                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
6796         return ret;
6797 }
6798
6799 struct btrfs_dio_private {
6800         struct inode *inode;
6801         u64 logical_offset;
6802         u64 disk_bytenr;
6803         u64 bytes;
6804         void *private;
6805
6806         /* number of bios pending for this dio */
6807         atomic_t pending_bios;
6808
6809         /* IO errors */
6810         int errors;
6811
6812         /* orig_bio is our btrfs_io_bio */
6813         struct bio *orig_bio;
6814
6815         /* dio_bio came from fs/direct-io.c */
6816         struct bio *dio_bio;
6817 };
6818
6819 static void btrfs_endio_direct_read(struct bio *bio, int err)
6820 {
6821         struct btrfs_dio_private *dip = bio->bi_private;
6822         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
6823         struct bio_vec *bvec = bio->bi_io_vec;
6824         struct inode *inode = dip->inode;
6825         struct btrfs_root *root = BTRFS_I(inode)->root;
6826         struct bio *dio_bio;
6827         u64 start;
6828
6829         start = dip->logical_offset;
6830         do {
6831                 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
6832                         struct page *page = bvec->bv_page;
6833                         char *kaddr;
6834                         u32 csum = ~(u32)0;
6835                         u64 private = ~(u32)0;
6836                         unsigned long flags;
6837
6838                         if (get_state_private(&BTRFS_I(inode)->io_tree,
6839                                               start, &private))
6840                                 goto failed;
6841                         local_irq_save(flags);
6842                         kaddr = kmap_atomic(page);
6843                         csum = btrfs_csum_data(kaddr + bvec->bv_offset,
6844                                                csum, bvec->bv_len);
6845                         btrfs_csum_final(csum, (char *)&csum);
6846                         kunmap_atomic(kaddr);
6847                         local_irq_restore(flags);
6848
6849                         flush_dcache_page(bvec->bv_page);
6850                         if (csum != private) {
6851 failed:
6852                                 btrfs_err(root->fs_info, "csum failed ino %llu off %llu csum %u private %u",
6853                                         (unsigned long long)btrfs_ino(inode),
6854                                         (unsigned long long)start,
6855                                         csum, (unsigned)private);
6856                                 err = -EIO;
6857                         }
6858                 }
6859
6860                 start += bvec->bv_len;
6861                 bvec++;
6862         } while (bvec <= bvec_end);
6863
6864         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
6865                       dip->logical_offset + dip->bytes - 1);
6866         dio_bio = dip->dio_bio;
6867
6868         kfree(dip);
6869
6870         /* If we had a csum failure make sure to clear the uptodate flag */
6871         if (err)
6872                 clear_bit(BIO_UPTODATE, &dio_bio->bi_flags);
6873         dio_end_io(dio_bio, err);
6874         bio_put(bio);
6875 }
6876
6877 static void btrfs_endio_direct_write(struct bio *bio, int err)
6878 {
6879         struct btrfs_dio_private *dip = bio->bi_private;
6880         struct inode *inode = dip->inode;
6881         struct btrfs_root *root = BTRFS_I(inode)->root;
6882         struct btrfs_ordered_extent *ordered = NULL;
6883         u64 ordered_offset = dip->logical_offset;
6884         u64 ordered_bytes = dip->bytes;
6885         struct bio *dio_bio;
6886         int ret;
6887
6888         if (err)
6889                 goto out_done;
6890 again:
6891         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
6892                                                    &ordered_offset,
6893                                                    ordered_bytes, !err);
6894         if (!ret)
6895                 goto out_test;
6896
6897         ordered->work.func = finish_ordered_fn;
6898         ordered->work.flags = 0;
6899         btrfs_queue_worker(&root->fs_info->endio_write_workers,
6900                            &ordered->work);
6901 out_test:
6902         /*
6903          * our bio might span multiple ordered extents.  If we haven't
6904          * completed the accounting for the whole dio, go back and try again
6905          */
6906         if (ordered_offset < dip->logical_offset + dip->bytes) {
6907                 ordered_bytes = dip->logical_offset + dip->bytes -
6908                         ordered_offset;
6909                 ordered = NULL;
6910                 goto again;
6911         }
6912 out_done:
6913         dio_bio = dip->dio_bio;
6914
6915         kfree(dip);
6916
6917         /* If we had an error make sure to clear the uptodate flag */
6918         if (err)
6919                 clear_bit(BIO_UPTODATE, &dio_bio->bi_flags);
6920         dio_end_io(dio_bio, err);
6921         bio_put(bio);
6922 }
6923
6924 static int __btrfs_submit_bio_start_direct_io(struct inode *inode, int rw,
6925                                     struct bio *bio, int mirror_num,
6926                                     unsigned long bio_flags, u64 offset)
6927 {
6928         int ret;
6929         struct btrfs_root *root = BTRFS_I(inode)->root;
6930         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
6931         BUG_ON(ret); /* -ENOMEM */
6932         return 0;
6933 }
6934
6935 static void btrfs_end_dio_bio(struct bio *bio, int err)
6936 {
6937         struct btrfs_dio_private *dip = bio->bi_private;
6938
6939         if (err) {
6940                 printk(KERN_ERR "btrfs direct IO failed ino %llu rw %lu "
6941                       "sector %#Lx len %u err no %d\n",
6942                       (unsigned long long)btrfs_ino(dip->inode), bio->bi_rw,
6943                       (unsigned long long)bio->bi_sector, bio->bi_size, err);
6944                 dip->errors = 1;
6945
6946                 /*
6947                  * before atomic variable goto zero, we must make sure
6948                  * dip->errors is perceived to be set.
6949                  */
6950                 smp_mb__before_atomic_dec();
6951         }
6952
6953         /* if there are more bios still pending for this dio, just exit */
6954         if (!atomic_dec_and_test(&dip->pending_bios))
6955                 goto out;
6956
6957         if (dip->errors) {
6958                 bio_io_error(dip->orig_bio);
6959         } else {
6960                 set_bit(BIO_UPTODATE, &dip->dio_bio->bi_flags);
6961                 bio_endio(dip->orig_bio, 0);
6962         }
6963 out:
6964         bio_put(bio);
6965 }
6966
6967 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
6968                                        u64 first_sector, gfp_t gfp_flags)
6969 {
6970         int nr_vecs = bio_get_nr_vecs(bdev);
6971         return btrfs_bio_alloc(bdev, first_sector, nr_vecs, gfp_flags);
6972 }
6973
6974 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
6975                                          int rw, u64 file_offset, int skip_sum,
6976                                          int async_submit)
6977 {
6978         int write = rw & REQ_WRITE;
6979         struct btrfs_root *root = BTRFS_I(inode)->root;
6980         int ret;
6981
6982         if (async_submit)
6983                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
6984
6985         bio_get(bio);
6986
6987         if (!write) {
6988                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
6989                 if (ret)
6990                         goto err;
6991         }
6992
6993         if (skip_sum)
6994                 goto map;
6995
6996         if (write && async_submit) {
6997                 ret = btrfs_wq_submit_bio(root->fs_info,
6998                                    inode, rw, bio, 0, 0,
6999                                    file_offset,
7000                                    __btrfs_submit_bio_start_direct_io,
7001                                    __btrfs_submit_bio_done);
7002                 goto err;
7003         } else if (write) {
7004                 /*
7005                  * If we aren't doing async submit, calculate the csum of the
7006                  * bio now.
