fs/btrfs/file.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/fs.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/highmem.h>
  22 #include <linux/time.h>
  23 #include <linux/init.h>
  24 #include <linux/string.h>
  25 #include <linux/backing-dev.h>
  26 #include <linux/mpage.h>
  27 #include <linux/falloc.h>
  28 #include <linux/swap.h>
  29 #include <linux/writeback.h>
  30 #include <linux/statfs.h>
  31 #include <linux/compat.h>
  32 #include <linux/slab.h>
  33 #include "ctree.h"
  34 #include "disk-io.h"
  35 #include "transaction.h"
  36 #include "btrfs_inode.h"
  37 #include "ioctl.h"
  38 #include "print-tree.h"
  39 #include "tree-log.h"
  40 #include "locking.h"
  41 #include "compat.h"
  42
  43
  44 /* simple helper to fault in pages and copy.  This should go away
  45  * and be replaced with calls into generic code.
  46  */
  47 static noinline int btrfs_copy_from_user(loff_t pos, int num_pages,
  48                                          int write_bytes,
  49                                          struct page **prepared_pages,
  50                                          struct iov_iter *i)
  51 {
  52         size_t copied = 0;
  53         int pg = 0;
  54         int offset = pos & (PAGE_CACHE_SIZE - 1);
  55         int total_copied = 0;
  56
  57         while (write_bytes > 0) {
  58                 size_t count = min_t(size_t,
  59                                      PAGE_CACHE_SIZE - offset, write_bytes);
  60                 struct page *page = prepared_pages[pg];
  61                 /*
  62                  * Copy data from userspace to the current page
  63                  *
  64                  * Disable pagefault to avoid recursive lock since
  65                  * the pages are already locked
  66                  */
  67                 pagefault_disable();
  68                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
  69                 pagefault_enable();
  70
  71                 /* Flush processor's dcache for this page */
  72                 flush_dcache_page(page);
  73
  74                 /*
  75                  * if we get a partial write, we can end up with
  76                  * partially up to date pages.  These add
  77                  * a lot of complexity, so make sure they don't
  78                  * happen by forcing this copy to be retried.
  79                  *
  80                  * The rest of the btrfs_file_write code will fall
  81                  * back to page at a time copies after we return 0.
  82                  */
  83                 if (!PageUptodate(page) && copied < count)
  84                         copied = 0;
  85
  86                 iov_iter_advance(i, copied);
  87                 write_bytes -= copied;
  88                 total_copied += copied;
  89
  90                 /* Return to btrfs_file_aio_write to fault page */
  91                 if (unlikely(copied == 0)) {
  92                         break;
  93                 }
  94
  95                 if (unlikely(copied < PAGE_CACHE_SIZE - offset)) {
  96                         offset += copied;
  97                 } else {
  98                         pg++;
  99                         offset = 0;
 100                 }
 101         }
 102         return total_copied;
 103 }
 104
 105 /*
 106  * unlocks pages after btrfs_file_write is done with them
 107  */
 108 static noinline void btrfs_drop_pages(struct page **pages, size_t num_pages)
 109 {
 110         size_t i;
 111         for (i = 0; i < num_pages; i++) {
 112                 if (!pages[i])
 113                         break;
 114                 /* page checked is some magic around finding pages that
 115                  * have been modified without going through btrfs_set_page_dirty
 116                  * clear it here
 117                  */
 118                 ClearPageChecked(pages[i]);
 119                 unlock_page(pages[i]);
 120                 mark_page_accessed(pages[i]);
 121                 page_cache_release(pages[i]);
 122         }
 123 }
 124
 125 /*
 126  * after copy_from_user, pages need to be dirtied and we need to make
 127  * sure holes are created between the current EOF and the start of
 128  * any next extents (if required).
 129  *
 130  * this also makes the decision about creating an inline extent vs
 131  * doing real data extents, marking pages dirty and delalloc as required.
 132  */
 133 static noinline int dirty_and_release_pages(struct btrfs_trans_handle *trans,
 134                                    struct btrfs_root *root,
 135                                    struct file *file,
 136                                    struct page **pages,
 137                                    size_t num_pages,
 138                                    loff_t pos,
 139                                    size_t write_bytes)
 140 {
 141         int err = 0;
 142         int i;
 143         struct inode *inode = fdentry(file)->d_inode;
 144         u64 num_bytes;
 145         u64 start_pos;
 146         u64 end_of_last_block;
 147         u64 end_pos = pos + write_bytes;
 148         loff_t isize = i_size_read(inode);
 149
 150         start_pos = pos & ~((u64)root->sectorsize - 1);
 151         num_bytes = (write_bytes + pos - start_pos +
 152                     root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
 153
 154         end_of_last_block = start_pos + num_bytes - 1;
 155         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 156                                         NULL);
 157         BUG_ON(err);
 158
 159         for (i = 0; i < num_pages; i++) {
 160                 struct page *p = pages[i];
 161                 SetPageUptodate(p);
 162                 ClearPageChecked(p);
 163                 set_page_dirty(p);
 164         }
 165         if (end_pos > isize) {
 166                 i_size_write(inode, end_pos);
 167                 /* we've only changed i_size in ram, and we haven't updated
 168                  * the disk i_size.  There is no need to log the inode
 169                  * at this time.
 170                  */
 171         }
 172         return 0;
 173 }
 174
 175 /*
 176  * this drops all the extents in the cache that intersect the range
 177  * [start, end].  Existing extents are split as required.
