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  * when auto defrag is enabled we
  45  * queue up these defrag structs to remember which
  46  * inodes need defragging passes
  47  */
  48 struct inode_defrag {
  49         struct rb_node rb_node;
  50         /* objectid */
  51         u64 ino;
  52         /*
  53          * transid where the defrag was added, we search for
  54          * extents newer than this
  55          */
  56         u64 transid;
  57
  58         /* root objectid */
  59         u64 root;
  60
  61         /* last offset we were able to defrag */
  62         u64 last_offset;
  63
  64         /* if we've wrapped around back to zero once already */
  65         int cycled;
  66 };
  67
  68 /* pop a record for an inode into the defrag tree.  The lock
  69  * must be held already
  70  *
  71  * If you're inserting a record for an older transid than an
  72  * existing record, the transid already in the tree is lowered
  73  *
  74  * If an existing record is found the defrag item you
  75  * pass in is freed
  76  */
  77 static void __btrfs_add_inode_defrag(struct inode *inode,
  78                                     struct inode_defrag *defrag)
  79 {
  80         struct btrfs_root *root = BTRFS_I(inode)->root;
  81         struct inode_defrag *entry;
  82         struct rb_node **p;
  83         struct rb_node *parent = NULL;
  84
  85         p = &root->fs_info->defrag_inodes.rb_node;
  86         while (*p) {
  87                 parent = *p;
  88                 entry = rb_entry(parent, struct inode_defrag, rb_node);
  89
  90                 if (defrag->ino < entry->ino)
  91                         p = &parent->rb_left;
  92                 else if (defrag->ino > entry->ino)
  93                         p = &parent->rb_right;
  94                 else {
  95                         /* if we're reinserting an entry for
  96                          * an old defrag run, make sure to
  97                          * lower the transid of our existing record
  98                          */
  99                         if (defrag->transid < entry->transid)
 100                                 entry->transid = defrag->transid;
 101                         if (defrag->last_offset > entry->last_offset)
 102                                 entry->last_offset = defrag->last_offset;
 103                         goto exists;
 104                 }
 105         }
 106         BTRFS_I(inode)->in_defrag = 1;
 107         rb_link_node(&defrag->rb_node, parent, p);
 108         rb_insert_color(&defrag->rb_node, &root->fs_info->defrag_inodes);
 109         return;
 110
 111 exists:
 112         kfree(defrag);
 113         return;
 114
 115 }
 116
 117 /*
 118  * insert a defrag record for this inode if auto defrag is
 119  * enabled
 120  */
 121 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 122                            struct inode *inode)
 123 {
 124         struct btrfs_root *root = BTRFS_I(inode)->root;
 125         struct inode_defrag *defrag;
 126         u64 transid;
 127
 128         if (!btrfs_test_opt(root, AUTO_DEFRAG))
 129                 return 0;
 130
 131         if (btrfs_fs_closing(root->fs_info))
 132                 return 0;
 133
 134         if (BTRFS_I(inode)->in_defrag)
 135                 return 0;
 136
 137         if (trans)
 138                 transid = trans->transid;
 139         else
 140                 transid = BTRFS_I(inode)->root->last_trans;
 141
 142         defrag = kzalloc(sizeof(*defrag), GFP_NOFS);
 143         if (!defrag)
 144                 return -ENOMEM;
 145
 146         defrag->ino = btrfs_ino(inode);
 147         defrag->transid = transid;
 148         defrag->root = root->root_key.objectid;
 149
 150         spin_lock(&root->fs_info->defrag_inodes_lock);
 151         if (!BTRFS_I(inode)->in_defrag)
 152                 __btrfs_add_inode_defrag(inode, defrag);
 153         else
 154                 kfree(defrag);
 155         spin_unlock(&root->fs_info->defrag_inodes_lock);
 156         return 0;
 157 }
 158
 159 /*
 160  * must be called with the defrag_inodes lock held
 161  */
 162 struct inode_defrag *btrfs_find_defrag_inode(struct btrfs_fs_info *info, u64 ino,
 163                                              struct rb_node **next)
 164 {
 165         struct inode_defrag *entry = NULL;
 166         struct rb_node *p;
 167         struct rb_node *parent = NULL;
 168
 169         p = info->defrag_inodes.rb_node;
 170         while (p) {
 171                 parent = p;
 172                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 173
 174                 if (ino < entry->ino)
 175                         p = parent->rb_left;
 176                 else if (ino > entry->ino)
 177                         p = parent->rb_right;
 178                 else
 179                         return entry;
 180         }
 181
 182         if (next) {
 183                 while (parent && ino > entry->ino) {
 184                         parent = rb_next(parent);
 185                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 186                 }
 187                 *next = parent;
 188         }
 189         return NULL;
 190 }
 191
 192 /*
 193  * run through the list of inodes in the FS that need
 194  * defragging
 195  */
 196 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 197 {
 198         struct inode_defrag *defrag;
 199         struct btrfs_root *inode_root;
 200         struct inode *inode;
 201         struct rb_node *n;
 202         struct btrfs_key key;
 203         struct btrfs_ioctl_defrag_range_args range;
 204         u64 first_ino = 0;
 205         int num_defrag;
 206         int defrag_batch = 1024;
 207
 208         memset(&range, 0, sizeof(range));
 209         range.len = (u64)-1;
 210
 211         atomic_inc(&fs_info->defrag_running);
 212         spin_lock(&fs_info->defrag_inodes_lock);
 213         while(1) {
 214                 n = NULL;
 215
 216                 /* find an inode to defrag */
 217                 defrag = btrfs_find_defrag_inode(fs_info, first_ino, &n);
 218                 if (!defrag) {
 219                         if (n)
 220                                 defrag = rb_entry(n, struct inode_defrag, rb_node);
 221                         else if (first_ino) {
 222                                 first_ino = 0;
 223                                 continue;
 224                         } else {
 225                                 break;
 226                         }
 227                 }
 228
 229                 /* remove it from the rbtree */
 230                 first_ino = defrag->ino + 1;
 231                 rb_erase(&defrag->rb_node, &fs_info->defrag_inodes);
 232
 233                 if (btrfs_fs_closing(fs_info))
 234                         goto next_free;
 235
 236                 spin_unlock(&fs_info->defrag_inodes_lock);
 237
 238                 /* get the inode */
 239                 key.objectid = defrag->root;
 240                 btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 241                 key.offset = (u64)-1;
 242                 inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 243                 if (IS_ERR(inode_root))
 244                         goto next;
 245
 246                 key.objectid = defrag->ino;
 247                 btrfs_set_key_type(&key, BTRFS_INODE_ITEM_KEY);
 248                 key.offset = 0;
 249
 250                 inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 251                 if (IS_ERR(inode))
 252                         goto next;
 253
 254                 /* do a chunk of defrag */
 255                 BTRFS_I(inode)->in_defrag = 0;
 256                 range.start = defrag->last_offset;
 257                 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 258                                                defrag_batch);
 259                 /*
 260                  * if we filled the whole defrag batch, there
 261                  * must be more work to do.  Queue this defrag
 262                  * again
 263                  */
 264                 if (num_defrag == defrag_batch) {
 265                         defrag->last_offset = range.start;
 266                         __btrfs_add_inode_defrag(inode, defrag);
 267                         /*
 268                          * we don't want to kfree defrag, we added it back to
 269                          * the rbtree
 270                          */
 271                         defrag = NULL;
 272                 } else if (defrag->last_offset && !defrag->cycled) {
 273                         /*
 274                          * we didn't fill our defrag batch, but
 275                          * we didn't start at zero.  Make sure we loop
 276                          * around to the start of the file.