7007                  */
7008                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
7009                 if (ret)
7010                         goto err;
7011         } else if (!skip_sum) {
7012                 ret = btrfs_lookup_bio_sums_dio(root, inode, bio, file_offset);
7013                 if (ret)
7014                         goto err;
7015         }
7016
7017 map:
7018         ret = btrfs_map_bio(root, rw, bio, 0, async_submit);
7019 err:
7020         bio_put(bio);
7021         return ret;
7022 }
7023
7024 static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
7025                                     int skip_sum)
7026 {
7027         struct inode *inode = dip->inode;
7028         struct btrfs_root *root = BTRFS_I(inode)->root;
7029         struct bio *bio;
7030         struct bio *orig_bio = dip->orig_bio;
7031         struct bio_vec *bvec = orig_bio->bi_io_vec;
7032         u64 start_sector = orig_bio->bi_sector;
7033         u64 file_offset = dip->logical_offset;
7034         u64 submit_len = 0;
7035         u64 map_length;
7036         int nr_pages = 0;
7037         int ret = 0;
7038         int async_submit = 0;
7039
7040         map_length = orig_bio->bi_size;
7041         ret = btrfs_map_block(root->fs_info, rw, start_sector << 9,
7042                               &map_length, NULL, 0);
7043         if (ret) {
7044                 bio_put(orig_bio);
7045                 return -EIO;
7046         }
7047         if (map_length >= orig_bio->bi_size) {
7048                 bio = orig_bio;
7049                 goto submit;
7050         }
7051
7052         /* async crcs make it difficult to collect full stripe writes. */
7053         if (btrfs_get_alloc_profile(root, 1) &
7054             (BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6))
7055                 async_submit = 0;
7056         else
7057                 async_submit = 1;
7058
7059         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
7060         if (!bio)
7061                 return -ENOMEM;
7062         bio->bi_private = dip;
7063         bio->bi_end_io = btrfs_end_dio_bio;
7064         atomic_inc(&dip->pending_bios);
7065
7066         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
7067                 if (unlikely(map_length < submit_len + bvec->bv_len ||
7068                     bio_add_page(bio, bvec->bv_page, bvec->bv_len,
7069                                  bvec->bv_offset) < bvec->bv_len)) {
7070                         /*
7071                          * inc the count before we submit the bio so
7072                          * we know the end IO handler won't happen before
7073                          * we inc the count. Otherwise, the dip might get freed
7074                          * before we're done setting it up
7075                          */
7076                         atomic_inc(&dip->pending_bios);
7077                         ret = __btrfs_submit_dio_bio(bio, inode, rw,
7078                                                      file_offset, skip_sum,
7079                                                      async_submit);
7080                         if (ret) {
7081                                 bio_put(bio);
7082                                 atomic_dec(&dip->pending_bios);
7083                                 goto out_err;
7084                         }
7085
7086                         start_sector += submit_len >> 9;
7087                         file_offset += submit_len;
7088
7089                         submit_len = 0;
7090                         nr_pages = 0;
7091
7092                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
7093                                                   start_sector, GFP_NOFS);
7094                         if (!bio)
7095                                 goto out_err;
7096                         bio->bi_private = dip;
7097                         bio->bi_end_io = btrfs_end_dio_bio;
7098
7099                         map_length = orig_bio->bi_size;
7100                         ret = btrfs_map_block(root->fs_info, rw,
7101                                               start_sector << 9,
7102                                               &map_length, NULL, 0);
7103                         if (ret) {
7104                                 bio_put(bio);
7105                                 goto out_err;
7106                         }
7107                 } else {
7108                         submit_len += bvec->bv_len;
7109                         nr_pages ++;
7110                         bvec++;
7111                 }
7112         }
7113
7114 submit:
7115         ret = __btrfs_submit_dio_bio(bio, inode, rw, file_offset, skip_sum,
7116                                      async_submit);
7117         if (!ret)
7118                 return 0;
7119
7120         bio_put(bio);
7121 out_err:
7122         dip->errors = 1;
7123         /*
7124          * before atomic variable goto zero, we must
7125          * make sure dip->errors is perceived to be set.
7126          */
7127         smp_mb__before_atomic_dec();
7128         if (atomic_dec_and_test(&dip->pending_bios))
7129                 bio_io_error(dip->orig_bio);
7130
7131         /* bio_end_io() will handle error, so we needn't return it */
7132         return 0;
7133 }
7134
7135 static void btrfs_submit_direct(int rw, struct bio *dio_bio,
7136                                 struct inode *inode, loff_t file_offset)
7137 {
7138         struct btrfs_root *root = BTRFS_I(inode)->root;
7139         struct btrfs_dio_private *dip;
7140         struct bio *io_bio;
7141         int skip_sum;
7142         int write = rw & REQ_WRITE;
7143         int ret = 0;
7144
7145         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
7146
7147         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
7148
7149         if (!io_bio) {
7150                 ret = -ENOMEM;
7151                 goto free_ordered;
7152         }
7153
7154         dip = kmalloc(sizeof(*dip), GFP_NOFS);
7155         if (!dip) {
7156                 ret = -ENOMEM;
7157                 goto free_io_bio;
7158         }
7159
7160         dip->private = dio_bio->bi_private;
7161         dip->inode = inode;
7162         dip->logical_offset = file_offset;
7163         dip->bytes = dio_bio->bi_size;
7164         dip->disk_bytenr = (u64)dio_bio->bi_sector << 9;
7165         io_bio->bi_private = dip;
7166         dip->errors = 0;
7167         dip->orig_bio = io_bio;
7168         dip->dio_bio = dio_bio;
7169         atomic_set(&dip->pending_bios, 0);
7170
7171         if (write)
7172                 io_bio->bi_end_io = btrfs_endio_direct_write;
7173         else
7174                 io_bio->bi_end_io = btrfs_endio_direct_read;
7175
7176         ret = btrfs_submit_direct_hook(rw, dip, skip_sum);
7177         if (!ret)
7178                 return;
7179
7180 free_io_bio:
7181         bio_put(io_bio);
7182
7183 free_ordered:
7184         /*
7185          * If this is a write, we need to clean up the reserved space and kill
7186          * the ordered extent.
7187          */
7188         if (write) {
7189                 struct btrfs_ordered_extent *ordered;
7190                 ordered = btrfs_lookup_ordered_extent(inode, file_offset);
7191                 if (!test_bit(BTRFS_ORDERED_PREALLOC, &ordered->flags) &&
7192                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered->flags))
7193                         btrfs_free_reserved_extent(root, ordered->start,
7194                                                    ordered->disk_len);
7195                 btrfs_put_ordered_extent(ordered);
7196                 btrfs_put_ordered_extent(ordered);
7197         }
7198         bio_endio(dio_bio, ret);
7199 }
7200
7201 static ssize_t check_direct_IO(struct btrfs_root *root, int rw, struct kiocb *iocb,
7202                         const struct iovec *iov, loff_t offset,
7203                         unsigned long nr_segs)
7204 {
7205         int seg;
7206         int i;
7207         size_t size;
7208         unsigned long addr;
7209         unsigned blocksize_mask = root->sectorsize - 1;
7210         ssize_t retval = -EINVAL;
7211         loff_t end = offset;
7212
7213         if (offset & blocksize_mask)
7214                 goto out;
7215
7216         /* Check the memory alignment.  Blocks cannot straddle pages */
7217         for (seg = 0; seg < nr_segs; seg++) {
7218                 addr = (unsigned long)iov[seg].iov_base;
7219                 size = iov[seg].iov_len;
7220                 end += size;
7221                 if ((addr & blocksize_mask) || (size & blocksize_mask))
7222                         goto out;
7223
7224                 /* If this is a write we don't need to check anymore */
7225                 if (rw & WRITE)
7226                         continue;
7227
7228                 /*
7229                  * Check to make sure we don't have duplicate iov_base's in this
7230                  * iovec, if so return EINVAL, otherwise we'll get csum errors
7231                  * when reading back.
7232                  */
7233                 for (i = seg + 1; i < nr_segs; i++) {
7234                         if (iov[seg].iov_base == iov[i].iov_base)
7235                                 goto out;
7236                 }
7237         }
7238         retval = 0;
7239 out:
7240         return retval;
7241 }
7242
7243 static ssize_t btrfs_direct_IO(int rw, struct kiocb *iocb,
7244                         const struct iovec *iov, loff_t offset,
7245                         unsigned long nr_segs)
7246 {
7247         struct file *file = iocb->ki_filp;
7248         struct inode *inode = file->f_mapping->host;
7249         size_t count = 0;
7250         int flags = 0;
7251         bool wakeup = true;
7252         bool relock = false;
7253         ssize_t ret;
7254
7255         if (check_direct_IO(BTRFS_I(inode)->root, rw, iocb, iov,
7256                             offset, nr_segs))
7257                 return 0;
7258
7259         atomic_inc(&inode->i_dio_count);
7260         smp_mb__after_atomic_inc();
7261
7262         /*
7263          * The generic stuff only does filemap_write_and_wait_range, which isn't
7264          * enough if we've written compressed pages to this area, so we need to
7265          * call btrfs_wait_ordered_range to make absolutely sure that any
7266          * outstanding dirty pages are on disk.
7267          */
7268         count = iov_length(iov, nr_segs);
7269         btrfs_wait_ordered_range(inode, offset, count);
7270
7271         if (rw & WRITE) {
7272                 /*
7273                  * If the write DIO is beyond the EOF, we need update
7274                  * the isize, but it is protected by i_mutex. So we can
7275                  * not unlock the i_mutex at this case.