 178  */
 179 int btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
 180                             int skip_pinned)
 181 {
 182         struct extent_map *em;
 183         struct extent_map *split = NULL;
 184         struct extent_map *split2 = NULL;
 185         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 186         u64 len = end - start + 1;
 187         int ret;
 188         int testend = 1;
 189         unsigned long flags;
 190         int compressed = 0;
 191
 192         WARN_ON(end < start);
 193         if (end == (u64)-1) {
 194                 len = (u64)-1;
 195                 testend = 0;
 196         }
 197         while (1) {
 198                 if (!split)
 199                         split = alloc_extent_map(GFP_NOFS);
 200                 if (!split2)
 201                         split2 = alloc_extent_map(GFP_NOFS);
 202                 BUG_ON(!split || !split2);
 203
 204                 write_lock(&em_tree->lock);
 205                 em = lookup_extent_mapping(em_tree, start, len);
 206                 if (!em) {
 207                         write_unlock(&em_tree->lock);
 208                         break;
 209                 }
 210                 flags = em->flags;
 211                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 212                         if (testend && em->start + em->len >= start + len) {
 213                                 free_extent_map(em);
 214                                 write_unlock(&em_tree->lock);
 215                                 break;
 216                         }
 217                         start = em->start + em->len;
 218                         if (testend)
 219                                 len = start + len - (em->start + em->len);
 220                         free_extent_map(em);
 221                         write_unlock(&em_tree->lock);
 222                         continue;
 223                 }
 224                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 225                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 226                 remove_extent_mapping(em_tree, em);
 227
 228                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
 229                     em->start < start) {
 230                         split->start = em->start;
 231                         split->len = start - em->start;
 232                         split->orig_start = em->orig_start;
 233                         split->block_start = em->block_start;
 234
 235                         if (compressed)
 236                                 split->block_len = em->block_len;
 237                         else
 238                                 split->block_len = split->len;
 239
 240                         split->bdev = em->bdev;
 241                         split->flags = flags;
 242                         split->compress_type = em->compress_type;
 243                         ret = add_extent_mapping(em_tree, split);
 244                         BUG_ON(ret);
 245                         free_extent_map(split);
 246                         split = split2;
 247                         split2 = NULL;
 248                 }
 249                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
 250                     testend && em->start + em->len > start + len) {
 251                         u64 diff = start + len - em->start;
 252
 253                         split->start = start + len;
 254                         split->len = em->start + em->len - (start + len);
 255                         split->bdev = em->bdev;
 256                         split->flags = flags;
 257                         split->compress_type = em->compress_type;
 258
 259                         if (compressed) {
 260                                 split->block_len = em->block_len;
 261                                 split->block_start = em->block_start;
 262                                 split->orig_start = em->orig_start;
 263                         } else {
 264                                 split->block_len = split->len;
 265                                 split->block_start = em->block_start + diff;
 266                                 split->orig_start = split->start;
 267                         }
 268
 269                         ret = add_extent_mapping(em_tree, split);
 270                         BUG_ON(ret);
 271                         free_extent_map(split);
 272                         split = NULL;
 273                 }
 274                 write_unlock(&em_tree->lock);
 275
 276                 /* once for us */
 277                 free_extent_map(em);
 278                 /* once for the tree*/
 279                 free_extent_map(em);
 280         }
 281         if (split)
 282                 free_extent_map(split);
 283         if (split2)
 284                 free_extent_map(split2);
 285         return 0;
 286 }
 287
 288 /*
 289  * this is very complex, but the basic idea is to drop all extents
 290  * in the range start - end.  hint_block is filled in with a block number
 291  * that would be a good hint to the block allocator for this file.
 292  *
 293  * If an extent intersects the range but is not entirely inside the range
 294  * it is either truncated or split.  Anything entirely inside the range
 295  * is deleted from the tree.
 296  */
 297 int btrfs_drop_extents(struct btrfs_trans_handle *trans, struct inode *inode,
 298                        u64 start, u64 end, u64 *hint_byte, int drop_cache)
 299 {
 300         struct btrfs_root *root = BTRFS_I(inode)->root;
 301         struct extent_buffer *leaf;
 302         struct btrfs_file_extent_item *fi;
 303         struct btrfs_path *path;
 304         struct btrfs_key key;
 305         struct btrfs_key new_key;
 306         u64 search_start = start;
 307         u64 disk_bytenr = 0;
 308         u64 num_bytes = 0;
 309         u64 extent_offset = 0;
 310         u64 extent_end = 0;
 311         int del_nr = 0;
 312         int del_slot = 0;
 313         int extent_type;
 314         int recow;
 315         int ret;
 316
 317         if (drop_cache)
 318                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 319
 320         path = btrfs_alloc_path();
 321         if (!path)
 322                 return -ENOMEM;
 323
 324         while (1) {
 325                 recow = 0;
 326                 ret = btrfs_lookup_file_extent(trans, root, path, inode->i_ino,
 327                                                search_start, -1);
 328                 if (ret < 0)
 329                         break;
 330                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 331                         leaf = path->nodes[0];
 332                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 333                         if (key.objectid == inode->i_ino &&
 334                             key.type == BTRFS_EXTENT_DATA_KEY)
 335                                 path->slots[0]--;
 336                 }
 337                 ret = 0;
 338 next_slot:
 339                 leaf = path->nodes[0];
 340                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 341                         BUG_ON(del_nr > 0);
 342                         ret = btrfs_next_leaf(root, path);
 343                         if (ret < 0)
 344                                 break;
 345                         if (ret > 0) {
 346                                 ret = 0;
 347                                 break;
 348                         }
 349                         leaf = path->nodes[0];
 350                         recow = 1;
 351                 }
 352
 353                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 354                 if (key.objectid > inode->i_ino ||
 355                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 356                         break;
 357
 358                 fi = btrfs_item_ptr(leaf, path->slots[0],
 359                                     struct btrfs_file_extent_item);
 360                 extent_type = btrfs_file_extent_type(leaf, fi);
 361
 362                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 363                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 364                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 365                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 366                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 367                         extent_end = key.offset +
 368                                 btrfs_file_extent_num_bytes(leaf, fi);
 369                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 370                         extent_end = key.offset +
 371                                 btrfs_file_extent_inline_len(leaf, fi);
 372                 } else {
 373                         WARN_ON(1);
 374                         extent_end = search_start;
 375                 }
 376
 377                 if (extent_end <= search_start) {
 378                         path->slots[0]++;
 379                         goto next_slot;
 380                 }