 277                          */
 278                         defrag->last_offset = 0;
 279                         defrag->cycled = 1;
 280                         __btrfs_add_inode_defrag(inode, defrag);
 281                         defrag = NULL;
 282                 }
 283
 284                 iput(inode);
 285 next:
 286                 spin_lock(&fs_info->defrag_inodes_lock);
 287 next_free:
 288                 kfree(defrag);
 289         }
 290         spin_unlock(&fs_info->defrag_inodes_lock);
 291
 292         atomic_dec(&fs_info->defrag_running);
 293
 294         /*
 295          * during unmount, we use the transaction_wait queue to
 296          * wait for the defragger to stop
 297          */
 298         wake_up(&fs_info->transaction_wait);
 299         return 0;
 300 }
 301
 302 /* simple helper to fault in pages and copy.  This should go away
 303  * and be replaced with calls into generic code.
 304  */
 305 static noinline int btrfs_copy_from_user(loff_t pos, int num_pages,
 306                                          size_t write_bytes,
 307                                          struct page **prepared_pages,
 308                                          struct iov_iter *i)
 309 {
 310         size_t copied = 0;
 311         size_t total_copied = 0;
 312         int pg = 0;
 313         int offset = pos & (PAGE_CACHE_SIZE - 1);
 314
 315         while (write_bytes > 0) {
 316                 size_t count = min_t(size_t,
 317                                      PAGE_CACHE_SIZE - offset, write_bytes);
 318                 struct page *page = prepared_pages[pg];
 319                 /*
 320                  * Copy data from userspace to the current page
 321                  *
 322                  * Disable pagefault to avoid recursive lock since
 323                  * the pages are already locked
 324                  */
 325                 pagefault_disable();
 326                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 327                 pagefault_enable();
 328
 329                 /* Flush processor's dcache for this page */
 330                 flush_dcache_page(page);
 331
 332                 /*
 333                  * if we get a partial write, we can end up with
 334                  * partially up to date pages.  These add
 335                  * a lot of complexity, so make sure they don't
 336                  * happen by forcing this copy to be retried.
 337                  *
 338                  * The rest of the btrfs_file_write code will fall
 339                  * back to page at a time copies after we return 0.
 340                  */
 341                 if (!PageUptodate(page) && copied < count)
 342                         copied = 0;
 343
 344                 iov_iter_advance(i, copied);
 345                 write_bytes -= copied;
 346                 total_copied += copied;
 347
 348                 /* Return to btrfs_file_aio_write to fault page */
 349                 if (unlikely(copied == 0))
 350                         break;
 351
 352                 if (unlikely(copied < PAGE_CACHE_SIZE - offset)) {
 353                         offset += copied;
 354                 } else {
 355                         pg++;
 356                         offset = 0;
 357                 }
 358         }
 359         return total_copied;
 360 }
 361
 362 /*
 363  * unlocks pages after btrfs_file_write is done with them
 364  */
 365 void btrfs_drop_pages(struct page **pages, size_t num_pages)
 366 {
 367         size_t i;
 368         for (i = 0; i < num_pages; i++) {
 369                 /* page checked is some magic around finding pages that
 370                  * have been modified without going through btrfs_set_page_dirty
 371                  * clear it here
 372                  */
 373                 ClearPageChecked(pages[i]);
 374                 unlock_page(pages[i]);
 375                 mark_page_accessed(pages[i]);
 376                 page_cache_release(pages[i]);
 377         }
 378 }
 379
 380 /*
 381  * after copy_from_user, pages need to be dirtied and we need to make
 382  * sure holes are created between the current EOF and the start of
 383  * any next extents (if required).
 384  *
 385  * this also makes the decision about creating an inline extent vs
 386  * doing real data extents, marking pages dirty and delalloc as required.
 387  */
 388 int btrfs_dirty_pages(struct btrfs_root *root, struct inode *inode,
 389                       struct page **pages, size_t num_pages,
 390                       loff_t pos, size_t write_bytes,
 391                       struct extent_state **cached)
 392 {
 393         int err = 0;
 394         int i;
 395         u64 num_bytes;
 396         u64 start_pos;
 397         u64 end_of_last_block;
 398         u64 end_pos = pos + write_bytes;
 399         loff_t isize = i_size_read(inode);
 400
 401         start_pos = pos & ~((u64)root->sectorsize - 1);
 402         num_bytes = (write_bytes + pos - start_pos +
 403                     root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
 404
 405         end_of_last_block = start_pos + num_bytes - 1;
 406         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 407                                         cached);
 408         if (err)
 409                 return err;
 410
 411         for (i = 0; i < num_pages; i++) {
 412                 struct page *p = pages[i];
 413                 SetPageUptodate(p);
 414                 ClearPageChecked(p);
 415                 set_page_dirty(p);
 416         }
 417
 418         /*
 419          * we've only changed i_size in ram, and we haven't updated
 420          * the disk i_size.  There is no need to log the inode
 421          * at this time.
 422          */
 423         if (end_pos > isize)
 424                 i_size_write(inode, end_pos);
 425         return 0;
 426 }
 427
 428 /*
 429  * this drops all the extents in the cache that intersect the range
 430  * [start, end].  Existing extents are split as required.
 431  */
 432 int btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
 433                             int skip_pinned)
 434 {
 435         struct extent_map *em;
 436         struct extent_map *split = NULL;
 437         struct extent_map *split2 = NULL;
 438         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 439         u64 len = end - start + 1;
 440         int ret;
 441         int testend = 1;
 442         unsigned long flags;
 443         int compressed = 0;
 444
 445         WARN_ON(end < start);
 446         if (end == (u64)-1) {
 447                 len = (u64)-1;
 448                 testend = 0;
 449         }
 450         while (1) {
 451                 if (!split)
 452                         split = alloc_extent_map();
 453                 if (!split2)
 454                         split2 = alloc_extent_map();
 455                 BUG_ON(!split || !split2);
 456
 457                 write_lock(&em_tree->lock);
 458                 em = lookup_extent_mapping(em_tree, start, len);
 459                 if (!em) {
 460                         write_unlock(&em_tree->lock);
 461                         break;
 462                 }
 463                 flags = em->flags;
 464                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 465                         if (testend && em->start + em->len >= start + len) {
 466                                 free_extent_map(em);
 467                                 write_unlock(&em_tree->lock);
 468                                 break;
 469                         }
 470                         start = em->start + em->len;
 471                         if (testend)
 472                                 len = start + len - (em->start + em->len);
 473                         free_extent_map(em);
 474                         write_unlock(&em_tree->lock);
 475                         continue;
 476                 }
 477                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 478                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 479                 remove_extent_mapping(em_tree, em);
 480
 481                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
 482                     em->start < start) {
 483                         split->start = em->start;
 484                         split->len = start - em->start;
 485                         split->orig_start = em->orig_start;
 486                         split->block_start = em->block_start;
 487
 488                         if (compressed)
 489                                 split->block_len = em->block_len;
 490                         else
 491                                 split->block_len = split->len;
 492
 493                         split->bdev = em->bdev;
 494                         split->flags = flags;
 495                         split->compress_type = em->compress_type;
 496                         ret = add_extent_mapping(em_tree, split);
 497                         BUG_ON(ret);
 498                         free_extent_map(split);
 499                         split = split2;
 500                         split2 = NULL;
 501                 }
 502                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
 503                     testend && em->start + em->len > start + len) {
 504                         u64 diff = start + len - em->start;
 505
 506                         split->start = start + len;
 507                         split->len = em->start + em->len - (start + len);
 508                         split->bdev = em->bdev;
 509                         split->flags = flags;
 510                         split->compress_type = em->compress_type;
 511
 512                         if (compressed) {
 513                                 split->block_len = em->block_len;
 514                                 split->block_start = em->block_start;
 515                                 split->orig_start = em->orig_start;
 516                         } else {
 517                                 split->block_len = split->len;
 518                                 split->block_start = em->block_start + diff;
 519                                 split->orig_start = split->start;
 520                         }
 521
 522                         ret = add_extent_mapping(em_tree, split);
 523                         BUG_ON(ret);
 524                         free_extent_map(split);
 525                         split = NULL;
 526                 }
 527                 write_unlock(&em_tree->lock);
 528
 529                 /* once for us */
 530                 free_extent_map(em);
 531                 /* once for the tree*/
 532                 free_extent_map(em);
 533         }
 534         if (split)
 535                 free_extent_map(split);
 536         if (split2)
 537                 free_extent_map(split2);
 538         return 0;
 539 }
 540
 541 /*
 542  * this is very complex, but the basic idea is to drop all extents
 543  * in the range start - end.  hint_block is filled in with a block number
 544  * that would be a good hint to the block allocator for this file.