7276                  */
7277                 if (offset + count <= inode->i_size) {
7278                         mutex_unlock(&inode->i_mutex);
7279                         relock = true;
7280                 }
7281                 ret = btrfs_delalloc_reserve_space(inode, count);
7282                 if (ret)
7283                         goto out;
7284         } else if (unlikely(test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
7285                                      &BTRFS_I(inode)->runtime_flags))) {
7286                 inode_dio_done(inode);
7287                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
7288                 wakeup = false;
7289         }
7290
7291         ret = __blockdev_direct_IO(rw, iocb, inode,
7292                         BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
7293                         iov, offset, nr_segs, btrfs_get_blocks_direct, NULL,
7294                         btrfs_submit_direct, flags);
7295         if (rw & WRITE) {
7296                 if (ret < 0 && ret != -EIOCBQUEUED)
7297                         btrfs_delalloc_release_space(inode, count);
7298                 else if (ret >= 0 && (size_t)ret < count)
7299                         btrfs_delalloc_release_space(inode,
7300                                                      count - (size_t)ret);
7301                 else
7302                         btrfs_delalloc_release_metadata(inode, 0);
7303         }
7304 out:
7305         if (wakeup)
7306                 inode_dio_done(inode);
7307         if (relock)
7308                 mutex_lock(&inode->i_mutex);
7309
7310         return ret;
7311 }
7312
7313 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
7314
7315 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
7316                 __u64 start, __u64 len)
7317 {
7318         int     ret;
7319
7320         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
7321         if (ret)
7322                 return ret;
7323
7324         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
7325 }
7326
7327 int btrfs_readpage(struct file *file, struct page *page)
7328 {
7329         struct extent_io_tree *tree;
7330         tree = &BTRFS_I(page->mapping->host)->io_tree;
7331         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
7332 }
7333
7334 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
7335 {
7336         struct extent_io_tree *tree;
7337
7338
7339         if (current->flags & PF_MEMALLOC) {
7340                 redirty_page_for_writepage(wbc, page);
7341                 unlock_page(page);
7342                 return 0;
7343         }
7344         tree = &BTRFS_I(page->mapping->host)->io_tree;
7345         return extent_write_full_page(tree, page, btrfs_get_extent, wbc);
7346 }
7347
7348 static int btrfs_writepages(struct address_space *mapping,
7349                             struct writeback_control *wbc)
7350 {
7351         struct extent_io_tree *tree;
7352
7353         tree = &BTRFS_I(mapping->host)->io_tree;
7354         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
7355 }
7356
7357 static int
7358 btrfs_readpages(struct file *file, struct address_space *mapping,
7359                 struct list_head *pages, unsigned nr_pages)
7360 {
7361         struct extent_io_tree *tree;
7362         tree = &BTRFS_I(mapping->host)->io_tree;
7363         return extent_readpages(tree, mapping, pages, nr_pages,
7364                                 btrfs_get_extent);
7365 }
7366 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
7367 {
7368         struct extent_io_tree *tree;
7369         struct extent_map_tree *map;
7370         int ret;
7371
7372         tree = &BTRFS_I(page->mapping->host)->io_tree;
7373         map = &BTRFS_I(page->mapping->host)->extent_tree;
7374         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
7375         if (ret == 1) {
7376                 ClearPagePrivate(page);
7377                 set_page_private(page, 0);
7378                 page_cache_release(page);
7379         }
7380         return ret;
7381 }
7382
7383 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
7384 {
7385         if (PageWriteback(page) || PageDirty(page))
7386                 return 0;
7387         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
7388 }
7389
7390 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
7391                                  unsigned int length)
7392 {
7393         struct inode *inode = page->mapping->host;
7394         struct extent_io_tree *tree;
7395         struct btrfs_ordered_extent *ordered;
7396         struct extent_state *cached_state = NULL;
7397         u64 page_start = page_offset(page);
7398         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
7399
7400         /*
7401          * we have the page locked, so new writeback can't start,
7402          * and the dirty bit won't be cleared while we are here.
7403          *
7404          * Wait for IO on this page so that we can safely clear
7405          * the PagePrivate2 bit and do ordered accounting
7406          */
7407         wait_on_page_writeback(page);
7408
7409         tree = &BTRFS_I(inode)->io_tree;
7410         if (offset) {
7411                 btrfs_releasepage(page, GFP_NOFS);
7412                 return;
7413         }
7414         lock_extent_bits(tree, page_start, page_end, 0, &cached_state);
7415         ordered = btrfs_lookup_ordered_extent(inode, page_offset(page));
7416         if (ordered) {
7417                 /*
7418                  * IO on this page will never be started, so we need
7419                  * to account for any ordered extents now
7420                  */
7421                 clear_extent_bit(tree, page_start, page_end,
7422                                  EXTENT_DIRTY | EXTENT_DELALLOC |
7423                                  EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
7424                                  EXTENT_DEFRAG, 1, 0, &cached_state, GFP_NOFS);
7425                 /*
7426                  * whoever cleared the private bit is responsible
7427                  * for the finish_ordered_io
7428                  */
7429                 if (TestClearPagePrivate2(page) &&
7430                     btrfs_dec_test_ordered_pending(inode, &ordered, page_start,
7431                                                    PAGE_CACHE_SIZE, 1)) {
7432                         btrfs_finish_ordered_io(ordered);
7433                 }
7434                 btrfs_put_ordered_extent(ordered);
7435                 cached_state = NULL;
7436                 lock_extent_bits(tree, page_start, page_end, 0, &cached_state);
7437         }
7438         clear_extent_bit(tree, page_start, page_end,
7439                  EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
7440                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
7441                  &cached_state, GFP_NOFS);
7442         __btrfs_releasepage(page, GFP_NOFS);
7443
7444         ClearPageChecked(page);
7445         if (PagePrivate(page)) {
7446                 ClearPagePrivate(page);
7447                 set_page_private(page, 0);
7448                 page_cache_release(page);
7449         }
7450 }
7451
7452 /*
7453  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
7454  * called from a page fault handler when a page is first dirtied. Hence we must
7455  * be careful to check for EOF conditions here. We set the page up correctly
7456  * for a written page which means we get ENOSPC checking when writing into
7457  * holes and correct delalloc and unwritten extent mapping on filesystems that
7458  * support these features.
7459  *
7460  * We are not allowed to take the i_mutex here so we have to play games to
7461  * protect against truncate races as the page could now be beyond EOF.  Because
7462  * vmtruncate() writes the inode size before removing pages, once we have the
7463  * page lock we can determine safely if the page is beyond EOF. If it is not
7464  * beyond EOF, then the page is guaranteed safe against truncation until we
7465  * unlock the page.
7466  */
7467 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
7468 {
7469         struct page *page = vmf->page;
7470         struct inode *inode = file_inode(vma->vm_file);
7471         struct btrfs_root *root = BTRFS_I(inode)->root;
7472         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7473         struct btrfs_ordered_extent *ordered;
7474         struct extent_state *cached_state = NULL;
7475         char *kaddr;
7476         unsigned long zero_start;
7477         loff_t size;
7478         int ret;
7479         int reserved = 0;
7480         u64 page_start;
7481         u64 page_end;
7482
7483         sb_start_pagefault(inode->i_sb);
7484         ret  = btrfs_delalloc_reserve_space(inode, PAGE_CACHE_SIZE);
7485         if (!ret) {
7486                 ret = file_update_time(vma->vm_file);
7487                 reserved = 1;
7488         }
7489         if (ret) {
7490                 if (ret == -ENOMEM)
7491                         ret = VM_FAULT_OOM;
7492                 else /* -ENOSPC, -EIO, etc */
7493                         ret = VM_FAULT_SIGBUS;
7494                 if (reserved)
7495                         goto out;
7496                 goto out_noreserve;
7497         }
7498
7499         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
7500 again:
7501         lock_page(page);
7502         size = i_size_read(inode);
7503         page_start = page_offset(page);
7504         page_end = page_start + PAGE_CACHE_SIZE - 1;
7505
7506         if ((page->mapping != inode->i_mapping) ||
7507             (page_start >= size)) {
7508                 /* page got truncated out from underneath us */
7509                 goto out_unlock;
7510         }
7511         wait_on_page_writeback(page);
7512
7513         lock_extent_bits(io_tree, page_start, page_end, 0, &cached_state);
7514         set_page_extent_mapped(page);
7515
7516         /*
7517          * we can't set the delalloc bits if there are pending ordered
7518          * extents.  Drop our locks and wait for them to finish
7519          */
7520         ordered = btrfs_lookup_ordered_extent(inode, page_start);
7521         if (ordered) {
7522                 unlock_extent_cached(io_tree, page_start, page_end,
7523                                      &cached_state, GFP_NOFS);
7524                 unlock_page(page);
7525                 btrfs_start_ordered_extent(inode, ordered, 1);
7526                 btrfs_put_ordered_extent(ordered);
7527                 goto again;
7528         }
7529
7530         /*
7531          * XXX - page_mkwrite gets called every time the page is dirtied, even
7532          * if it was already dirty, so for space accounting reasons we need to
7533          * clear any delalloc bits for the range we are fixing to save.  There
7534          * is probably a better way to do this, but for now keep consistent with
7535          * prepare_pages in the normal write path.