 381
 382                 search_start = max(key.offset, start);
 383                 if (recow) {
 384                         btrfs_release_path(root, path);
 385                         continue;
 386                 }
 387
 388                 /*
 389                  *     | - range to drop - |
 390                  *  | -------- extent -------- |
 391                  */
 392                 if (start > key.offset && end < extent_end) {
 393                         BUG_ON(del_nr > 0);
 394                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 395
 396                         memcpy(&new_key, &key, sizeof(new_key));
 397                         new_key.offset = start;
 398                         ret = btrfs_duplicate_item(trans, root, path,
 399                                                    &new_key);
 400                         if (ret == -EAGAIN) {
 401                                 btrfs_release_path(root, path);
 402                                 continue;
 403                         }
 404                         if (ret < 0)
 405                                 break;
 406
 407                         leaf = path->nodes[0];
 408                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 409                                             struct btrfs_file_extent_item);
 410                         btrfs_set_file_extent_num_bytes(leaf, fi,
 411                                                         start - key.offset);
 412
 413                         fi = btrfs_item_ptr(leaf, path->slots[0],
 414                                             struct btrfs_file_extent_item);
 415
 416                         extent_offset += start - key.offset;
 417                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 418                         btrfs_set_file_extent_num_bytes(leaf, fi,
 419                                                         extent_end - start);
 420                         btrfs_mark_buffer_dirty(leaf);
 421
 422                         if (disk_bytenr > 0) {
 423                                 ret = btrfs_inc_extent_ref(trans, root,
 424                                                 disk_bytenr, num_bytes, 0,
 425                                                 root->root_key.objectid,
 426                                                 new_key.objectid,
 427                                                 start - extent_offset);
 428                                 BUG_ON(ret);
 429                                 *hint_byte = disk_bytenr;
 430                         }
 431                         key.offset = start;
 432                 }
 433                 /*
 434                  *  | ---- range to drop ----- |
 435                  *      | -------- extent -------- |
 436                  */
 437                 if (start <= key.offset && end < extent_end) {
 438                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 439
 440                         memcpy(&new_key, &key, sizeof(new_key));
 441                         new_key.offset = end;
 442                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 443
 444                         extent_offset += end - key.offset;
 445                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 446                         btrfs_set_file_extent_num_bytes(leaf, fi,
 447                                                         extent_end - end);
 448                         btrfs_mark_buffer_dirty(leaf);
 449                         if (disk_bytenr > 0) {
 450                                 inode_sub_bytes(inode, end - key.offset);
 451                                 *hint_byte = disk_bytenr;
 452                         }
 453                         break;
 454                 }
 455
 456                 search_start = extent_end;
 457                 /*
 458                  *       | ---- range to drop ----- |
 459                  *  | -------- extent -------- |
 460                  */
 461                 if (start > key.offset && end >= extent_end) {
 462                         BUG_ON(del_nr > 0);
 463                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 464
 465                         btrfs_set_file_extent_num_bytes(leaf, fi,
 466                                                         start - key.offset);
 467                         btrfs_mark_buffer_dirty(leaf);
 468                         if (disk_bytenr > 0) {
 469                                 inode_sub_bytes(inode, extent_end - start);
 470                                 *hint_byte = disk_bytenr;
 471                         }
 472                         if (end == extent_end)
 473                                 break;
 474
 475                         path->slots[0]++;
 476                         goto next_slot;
 477                 }
 478
 479                 /*
 480                  *  | ---- range to drop ----- |
 481                  *    | ------ extent ------ |
 482                  */
 483                 if (start <= key.offset && end >= extent_end) {
 484                         if (del_nr == 0) {
 485                                 del_slot = path->slots[0];
 486                                 del_nr = 1;
 487                         } else {
 488                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 489                                 del_nr++;
 490                         }
 491
 492                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 493                                 inode_sub_bytes(inode,
 494                                                 extent_end - key.offset);
 495                                 extent_end = ALIGN(extent_end,
 496                                                    root->sectorsize);
 497                         } else if (disk_bytenr > 0) {
 498                                 ret = btrfs_free_extent(trans, root,
 499                                                 disk_bytenr, num_bytes, 0,
 500                                                 root->root_key.objectid,
 501                                                 key.objectid, key.offset -
 502                                                 extent_offset);
 503                                 BUG_ON(ret);
 504                                 inode_sub_bytes(inode,
 505                                                 extent_end - key.offset);
 506                                 *hint_byte = disk_bytenr;
 507                         }
 508
 509                         if (end == extent_end)
 510                                 break;
 511
 512                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 513                                 path->slots[0]++;
 514                                 goto next_slot;
 515                         }
 516
 517                         ret = btrfs_del_items(trans, root, path, del_slot,
 518                                               del_nr);
 519                         BUG_ON(ret);
 520
 521                         del_nr = 0;
 522                         del_slot = 0;
 523
 524                         btrfs_release_path(root, path);
 525                         continue;
 526                 }
 527
 528                 BUG_ON(1);
 529         }
 530
 531         if (del_nr > 0) {
 532                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 533                 BUG_ON(ret);
 534         }
 535
 536         btrfs_free_path(path);
 537         return ret;
 538 }
 539
 540 static int extent_mergeable(struct extent_buffer *leaf, int slot,
 541                             u64 objectid, u64 bytenr, u64 orig_offset,
 542                             u64 *start, u64 *end)
 543 {
 544         struct btrfs_file_extent_item *fi;
 545         struct btrfs_key key;
 546         u64 extent_end;
 547
 548         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
 549                 return 0;
 550
 551         btrfs_item_key_to_cpu(leaf, &key, slot);
 552         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
 553                 return 0;
 554
 555         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
 556         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
 557             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
 558             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
 559             btrfs_file_extent_compression(leaf, fi) ||
 560             btrfs_file_extent_encryption(leaf, fi) ||
 561             btrfs_file_extent_other_encoding(leaf, fi))
 562                 return 0;
 563
 564         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 565         if ((*start && *start != key.offset) || (*end && *end != extent_end))
 566                 return 0;
 567
 568         *start = key.offset;
 569         *end = extent_end;
 570         return 1;
 571 }
 572
 573 /*
 574  * Mark extent in the range start - end as written.
 575  *
 576  * This changes extent type from 'pre-allocated' to 'regular'. If only
 577  * part of extent is marked as written, the extent will be split into
 578  * two or three.