 545  *
 546  * If an extent intersects the range but is not entirely inside the range
 547  * it is either truncated or split.  Anything entirely inside the range
 548  * is deleted from the tree.
 549  */
 550 int btrfs_drop_extents(struct btrfs_trans_handle *trans, struct inode *inode,
 551                        u64 start, u64 end, u64 *hint_byte, int drop_cache)
 552 {
 553         struct btrfs_root *root = BTRFS_I(inode)->root;
 554         struct extent_buffer *leaf;
 555         struct btrfs_file_extent_item *fi;
 556         struct btrfs_path *path;
 557         struct btrfs_key key;
 558         struct btrfs_key new_key;
 559         u64 ino = btrfs_ino(inode);
 560         u64 search_start = start;
 561         u64 disk_bytenr = 0;
 562         u64 num_bytes = 0;
 563         u64 extent_offset = 0;
 564         u64 extent_end = 0;
 565         int del_nr = 0;
 566         int del_slot = 0;
 567         int extent_type;
 568         int recow;
 569         int ret;
 570
 571         if (drop_cache)
 572                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 573
 574         path = btrfs_alloc_path();
 575         if (!path)
 576                 return -ENOMEM;
 577
 578         while (1) {
 579                 recow = 0;
 580                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 581                                                search_start, -1);
 582                 if (ret < 0)
 583                         break;
 584                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 585                         leaf = path->nodes[0];
 586                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 587                         if (key.objectid == ino &&
 588                             key.type == BTRFS_EXTENT_DATA_KEY)
 589                                 path->slots[0]--;
 590                 }
 591                 ret = 0;
 592 next_slot:
 593                 leaf = path->nodes[0];
 594                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 595                         BUG_ON(del_nr > 0);
 596                         ret = btrfs_next_leaf(root, path);
 597                         if (ret < 0)
 598                                 break;
 599                         if (ret > 0) {
 600                                 ret = 0;
 601                                 break;
 602                         }
 603                         leaf = path->nodes[0];
 604                         recow = 1;
 605                 }
 606
 607                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 608                 if (key.objectid > ino ||
 609                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 610                         break;
 611
 612                 fi = btrfs_item_ptr(leaf, path->slots[0],
 613                                     struct btrfs_file_extent_item);
 614                 extent_type = btrfs_file_extent_type(leaf, fi);
 615
 616                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 617                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 618                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 619                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 620                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 621                         extent_end = key.offset +
 622                                 btrfs_file_extent_num_bytes(leaf, fi);
 623                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 624                         extent_end = key.offset +
 625                                 btrfs_file_extent_inline_len(leaf,
 626                                                      path->slots[0], fi);
 627                 } else {
 628                         WARN_ON(1);
 629                         extent_end = search_start;
 630                 }
 631
 632                 if (extent_end <= search_start) {
 633                         path->slots[0]++;
 634                         goto next_slot;
 635                 }
 636
 637                 search_start = max(key.offset, start);
 638                 if (recow) {
 639                         btrfs_release_path(path);
 640                         continue;
 641                 }
 642
 643                 /*
 644                  *     | - range to drop - |
 645                  *  | -------- extent -------- |
 646                  */
 647                 if (start > key.offset && end < extent_end) {
 648                         BUG_ON(del_nr > 0);
 649                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 650
 651                         memcpy(&new_key, &key, sizeof(new_key));
 652                         new_key.offset = start;
 653                         ret = btrfs_duplicate_item(trans, root, path,
 654                                                    &new_key);
 655                         if (ret == -EAGAIN) {
 656                                 btrfs_release_path(path);
 657                                 continue;
 658                         }
 659                         if (ret < 0)
 660                                 break;
 661
 662                         leaf = path->nodes[0];
 663                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 664                                             struct btrfs_file_extent_item);
 665                         btrfs_set_file_extent_num_bytes(leaf, fi,
 666                                                         start - key.offset);
 667
 668                         fi = btrfs_item_ptr(leaf, path->slots[0],
 669                                             struct btrfs_file_extent_item);
 670
 671                         extent_offset += start - key.offset;
 672                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 673                         btrfs_set_file_extent_num_bytes(leaf, fi,
 674                                                         extent_end - start);
 675                         btrfs_mark_buffer_dirty(leaf);
 676
 677                         if (disk_bytenr > 0) {
 678                                 ret = btrfs_inc_extent_ref(trans, root,
 679                                                 disk_bytenr, num_bytes, 0,
 680                                                 root->root_key.objectid,
 681                                                 new_key.objectid,
 682                                                 start - extent_offset);
 683                                 BUG_ON(ret);
 684                                 *hint_byte = disk_bytenr;
 685                         }
 686                         key.offset = start;
 687                 }
 688                 /*
 689                  *  | ---- range to drop ----- |
 690                  *      | -------- extent -------- |
 691                  */
 692                 if (start <= key.offset && end < extent_end) {
 693                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 694
 695                         memcpy(&new_key, &key, sizeof(new_key));
 696                         new_key.offset = end;
 697                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 698
 699                         extent_offset += end - key.offset;
 700                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 701                         btrfs_set_file_extent_num_bytes(leaf, fi,
 702                                                         extent_end - end);
 703                         btrfs_mark_buffer_dirty(leaf);
 704                         if (disk_bytenr > 0) {
 705                                 inode_sub_bytes(inode, end - key.offset);
 706                                 *hint_byte = disk_bytenr;
 707                         }
 708                         break;
 709                 }
 710
 711                 search_start = extent_end;
 712                 /*
 713                  *       | ---- range to drop ----- |
 714                  *  | -------- extent -------- |
 715                  */
 716                 if (start > key.offset && end >= extent_end) {
 717                         BUG_ON(del_nr > 0);
 718                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 719
 720                         btrfs_set_file_extent_num_bytes(leaf, fi,
 721                                                         start - key.offset);
 722                         btrfs_mark_buffer_dirty(leaf);
 723                         if (disk_bytenr > 0) {
 724                                 inode_sub_bytes(inode, extent_end - start);
 725                                 *hint_byte = disk_bytenr;
 726                         }
 727                         if (end == extent_end)
 728                                 break;
 729
 730                         path->slots[0]++;
 731                         goto next_slot;
 732                 }
 733
 734                 /*
 735                  *  | ---- range to drop ----- |
 736                  *    | ------ extent ------ |
 737                  */
 738                 if (start <= key.offset && end >= extent_end) {
 739                         if (del_nr == 0) {
 740                                 del_slot = path->slots[0];
 741                                 del_nr = 1;
 742                         } else {
 743                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 744                                 del_nr++;
 745                         }
 746
 747                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 748                                 inode_sub_bytes(inode,
 749                                                 extent_end - key.offset);
 750                                 extent_end = ALIGN(extent_end,
 751                                                    root->sectorsize);
 752                         } else if (disk_bytenr > 0) {
 753                                 ret = btrfs_free_extent(trans, root,
 754                                                 disk_bytenr, num_bytes, 0,
 755                                                 root->root_key.objectid,
 756                                                 key.objectid, key.offset -
 757                                                 extent_offset);
 758                                 BUG_ON(ret);
 759                                 inode_sub_bytes(inode,
 760                                                 extent_end - key.offset);
 761                                 *hint_byte = disk_bytenr;
 762                         }
 763
 764                         if (end == extent_end)
 765                                 break;
 766
 767                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 768                                 path->slots[0]++;
 769                                 goto next_slot;
 770                         }
 771
 772                         ret = btrfs_del_items(trans, root, path, del_slot,
 773                                               del_nr);
 774                         BUG_ON(ret);
 775
 776                         del_nr = 0;
 777                         del_slot = 0;
 778
 779                         btrfs_release_path(path);
 780                         continue;
 781                 }
 782
 783                 BUG_ON(1);
 784         }
 785
 786         if (del_nr > 0) {
 787                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 788                 BUG_ON(ret);
 789         }
 790
 791         btrfs_free_path(path);
 792         return ret;
 793 }
 794
 795 static int extent_mergeable(struct extent_buffer *leaf, int slot,
 796                             u64 objectid, u64 bytenr, u64 orig_offset,
 797                             u64 *start, u64 *end)
 798 {
 799         struct btrfs_file_extent_item *fi;
 800         struct btrfs_key key;
 801         u64 extent_end;
 802
 803         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
 804                 return 0;
 805
 806         btrfs_item_key_to_cpu(leaf, &key, slot);
 807         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
 808                 return 0;
 809
 810         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
 811         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
 812             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
 813             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
 814             btrfs_file_extent_compression(leaf, fi) ||
 815             btrfs_file_extent_encryption(leaf, fi) ||
 816             btrfs_file_extent_other_encoding(leaf, fi))
 817                 return 0;
 818
 819         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 820         if ((*start && *start != key.offset) || (*end && *end != extent_end))
 821                 return 0;
 822
 823         *start = key.offset;
 824         *end = extent_end;
 825         return 1;
 826 }
 827
 828 /*
 829  * Mark extent in the range start - end as written.