7536          */
7537         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
7538                           EXTENT_DIRTY | EXTENT_DELALLOC |
7539                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
7540                           0, 0, &cached_state, GFP_NOFS);
7541
7542         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
7543                                         &cached_state);
7544         if (ret) {
7545                 unlock_extent_cached(io_tree, page_start, page_end,
7546                                      &cached_state, GFP_NOFS);
7547                 ret = VM_FAULT_SIGBUS;
7548                 goto out_unlock;
7549         }
7550         ret = 0;
7551
7552         /* page is wholly or partially inside EOF */
7553         if (page_start + PAGE_CACHE_SIZE > size)
7554                 zero_start = size & ~PAGE_CACHE_MASK;
7555         else
7556                 zero_start = PAGE_CACHE_SIZE;
7557
7558         if (zero_start != PAGE_CACHE_SIZE) {
7559                 kaddr = kmap(page);
7560                 memset(kaddr + zero_start, 0, PAGE_CACHE_SIZE - zero_start);
7561                 flush_dcache_page(page);
7562                 kunmap(page);
7563         }
7564         ClearPageChecked(page);
7565         set_page_dirty(page);
7566         SetPageUptodate(page);
7567
7568         BTRFS_I(inode)->last_trans = root->fs_info->generation;
7569         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
7570         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
7571
7572         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
7573
7574 out_unlock:
7575         if (!ret) {
7576                 sb_end_pagefault(inode->i_sb);
7577                 return VM_FAULT_LOCKED;
7578         }
7579         unlock_page(page);
7580 out:
7581         btrfs_delalloc_release_space(inode, PAGE_CACHE_SIZE);
7582 out_noreserve:
7583         sb_end_pagefault(inode->i_sb);
7584         return ret;
7585 }
7586
7587 static int btrfs_truncate(struct inode *inode)
7588 {
7589         struct btrfs_root *root = BTRFS_I(inode)->root;
7590         struct btrfs_block_rsv *rsv;
7591         int ret = 0;
7592         int err = 0;
7593         struct btrfs_trans_handle *trans;
7594         u64 mask = root->sectorsize - 1;
7595         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
7596
7597         btrfs_wait_ordered_range(inode, inode->i_size & (~mask), (u64)-1);
7598         btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
7599
7600         /*
7601          * Yes ladies and gentelment, this is indeed ugly.  The fact is we have
7602          * 3 things going on here
7603          *
7604          * 1) We need to reserve space for our orphan item and the space to
7605          * delete our orphan item.  Lord knows we don't want to have a dangling
7606          * orphan item because we didn't reserve space to remove it.
7607          *
7608          * 2) We need to reserve space to update our inode.
7609          *
7610          * 3) We need to have something to cache all the space that is going to
7611          * be free'd up by the truncate operation, but also have some slack
7612          * space reserved in case it uses space during the truncate (thank you
7613          * very much snapshotting).
7614          *
7615          * And we need these to all be seperate.  The fact is we can use alot of
7616          * space doing the truncate, and we have no earthly idea how much space
7617          * we will use, so we need the truncate reservation to be seperate so it
7618          * doesn't end up using space reserved for updating the inode or
7619          * removing the orphan item.  We also need to be able to stop the
7620          * transaction and start a new one, which means we need to be able to
7621          * update the inode several times, and we have no idea of knowing how
7622          * many times that will be, so we can't just reserve 1 item for the
7623          * entirety of the opration, so that has to be done seperately as well.
7624          * Then there is the orphan item, which does indeed need to be held on
7625          * to for the whole operation, and we need nobody to touch this reserved
7626          * space except the orphan code.
7627          *
7628          * So that leaves us with
7629          *
7630          * 1) root->orphan_block_rsv - for the orphan deletion.
7631          * 2) rsv - for the truncate reservation, which we will steal from the
7632          * transaction reservation.
7633          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
7634          * updating the inode.
7635          */
7636         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
7637         if (!rsv)
7638                 return -ENOMEM;
7639         rsv->size = min_size;
7640         rsv->failfast = 1;
7641
7642         /*
7643          * 1 for the truncate slack space
7644          * 1 for updating the inode.
7645          */
7646         trans = btrfs_start_transaction(root, 2);
7647         if (IS_ERR(trans)) {
7648                 err = PTR_ERR(trans);
7649                 goto out;
7650         }
7651
7652         /* Migrate the slack space for the truncate to our reserve */
7653         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
7654                                       min_size);
7655         BUG_ON(ret);
7656
7657         /*
7658          * setattr is responsible for setting the ordered_data_close flag,
7659          * but that is only tested during the last file release.  That
7660          * could happen well after the next commit, leaving a great big
7661          * window where new writes may get lost if someone chooses to write
7662          * to this file after truncating to zero
7663          *
7664          * The inode doesn't have any dirty data here, and so if we commit
7665          * this is a noop.  If someone immediately starts writing to the inode
7666          * it is very likely we'll catch some of their writes in this
7667          * transaction, and the commit will find this file on the ordered
7668          * data list with good things to send down.
7669          *
7670          * This is a best effort solution, there is still a window where
7671          * using truncate to replace the contents of the file will
7672          * end up with a zero length file after a crash.
7673          */
7674         if (inode->i_size == 0 && test_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
7675                                            &BTRFS_I(inode)->runtime_flags))
7676                 btrfs_add_ordered_operation(trans, root, inode);
7677
7678         /*
7679          * So if we truncate and then write and fsync we normally would just
7680          * write the extents that changed, which is a problem if we need to
7681          * first truncate that entire inode.  So set this flag so we write out
7682          * all of the extents in the inode to the sync log so we're completely
7683          * safe.
7684          */
7685         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
7686         trans->block_rsv = rsv;
7687
7688         while (1) {
7689                 ret = btrfs_truncate_inode_items(trans, root, inode,
7690                                                  inode->i_size,
7691                                                  BTRFS_EXTENT_DATA_KEY);
7692                 if (ret != -ENOSPC) {
7693                         err = ret;
7694                         break;
7695                 }
7696
7697                 trans->block_rsv = &root->fs_info->trans_block_rsv;
7698                 ret = btrfs_update_inode(trans, root, inode);
7699                 if (ret) {
7700                         err = ret;
7701                         break;
7702                 }
7703
7704                 btrfs_end_transaction(trans, root);
7705                 btrfs_btree_balance_dirty(root);
7706
7707                 trans = btrfs_start_transaction(root, 2);
7708                 if (IS_ERR(trans)) {
7709                         ret = err = PTR_ERR(trans);
7710                         trans = NULL;
7711                         break;
7712                 }
7713
7714                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
7715                                               rsv, min_size);
7716                 BUG_ON(ret);    /* shouldn't happen */
7717                 trans->block_rsv = rsv;
7718         }
7719
7720         if (ret == 0 && inode->i_nlink > 0) {
7721                 trans->block_rsv = root->orphan_block_rsv;
7722                 ret = btrfs_orphan_del(trans, inode);
7723                 if (ret)
7724                         err = ret;
7725         }
7726
7727         if (trans) {
7728                 trans->block_rsv = &root->fs_info->trans_block_rsv;
7729                 ret = btrfs_update_inode(trans, root, inode);
7730                 if (ret && !err)
7731                         err = ret;
7732
7733                 ret = btrfs_end_transaction(trans, root);
7734                 btrfs_btree_balance_dirty(root);
7735         }
7736
7737 out:
7738         btrfs_free_block_rsv(root, rsv);
7739
7740         if (ret && !err)
7741                 err = ret;
7742
7743         return err;
7744 }
7745
7746 /*
7747  * create a new subvolume directory/inode (helper for the ioctl).
7748  */
7749 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
7750                              struct btrfs_root *new_root, u64 new_dirid)
7751 {
7752         struct inode *inode;
7753         int err;
7754         u64 index = 0;
7755
7756         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
7757                                 new_dirid, new_dirid,
7758                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
7759                                 &index);
7760         if (IS_ERR(inode))
7761                 return PTR_ERR(inode);
7762         inode->i_op = &btrfs_dir_inode_operations;
7763         inode->i_fop = &btrfs_dir_file_operations;
7764
7765         set_nlink(inode, 1);
7766         btrfs_i_size_write(inode, 0);
7767
7768         err = btrfs_update_inode(trans, new_root, inode);
7769
7770         iput(inode);
7771         return err;
7772 }
7773
7774 struct inode *btrfs_alloc_inode(struct super_block *sb)
7775 {
7776         struct btrfs_inode *ei;
7777         struct inode *inode;
7778
7779         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
7780         if (!ei)
7781                 return NULL;
7782
7783         ei->root = NULL;
7784         ei->generation = 0;
7785         ei->last_trans = 0;
7786         ei->last_sub_trans = 0;
7787         ei->logged_trans = 0;
7788         ei->delalloc_bytes = 0;
7789         ei->disk_i_size = 0;
7790         ei->flags = 0;
7791         ei->csum_bytes = 0;
7792         ei->index_cnt = (u64)-1;
7793         ei->last_unlink_trans = 0;
7794         ei->last_log_commit = 0;
7795
7796         spin_lock_init(&ei->lock);
7797         ei->outstanding_extents = 0;
7798         ei->reserved_extents = 0;
7799
7800         ei->runtime_flags = 0;
7801         ei->force_compress = BTRFS_COMPRESS_NONE;
7802
7803         ei->delayed_node = NULL;
7804
7805         inode = &ei->vfs_inode;
7806         extent_map_tree_init(&ei->extent_tree);
7807         extent_io_tree_init(&ei->io_tree, &inode->i_data);
7808         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
7809         ei->io_tree.track_uptodate = 1;
7810         ei->io_failure_tree.track_uptodate = 1;
7811         atomic_set(&ei->sync_writers, 0);
7812         mutex_init(&ei->log_mutex);
7813         mutex_init(&ei->delalloc_mutex);
7814         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
7815         INIT_LIST_HEAD(&ei->delalloc_inodes);
7816         INIT_LIST_HEAD(&ei->ordered_operations);
7817         RB_CLEAR_NODE(&ei->rb_node);
7818
7819         return inode;
7820 }
7821
7822 static void btrfs_i_callback(struct rcu_head *head)
7823 {
7824         struct inode *inode = container_of(head, struct inode, i_rcu);
7825         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
7826 }
7827
7828 void btrfs_destroy_inode(struct inode *inode)
7829 {
7830         struct btrfs_ordered_extent *ordered;
7831         struct btrfs_root *root = BTRFS_I(inode)->root;
7832
7833         WARN_ON(!hlist_empty(&inode->i_dentry));
7834         WARN_ON(inode->i_data.nrpages);
7835         WARN_ON(BTRFS_I(inode)->outstanding_extents);
7836         WARN_ON(BTRFS_I(inode)->reserved_extents);
7837         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
7838         WARN_ON(BTRFS_I(inode)->csum_bytes);
7839
7840         /*
7841          * This can happen where we create an inode, but somebody else also
7842          * created the same inode and we need to destroy the one we already
7843          * created.