 579  */
 580 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
 581                               struct inode *inode, u64 start, u64 end)
 582 {
 583         struct btrfs_root *root = BTRFS_I(inode)->root;
 584         struct extent_buffer *leaf;
 585         struct btrfs_path *path;
 586         struct btrfs_file_extent_item *fi;
 587         struct btrfs_key key;
 588         struct btrfs_key new_key;
 589         u64 bytenr;
 590         u64 num_bytes;
 591         u64 extent_end;
 592         u64 orig_offset;
 593         u64 other_start;
 594         u64 other_end;
 595         u64 split;
 596         int del_nr = 0;
 597         int del_slot = 0;
 598         int recow;
 599         int ret;
 600
 601         btrfs_drop_extent_cache(inode, start, end - 1, 0);
 602
 603         path = btrfs_alloc_path();
 604         BUG_ON(!path);
 605 again:
 606         recow = 0;
 607         split = start;
 608         key.objectid = inode->i_ino;
 609         key.type = BTRFS_EXTENT_DATA_KEY;
 610         key.offset = split;
 611
 612         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 613         if (ret > 0 && path->slots[0] > 0)
 614                 path->slots[0]--;
 615
 616         leaf = path->nodes[0];
 617         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 618         BUG_ON(key.objectid != inode->i_ino ||
 619                key.type != BTRFS_EXTENT_DATA_KEY);
 620         fi = btrfs_item_ptr(leaf, path->slots[0],
 621                             struct btrfs_file_extent_item);
 622         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
 623                BTRFS_FILE_EXTENT_PREALLOC);
 624         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 625         BUG_ON(key.offset > start || extent_end < end);
 626
 627         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 628         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 629         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
 630         memcpy(&new_key, &key, sizeof(new_key));
 631
 632         if (start == key.offset && end < extent_end) {
 633                 other_start = 0;
 634                 other_end = start;
 635                 if (extent_mergeable(leaf, path->slots[0] - 1,
 636                                      inode->i_ino, bytenr, orig_offset,
 637                                      &other_start, &other_end)) {
 638                         new_key.offset = end;
 639                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 640                         fi = btrfs_item_ptr(leaf, path->slots[0],
 641                                             struct btrfs_file_extent_item);
 642                         btrfs_set_file_extent_num_bytes(leaf, fi,
 643                                                         extent_end - end);
 644                         btrfs_set_file_extent_offset(leaf, fi,
 645                                                      end - orig_offset);
 646                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 647                                             struct btrfs_file_extent_item);
 648                         btrfs_set_file_extent_num_bytes(leaf, fi,
 649                                                         end - other_start);
 650                         btrfs_mark_buffer_dirty(leaf);
 651                         goto out;
 652                 }
 653         }
 654
 655         if (start > key.offset && end == extent_end) {
 656                 other_start = end;
 657                 other_end = 0;
 658                 if (extent_mergeable(leaf, path->slots[0] + 1,
 659                                      inode->i_ino, bytenr, orig_offset,
 660                                      &other_start, &other_end)) {
 661                         fi = btrfs_item_ptr(leaf, path->slots[0],
 662                                             struct btrfs_file_extent_item);
 663                         btrfs_set_file_extent_num_bytes(leaf, fi,
 664                                                         start - key.offset);
 665                         path->slots[0]++;
 666                         new_key.offset = start;
 667                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 668
 669                         fi = btrfs_item_ptr(leaf, path->slots[0],
 670                                             struct btrfs_file_extent_item);
 671                         btrfs_set_file_extent_num_bytes(leaf, fi,
 672                                                         other_end - start);
 673                         btrfs_set_file_extent_offset(leaf, fi,
 674                                                      start - orig_offset);
 675                         btrfs_mark_buffer_dirty(leaf);
 676                         goto out;
 677                 }
 678         }
 679
 680         while (start > key.offset || end < extent_end) {
 681                 if (key.offset == start)
 682                         split = end;
 683
 684                 new_key.offset = split;
 685                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
 686                 if (ret == -EAGAIN) {
 687                         btrfs_release_path(root, path);
 688                         goto again;
 689                 }
 690                 BUG_ON(ret < 0);
 691
 692                 leaf = path->nodes[0];
 693                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 694                                     struct btrfs_file_extent_item);
 695                 btrfs_set_file_extent_num_bytes(leaf, fi,
 696                                                 split - key.offset);
 697
 698                 fi = btrfs_item_ptr(leaf, path->slots[0],
 699                                     struct btrfs_file_extent_item);
 700
 701                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
 702                 btrfs_set_file_extent_num_bytes(leaf, fi,
 703                                                 extent_end - split);
 704                 btrfs_mark_buffer_dirty(leaf);
 705
 706                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
 707                                            root->root_key.objectid,
 708                                            inode->i_ino, orig_offset);
 709                 BUG_ON(ret);
 710
 711                 if (split == start) {
 712                         key.offset = start;
 713                 } else {
 714                         BUG_ON(start != key.offset);
 715                         path->slots[0]--;
 716                         extent_end = end;
 717                 }
 718                 recow = 1;
 719         }
 720
 721         other_start = end;
 722         other_end = 0;
 723         if (extent_mergeable(leaf, path->slots[0] + 1,
 724                              inode->i_ino, bytenr, orig_offset,
 725                              &other_start, &other_end)) {
 726                 if (recow) {
 727                         btrfs_release_path(root, path);
 728                         goto again;
 729                 }
 730                 extent_end = other_end;
 731                 del_slot = path->slots[0] + 1;
 732                 del_nr++;
 733                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
 734                                         0, root->root_key.objectid,
 735                                         inode->i_ino, orig_offset);
 736                 BUG_ON(ret);
 737         }
 738         other_start = 0;
 739         other_end = start;
 740         if (extent_mergeable(leaf, path->slots[0] - 1,
 741                              inode->i_ino, bytenr, orig_offset,
 742                              &other_start, &other_end)) {
 743                 if (recow) {
 744                         btrfs_release_path(root, path);
 745                         goto again;
 746                 }
 747                 key.offset = other_start;
 748                 del_slot = path->slots[0];
 749                 del_nr++;
 750                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
 751                                         0, root->root_key.objectid,
 752                                         inode->i_ino, orig_offset);
 753                 BUG_ON(ret);
 754         }