 830  *
 831  * This changes extent type from 'pre-allocated' to 'regular'. If only
 832  * part of extent is marked as written, the extent will be split into
 833  * two or three.
 834  */
 835 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
 836                               struct inode *inode, u64 start, u64 end)
 837 {
 838         struct btrfs_root *root = BTRFS_I(inode)->root;
 839         struct extent_buffer *leaf;
 840         struct btrfs_path *path;
 841         struct btrfs_file_extent_item *fi;
 842         struct btrfs_key key;
 843         struct btrfs_key new_key;
 844         u64 bytenr;
 845         u64 num_bytes;
 846         u64 extent_end;
 847         u64 orig_offset;
 848         u64 other_start;
 849         u64 other_end;
 850         u64 split;
 851         int del_nr = 0;
 852         int del_slot = 0;
 853         int recow;
 854         int ret;
 855         u64 ino = btrfs_ino(inode);
 856
 857         btrfs_drop_extent_cache(inode, start, end - 1, 0);
 858
 859         path = btrfs_alloc_path();
 860         if (!path)
 861                 return -ENOMEM;
 862 again:
 863         recow = 0;
 864         split = start;
 865         key.objectid = ino;
 866         key.type = BTRFS_EXTENT_DATA_KEY;
 867         key.offset = split;
 868
 869         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 870         if (ret < 0)
 871                 goto out;
 872         if (ret > 0 && path->slots[0] > 0)
 873                 path->slots[0]--;
 874
 875         leaf = path->nodes[0];
 876         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 877         BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
 878         fi = btrfs_item_ptr(leaf, path->slots[0],
 879                             struct btrfs_file_extent_item);
 880         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
 881                BTRFS_FILE_EXTENT_PREALLOC);
 882         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 883         BUG_ON(key.offset > start || extent_end < end);
 884
 885         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 886         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 887         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
 888         memcpy(&new_key, &key, sizeof(new_key));
 889
 890         if (start == key.offset && end < extent_end) {
 891                 other_start = 0;
 892                 other_end = start;
 893                 if (extent_mergeable(leaf, path->slots[0] - 1,
 894                                      ino, bytenr, orig_offset,
 895                                      &other_start, &other_end)) {
 896                         new_key.offset = end;
 897                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 898                         fi = btrfs_item_ptr(leaf, path->slots[0],
 899                                             struct btrfs_file_extent_item);
 900                         btrfs_set_file_extent_num_bytes(leaf, fi,
 901                                                         extent_end - end);
 902                         btrfs_set_file_extent_offset(leaf, fi,
 903                                                      end - orig_offset);
 904                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 905                                             struct btrfs_file_extent_item);
 906                         btrfs_set_file_extent_num_bytes(leaf, fi,
 907                                                         end - other_start);
 908                         btrfs_mark_buffer_dirty(leaf);
 909                         goto out;
 910                 }
 911         }
 912
 913         if (start > key.offset && end == extent_end) {
 914                 other_start = end;
 915                 other_end = 0;
 916                 if (extent_mergeable(leaf, path->slots[0] + 1,
 917                                      ino, bytenr, orig_offset,
 918                                      &other_start, &other_end)) {
 919                         fi = btrfs_item_ptr(leaf, path->slots[0],
 920                                             struct btrfs_file_extent_item);
 921                         btrfs_set_file_extent_num_bytes(leaf, fi,
 922                                                         start - key.offset);
 923                         path->slots[0]++;
 924                         new_key.offset = start;
 925                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 926
 927                         fi = btrfs_item_ptr(leaf, path->slots[0],
 928                                             struct btrfs_file_extent_item);
 929                         btrfs_set_file_extent_num_bytes(leaf, fi,
 930                                                         other_end - start);
 931                         btrfs_set_file_extent_offset(leaf, fi,
 932                                                      start - orig_offset);
 933                         btrfs_mark_buffer_dirty(leaf);
 934                         goto out;
 935                 }
 936         }
 937
 938         while (start > key.offset || end < extent_end) {
 939                 if (key.offset == start)
 940                         split = end;
 941
 942                 new_key.offset = split;
 943                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
 944                 if (ret == -EAGAIN) {
 945                         btrfs_release_path(path);
 946                         goto again;
 947                 }
 948                 BUG_ON(ret < 0);
 949
 950                 leaf = path->nodes[0];
 951                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 952                                     struct btrfs_file_extent_item);
 953                 btrfs_set_file_extent_num_bytes(leaf, fi,
 954                                                 split - key.offset);
 955
 956                 fi = btrfs_item_ptr(leaf, path->slots[0],
 957                                     struct btrfs_file_extent_item);
 958
 959                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
 960                 btrfs_set_file_extent_num_bytes(leaf, fi,
 961                                                 extent_end - split);
 962                 btrfs_mark_buffer_dirty(leaf);
 963
 964                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
 965                                            root->root_key.objectid,
 966                                            ino, orig_offset);
 967                 BUG_ON(ret);
 968
 969                 if (split == start) {
 970                         key.offset = start;
 971                 } else {
 972                         BUG_ON(start != key.offset);
 973                         path->slots[0]--;
 974                         extent_end = end;
 975                 }
 976                 recow = 1;
 977         }
 978
 979         other_start = end;
 980         other_end = 0;
 981         if (extent_mergeable(leaf, path->slots[0] + 1,
 982                              ino, bytenr, orig_offset,
 983                              &other_start, &other_end)) {
 984                 if (recow) {
 985                         btrfs_release_path(path);
 986                         goto again;
 987                 }
 988                 extent_end = other_end;
 989                 del_slot = path->slots[0] + 1;
 990                 del_nr++;
 991                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
 992                                         0, root->root_key.objectid,
 993                                         ino, orig_offset);
 994                 BUG_ON(ret);
 995         }
 996         other_start = 0;
 997         other_end = start;
 998         if (extent_mergeable(leaf, path->slots[0] - 1,
 999                              ino, bytenr, orig_offset,
1000                              &other_start, &other_end)) {
1001                 if (recow) {
1002                         btrfs_release_path(path);
1003                         goto again;
1004                 }
1005                 key.offset = other_start;
1006                 del_slot = path->slots[0];
1007                 del_nr++;
1008                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1009                                         0, root->root_key.objectid,
1010                                         ino, orig_offset);
1011                 BUG_ON(ret);
1012         }
1013         if (del_nr == 0) {
1014                 fi = btrfs_item_ptr(leaf, path->slots[0],
1015                            struct btrfs_file_extent_item);
1016                 btrfs_set_file_extent_type(leaf, fi,
1017                                            BTRFS_FILE_EXTENT_REG);
1018                 btrfs_mark_buffer_dirty(leaf);
1019         } else {
1020                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1021                            struct btrfs_file_extent_item);
1022                 btrfs_set_file_extent_type(leaf, fi,
1023                                            BTRFS_FILE_EXTENT_REG);
1024                 btrfs_set_file_extent_num_bytes(leaf, fi,
1025                                                 extent_end - key.offset);
1026                 btrfs_mark_buffer_dirty(leaf);
1027
1028                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1029                 BUG_ON(ret);
1030         }
1031 out:
1032         btrfs_free_path(path);
1033         return 0;
1034 }
1035
1036 /*
1037  * on error we return an unlocked page and the error value
1038  * on success we return a locked page and 0
1039  */
1040 static int prepare_uptodate_page(struct page *page, u64 pos,
1041                                  bool force_uptodate)
1042 {
1043         int ret = 0;
1044
1045         if (((pos & (PAGE_CACHE_SIZE - 1)) || force_uptodate) &&
1046             !PageUptodate(page)) {
1047                 ret = btrfs_readpage(NULL, page);
1048                 if (ret)
1049                         return ret;
1050                 lock_page(page);
1051                 if (!PageUptodate(page)) {
1052                         unlock_page(page);
1053                         return -EIO;
1054                 }
1055         }
1056         return 0;
1057 }
1058
1059 /*
1060  * this gets pages into the page cache and locks them down, it also properly
1061  * waits for data=ordered extents to finish before allowing the pages to be
1062  * modified.