7844          */
7845         if (!root)
7846                 goto free;
7847
7848         /*
7849          * Make sure we're properly removed from the ordered operation
7850          * lists.
7851          */
7852         smp_mb();
7853         if (!list_empty(&BTRFS_I(inode)->ordered_operations)) {
7854                 spin_lock(&root->fs_info->ordered_root_lock);
7855                 list_del_init(&BTRFS_I(inode)->ordered_operations);
7856                 spin_unlock(&root->fs_info->ordered_root_lock);
7857         }
7858
7859         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
7860                      &BTRFS_I(inode)->runtime_flags)) {
7861                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
7862                         (unsigned long long)btrfs_ino(inode));
7863                 atomic_dec(&root->orphan_inodes);
7864         }
7865
7866         while (1) {
7867                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
7868                 if (!ordered)
7869                         break;
7870                 else {
7871                         btrfs_err(root->fs_info, "found ordered extent %llu %llu on inode cleanup",
7872                                 (unsigned long long)ordered->file_offset,
7873                                 (unsigned long long)ordered->len);
7874                         btrfs_remove_ordered_extent(inode, ordered);
7875                         btrfs_put_ordered_extent(ordered);
7876                         btrfs_put_ordered_extent(ordered);
7877                 }
7878         }
7879         inode_tree_del(inode);
7880         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
7881 free:
7882         call_rcu(&inode->i_rcu, btrfs_i_callback);
7883 }
7884
7885 int btrfs_drop_inode(struct inode *inode)
7886 {
7887         struct btrfs_root *root = BTRFS_I(inode)->root;
7888
7889         if (root == NULL)
7890                 return 1;
7891
7892         /* the snap/subvol tree is on deleting */
7893         if (btrfs_root_refs(&root->root_item) == 0 &&
7894             root != root->fs_info->tree_root)
7895                 return 1;
7896         else
7897                 return generic_drop_inode(inode);
7898 }
7899
7900 static void init_once(void *foo)
7901 {
7902         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
7903
7904         inode_init_once(&ei->vfs_inode);
7905 }
7906
7907 void btrfs_destroy_cachep(void)
7908 {
7909         /*
7910          * Make sure all delayed rcu free inodes are flushed before we
7911          * destroy cache.
7912          */
7913         rcu_barrier();
7914         if (btrfs_inode_cachep)
7915                 kmem_cache_destroy(btrfs_inode_cachep);
7916         if (btrfs_trans_handle_cachep)
7917                 kmem_cache_destroy(btrfs_trans_handle_cachep);
7918         if (btrfs_transaction_cachep)
7919                 kmem_cache_destroy(btrfs_transaction_cachep);
7920         if (btrfs_path_cachep)
7921                 kmem_cache_destroy(btrfs_path_cachep);
7922         if (btrfs_free_space_cachep)
7923                 kmem_cache_destroy(btrfs_free_space_cachep);
7924         if (btrfs_delalloc_work_cachep)
7925                 kmem_cache_destroy(btrfs_delalloc_work_cachep);
7926 }
7927
7928 int btrfs_init_cachep(void)
7929 {
7930         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
7931                         sizeof(struct btrfs_inode), 0,
7932                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, init_once);
7933         if (!btrfs_inode_cachep)
7934                 goto fail;
7935
7936         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
7937                         sizeof(struct btrfs_trans_handle), 0,
7938                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
7939         if (!btrfs_trans_handle_cachep)
7940                 goto fail;
7941
7942         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
7943                         sizeof(struct btrfs_transaction), 0,
7944                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
7945         if (!btrfs_transaction_cachep)
7946                 goto fail;
7947
7948         btrfs_path_cachep = kmem_cache_create("btrfs_path",
7949                         sizeof(struct btrfs_path), 0,
7950                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
7951         if (!btrfs_path_cachep)
7952                 goto fail;
7953
7954         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
7955                         sizeof(struct btrfs_free_space), 0,
7956                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
7957         if (!btrfs_free_space_cachep)
7958                 goto fail;
7959
7960         btrfs_delalloc_work_cachep = kmem_cache_create("btrfs_delalloc_work",
7961                         sizeof(struct btrfs_delalloc_work), 0,
7962                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
7963                         NULL);
7964         if (!btrfs_delalloc_work_cachep)
7965                 goto fail;
7966
7967         return 0;
7968 fail:
7969         btrfs_destroy_cachep();
7970         return -ENOMEM;
7971 }
7972
7973 static int btrfs_getattr(struct vfsmount *mnt,
7974                          struct dentry *dentry, struct kstat *stat)
7975 {
7976         u64 delalloc_bytes;
7977         struct inode *inode = dentry->d_inode;
7978         u32 blocksize = inode->i_sb->s_blocksize;
7979
7980         generic_fillattr(inode, stat);
7981         stat->dev = BTRFS_I(inode)->root->anon_dev;
7982         stat->blksize = PAGE_CACHE_SIZE;
7983
7984         spin_lock(&BTRFS_I(inode)->lock);
7985         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
7986         spin_unlock(&BTRFS_I(inode)->lock);
7987         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
7988                         ALIGN(delalloc_bytes, blocksize)) >> 9;
7989         return 0;
7990 }
7991
7992 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
7993                            struct inode *new_dir, struct dentry *new_dentry)
7994 {
7995         struct btrfs_trans_handle *trans;
7996         struct btrfs_root *root = BTRFS_I(old_dir)->root;
7997         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
7998         struct inode *new_inode = new_dentry->d_inode;
7999         struct inode *old_inode = old_dentry->d_inode;
8000         struct timespec ctime = CURRENT_TIME;
8001         u64 index = 0;
8002         u64 root_objectid;
8003         int ret;
8004         u64 old_ino = btrfs_ino(old_inode);
8005
8006         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
8007                 return -EPERM;
8008
8009         /* we only allow rename subvolume link between subvolumes */
8010         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
8011                 return -EXDEV;
8012
8013         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
8014             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
8015                 return -ENOTEMPTY;
8016
8017         if (S_ISDIR(old_inode->i_mode) && new_inode &&
8018             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
8019                 return -ENOTEMPTY;
8020
8021
8022         /* check for collisions, even if the  name isn't there */
8023         ret = btrfs_check_dir_item_collision(root, new_dir->i_ino,
8024                              new_dentry->d_name.name,
8025                              new_dentry->d_name.len);
8026
8027         if (ret) {
8028                 if (ret == -EEXIST) {
8029                         /* we shouldn't get
8030                          * eexist without a new_inode */
8031                         if (!new_inode) {
8032                                 WARN_ON(1);
8033                                 return ret;
8034                         }
8035                 } else {
8036                         /* maybe -EOVERFLOW */
8037                         return ret;
8038                 }
8039         }
8040         ret = 0;
8041
8042         /*
8043          * we're using rename to replace one file with another.
8044          * and the replacement file is large.  Start IO on it now so
8045          * we don't add too much work to the end of the transaction
8046          */
8047         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size &&
8048             old_inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
8049                 filemap_flush(old_inode->i_mapping);
8050
8051         /* close the racy window with snapshot create/destroy ioctl */
8052         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
8053                 down_read(&root->fs_info->subvol_sem);
8054         /*
8055          * We want to reserve the absolute worst case amount of items.  So if
8056          * both inodes are subvols and we need to unlink them then that would
8057          * require 4 item modifications, but if they are both normal inodes it
8058          * would require 5 item modifications, so we'll assume their normal
8059          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
8060          * should cover the worst case number of items we'll modify.
