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
 609                 if (key.objectid > ino)
 610                         break;
 611                 if (WARN_ON_ONCE(key.objectid < ino) ||
 612                     key.type < BTRFS_EXTENT_DATA_KEY) {
 613                         path->slots[0]++;
 614                         goto next_slot;
 615                 }
 616                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 617                         break;
 618
 619                 fi = btrfs_item_ptr(leaf, path->slots[0],
 620                                     struct btrfs_file_extent_item);
 621                 extent_type = btrfs_file_extent_type(leaf, fi);
 622
 623                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 624                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 625                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 626                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 627                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 628                         extent_end = key.offset +
 629                                 btrfs_file_extent_num_bytes(leaf, fi);
 630                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 631                         extent_end = key.offset +
 632                                 btrfs_file_extent_inline_len(leaf,
 633                                                      path->slots[0], fi);
 634                 } else {
 635                         /* can't happen */
 636                         BUG();
 637                 }
 638
 639                 if (extent_end <= search_start) {
 640                         path->slots[0]++;
 641                         goto next_slot;
 642                 }
 643
 644                 search_start = max(key.offset, start);
 645                 if (recow) {
 646                         btrfs_release_path(path);
 647                         continue;
 648                 }
 649
 650                 /*
 651                  *     | - range to drop - |
 652                  *  | -------- extent -------- |
 653                  */
 654                 if (start > key.offset && end < extent_end) {
 655                         BUG_ON(del_nr > 0);
 656                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 657
 658                         memcpy(&new_key, &key, sizeof(new_key));
 659                         new_key.offset = start;
 660                         ret = btrfs_duplicate_item(trans, root, path,
 661                                                    &new_key);
 662                         if (ret == -EAGAIN) {
 663                                 btrfs_release_path(path);
 664                                 continue;
 665                         }
 666                         if (ret < 0)
 667                                 break;
 668
 669                         leaf = path->nodes[0];
 670                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 671                                             struct btrfs_file_extent_item);
 672                         btrfs_set_file_extent_num_bytes(leaf, fi,
 673                                                         start - key.offset);
 674
 675                         fi = btrfs_item_ptr(leaf, path->slots[0],
 676                                             struct btrfs_file_extent_item);
 677
 678                         extent_offset += start - key.offset;
 679                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 680                         btrfs_set_file_extent_num_bytes(leaf, fi,
 681                                                         extent_end - start);
 682                         btrfs_mark_buffer_dirty(leaf);
 683
 684                         if (disk_bytenr > 0) {
 685                                 ret = btrfs_inc_extent_ref(trans, root,
 686                                                 disk_bytenr, num_bytes, 0,
 687                                                 root->root_key.objectid,
 688                                                 new_key.objectid,
 689                                                 start - extent_offset);
 690                                 BUG_ON(ret);
 691                                 *hint_byte = disk_bytenr;
 692                         }
 693                         key.offset = start;
 694                 }
 695                 /*
 696                  *  | ---- range to drop ----- |
 697                  *      | -------- extent -------- |
 698                  */
 699                 if (start <= key.offset && end < extent_end) {
 700                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 701
 702                         memcpy(&new_key, &key, sizeof(new_key));
 703                         new_key.offset = end;
 704                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 705
 706                         extent_offset += end - key.offset;
 707                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 708                         btrfs_set_file_extent_num_bytes(leaf, fi,
 709                                                         extent_end - end);
 710                         btrfs_mark_buffer_dirty(leaf);
 711                         if (disk_bytenr > 0) {
 712                                 inode_sub_bytes(inode, end - key.offset);
 713                                 *hint_byte = disk_bytenr;