1063  */
1064 static noinline int prepare_pages(struct btrfs_root *root, struct file *file,
1065                          struct page **pages, size_t num_pages,
1066                          loff_t pos, unsigned long first_index,
1067                          size_t write_bytes, bool force_uptodate)
1068 {
1069         struct extent_state *cached_state = NULL;
1070         int i;
1071         unsigned long index = pos >> PAGE_CACHE_SHIFT;
1072         struct inode *inode = fdentry(file)->d_inode;
1073         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1074         int err = 0;
1075         int faili = 0;
1076         u64 start_pos;
1077         u64 last_pos;
1078
1079         start_pos = pos & ~((u64)root->sectorsize - 1);
1080         last_pos = ((u64)index + num_pages) << PAGE_CACHE_SHIFT;
1081
1082 again:
1083         for (i = 0; i < num_pages; i++) {
1084                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1085                                                mask);
1086                 if (!pages[i]) {
1087                         faili = i - 1;
1088                         err = -ENOMEM;
1089                         goto fail;
1090                 }
1091
1092                 if (i == 0)
1093                         err = prepare_uptodate_page(pages[i], pos,
1094                                                     force_uptodate);
1095                 if (i == num_pages - 1)
1096                         err = prepare_uptodate_page(pages[i],
1097                                                     pos + write_bytes, false);
1098                 if (err) {
1099                         page_cache_release(pages[i]);
1100                         faili = i - 1;
1101                         goto fail;
1102                 }
1103                 wait_on_page_writeback(pages[i]);
1104         }
1105         err = 0;
1106         if (start_pos < inode->i_size) {
1107                 struct btrfs_ordered_extent *ordered;
1108                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1109                                  start_pos, last_pos - 1, 0, &cached_state,
1110                                  GFP_NOFS);
1111                 ordered = btrfs_lookup_first_ordered_extent(inode,
1112                                                             last_pos - 1);
1113                 if (ordered &&
1114                     ordered->file_offset + ordered->len > start_pos &&
1115                     ordered->file_offset < last_pos) {
1116                         btrfs_put_ordered_extent(ordered);
1117                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1118                                              start_pos, last_pos - 1,
1119                                              &cached_state, GFP_NOFS);
1120                         for (i = 0; i < num_pages; i++) {
1121                                 unlock_page(pages[i]);
1122                                 page_cache_release(pages[i]);
1123                         }
1124                         btrfs_wait_ordered_range(inode, start_pos,
1125                                                  last_pos - start_pos);
1126                         goto again;
1127                 }
1128                 if (ordered)
1129                         btrfs_put_ordered_extent(ordered);
1130
1131                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1132                                   last_pos - 1, EXTENT_DIRTY | EXTENT_DELALLOC |
1133                                   EXTENT_DO_ACCOUNTING, 0, 0, &cached_state,
1134                                   GFP_NOFS);
1135                 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1136                                      start_pos, last_pos - 1, &cached_state,
1137                                      GFP_NOFS);
1138         }
1139         for (i = 0; i < num_pages; i++) {
1140                 clear_page_dirty_for_io(pages[i]);
1141                 set_page_extent_mapped(pages[i]);
1142                 WARN_ON(!PageLocked(pages[i]));
1143         }
1144         return 0;
1145 fail:
1146         while (faili >= 0) {
1147                 unlock_page(pages[faili]);
1148                 page_cache_release(pages[faili]);
1149                 faili--;
1150         }
1151         return err;
1152
1153 }
1154
1155 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1156                                                struct iov_iter *i,
1157                                                loff_t pos)
1158 {
1159         struct inode *inode = fdentry(file)->d_inode;
1160         struct btrfs_root *root = BTRFS_I(inode)->root;
1161         struct page **pages = NULL;
1162         unsigned long first_index;
1163         size_t num_written = 0;
1164         int nrptrs;
1165         int ret = 0;
1166         bool force_page_uptodate = false;
1167
1168         nrptrs = min((iov_iter_count(i) + PAGE_CACHE_SIZE - 1) /
1169                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1170                      (sizeof(struct page *)));
1171         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1172         nrptrs = max(nrptrs, 8);
1173         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1174         if (!pages)
1175                 return -ENOMEM;
1176
1177         first_index = pos >> PAGE_CACHE_SHIFT;
1178
1179         while (iov_iter_count(i) > 0) {
1180                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1181                 size_t write_bytes = min(iov_iter_count(i),
1182                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1183                                          offset);
1184                 size_t num_pages = (write_bytes + offset +
1185                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1186                 size_t dirty_pages;
1187                 size_t copied;
1188
1189                 WARN_ON(num_pages > nrptrs);
1190
1191                 /*
1192                  * Fault pages before locking them in prepare_pages
1193                  * to avoid recursive lock
1194                  */
1195                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1196                         ret = -EFAULT;
1197                         break;
1198                 }
1199
1200                 ret = btrfs_delalloc_reserve_space(inode,
1201                                         num_pages << PAGE_CACHE_SHIFT);
1202                 if (ret)
1203                         break;
1204
1205                 /*
1206                  * This is going to setup the pages array with the number of
1207                  * pages we want, so we don't really need to worry about the
1208                  * contents of pages from loop to loop
1209                  */
1210                 ret = prepare_pages(root, file, pages, num_pages,
1211                                     pos, first_index, write_bytes,
1212                                     force_page_uptodate);
1213                 if (ret) {
1214                         btrfs_delalloc_release_space(inode,
1215                                         num_pages << PAGE_CACHE_SHIFT);
1216                         break;
1217                 }
1218
1219                 copied = btrfs_copy_from_user(pos, num_pages,
1220                                            write_bytes, pages, i);
1221
1222                 /*
1223                  * if we have trouble faulting in the pages, fall
1224                  * back to one page at a time
1225                  */
1226                 if (copied < write_bytes)
1227                         nrptrs = 1;
1228
1229                 if (copied == 0) {
1230                         force_page_uptodate = true;
1231                         dirty_pages = 0;
1232                 } else {
1233                         force_page_uptodate = false;
1234                         dirty_pages = (copied + offset +
1235                                        PAGE_CACHE_SIZE - 1) >>
1236                                        PAGE_CACHE_SHIFT;
1237                 }
1238
1239                 /*
1240                  * If we had a short copy we need to release the excess delaloc
1241                  * bytes we reserved.  We need to increment outstanding_extents
1242                  * because btrfs_delalloc_release_space will decrement it, but
1243                  * we still have an outstanding extent for the chunk we actually
1244                  * managed to copy.