8061          */
8062         trans = btrfs_start_transaction(root, 11);
8063         if (IS_ERR(trans)) {
8064                 ret = PTR_ERR(trans);
8065                 goto out_notrans;
8066         }
8067
8068         if (dest != root)
8069                 btrfs_record_root_in_trans(trans, dest);
8070
8071         ret = btrfs_set_inode_index(new_dir, &index);
8072         if (ret)
8073                 goto out_fail;
8074
8075         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
8076                 /* force full log commit if subvolume involved. */
8077                 root->fs_info->last_trans_log_full_commit = trans->transid;
8078         } else {
8079                 ret = btrfs_insert_inode_ref(trans, dest,
8080                                              new_dentry->d_name.name,
8081                                              new_dentry->d_name.len,
8082                                              old_ino,
8083                                              btrfs_ino(new_dir), index);
8084                 if (ret)
8085                         goto out_fail;
8086                 /*
8087                  * this is an ugly little race, but the rename is required
8088                  * to make sure that if we crash, the inode is either at the
8089                  * old name or the new one.  pinning the log transaction lets
8090                  * us make sure we don't allow a log commit to come in after
8091                  * we unlink the name but before we add the new name back in.
8092                  */
8093                 btrfs_pin_log_trans(root);
8094         }
8095         /*
8096          * make sure the inode gets flushed if it is replacing
8097          * something.
8098          */
8099         if (new_inode && new_inode->i_size && S_ISREG(old_inode->i_mode))
8100                 btrfs_add_ordered_operation(trans, root, old_inode);
8101
8102         inode_inc_iversion(old_dir);
8103         inode_inc_iversion(new_dir);
8104         inode_inc_iversion(old_inode);
8105         old_dir->i_ctime = old_dir->i_mtime = ctime;
8106         new_dir->i_ctime = new_dir->i_mtime = ctime;
8107         old_inode->i_ctime = ctime;
8108
8109         if (old_dentry->d_parent != new_dentry->d_parent)
8110                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
8111
8112         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
8113                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
8114                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
8115                                         old_dentry->d_name.name,
8116                                         old_dentry->d_name.len);
8117         } else {
8118                 ret = __btrfs_unlink_inode(trans, root, old_dir,
8119                                         old_dentry->d_inode,
8120                                         old_dentry->d_name.name,
8121                                         old_dentry->d_name.len);
8122                 if (!ret)
8123                         ret = btrfs_update_inode(trans, root, old_inode);
8124         }
8125         if (ret) {
8126                 btrfs_abort_transaction(trans, root, ret);
8127                 goto out_fail;
8128         }
8129
8130         if (new_inode) {
8131                 inode_inc_iversion(new_inode);
8132                 new_inode->i_ctime = CURRENT_TIME;
8133                 if (unlikely(btrfs_ino(new_inode) ==
8134                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
8135                         root_objectid = BTRFS_I(new_inode)->location.objectid;
8136                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
8137                                                 root_objectid,
8138                                                 new_dentry->d_name.name,
8139                                                 new_dentry->d_name.len);
8140                         BUG_ON(new_inode->i_nlink == 0);
8141                 } else {
8142                         ret = btrfs_unlink_inode(trans, dest, new_dir,
8143                                                  new_dentry->d_inode,
8144                                                  new_dentry->d_name.name,
8145                                                  new_dentry->d_name.len);
8146                 }
8147                 if (!ret && new_inode->i_nlink == 0) {
8148                         ret = btrfs_orphan_add(trans, new_dentry->d_inode);
8149                         BUG_ON(ret);
8150                 }
8151                 if (ret) {
8152                         btrfs_abort_transaction(trans, root, ret);
8153                         goto out_fail;
8154                 }
8155         }
8156
8157         ret = btrfs_add_link(trans, new_dir, old_inode,
8158                              new_dentry->d_name.name,
8159                              new_dentry->d_name.len, 0, index);
8160         if (ret) {
8161                 btrfs_abort_transaction(trans, root, ret);
8162                 goto out_fail;
8163         }
8164
8165         if (old_ino != BTRFS_FIRST_FREE_OBJECTID) {
8166                 struct dentry *parent = new_dentry->d_parent;
8167                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
8168                 btrfs_end_log_trans(root);
8169         }
8170 out_fail:
8171         btrfs_end_transaction(trans, root);
8172 out_notrans:
8173         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
8174                 up_read(&root->fs_info->subvol_sem);
8175
8176         return ret;
8177 }
8178
8179 static void btrfs_run_delalloc_work(struct btrfs_work *work)
8180 {
8181         struct btrfs_delalloc_work *delalloc_work;
8182
8183         delalloc_work = container_of(work, struct btrfs_delalloc_work,
8184                                      work);
8185         if (delalloc_work->wait)
8186                 btrfs_wait_ordered_range(delalloc_work->inode, 0, (u64)-1);
8187         else
8188                 filemap_flush(delalloc_work->inode->i_mapping);
8189
8190         if (delalloc_work->delay_iput)
8191                 btrfs_add_delayed_iput(delalloc_work->inode);
8192         else
8193                 iput(delalloc_work->inode);
8194         complete(&delalloc_work->completion);
8195 }
8196
8197 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
8198                                                     int wait, int delay_iput)
8199 {
8200         struct btrfs_delalloc_work *work;
8201
8202         work = kmem_cache_zalloc(btrfs_delalloc_work_cachep, GFP_NOFS);
8203         if (!work)
8204                 return NULL;
8205
8206         init_completion(&work->completion);
8207         INIT_LIST_HEAD(&work->list);
8208         work->inode = inode;
8209         work->wait = wait;
8210         work->delay_iput = delay_iput;
8211         work->work.func = btrfs_run_delalloc_work;
8212
8213         return work;
8214 }
8215
8216 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
8217 {
8218         wait_for_completion(&work->completion);
8219         kmem_cache_free(btrfs_delalloc_work_cachep, work);
8220 }
8221
8222 /*
8223  * some fairly slow code that needs optimization. This walks the list
8224  * of all the inodes with pending delalloc and forces them to disk.