 755         if (del_nr == 0) {
 756                 fi = btrfs_item_ptr(leaf, path->slots[0],
 757                            struct btrfs_file_extent_item);
 758                 btrfs_set_file_extent_type(leaf, fi,
 759                                            BTRFS_FILE_EXTENT_REG);
 760                 btrfs_mark_buffer_dirty(leaf);
 761         } else {
 762                 fi = btrfs_item_ptr(leaf, del_slot - 1,
 763                            struct btrfs_file_extent_item);
 764                 btrfs_set_file_extent_type(leaf, fi,
 765                                            BTRFS_FILE_EXTENT_REG);
 766                 btrfs_set_file_extent_num_bytes(leaf, fi,
 767                                                 extent_end - key.offset);
 768                 btrfs_mark_buffer_dirty(leaf);
 769
 770                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 771                 BUG_ON(ret);
 772         }
 773 out:
 774         btrfs_free_path(path);
 775         return 0;
 776 }
 777
 778 /*
 779  * on error we return an unlocked page and the error value
 780  * on success we return a locked page and 0
 781  */
 782 static int prepare_uptodate_page(struct page *page, u64 pos)
 783 {
 784         int ret = 0;
 785
 786         if ((pos & (PAGE_CACHE_SIZE - 1)) && !PageUptodate(page)) {
 787                 ret = btrfs_readpage(NULL, page);
 788                 if (ret)
 789                         return ret;
 790                 lock_page(page);
 791                 if (!PageUptodate(page)) {
 792                         unlock_page(page);
 793                         return -EIO;
 794                 }
 795         }
 796         return 0;
 797 }
 798
 799 /*
 800  * this gets pages into the page cache and locks them down, it also properly
 801  * waits for data=ordered extents to finish before allowing the pages to be
 802  * modified.
 803  */
 804 static noinline int prepare_pages(struct btrfs_root *root, struct file *file,
 805                          struct page **pages, size_t num_pages,
 806                          loff_t pos, unsigned long first_index,
 807                          unsigned long last_index, size_t write_bytes)
 808 {
 809         struct extent_state *cached_state = NULL;
 810         int i;
 811         unsigned long index = pos >> PAGE_CACHE_SHIFT;
 812         struct inode *inode = fdentry(file)->d_inode;
 813         int err = 0;
 814         int faili = 0;
 815         u64 start_pos;
 816         u64 last_pos;
 817
 818         start_pos = pos & ~((u64)root->sectorsize - 1);
 819         last_pos = ((u64)index + num_pages) << PAGE_CACHE_SHIFT;
 820
 821         if (start_pos > inode->i_size) {
 822                 err = btrfs_cont_expand(inode, start_pos);
 823                 if (err)
 824                         return err;
 825         }
 826
 827         memset(pages, 0, num_pages * sizeof(struct page *));
 828 again:
 829         for (i = 0; i < num_pages; i++) {
 830                 pages[i] = grab_cache_page(inode->i_mapping, index + i);
 831                 if (!pages[i]) {
 832                         faili = i - 1;
 833                         err = -ENOMEM;
 834                         goto fail;
 835                 }
 836
 837                 if (i == 0)
 838                         err = prepare_uptodate_page(pages[i], pos);
 839                 if (i == num_pages - 1)
 840                         err = prepare_uptodate_page(pages[i],
 841                                                     pos + write_bytes);
 842                 if (err) {
 843                         page_cache_release(pages[i]);
 844                         faili = i - 1;
 845                         goto fail;
 846                 }
 847                 wait_on_page_writeback(pages[i]);
 848         }
 849         err = 0;
 850         if (start_pos < inode->i_size) {
 851                 struct btrfs_ordered_extent *ordered;
 852                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
 853                                  start_pos, last_pos - 1, 0, &cached_state,
 854                                  GFP_NOFS);
 855                 ordered = btrfs_lookup_first_ordered_extent(inode,
 856                                                             last_pos - 1);
 857                 if (ordered &&
 858                     ordered->file_offset + ordered->len > start_pos &&
 859                     ordered->file_offset < last_pos) {
 860                         btrfs_put_ordered_extent(ordered);
 861                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
 862                                              start_pos, last_pos - 1,
 863                                              &cached_state, GFP_NOFS);
 864                         for (i = 0; i < num_pages; i++) {
 865                                 unlock_page(pages[i]);
 866                                 page_cache_release(pages[i]);
 867                         }
 868                         btrfs_wait_ordered_range(inode, start_pos,
 869                                                  last_pos - start_pos);
 870                         goto again;
 871                 }
 872                 if (ordered)
 873                         btrfs_put_ordered_extent(ordered);
 874
 875                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
 876                                   last_pos - 1, EXTENT_DIRTY | EXTENT_DELALLOC |
 877                                   EXTENT_DO_ACCOUNTING, 0, 0, &cached_state,
 878                                   GFP_NOFS);
 879                 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
 880                                      start_pos, last_pos - 1, &cached_state,
 881                                      GFP_NOFS);
 882         }
 883         for (i = 0; i < num_pages; i++) {
 884                 clear_page_dirty_for_io(pages[i]);
 885                 set_page_extent_mapped(pages[i]);
 886                 WARN_ON(!PageLocked(pages[i]));
 887         }
 888         return 0;
 889 fail:
 890         while (faili >= 0) {
 891                 unlock_page(pages[faili]);
 892                 page_cache_release(pages[faili]);
 893                 faili--;
 894         }
 895         return err;
 896
 897 }
 898
 899 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
 900                                     const struct iovec *iov,
 901                                     unsigned long nr_segs, loff_t pos)
 902 {
 903         struct file *file = iocb->ki_filp;
 904         struct inode *inode = fdentry(file)->d_inode;
 905         struct btrfs_root *root = BTRFS_I(inode)->root;
 906         struct page **pages = NULL;
 907         struct iov_iter i;
 908         loff_t *ppos = &iocb->ki_pos;
 909         loff_t start_pos;
 910         ssize_t num_written = 0;
 911         ssize_t err = 0;
 912         size_t count;
 913         size_t ocount;
 914         int ret = 0;
 915         int nrptrs;
 916         unsigned long first_index;
 917         unsigned long last_index;
 918         int will_write;
 919         int buffered = 0;
 920         int copied = 0;
 921         int dirty_pages = 0;
 922
 923         will_write = ((file->f_flags & O_DSYNC) || IS_SYNC(inode) ||
 924                       (file->f_flags & O_DIRECT));
 925
 926         start_pos = pos;
 927
 928         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
 929
 930         mutex_lock(&inode->i_mutex);
 931
 932         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
 933         if (err)
 934                 goto out;
 935         count = ocount;
 936
 937         current->backing_dev_info = inode->i_mapping->backing_dev_info;
 938         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
 939         if (err)
 940                 goto out;
 941
 942         if (count == 0)
 943                 goto out;
 944
 945         err = file_remove_suid(file);
 946         if (err)
 947                 goto out;
 948
 949         /*
 950          * If BTRFS flips readonly due to some impossible error
 951          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
 952          * although we have opened a file as writable, we have
 953          * to stop this write operation to ensure FS consistency.