1245                  */
1246                 if (num_pages > dirty_pages) {
1247                         if (copied > 0) {
1248                                 spin_lock(&BTRFS_I(inode)->lock);
1249                                 BTRFS_I(inode)->outstanding_extents++;
1250                                 spin_unlock(&BTRFS_I(inode)->lock);
1251                         }
1252                         btrfs_delalloc_release_space(inode,
1253                                         (num_pages - dirty_pages) <<
1254                                         PAGE_CACHE_SHIFT);
1255                 }
1256
1257                 if (copied > 0) {
1258                         ret = btrfs_dirty_pages(root, inode, pages,
1259                                                 dirty_pages, pos, copied,
1260                                                 NULL);
1261                         if (ret) {
1262                                 btrfs_delalloc_release_space(inode,
1263                                         dirty_pages << PAGE_CACHE_SHIFT);
1264                                 btrfs_drop_pages(pages, num_pages);
1265                                 break;
1266                         }
1267                 }
1268
1269                 btrfs_drop_pages(pages, num_pages);
1270
1271                 cond_resched();
1272
1273                 balance_dirty_pages_ratelimited_nr(inode->i_mapping,
1274                                                    dirty_pages);
1275                 if (dirty_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1276                         btrfs_btree_balance_dirty(root, 1);
1277                 btrfs_throttle(root);
1278
1279                 pos += copied;
1280                 num_written += copied;
1281         }
1282
1283         kfree(pages);
1284
1285         return num_written ? num_written : ret;
1286 }
1287
1288 static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1289                                     const struct iovec *iov,
1290                                     unsigned long nr_segs, loff_t pos,
1291                                     loff_t *ppos, size_t count, size_t ocount)
1292 {
1293         struct file *file = iocb->ki_filp;
1294         struct inode *inode = fdentry(file)->d_inode;
1295         struct iov_iter i;
1296         ssize_t written;
1297         ssize_t written_buffered;
1298         loff_t endbyte;
1299         int err;
1300
1301         written = generic_file_direct_write(iocb, iov, &nr_segs, pos, ppos,
1302                                             count, ocount);
1303
1304         /*
1305          * the generic O_DIRECT will update in-memory i_size after the
1306          * DIOs are done.  But our endio handlers that update the on
1307          * disk i_size never update past the in memory i_size.  So we
1308          * need one more update here to catch any additions to the
1309          * file
1310          */
1311         if (inode->i_size != BTRFS_I(inode)->disk_i_size) {
1312                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
1313                 mark_inode_dirty(inode);
1314         }
1315
1316         if (written < 0 || written == count)
1317                 return written;
1318
1319         pos += written;
1320         count -= written;
1321         iov_iter_init(&i, iov, nr_segs, count, written);
1322         written_buffered = __btrfs_buffered_write(file, &i, pos);
1323         if (written_buffered < 0) {
1324                 err = written_buffered;
1325                 goto out;
1326         }
1327         endbyte = pos + written_buffered - 1;
1328         err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
1329         if (err)
1330                 goto out;
1331         written += written_buffered;
1332         *ppos = pos + written_buffered;
1333         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_CACHE_SHIFT,
1334                                  endbyte >> PAGE_CACHE_SHIFT);
1335 out:
1336         return written ? written : err;
1337 }
1338
1339 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
1340                                     const struct iovec *iov,
1341                                     unsigned long nr_segs, loff_t pos)
1342 {
1343         struct file *file = iocb->ki_filp;
1344         struct inode *inode = fdentry(file)->d_inode;
1345         struct btrfs_root *root = BTRFS_I(inode)->root;
1346         loff_t *ppos = &iocb->ki_pos;
1347         u64 start_pos;
1348         ssize_t num_written = 0;
1349         ssize_t err = 0;
1350         size_t count, ocount;
1351
1352         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1353
1354         mutex_lock(&inode->i_mutex);
1355
1356         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1357         if (err) {
1358                 mutex_unlock(&inode->i_mutex);
1359                 goto out;
1360         }
1361         count = ocount;
1362
1363         current->backing_dev_info = inode->i_mapping->backing_dev_info;
1364         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1365         if (err) {
1366                 mutex_unlock(&inode->i_mutex);
1367                 goto out;
1368         }
1369
1370         if (count == 0) {
1371                 mutex_unlock(&inode->i_mutex);
1372                 goto out;
1373         }
1374
1375         err = file_remove_suid(file);
1376         if (err) {
1377                 mutex_unlock(&inode->i_mutex);
1378                 goto out;
1379         }
1380
1381         /*
1382          * If BTRFS flips readonly due to some impossible error
1383          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1384          * although we have opened a file as writable, we have
1385          * to stop this write operation to ensure FS consistency.
1386          */
1387         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
1388                 mutex_unlock(&inode->i_mutex);
1389                 err = -EROFS;
1390                 goto out;
1391         }
1392
1393         err = btrfs_update_time(file);
1394         if (err) {
1395                 mutex_unlock(&inode->i_mutex);
1396                 goto out;
1397         }
1398         BTRFS_I(inode)->sequence++;
1399
1400         start_pos = round_down(pos, root->sectorsize);
1401         if (start_pos > i_size_read(inode)) {
1402                 err = btrfs_cont_expand(inode, i_size_read(inode), start_pos);
1403                 if (err) {
1404                         mutex_unlock(&inode->i_mutex);
1405                         goto out;
1406                 }
1407         }
1408
1409         if (unlikely(file->f_flags & O_DIRECT)) {
1410                 num_written = __btrfs_direct_write(iocb, iov, nr_segs,
1411                                                    pos, ppos, count, ocount);
1412         } else {
1413                 struct iov_iter i;
1414
1415                 iov_iter_init(&i, iov, nr_segs, count, num_written);
1416
1417                 num_written = __btrfs_buffered_write(file, &i, pos);
1418                 if (num_written > 0)
1419                         *ppos = pos + num_written;
1420         }
1421
1422         mutex_unlock(&inode->i_mutex);
1423
1424         /*
1425          * we want to make sure fsync finds this change
1426          * but we haven't joined a transaction running right now.
1427          *
1428          * Later on, someone is sure to update the inode and get the
1429          * real transid recorded.
1430          *
1431          * We set last_trans now to the fs_info generation + 1,
1432          * this will either be one more than the running transaction
1433          * or the generation used for the next transaction if there isn't
1434          * one running right now.
1435          */
1436         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1437         if (num_written > 0 || num_written == -EIOCBQUEUED) {
1438                 err = generic_write_sync(file, pos, num_written);
1439                 if (err < 0 && num_written > 0)
1440                         num_written = err;
1441         }
1442 out:
1443         current->backing_dev_info = NULL;
1444         return num_written ? num_written : err;
1445 }
1446
1447 int btrfs_release_file(struct inode *inode, struct file *filp)
1448 {
1449         /*
1450          * ordered_data_close is set by settattr when we are about to truncate
1451          * a file from a non-zero size to a zero size.  This tries to
1452          * flush down new bytes that may have been written if the
1453          * application were using truncate to replace a file in place.