8225  */
8226 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
8227 {
8228         struct btrfs_inode *binode;
8229         struct inode *inode;
8230         struct btrfs_delalloc_work *work, *next;
8231         struct list_head works;
8232         struct list_head splice;
8233         int ret = 0;
8234
8235         INIT_LIST_HEAD(&works);
8236         INIT_LIST_HEAD(&splice);
8237
8238         spin_lock(&root->delalloc_lock);
8239         list_splice_init(&root->delalloc_inodes, &splice);
8240         while (!list_empty(&splice)) {
8241                 binode = list_entry(splice.next, struct btrfs_inode,
8242                                     delalloc_inodes);
8243
8244                 list_move_tail(&binode->delalloc_inodes,
8245                                &root->delalloc_inodes);
8246                 inode = igrab(&binode->vfs_inode);
8247                 if (!inode) {
8248                         cond_resched_lock(&root->delalloc_lock);
8249                         continue;
8250                 }
8251                 spin_unlock(&root->delalloc_lock);
8252
8253                 work = btrfs_alloc_delalloc_work(inode, 0, delay_iput);
8254                 if (unlikely(!work)) {
8255                         ret = -ENOMEM;
8256                         goto out;
8257                 }
8258                 list_add_tail(&work->list, &works);
8259                 btrfs_queue_worker(&root->fs_info->flush_workers,
8260                                    &work->work);
8261
8262                 cond_resched();
8263                 spin_lock(&root->delalloc_lock);
8264         }
8265         spin_unlock(&root->delalloc_lock);
8266
8267         list_for_each_entry_safe(work, next, &works, list) {
8268                 list_del_init(&work->list);
8269                 btrfs_wait_and_free_delalloc_work(work);
8270         }
8271         return 0;
8272 out:
8273         list_for_each_entry_safe(work, next, &works, list) {
8274                 list_del_init(&work->list);
8275                 btrfs_wait_and_free_delalloc_work(work);
8276         }
8277
8278         if (!list_empty_careful(&splice)) {
8279                 spin_lock(&root->delalloc_lock);
8280                 list_splice_tail(&splice, &root->delalloc_inodes);
8281                 spin_unlock(&root->delalloc_lock);
8282         }
8283         return ret;
8284 }
8285
8286 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
8287 {
8288         int ret;
8289
8290         if (root->fs_info->sb->s_flags & MS_RDONLY)
8291                 return -EROFS;
8292
8293         ret = __start_delalloc_inodes(root, delay_iput);
8294         /*
8295          * the filemap_flush will queue IO into the worker threads, but
8296          * we have to make sure the IO is actually started and that
8297          * ordered extents get created before we return
8298          */
8299         atomic_inc(&root->fs_info->async_submit_draining);
8300         while (atomic_read(&root->fs_info->nr_async_submits) ||
8301               atomic_read(&root->fs_info->async_delalloc_pages)) {
8302                 wait_event(root->fs_info->async_submit_wait,
8303                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
8304                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
8305         }
8306         atomic_dec(&root->fs_info->async_submit_draining);
8307         return ret;
8308 }
8309
8310 int btrfs_start_all_delalloc_inodes(struct btrfs_fs_info *fs_info,
8311                                     int delay_iput)
8312 {
8313         struct btrfs_root *root;
8314         struct list_head splice;
8315         int ret;
8316
8317         if (fs_info->sb->s_flags & MS_RDONLY)
8318                 return -EROFS;
8319
8320         INIT_LIST_HEAD(&splice);
8321
8322         spin_lock(&fs_info->delalloc_root_lock);
8323         list_splice_init(&fs_info->delalloc_roots, &splice);
8324         while (!list_empty(&splice)) {
8325                 root = list_first_entry(&splice, struct btrfs_root,
8326                                         delalloc_root);
8327                 root = btrfs_grab_fs_root(root);
8328                 BUG_ON(!root);
8329                 list_move_tail(&root->delalloc_root,
8330                                &fs_info->delalloc_roots);
8331                 spin_unlock(&fs_info->delalloc_root_lock);
8332
8333                 ret = __start_delalloc_inodes(root, delay_iput);
8334                 btrfs_put_fs_root(root);
8335                 if (ret)
8336                         goto out;
8337
8338                 spin_lock(&fs_info->delalloc_root_lock);
8339         }
8340         spin_unlock(&fs_info->delalloc_root_lock);
8341
8342         atomic_inc(&fs_info->async_submit_draining);
8343         while (atomic_read(&fs_info->nr_async_submits) ||
8344               atomic_read(&fs_info->async_delalloc_pages)) {
8345                 wait_event(fs_info->async_submit_wait,
8346                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
8347                     atomic_read(&fs_info->async_delalloc_pages) == 0));
8348         }
8349         atomic_dec(&fs_info->async_submit_draining);
8350         return 0;
8351 out:
8352         if (!list_empty_careful(&splice)) {
8353                 spin_lock(&fs_info->delalloc_root_lock);
8354                 list_splice_tail(&splice, &fs_info->delalloc_roots);
8355                 spin_unlock(&fs_info->delalloc_root_lock);
8356         }
8357         return ret;
8358 }
8359
8360 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
8361                          const char *symname)
8362 {
8363         struct btrfs_trans_handle *trans;
8364         struct btrfs_root *root = BTRFS_I(dir)->root;
8365         struct btrfs_path *path;
8366         struct btrfs_key key;
8367         struct inode *inode = NULL;
8368         int err;
8369         int drop_inode = 0;
8370         u64 objectid;
8371         u64 index = 0 ;
8372         int name_len;
8373         int datasize;
8374         unsigned long ptr;
8375         struct btrfs_file_extent_item *ei;
8376         struct extent_buffer *leaf;
8377
8378         name_len = strlen(symname) + 1;
8379         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
8380                 return -ENAMETOOLONG;
8381
8382         /*
8383          * 2 items for inode item and ref
8384          * 2 items for dir items
8385          * 1 item for xattr if selinux is on
8386          */
8387         trans = btrfs_start_transaction(root, 5);
8388         if (IS_ERR(trans))
8389                 return PTR_ERR(trans);
8390
8391         err = btrfs_find_free_ino(root, &objectid);
8392         if (err)
8393                 goto out_unlock;
8394
8395         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
8396                                 dentry->d_name.len, btrfs_ino(dir), objectid,
8397                                 S_IFLNK|S_IRWXUGO, &index);
8398         if (IS_ERR(inode)) {
8399                 err = PTR_ERR(inode);
8400                 goto out_unlock;
8401         }
8402
8403         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
8404         if (err) {
8405                 drop_inode = 1;
8406                 goto out_unlock;
8407         }
8408
8409         /*
8410         * If the active LSM wants to access the inode during
8411         * d_instantiate it needs these. Smack checks to see
8412         * if the filesystem supports xattrs by looking at the
8413         * ops vector.
8414         */
8415         inode->i_fop = &btrfs_file_operations;
8416         inode->i_op = &btrfs_file_inode_operations;
8417
8418         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
8419         if (err)
8420                 drop_inode = 1;
8421         else {
8422                 inode->i_mapping->a_ops = &btrfs_aops;
8423                 inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
8424                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
8425         }
8426         if (drop_inode)
8427                 goto out_unlock;
8428
8429         path = btrfs_alloc_path();
8430         if (!path) {
8431                 err = -ENOMEM;
8432                 drop_inode = 1;
8433                 goto out_unlock;
8434         }
8435         key.objectid = btrfs_ino(inode);
8436         key.offset = 0;
8437         btrfs_set_key_type(&key, BTRFS_EXTENT_DATA_KEY);
8438         datasize = btrfs_file_extent_calc_inline_size(name_len);
8439         err = btrfs_insert_empty_item(trans, root, path, &key,
8440                                       datasize);
8441         if (err) {
8442                 drop_inode = 1;
8443                 btrfs_free_path(path);
8444                 goto out_unlock;
8445         }
8446         leaf = path->nodes[0];
8447         ei = btrfs_item_ptr(leaf, path->slots[0],
8448                             struct btrfs_file_extent_item);
8449         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
8450         btrfs_set_file_extent_type(leaf, ei,
8451                                    BTRFS_FILE_EXTENT_INLINE);
8452         btrfs_set_file_extent_encryption(leaf, ei, 0);
8453         btrfs_set_file_extent_compression(leaf, ei, 0);
8454         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
8455         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
8456
8457         ptr = btrfs_file_extent_inline_start(ei);
8458         write_extent_buffer(leaf, symname, ptr, name_len);
8459         btrfs_mark_buffer_dirty(leaf);
8460         btrfs_free_path(path);
8461
8462         inode->i_op = &btrfs_symlink_inode_operations;
8463         inode->i_mapping->a_ops = &btrfs_symlink_aops;
8464         inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
8465         inode_set_bytes(inode, name_len);
8466         btrfs_i_size_write(inode, name_len - 1);
8467         err = btrfs_update_inode(trans, root, inode);
8468         if (err)
8469                 drop_inode = 1;
8470
8471 out_unlock:
8472         if (!err)
8473                 d_instantiate(dentry, inode);
8474         btrfs_end_transaction(trans, root);
8475         if (drop_inode) {
8476                 inode_dec_link_count(inode);
8477                 iput(inode);
8478         }
8479         btrfs_btree_balance_dirty(root);
8480         return err;
8481 }
8482
8483 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
8484                                        u64 start, u64 num_bytes, u64 min_size,
8485                                        loff_t actual_len, u64 *alloc_hint,
8486                                        struct btrfs_trans_handle *trans)
8487 {
8488         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
8489         struct extent_map *em;
8490         struct btrfs_root *root = BTRFS_I(inode)->root;
8491         struct btrfs_key ins;
8492         u64 cur_offset = start;
8493         u64 i_size;
8494         u64 cur_bytes;
8495         int ret = 0;
8496         bool own_trans = true;
8497
8498         if (trans)
8499                 own_trans = false;
8500         while (num_bytes > 0) {
8501                 if (own_trans) {
8502                         trans = btrfs_start_transaction(root, 3);
8503                         if (IS_ERR(trans)) {
8504                                 ret = PTR_ERR(trans);
8505                                 break;