 714                         }
 715                         break;
 716                 }
 717
 718                 search_start = extent_end;
 719                 /*
 720                  *       | ---- range to drop ----- |
 721                  *  | -------- extent -------- |
 722                  */
 723                 if (start > key.offset && end >= extent_end) {
 724                         BUG_ON(del_nr > 0);
 725                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 726
 727                         btrfs_set_file_extent_num_bytes(leaf, fi,
 728                                                         start - key.offset);
 729                         btrfs_mark_buffer_dirty(leaf);
 730                         if (disk_bytenr > 0) {
 731                                 inode_sub_bytes(inode, extent_end - start);
 732                                 *hint_byte = disk_bytenr;
 733                         }
 734                         if (end == extent_end)
 735                                 break;
 736
 737                         path->slots[0]++;
 738                         goto next_slot;
 739                 }
 740
 741                 /*
 742                  *  | ---- range to drop ----- |
 743                  *    | ------ extent ------ |
 744                  */
 745                 if (start <= key.offset && end >= extent_end) {
 746                         if (del_nr == 0) {
 747                                 del_slot = path->slots[0];
 748                                 del_nr = 1;
 749                         } else {
 750                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 751                                 del_nr++;
 752                         }
 753
 754                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 755                                 inode_sub_bytes(inode,
 756                                                 extent_end - key.offset);
 757                                 extent_end = ALIGN(extent_end,
 758                                                    root->sectorsize);
 759                         } else if (disk_bytenr > 0) {
 760                                 ret = btrfs_free_extent(trans, root,
 761                                                 disk_bytenr, num_bytes, 0,
 762                                                 root->root_key.objectid,
 763                                                 key.objectid, key.offset -
 764                                                 extent_offset);
 765                                 BUG_ON(ret);
 766                                 inode_sub_bytes(inode,
 767                                                 extent_end - key.offset);
 768                                 *hint_byte = disk_bytenr;
 769                         }
 770
 771                         if (end == extent_end)
 772                                 break;
 773
 774                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 775                                 path->slots[0]++;
 776                                 goto next_slot;
 777                         }
 778
 779                         ret = btrfs_del_items(trans, root, path, del_slot,
 780                                               del_nr);
 781                         BUG_ON(ret);
 782
 783                         del_nr = 0;
 784                         del_slot = 0;
 785
 786                         btrfs_release_path(path);
 787                         continue;
 788                 }
 789
 790                 BUG_ON(1);
 791         }
 792
 793         if (del_nr > 0) {
 794                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 795                 BUG_ON(ret);
 796         }
 797
 798         btrfs_free_path(path);
 799         return ret;
 800 }
 801
 802 static int extent_mergeable(struct extent_buffer *leaf, int slot,
 803                             u64 objectid, u64 bytenr, u64 orig_offset,
 804                             u64 *start, u64 *end)
 805 {
 806         struct btrfs_file_extent_item *fi;
 807         struct btrfs_key key;
 808         u64 extent_end;
 809
 810         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
 811                 return 0;
 812
 813         btrfs_item_key_to_cpu(leaf, &key, slot);
 814         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
 815                 return 0;
 816
 817         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
 818         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
 819             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
 820             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
 821             btrfs_file_extent_compression(leaf, fi) ||
 822             btrfs_file_extent_encryption(leaf, fi) ||
 823             btrfs_file_extent_other_encoding(leaf, fi))
 824                 return 0;
 825
 826         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 827         if ((*start && *start != key.offset) || (*end && *end != extent_end))
 828                 return 0;
 829
 830         *start = key.offset;
 831         *end = extent_end;
 832         return 1;
 833 }
 834
 835 /*
 836  * Mark extent in the range start - end as written.
 837  *
 838  * This changes extent type from 'pre-allocated' to 'regular'. If only
 839  * part of extent is marked as written, the extent will be split into
 840  * two or three.