1454          */
1455         if (BTRFS_I(inode)->ordered_data_close) {
1456                 BTRFS_I(inode)->ordered_data_close = 0;
1457                 btrfs_add_ordered_operation(NULL, BTRFS_I(inode)->root, inode);
1458                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1459                         filemap_flush(inode->i_mapping);
1460         }
1461         if (filp->private_data)
1462                 btrfs_ioctl_trans_end(filp);
1463         return 0;
1464 }
1465
1466 /*
1467  * fsync call for both files and directories.  This logs the inode into
1468  * the tree log instead of forcing full commits whenever possible.
1469  *
1470  * It needs to call filemap_fdatawait so that all ordered extent updates are
1471  * in the metadata btree are up to date for copying to the log.
1472  *
1473  * It drops the inode mutex before doing the tree log commit.  This is an
1474  * important optimization for directories because holding the mutex prevents
1475  * new operations on the dir while we write to disk.
1476  */
1477 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1478 {
1479         struct dentry *dentry = file->f_path.dentry;
1480         struct inode *inode = dentry->d_inode;
1481         struct btrfs_root *root = BTRFS_I(inode)->root;
1482         int ret = 0;
1483         struct btrfs_trans_handle *trans;
1484
1485         trace_btrfs_sync_file(file, datasync);
1486
1487         ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1488         if (ret)
1489                 return ret;
1490         mutex_lock(&inode->i_mutex);
1491
1492         /* we wait first, since the writeback may change the inode */
1493         root->log_batch++;
1494         btrfs_wait_ordered_range(inode, 0, (u64)-1);
1495         root->log_batch++;
1496
1497         /*
1498          * check the transaction that last modified this inode
1499          * and see if its already been committed
1500          */
1501         if (!BTRFS_I(inode)->last_trans) {
1502                 mutex_unlock(&inode->i_mutex);
1503                 goto out;
1504         }
1505
1506         /*
1507          * if the last transaction that changed this file was before
1508          * the current transaction, we can bail out now without any
1509          * syncing
1510          */
1511         smp_mb();
1512         if (BTRFS_I(inode)->last_trans <=
1513             root->fs_info->last_trans_committed) {
1514                 BTRFS_I(inode)->last_trans = 0;
1515                 mutex_unlock(&inode->i_mutex);
1516                 goto out;
1517         }
1518
1519         /*
1520          * ok we haven't committed the transaction yet, lets do a commit
1521          */
1522         if (file->private_data)
1523                 btrfs_ioctl_trans_end(file);
1524
1525         trans = btrfs_start_transaction(root, 0);
1526         if (IS_ERR(trans)) {
1527                 ret = PTR_ERR(trans);
1528                 mutex_unlock(&inode->i_mutex);
1529                 goto out;
1530         }
1531
1532         ret = btrfs_log_dentry_safe(trans, root, dentry);
1533         if (ret < 0) {
1534                 mutex_unlock(&inode->i_mutex);
1535                 goto out;
1536         }
1537
1538         /* we've logged all the items and now have a consistent
1539          * version of the file in the log.  It is possible that
1540          * someone will come in and modify the file, but that's
1541          * fine because the log is consistent on disk, and we
1542          * have references to all of the file's extents
1543          *
1544          * It is possible that someone will come in and log the
1545          * file again, but that will end up using the synchronization
1546          * inside btrfs_sync_log to keep things safe.
1547          */
1548         mutex_unlock(&inode->i_mutex);
1549
1550         if (ret != BTRFS_NO_LOG_SYNC) {
1551                 if (ret > 0) {
1552                         ret = btrfs_commit_transaction(trans, root);
1553                 } else {
1554                         ret = btrfs_sync_log(trans, root);
1555                         if (ret == 0)
1556                                 ret = btrfs_end_transaction(trans, root);
1557                         else
1558                                 ret = btrfs_commit_transaction(trans, root);
1559                 }
1560         } else {
1561                 ret = btrfs_end_transaction(trans, root);
1562         }
1563 out:
1564         return ret > 0 ? -EIO : ret;
1565 }
1566
1567 static const struct vm_operations_struct btrfs_file_vm_ops = {
1568         .fault          = filemap_fault,
1569         .page_mkwrite   = btrfs_page_mkwrite,
1570 };
1571
1572 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
1573 {
1574         struct address_space *mapping = filp->f_mapping;
1575
1576         if (!mapping->a_ops->readpage)
1577                 return -ENOEXEC;
1578
1579         file_accessed(filp);
1580         vma->vm_ops = &btrfs_file_vm_ops;
1581         vma->vm_flags |= VM_CAN_NONLINEAR;
1582
1583         return 0;
1584 }
1585
1586 static long btrfs_fallocate(struct file *file, int mode,
1587                             loff_t offset, loff_t len)
1588 {
1589         struct inode *inode = file->f_path.dentry->d_inode;
1590         struct extent_state *cached_state = NULL;
1591         u64 cur_offset;
1592         u64 last_byte;
1593         u64 alloc_start;
1594         u64 alloc_end;
1595         u64 alloc_hint = 0;
1596         u64 locked_end;
1597         u64 mask = BTRFS_I(inode)->root->sectorsize - 1;
1598         struct extent_map *em;
1599         int ret;
1600
1601         alloc_start = offset & ~mask;
1602         alloc_end =  (offset + len + mask) & ~mask;
1603
1604         /* We only support the FALLOC_FL_KEEP_SIZE mode */
1605         if (mode & ~FALLOC_FL_KEEP_SIZE)
1606                 return -EOPNOTSUPP;
1607
1608         /*
1609          * wait for ordered IO before we have any locks.  We'll loop again
1610          * below with the locks held.
1611          */
1612         btrfs_wait_ordered_range(inode, alloc_start, alloc_end - alloc_start);
1613
1614         mutex_lock(&inode->i_mutex);
1615         ret = inode_newsize_ok(inode, alloc_end);
1616         if (ret)
1617                 goto out;
1618
1619         if (alloc_start > inode->i_size) {
1620                 ret = btrfs_cont_expand(inode, i_size_read(inode),
1621                                         alloc_start);
1622                 if (ret)
1623                         goto out;
1624         }
1625
1626         locked_end = alloc_end - 1;
1627         while (1) {
1628                 struct btrfs_ordered_extent *ordered;
1629
1630                 /* the extent lock is ordered inside the running
1631                  * transaction
1632                  */
1633                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
1634                                  locked_end, 0, &cached_state, GFP_NOFS);
1635                 ordered = btrfs_lookup_first_ordered_extent(inode,
1636                                                             alloc_end - 1);
1637                 if (ordered &&
1638                     ordered->file_offset + ordered->len > alloc_start &&
1639                     ordered->file_offset < alloc_end) {
1640                         btrfs_put_ordered_extent(ordered);
1641                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1642                                              alloc_start, locked_end,
1643                                              &cached_state, GFP_NOFS);
1644                         /*
1645                          * we can't wait on the range with the transaction
1646                          * running or with the extent lock held
1647                          */
1648                         btrfs_wait_ordered_range(inode, alloc_start,
1649                                                  alloc_end - alloc_start);
1650                 } else {
1651                         if (ordered)
1652                                 btrfs_put_ordered_extent(ordered);
1653                         break;
1654                 }
1655         }
1656
1657         cur_offset = alloc_start;
1658         while (1) {
1659                 u64 actual_end;
1660
1661                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
1662                                       alloc_end - cur_offset, 0);
1663                 BUG_ON(IS_ERR_OR_NULL(em));
1664                 last_byte = min(extent_map_end(em), alloc_end);
1665                 actual_end = min_t(u64, extent_map_end(em), offset + len);
1666                 last_byte = (last_byte + mask) & ~mask;
1667
1668                 if (em->block_start == EXTENT_MAP_HOLE ||
1669                     (cur_offset >= inode->i_size &&
1670                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
1671
1672                         /*
1673                          * Make sure we have enough space before we do the
1674                          * allocation.