 954          */
 955         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
 956                 err = -EROFS;
 957                 goto out;
 958         }
 959
 960         file_update_time(file);
 961         BTRFS_I(inode)->sequence++;
 962
 963         if (unlikely(file->f_flags & O_DIRECT)) {
 964                 num_written = generic_file_direct_write(iocb, iov, &nr_segs,
 965                                                         pos, ppos, count,
 966                                                         ocount);
 967                 /*
 968                  * the generic O_DIRECT will update in-memory i_size after the
 969                  * DIOs are done.  But our endio handlers that update the on
 970                  * disk i_size never update past the in memory i_size.  So we
 971                  * need one more update here to catch any additions to the
 972                  * file
 973                  */
 974                 if (inode->i_size != BTRFS_I(inode)->disk_i_size) {
 975                         btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
 976                         mark_inode_dirty(inode);
 977                 }
 978
 979                 if (num_written < 0) {
 980                         ret = num_written;
 981                         num_written = 0;
 982                         goto out;
 983                 } else if (num_written == count) {
 984                         /* pick up pos changes done by the generic code */
 985                         pos = *ppos;
 986                         goto out;
 987                 }
 988                 /*
 989                  * We are going to do buffered for the rest of the range, so we
 990                  * need to make sure to invalidate the buffered pages when we're
 991                  * done.
 992                  */
 993                 buffered = 1;
 994                 pos += num_written;
 995         }
 996
 997         iov_iter_init(&i, iov, nr_segs, count, num_written);
 998         nrptrs = min((iov_iter_count(&i) + PAGE_CACHE_SIZE - 1) /
 999                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1000                      (sizeof(struct page *)));
1001         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1002         if (!pages) {
1003                 ret = -ENOMEM;
1004                 goto out;
1005         }
1006
1007         /* generic_write_checks can change our pos */
1008         start_pos = pos;
1009
1010         first_index = pos >> PAGE_CACHE_SHIFT;
1011         last_index = (pos + iov_iter_count(&i)) >> PAGE_CACHE_SHIFT;
1012
1013         while (iov_iter_count(&i) > 0) {
1014                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1015                 size_t write_bytes = min(iov_iter_count(&i),
1016                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1017                                          offset);
1018                 size_t num_pages = (write_bytes + offset +
1019                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1020
1021                 WARN_ON(num_pages > nrptrs);
1022                 memset(pages, 0, sizeof(struct page *) * nrptrs);
1023
1024                 /*
1025                  * Fault pages before locking them in prepare_pages
1026                  * to avoid recursive lock
1027                  */
1028                 if (unlikely(iov_iter_fault_in_readable(&i, write_bytes))) {
1029                         ret = -EFAULT;
1030                         goto out;
1031                 }
1032
1033                 ret = btrfs_delalloc_reserve_space(inode,
1034                                         num_pages << PAGE_CACHE_SHIFT);
1035                 if (ret)
1036                         goto out;
1037
1038                 ret = prepare_pages(root, file, pages, num_pages,
1039                                     pos, first_index, last_index,
1040                                     write_bytes);
1041                 if (ret) {
1042                         btrfs_delalloc_release_space(inode,
1043                                         num_pages << PAGE_CACHE_SHIFT);
1044                         goto out;
1045                 }
1046
1047                 copied = btrfs_copy_from_user(pos, num_pages,
1048                                            write_bytes, pages, &i);
1049
1050                 /*
1051                  * if we have trouble faulting in the pages, fall
1052                  * back to one page at a time
1053                  */
1054                 if (copied < write_bytes)
1055                         nrptrs = 1;
1056
1057                 if (copied == 0)
1058                         dirty_pages = 0;
1059                 else
1060                         dirty_pages = (copied + offset +
1061                                        PAGE_CACHE_SIZE - 1) >>
1062                                        PAGE_CACHE_SHIFT;
1063
1064                 if (num_pages > dirty_pages) {
1065                         if (copied > 0)
1066                                 atomic_inc(
1067                                         &BTRFS_I(inode)->outstanding_extents);
1068                         btrfs_delalloc_release_space(inode,
1069                                         (num_pages - dirty_pages) <<
1070                                         PAGE_CACHE_SHIFT);
1071                 }
1072
1073                 if (copied > 0) {
1074                         dirty_and_release_pages(NULL, root, file, pages,
1075                                                 dirty_pages, pos, copied);
1076                 }
1077
1078                 btrfs_drop_pages(pages, num_pages);
1079
1080                 if (copied > 0) {
1081                         if (will_write) {
1082                                 filemap_fdatawrite_range(inode->i_mapping, pos,
1083                                                          pos + copied - 1);
1084                         } else {
1085                                 balance_dirty_pages_ratelimited_nr(
1086                                                         inode->i_mapping,
1087                                                         dirty_pages);
1088                                 if (dirty_pages <
1089                                 (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1090                                         btrfs_btree_balance_dirty(root, 1);
1091                                 btrfs_throttle(root);
1092                         }
1093                 }
1094
1095                 pos += copied;
1096                 num_written += copied;
1097
1098                 cond_resched();
1099         }
1100 out:
1101         mutex_unlock(&inode->i_mutex);
1102         if (ret)
1103                 err = ret;
1104
1105         kfree(pages);
1106         *ppos = pos;
1107
1108         /*
1109          * we want to make sure fsync finds this change
1110          * but we haven't joined a transaction running right now.