 841  */
 842 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
 843                               struct inode *inode, u64 start, u64 end)
 844 {
 845         struct btrfs_root *root = BTRFS_I(inode)->root;
 846         struct extent_buffer *leaf;
 847         struct btrfs_path *path;
 848         struct btrfs_file_extent_item *fi;
 849         struct btrfs_key key;
 850         struct btrfs_key new_key;
 851         u64 bytenr;
 852         u64 num_bytes;
 853         u64 extent_end;
 854         u64 orig_offset;
 855         u64 other_start;
 856         u64 other_end;
 857         u64 split;
 858         int del_nr = 0;
 859         int del_slot = 0;
 860         int recow;
 861         int ret;
 862         u64 ino = btrfs_ino(inode);
 863
 864         btrfs_drop_extent_cache(inode, start, end - 1, 0);
 865
 866         path = btrfs_alloc_path();
 867         if (!path)
 868                 return -ENOMEM;
 869 again:
 870         recow = 0;
 871         split = start;
 872         key.objectid = ino;
 873         key.type = BTRFS_EXTENT_DATA_KEY;
 874         key.offset = split;
 875
 876         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 877         if (ret < 0)
 878                 goto out;
 879         if (ret > 0 && path->slots[0] > 0)
 880                 path->slots[0]--;
 881
 882         leaf = path->nodes[0];
 883         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 884         BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
 885         fi = btrfs_item_ptr(leaf, path->slots[0],
 886                             struct btrfs_file_extent_item);
 887         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
 888                BTRFS_FILE_EXTENT_PREALLOC);
 889         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 890         BUG_ON(key.offset > start || extent_end < end);
 891
 892         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 893         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 894         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
 895         memcpy(&new_key, &key, sizeof(new_key));
 896
 897         if (start == key.offset && end < extent_end) {
 898                 other_start = 0;
 899                 other_end = start;
 900                 if (extent_mergeable(leaf, path->slots[0] - 1,
 901                                      ino, bytenr, orig_offset,
 902                                      &other_start, &other_end)) {
 903                         new_key.offset = end;
 904                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 905                         fi = btrfs_item_ptr(leaf, path->slots[0],
 906                                             struct btrfs_file_extent_item);
 907                         btrfs_set_file_extent_num_bytes(leaf, fi,
 908                                                         extent_end - end);
 909                         btrfs_set_file_extent_offset(leaf, fi,
 910                                                      end - orig_offset);
 911                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 912                                             struct btrfs_file_extent_item);
 913                         btrfs_set_file_extent_num_bytes(leaf, fi,
 914                                                         end - other_start);
 915                         btrfs_mark_buffer_dirty(leaf);
 916                         goto out;
 917                 }
 918         }
 919
 920         if (start > key.offset && end == extent_end) {
 921                 other_start = end;
 922                 other_end = 0;
 923                 if (extent_mergeable(leaf, path->slots[0] + 1,
 924                                      ino, bytenr, orig_offset,
 925                                      &other_start, &other_end)) {
 926                         fi = btrfs_item_ptr(leaf, path->slots[0],
 927                                             struct btrfs_file_extent_item);
 928                         btrfs_set_file_extent_num_bytes(leaf, fi,
 929                                                         start - key.offset);
 930                         path->slots[0]++;
 931                         new_key.offset = start;
 932                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 933
 934                         fi = btrfs_item_ptr(leaf, path->slots[0],
 935                                             struct btrfs_file_extent_item);
 936                         btrfs_set_file_extent_num_bytes(leaf, fi,
 937                                                         other_end - start);
 938                         btrfs_set_file_extent_offset(leaf, fi,
 939                                                      start - orig_offset);
 940                         btrfs_mark_buffer_dirty(leaf);
 941                         goto out;
 942                 }
 943         }
 944
 945         while (start > key.offset || end < extent_end) {
 946                 if (key.offset == start)
 947                         split = end;
 948
 949                 new_key.offset = split;
 950                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
 951                 if (ret == -EAGAIN) {
 952                         btrfs_release_path(path);
 953                         goto again;
 954                 }
 955                 BUG_ON(ret < 0);
 956
 957                 leaf = path->nodes[0];
 958                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 959                                     struct btrfs_file_extent_item);
 960                 btrfs_set_file_extent_num_bytes(leaf, fi,
 961                                                 split - key.offset);
 962
 963                 fi = btrfs_item_ptr(leaf, path->slots[0],
 964                                     struct btrfs_file_extent_item);
 965
 966                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
 967                 btrfs_set_file_extent_num_bytes(leaf, fi,
 968                                                 extent_end - split);
 969                 btrfs_mark_buffer_dirty(leaf);