1675                          */
1676                         ret = btrfs_check_data_free_space(inode, last_byte -
1677                                                           cur_offset);
1678                         if (ret) {
1679                                 free_extent_map(em);
1680                                 break;
1681                         }
1682
1683                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
1684                                                         last_byte - cur_offset,
1685                                                         1 << inode->i_blkbits,
1686                                                         offset + len,
1687                                                         &alloc_hint);
1688
1689                         /* Let go of our reservation. */
1690                         btrfs_free_reserved_data_space(inode, last_byte -
1691                                                        cur_offset);
1692                         if (ret < 0) {
1693                                 free_extent_map(em);
1694                                 break;
1695                         }
1696                 } else if (actual_end > inode->i_size &&
1697                            !(mode & FALLOC_FL_KEEP_SIZE)) {
1698                         /*
1699                          * We didn't need to allocate any more space, but we
1700                          * still extended the size of the file so we need to
1701                          * update i_size.
1702                          */
1703                         inode->i_ctime = CURRENT_TIME;
1704                         i_size_write(inode, actual_end);
1705                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
1706                 }
1707                 free_extent_map(em);
1708
1709                 cur_offset = last_byte;
1710                 if (cur_offset >= alloc_end) {
1711                         ret = 0;
1712                         break;
1713                 }
1714         }
1715         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
1716                              &cached_state, GFP_NOFS);
1717 out:
1718         mutex_unlock(&inode->i_mutex);
1719         return ret;
1720 }
1721
1722 static int find_desired_extent(struct inode *inode, loff_t *offset, int origin)
1723 {
1724         struct btrfs_root *root = BTRFS_I(inode)->root;
1725         struct extent_map *em;
1726         struct extent_state *cached_state = NULL;
1727         u64 lockstart = *offset;
1728         u64 lockend = i_size_read(inode);
1729         u64 start = *offset;
1730         u64 orig_start = *offset;
1731         u64 len = i_size_read(inode);
1732         u64 last_end = 0;
1733         int ret = 0;
1734
1735         lockend = max_t(u64, root->sectorsize, lockend);
1736         if (lockend <= lockstart)
1737                 lockend = lockstart + root->sectorsize;
1738
1739         len = lockend - lockstart + 1;
1740
1741         len = max_t(u64, len, root->sectorsize);
1742         if (inode->i_size == 0)
1743                 return -ENXIO;
1744
1745         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend, 0,
1746                          &cached_state, GFP_NOFS);
1747
1748         /*
1749          * Delalloc is such a pain.  If we have a hole and we have pending
1750          * delalloc for a portion of the hole we will get back a hole that
1751          * exists for the entire range since it hasn't been actually written
1752          * yet.  So to take care of this case we need to look for an extent just
1753          * before the position we want in case there is outstanding delalloc
1754          * going on here.
1755          */
1756         if (origin == SEEK_HOLE && start != 0) {
1757                 if (start <= root->sectorsize)
1758                         em = btrfs_get_extent_fiemap(inode, NULL, 0, 0,
1759                                                      root->sectorsize, 0);
1760                 else
1761                         em = btrfs_get_extent_fiemap(inode, NULL, 0,
1762                                                      start - root->sectorsize,
1763                                                      root->sectorsize, 0);
1764                 if (IS_ERR(em)) {
1765                         ret = -ENXIO;
1766                         goto out;
1767                 }
1768                 last_end = em->start + em->len;
1769                 if (em->block_start == EXTENT_MAP_DELALLOC)
1770                         last_end = min_t(u64, last_end, inode->i_size);
1771                 free_extent_map(em);
1772         }
1773
1774         while (1) {
1775                 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
1776                 if (IS_ERR(em)) {
1777                         ret = -ENXIO;
1778                         break;
1779                 }
1780
1781                 if (em->block_start == EXTENT_MAP_HOLE) {
1782                         if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
1783                                 if (last_end <= orig_start) {
1784                                         free_extent_map(em);
1785                                         ret = -ENXIO;
1786                                         break;
1787                                 }
1788                         }
1789
1790                         if (origin == SEEK_HOLE) {
1791                                 *offset = start;
1792                                 free_extent_map(em);
1793                                 break;
1794                         }
1795                 } else {
1796                         if (origin == SEEK_DATA) {
1797                                 if (em->block_start == EXTENT_MAP_DELALLOC) {
1798                                         if (start >= inode->i_size) {
1799                                                 free_extent_map(em);
1800                                                 ret = -ENXIO;
1801                                                 break;
1802                                         }
1803                                 }
1804
1805                                 *offset = start;
1806                                 free_extent_map(em);
1807                                 break;
1808                         }
1809                 }
1810
1811                 start = em->start + em->len;
1812                 last_end = em->start + em->len;
1813
1814                 if (em->block_start == EXTENT_MAP_DELALLOC)
1815                         last_end = min_t(u64, last_end, inode->i_size);
1816
1817                 if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
1818                         free_extent_map(em);
1819                         ret = -ENXIO;
1820                         break;
1821                 }
1822                 free_extent_map(em);
1823                 cond_resched();
1824         }
1825         if (!ret)
1826                 *offset = min(*offset, inode->i_size);
1827 out:
1828         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
1829                              &cached_state, GFP_NOFS);
1830         return ret;
1831 }
1832
1833 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int origin)
1834 {
1835         struct inode *inode = file->f_mapping->host;
1836         int ret;
1837
1838         mutex_lock(&inode->i_mutex);
1839         switch (origin) {
1840         case SEEK_END:
1841         case SEEK_CUR:
1842                 offset = generic_file_llseek(file, offset, origin);
1843                 goto out;
1844         case SEEK_DATA:
1845         case SEEK_HOLE:
1846                 if (offset >= i_size_read(inode)) {
1847                         mutex_unlock(&inode->i_mutex);
1848                         return -ENXIO;
1849                 }
1850
1851                 ret = find_desired_extent(inode, &offset, origin);
1852                 if (ret) {
1853                         mutex_unlock(&inode->i_mutex);
1854                         return ret;
1855                 }
1856         }
1857
1858         if (offset < 0 && !(file->f_mode & FMODE_UNSIGNED_OFFSET)) {
1859                 offset = -EINVAL;
1860                 goto out;
1861         }
1862         if (offset > inode->i_sb->s_maxbytes) {
1863                 offset = -EINVAL;
1864                 goto out;
1865         }
1866
1867         /* Special lock needed here? */
1868         if (offset != file->f_pos) {
1869                 file->f_pos = offset;
1870                 file->f_version = 0;
1871         }
1872 out:
1873         mutex_unlock(&inode->i_mutex);
1874         return offset;
1875 }
1876
1877 const struct file_operations btrfs_file_operations = {
1878         .llseek         = btrfs_file_llseek,
1879         .read           = do_sync_read,
1880         .write          = do_sync_write,
1881         .aio_read       = generic_file_aio_read,
1882         .splice_read    = generic_file_splice_read,
1883         .aio_write      = btrfs_file_aio_write,
1884         .mmap           = btrfs_file_mmap,
1885         .open           = generic_file_open,
1886         .release        = btrfs_release_file,
1887         .fsync          = btrfs_sync_file,
1888         .fallocate      = btrfs_fallocate,
1889         .unlocked_ioctl = btrfs_ioctl,
1890 #ifdef CONFIG_COMPAT
1891         .compat_ioctl   = btrfs_ioctl,
1892 #endif
1893 };