1111          *
1112          * Later on, someone is sure to update the inode and get the
1113          * real transid recorded.
1114          *
1115          * We set last_trans now to the fs_info generation + 1,
1116          * this will either be one more than the running transaction
1117          * or the generation used for the next transaction if there isn't
1118          * one running right now.
1119          */
1120         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1121
1122         if (num_written > 0 && will_write) {
1123                 struct btrfs_trans_handle *trans;
1124
1125                 err = btrfs_wait_ordered_range(inode, start_pos, num_written);
1126                 if (err)
1127                         num_written = err;
1128
1129                 if ((file->f_flags & O_DSYNC) || IS_SYNC(inode)) {
1130                         trans = btrfs_start_transaction(root, 0);
1131                         if (IS_ERR(trans)) {
1132                                 num_written = PTR_ERR(trans);
1133                                 goto done;
1134                         }
1135                         mutex_lock(&inode->i_mutex);
1136                         ret = btrfs_log_dentry_safe(trans, root,
1137                                                     file->f_dentry);
1138                         mutex_unlock(&inode->i_mutex);
1139                         if (ret == 0) {
1140                                 ret = btrfs_sync_log(trans, root);
1141                                 if (ret == 0)
1142                                         btrfs_end_transaction(trans, root);
1143                                 else
1144                                         btrfs_commit_transaction(trans, root);
1145                         } else if (ret != BTRFS_NO_LOG_SYNC) {
1146                                 btrfs_commit_transaction(trans, root);
1147                         } else {
1148                                 btrfs_end_transaction(trans, root);
1149                         }
1150                 }
1151                 if (file->f_flags & O_DIRECT && buffered) {
1152                         invalidate_mapping_pages(inode->i_mapping,
1153                               start_pos >> PAGE_CACHE_SHIFT,
1154                              (start_pos + num_written - 1) >> PAGE_CACHE_SHIFT);
1155                 }
1156         }
1157 done:
1158         current->backing_dev_info = NULL;
1159         return num_written ? num_written : err;
1160 }
1161
1162 int btrfs_release_file(struct inode *inode, struct file *filp)
1163 {
1164         /*
1165          * ordered_data_close is set by settattr when we are about to truncate
1166          * a file from a non-zero size to a zero size.  This tries to
1167          * flush down new bytes that may have been written if the
1168          * application were using truncate to replace a file in place.
1169          */
1170         if (BTRFS_I(inode)->ordered_data_close) {
1171                 BTRFS_I(inode)->ordered_data_close = 0;
1172                 btrfs_add_ordered_operation(NULL, BTRFS_I(inode)->root, inode);
1173                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1174                         filemap_flush(inode->i_mapping);
1175         }
1176         if (filp->private_data)
1177                 btrfs_ioctl_trans_end(filp);
1178         return 0;
1179 }
1180
1181 /*
1182  * fsync call for both files and directories.  This logs the inode into
1183  * the tree log instead of forcing full commits whenever possible.
1184  *
1185  * It needs to call filemap_fdatawait so that all ordered extent updates are
1186  * in the metadata btree are up to date for copying to the log.
1187  *
1188  * It drops the inode mutex before doing the tree log commit.  This is an
1189  * important optimization for directories because holding the mutex prevents
1190  * new operations on the dir while we write to disk.
1191  */
1192 int btrfs_sync_file(struct file *file, int datasync)
1193 {
1194         struct dentry *dentry = file->f_path.dentry;
1195         struct inode *inode = dentry->d_inode;
1196         struct btrfs_root *root = BTRFS_I(inode)->root;
1197         int ret = 0;
1198         struct btrfs_trans_handle *trans;
1199
1200
1201         /* we wait first, since the writeback may change the inode */
1202         root->log_batch++;
1203         /* the VFS called filemap_fdatawrite for us */
1204         btrfs_wait_ordered_range(inode, 0, (u64)-1);
1205         root->log_batch++;
1206
1207         /*
1208          * check the transaction that last modified this inode
1209          * and see if its already been committed
1210          */
1211         if (!BTRFS_I(inode)->last_trans)
1212                 goto out;
1213
1214         /*
1215          * if the last transaction that changed this file was before
1216          * the current transaction, we can bail out now without any
1217          * syncing
1218          */
1219         mutex_lock(&root->fs_info->trans_mutex);
1220         if (BTRFS_I(inode)->last_trans <=
1221             root->fs_info->last_trans_committed) {
1222                 BTRFS_I(inode)->last_trans = 0;
1223                 mutex_unlock(&root->fs_info->trans_mutex);
1224                 goto out;
1225         }
1226         mutex_unlock(&root->fs_info->trans_mutex);
1227
1228         /*
1229          * ok we haven't committed the transaction yet, lets do a commit
1230          */
1231         if (file->private_data)
1232                 btrfs_ioctl_trans_end(file);
1233
1234         trans = btrfs_start_transaction(root, 0);
1235         if (IS_ERR(trans)) {
1236                 ret = PTR_ERR(trans);
1237                 goto out;
1238         }
1239
1240         ret = btrfs_log_dentry_safe(trans, root, dentry);
1241         if (ret < 0)
1242                 goto out;
1243
1244         /* we've logged all the items and now have a consistent
1245          * version of the file in the log.  It is possible that
1246          * someone will come in and modify the file, but that's
1247          * fine because the log is consistent on disk, and we
1248          * have references to all of the file's extents
1249          *
1250          * It is possible that someone will come in and log the
1251          * file again, but that will end up using the synchronization
1252          * inside btrfs_sync_log to keep things safe.