 970
 971                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
 972                                            root->root_key.objectid,
 973                                            ino, orig_offset);
 974                 BUG_ON(ret);
 975
 976                 if (split == start) {
 977                         key.offset = start;
 978                 } else {
 979                         BUG_ON(start != key.offset);
 980                         path->slots[0]--;
 981                         extent_end = end;
 982                 }
 983                 recow = 1;
 984         }
 985
 986         other_start = end;
 987         other_end = 0;
 988         if (extent_mergeable(leaf, path->slots[0] + 1,
 989                              ino, bytenr, orig_offset,
 990                              &other_start, &other_end)) {
 991                 if (recow) {
 992                         btrfs_release_path(path);
 993                         goto again;
 994                 }
 995                 extent_end = other_end;
 996                 del_slot = path->slots[0] + 1;
 997                 del_nr++;
 998                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
 999                                         0, root->root_key.objectid,
1000                                         ino, orig_offset);
1001                 BUG_ON(ret);
1002         }
1003         other_start = 0;
1004         other_end = start;
1005         if (extent_mergeable(leaf, path->slots[0] - 1,
1006                              ino, bytenr, orig_offset,
1007                              &other_start, &other_end)) {
1008                 if (recow) {
1009                         btrfs_release_path(path);
1010                         goto again;
1011                 }
1012                 key.offset = other_start;
1013                 del_slot = path->slots[0];
1014                 del_nr++;
1015                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1016                                         0, root->root_key.objectid,
1017                                         ino, orig_offset);
1018                 BUG_ON(ret);
1019         }
1020         if (del_nr == 0) {
1021                 fi = btrfs_item_ptr(leaf, path->slots[0],
1022                            struct btrfs_file_extent_item);
1023                 btrfs_set_file_extent_type(leaf, fi,
1024                                            BTRFS_FILE_EXTENT_REG);
1025                 btrfs_mark_buffer_dirty(leaf);
1026         } else {
1027                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1028                            struct btrfs_file_extent_item);
1029                 btrfs_set_file_extent_type(leaf, fi,
1030                                            BTRFS_FILE_EXTENT_REG);
1031                 btrfs_set_file_extent_num_bytes(leaf, fi,
1032                                                 extent_end - key.offset);
1033                 btrfs_mark_buffer_dirty(leaf);
1034
1035                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1036                 BUG_ON(ret);
1037         }
1038 out:
1039         btrfs_free_path(path);
1040         return 0;
1041 }
1042
1043 /*
1044  * on error we return an unlocked page and the error value
1045  * on success we return a locked page and 0
1046  */
1047 static int prepare_uptodate_page(struct page *page, u64 pos,
1048                                  bool force_uptodate)
1049 {
1050         int ret = 0;
1051
1052         if (((pos & (PAGE_CACHE_SIZE - 1)) || force_uptodate) &&
1053             !PageUptodate(page)) {
1054                 ret = btrfs_readpage(NULL, page);
1055                 if (ret)
1056                         return ret;
1057                 lock_page(page);
1058                 if (!PageUptodate(page)) {
1059                         unlock_page(page);
1060                         return -EIO;
1061                 }
1062         }
1063         return 0;
1064 }
1065
1066 /*
1067  * this gets pages into the page cache and locks them down, it also properly
1068  * waits for data=ordered extents to finish before allowing the pages to be
1069  * modified.
1070  */
1071 static noinline int prepare_pages(struct btrfs_root *root, struct file *file,
1072                          struct page **pages, size_t num_pages,
1073                          loff_t pos, unsigned long first_index,
1074                          size_t write_bytes, bool force_uptodate)
1075 {
1076         struct extent_state *cached_state = NULL;
1077         int i;
1078         unsigned long index = pos >> PAGE_CACHE_SHIFT;
1079         struct inode *inode = fdentry(file)->d_inode;
1080         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1081         int err = 0;
1082         int faili = 0;
1083         u64 start_pos;
1084         u64 last_pos;
1085
1086         start_pos = pos & ~((u64)root->sectorsize - 1);
1087         last_pos = ((u64)index + num_pages) << PAGE_CACHE_SHIFT;
1088
1089 again:
1090         for (i = 0; i < num_pages; i++) {
1091                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1092                                                mask);
1093                 if (!pages[i]) {
1094                         faili = i - 1;
1095                         err = -ENOMEM;
1096                         goto fail;
1097                 }
1098
1099                 if (i == 0)
1100                         err = prepare_uptodate_page(pages[i], pos,
1101                                                     force_uptodate);
1102                 if (i == num_pages - 1)
1103                         err = prepare_uptodate_page(pages[i],
1104                                                     pos + write_bytes, false);
1105                 if (err) {
1106                         page_cache_release(pages[i]);
1107                         faili = i - 1;
1108                         goto fail;
1109                 }
1110                 wait_on_page_writeback(pages[i]);
1111         }
1112         err = 0;
1113         if (start_pos < inode->i_size) {