8506                         }
8507                 }
8508
8509                 cur_bytes = min(num_bytes, 256ULL * 1024 * 1024);
8510                 cur_bytes = max(cur_bytes, min_size);
8511                 ret = btrfs_reserve_extent(trans, root, cur_bytes,
8512                                            min_size, 0, *alloc_hint, &ins, 1);
8513                 if (ret) {
8514                         if (own_trans)
8515                                 btrfs_end_transaction(trans, root);
8516                         break;
8517                 }
8518
8519                 ret = insert_reserved_file_extent(trans, inode,
8520                                                   cur_offset, ins.objectid,
8521                                                   ins.offset, ins.offset,
8522                                                   ins.offset, 0, 0, 0,
8523                                                   BTRFS_FILE_EXTENT_PREALLOC);
8524                 if (ret) {
8525                         btrfs_abort_transaction(trans, root, ret);
8526                         if (own_trans)
8527                                 btrfs_end_transaction(trans, root);
8528                         break;
8529                 }
8530                 btrfs_drop_extent_cache(inode, cur_offset,
8531                                         cur_offset + ins.offset -1, 0);
8532
8533                 em = alloc_extent_map();
8534                 if (!em) {
8535                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
8536                                 &BTRFS_I(inode)->runtime_flags);
8537                         goto next;
8538                 }
8539
8540                 em->start = cur_offset;
8541                 em->orig_start = cur_offset;
8542                 em->len = ins.offset;
8543                 em->block_start = ins.objectid;
8544                 em->block_len = ins.offset;
8545                 em->orig_block_len = ins.offset;
8546                 em->ram_bytes = ins.offset;
8547                 em->bdev = root->fs_info->fs_devices->latest_bdev;
8548                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
8549                 em->generation = trans->transid;
8550
8551                 while (1) {
8552                         write_lock(&em_tree->lock);
8553                         ret = add_extent_mapping(em_tree, em, 1);
8554                         write_unlock(&em_tree->lock);
8555                         if (ret != -EEXIST)
8556                                 break;
8557                         btrfs_drop_extent_cache(inode, cur_offset,
8558                                                 cur_offset + ins.offset - 1,
8559                                                 0);
8560                 }
8561                 free_extent_map(em);
8562 next:
8563                 num_bytes -= ins.offset;
8564                 cur_offset += ins.offset;
8565                 *alloc_hint = ins.objectid + ins.offset;
8566
8567                 inode_inc_iversion(inode);
8568                 inode->i_ctime = CURRENT_TIME;
8569                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
8570                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
8571                     (actual_len > inode->i_size) &&
8572                     (cur_offset > inode->i_size)) {
8573                         if (cur_offset > actual_len)
8574                                 i_size = actual_len;
8575                         else
8576                                 i_size = cur_offset;
8577                         i_size_write(inode, i_size);
8578                         btrfs_ordered_update_i_size(inode, i_size, NULL);
8579                 }
8580
8581                 ret = btrfs_update_inode(trans, root, inode);
8582
8583                 if (ret) {
8584                         btrfs_abort_transaction(trans, root, ret);
8585                         if (own_trans)
8586                                 btrfs_end_transaction(trans, root);
8587                         break;
8588                 }
8589
8590                 if (own_trans)
8591                         btrfs_end_transaction(trans, root);
8592         }
8593         return ret;
8594 }
8595
8596 int btrfs_prealloc_file_range(struct inode *inode, int mode,
8597                               u64 start, u64 num_bytes, u64 min_size,
8598                               loff_t actual_len, u64 *alloc_hint)
8599 {
8600         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
8601                                            min_size, actual_len, alloc_hint,
8602                                            NULL);
8603 }
8604
8605 int btrfs_prealloc_file_range_trans(struct inode *inode,
8606                                     struct btrfs_trans_handle *trans, int mode,
8607                                     u64 start, u64 num_bytes, u64 min_size,
8608                                     loff_t actual_len, u64 *alloc_hint)
8609 {
8610         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
8611                                            min_size, actual_len, alloc_hint, trans);
8612 }
8613
8614 static int btrfs_set_page_dirty(struct page *page)
8615 {
8616         return __set_page_dirty_nobuffers(page);
8617 }
8618
8619 static int btrfs_permission(struct inode *inode, int mask)
8620 {
8621         struct btrfs_root *root = BTRFS_I(inode)->root;
8622         umode_t mode = inode->i_mode;
8623
8624         if (mask & MAY_WRITE &&
8625             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
8626                 if (btrfs_root_readonly(root))
8627                         return -EROFS;
8628                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
8629                         return -EACCES;
8630         }
8631         return generic_permission(inode, mask);
8632 }
8633
8634 static const struct inode_operations btrfs_dir_inode_operations = {
8635         .getattr        = btrfs_getattr,
8636         .lookup         = btrfs_lookup,
8637         .create         = btrfs_create,
8638         .unlink         = btrfs_unlink,
8639         .link           = btrfs_link,
8640         .mkdir          = btrfs_mkdir,
8641         .rmdir          = btrfs_rmdir,
8642         .rename         = btrfs_rename,
8643         .symlink        = btrfs_symlink,
8644         .setattr        = btrfs_setattr,
8645         .mknod          = btrfs_mknod,
8646         .setxattr       = btrfs_setxattr,
8647         .getxattr       = btrfs_getxattr,
8648         .listxattr      = btrfs_listxattr,
8649         .removexattr    = btrfs_removexattr,
8650         .permission     = btrfs_permission,
8651         .get_acl        = btrfs_get_acl,
8652 };
8653 static const struct inode_operations btrfs_dir_ro_inode_operations = {
8654         .lookup         = btrfs_lookup,
8655         .permission     = btrfs_permission,
8656         .get_acl        = btrfs_get_acl,
8657 };
8658
8659 static const struct file_operations btrfs_dir_file_operations = {
8660         .llseek         = generic_file_llseek,
8661         .read           = generic_read_dir,
8662         .iterate        = btrfs_real_readdir,
8663         .unlocked_ioctl = btrfs_ioctl,
8664 #ifdef CONFIG_COMPAT
8665         .compat_ioctl   = btrfs_ioctl,
8666 #endif
8667         .release        = btrfs_release_file,
8668         .fsync          = btrfs_sync_file,
8669 };
8670
8671 static struct extent_io_ops btrfs_extent_io_ops = {
8672         .fill_delalloc = run_delalloc_range,
8673         .submit_bio_hook = btrfs_submit_bio_hook,
8674         .merge_bio_hook = btrfs_merge_bio_hook,
8675         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
8676         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
8677         .writepage_start_hook = btrfs_writepage_start_hook,
8678         .set_bit_hook = btrfs_set_bit_hook,
8679         .clear_bit_hook = btrfs_clear_bit_hook,
8680         .merge_extent_hook = btrfs_merge_extent_hook,
8681         .split_extent_hook = btrfs_split_extent_hook,
8682 };
8683
8684 /*
8685  * btrfs doesn't support the bmap operation because swapfiles
8686  * use bmap to make a mapping of extents in the file.  They assume
8687  * these extents won't change over the life of the file and they
8688  * use the bmap result to do IO directly to the drive.
8689  *
8690  * the btrfs bmap call would return logical addresses that aren't
8691  * suitable for IO and they also will change frequently as COW
8692  * operations happen.  So, swapfile + btrfs == corruption.
8693  *
8694  * For now we're avoiding this by dropping bmap.
8695  */
8696 static const struct address_space_operations btrfs_aops = {
8697         .readpage       = btrfs_readpage,
8698         .writepage      = btrfs_writepage,
8699         .writepages     = btrfs_writepages,
8700         .readpages      = btrfs_readpages,
8701         .direct_IO      = btrfs_direct_IO,
8702         .invalidatepage = btrfs_invalidatepage,
8703         .releasepage    = btrfs_releasepage,
8704         .set_page_dirty = btrfs_set_page_dirty,
8705         .error_remove_page = generic_error_remove_page,
8706 };
8707
8708 static const struct address_space_operations btrfs_symlink_aops = {
8709         .readpage       = btrfs_readpage,
8710         .writepage      = btrfs_writepage,
8711         .invalidatepage = btrfs_invalidatepage,
8712         .releasepage    = btrfs_releasepage,
8713 };
8714
8715 static const struct inode_operations btrfs_file_inode_operations = {
8716         .getattr        = btrfs_getattr,
8717         .setattr        = btrfs_setattr,
8718         .setxattr       = btrfs_setxattr,
8719         .getxattr       = btrfs_getxattr,
8720         .listxattr      = btrfs_listxattr,
8721         .removexattr    = btrfs_removexattr,
8722         .permission     = btrfs_permission,
8723         .fiemap         = btrfs_fiemap,
8724         .get_acl        = btrfs_get_acl,
8725         .update_time    = btrfs_update_time,
8726 };
8727 static const struct inode_operations btrfs_special_inode_operations = {
8728         .getattr        = btrfs_getattr,
8729         .setattr        = btrfs_setattr,
8730         .permission     = btrfs_permission,
8731         .setxattr       = btrfs_setxattr,
8732         .getxattr       = btrfs_getxattr,
8733         .listxattr      = btrfs_listxattr,
8734         .removexattr    = btrfs_removexattr,
8735         .get_acl        = btrfs_get_acl,
8736         .update_time    = btrfs_update_time,
8737 };
8738 static const struct inode_operations btrfs_symlink_inode_operations = {
8739         .readlink       = generic_readlink,
8740         .follow_link    = page_follow_link_light,
8741         .put_link       = page_put_link,
8742         .getattr        = btrfs_getattr,
8743         .setattr        = btrfs_setattr,
8744         .permission     = btrfs_permission,
8745         .setxattr       = btrfs_setxattr,
8746         .getxattr       = btrfs_getxattr,
8747         .listxattr      = btrfs_listxattr,
8748         .removexattr    = btrfs_removexattr,
8749         .get_acl        = btrfs_get_acl,
8750         .update_time    = btrfs_update_time,
8751 };
8752
8753 const struct dentry_operations btrfs_dentry_operations = {
8754         .d_delete       = btrfs_dentry_delete,
8755         .d_release      = btrfs_dentry_release,
8756 };