1253          */
1254         mutex_unlock(&dentry->d_inode->i_mutex);
1255
1256         if (ret != BTRFS_NO_LOG_SYNC) {
1257                 if (ret > 0) {
1258                         ret = btrfs_commit_transaction(trans, root);
1259                 } else {
1260                         ret = btrfs_sync_log(trans, root);
1261                         if (ret == 0)
1262                                 ret = btrfs_end_transaction(trans, root);
1263                         else
1264                                 ret = btrfs_commit_transaction(trans, root);
1265                 }
1266         } else {
1267                 ret = btrfs_end_transaction(trans, root);
1268         }
1269         mutex_lock(&dentry->d_inode->i_mutex);
1270 out:
1271         return ret > 0 ? -EIO : ret;
1272 }
1273
1274 static const struct vm_operations_struct btrfs_file_vm_ops = {
1275         .fault          = filemap_fault,
1276         .page_mkwrite   = btrfs_page_mkwrite,
1277 };
1278
1279 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
1280 {
1281         struct address_space *mapping = filp->f_mapping;
1282
1283         if (!mapping->a_ops->readpage)
1284                 return -ENOEXEC;
1285
1286         file_accessed(filp);
1287         vma->vm_ops = &btrfs_file_vm_ops;
1288         vma->vm_flags |= VM_CAN_NONLINEAR;
1289
1290         return 0;
1291 }
1292
1293 static long btrfs_fallocate(struct file *file, int mode,
1294                             loff_t offset, loff_t len)
1295 {
1296         struct inode *inode = file->f_path.dentry->d_inode;
1297         struct extent_state *cached_state = NULL;
1298         u64 cur_offset;
1299         u64 last_byte;
1300         u64 alloc_start;
1301         u64 alloc_end;
1302         u64 alloc_hint = 0;
1303         u64 locked_end;
1304         u64 mask = BTRFS_I(inode)->root->sectorsize - 1;
1305         struct extent_map *em;
1306         int ret;
1307
1308         alloc_start = offset & ~mask;
1309         alloc_end =  (offset + len + mask) & ~mask;
1310
1311         /* We only support the FALLOC_FL_KEEP_SIZE mode */
1312         if (mode & ~FALLOC_FL_KEEP_SIZE)
1313                 return -EOPNOTSUPP;
1314
1315         /*
1316          * wait for ordered IO before we have any locks.  We'll loop again
1317          * below with the locks held.
1318          */
1319         btrfs_wait_ordered_range(inode, alloc_start, alloc_end - alloc_start);
1320
1321         mutex_lock(&inode->i_mutex);
1322         ret = inode_newsize_ok(inode, alloc_end);
1323         if (ret)
1324                 goto out;
1325
1326         if (alloc_start > inode->i_size) {
1327                 ret = btrfs_cont_expand(inode, alloc_start);
1328                 if (ret)
1329                         goto out;
1330         }
1331
1332         ret = btrfs_check_data_free_space(inode, alloc_end - alloc_start);
1333         if (ret)
1334                 goto out;
1335
1336         locked_end = alloc_end - 1;
1337         while (1) {
1338                 struct btrfs_ordered_extent *ordered;
1339
1340                 /* the extent lock is ordered inside the running
1341                  * transaction
1342                  */
1343                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
1344                                  locked_end, 0, &cached_state, GFP_NOFS);
1345                 ordered = btrfs_lookup_first_ordered_extent(inode,
1346                                                             alloc_end - 1);
1347                 if (ordered &&
1348                     ordered->file_offset + ordered->len > alloc_start &&
1349                     ordered->file_offset < alloc_end) {
1350                         btrfs_put_ordered_extent(ordered);
1351                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1352                                              alloc_start, locked_end,
1353                                              &cached_state, GFP_NOFS);
1354                         /*
1355                          * we can't wait on the range with the transaction
1356                          * running or with the extent lock held
1357                          */
1358                         btrfs_wait_ordered_range(inode, alloc_start,
1359                                                  alloc_end - alloc_start);
1360                 } else {
1361                         if (ordered)
1362                                 btrfs_put_ordered_extent(ordered);
1363                         break;
1364                 }
1365         }
1366
1367         cur_offset = alloc_start;
1368         while (1) {
1369                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
1370                                       alloc_end - cur_offset, 0);
1371                 BUG_ON(IS_ERR(em) || !em);
1372                 last_byte = min(extent_map_end(em), alloc_end);
1373                 last_byte = (last_byte + mask) & ~mask;
1374                 if (em->block_start == EXTENT_MAP_HOLE ||
1375                     (cur_offset >= inode->i_size &&
1376                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
1377                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
1378                                                         last_byte - cur_offset,
1379                                                         1 << inode->i_blkbits,
1380                                                         offset + len,
1381                                                         &alloc_hint);
1382                         if (ret < 0) {
1383                                 free_extent_map(em);
1384                                 break;
1385                         }
1386                 }
1387                 free_extent_map(em);
1388
1389                 cur_offset = last_byte;
1390                 if (cur_offset >= alloc_end) {
1391                         ret = 0;
1392                         break;
1393                 }
1394         }
1395         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
1396                              &cached_state, GFP_NOFS);
1397
1398         btrfs_free_reserved_data_space(inode, alloc_end - alloc_start);
1399 out:
1400         mutex_unlock(&inode->i_mutex);
1401         return ret;
1402 }
1403
1404 const struct file_operations btrfs_file_operations = {
1405         .llseek         = generic_file_llseek,
1406         .read           = do_sync_read,
1407         .write          = do_sync_write,
1408         .aio_read       = generic_file_aio_read,
1409         .splice_read    = generic_file_splice_read,
1410         .aio_write      = btrfs_file_aio_write,
1411         .mmap           = btrfs_file_mmap,
1412         .open           = generic_file_open,
1413         .release        = btrfs_release_file,
1414         .fsync          = btrfs_sync_file,
1415         .fallocate      = btrfs_fallocate,
1416         .unlocked_ioctl = btrfs_ioctl,
1417 #ifdef CONFIG_COMPAT
1418         .compat_ioctl   = btrfs_ioctl,
1419 #endif
1420 };