1114                 struct btrfs_ordered_extent *ordered;
1115                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1116                                  start_pos, last_pos - 1, 0, &cached_state,
1117                                  GFP_NOFS);
1118                 ordered = btrfs_lookup_first_ordered_extent(inode,
1119                                                             last_pos - 1);
1120                 if (ordered &&
1121                     ordered->file_offset + ordered->len > start_pos &&
1122                     ordered->file_offset < last_pos) {
1123                         btrfs_put_ordered_extent(ordered);
1124                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1125                                              start_pos, last_pos - 1,
1126                                              &cached_state, GFP_NOFS);
1127                         for (i = 0; i < num_pages; i++) {
1128                                 unlock_page(pages[i]);
1129                                 page_cache_release(pages[i]);
1130                         }
1131                         btrfs_wait_ordered_range(inode, start_pos,
1132                                                  last_pos - start_pos);
1133                         goto again;
1134                 }
1135                 if (ordered)
1136                         btrfs_put_ordered_extent(ordered);
1137
1138                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1139                                   last_pos - 1, EXTENT_DIRTY | EXTENT_DELALLOC |
1140                                   EXTENT_DO_ACCOUNTING, 0, 0, &cached_state,
1141                                   GFP_NOFS);
1142                 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1143                                      start_pos, last_pos - 1, &cached_state,
1144                                      GFP_NOFS);
1145         }
1146         for (i = 0; i < num_pages; i++) {
1147                 clear_page_dirty_for_io(pages[i]);
1148                 set_page_extent_mapped(pages[i]);
1149                 WARN_ON(!PageLocked(pages[i]));
1150         }
1151         return 0;
1152 fail:
1153         while (faili >= 0) {
1154                 unlock_page(pages[faili]);
1155                 page_cache_release(pages[faili]);
1156                 faili--;
1157         }
1158         return err;
1159
1160 }
1161
1162 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1163                                                struct iov_iter *i,
1164                                                loff_t pos)
1165 {
1166         struct inode *inode = fdentry(file)->d_inode;
1167         struct btrfs_root *root = BTRFS_I(inode)->root;
1168         struct page **pages = NULL;
1169         unsigned long first_index;
1170         size_t num_written = 0;
1171         int nrptrs;
1172         int ret = 0;
1173         bool force_page_uptodate = false;
1174
1175         nrptrs = min((iov_iter_count(i) + PAGE_CACHE_SIZE - 1) /
1176                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1177                      (sizeof(struct page *)));
1178         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1179         nrptrs = max(nrptrs, 8);
1180         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1181         if (!pages)
1182                 return -ENOMEM;
1183
1184         first_index = pos >> PAGE_CACHE_SHIFT;
1185
1186         while (iov_iter_count(i) > 0) {
1187                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1188                 size_t write_bytes = min(iov_iter_count(i),
1189                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1190                                          offset);
1191                 size_t num_pages = (write_bytes + offset +
1192                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1193                 size_t dirty_pages;
1194                 size_t copied;
1195
1196                 WARN_ON(num_pages > nrptrs);
1197
1198                 /*
1199                  * Fault pages before locking them in prepare_pages
1200                  * to avoid recursive lock
1201                  */
1202                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1203                         ret = -EFAULT;
1204                         break;
1205                 }
1206
1207                 ret = btrfs_delalloc_reserve_space(inode,
1208                                         num_pages << PAGE_CACHE_SHIFT);
1209                 if (ret)
1210                         break;
1211
1212                 /*
1213                  * This is going to setup the pages array with the number of
1214                  * pages we want, so we don't really need to worry about the
1215                  * contents of pages from loop to loop
1216                  */
1217                 ret = prepare_pages(root, file, pages, num_pages,
1218                                     pos, first_index, write_bytes,
1219                                     force_page_uptodate);
1220                 if (ret) {
1221                         btrfs_delalloc_release_space(inode,
1222                                         num_pages << PAGE_CACHE_SHIFT);
1223                         break;
1224                 }
1225
1226                 copied = btrfs_copy_from_user(pos, num_pages,
1227                                            write_bytes, pages, i);
1228
1229                 /*
1230                  * if we have trouble faulting in the pages, fall
1231                  * back to one page at a time
1232                  */
1233                 if (copied < write_bytes)
1234                         nrptrs = 1;
1235
1236                 if (copied == 0) {
1237                         force_page_uptodate = true;
1238                         dirty_pages = 0;
1239                 } else {
1240                         force_page_uptodate = false;
1241                         dirty_pages = (copied + offset +
1242                                        PAGE_CACHE_SIZE - 1) >>
1243                                        PAGE_CACHE_SHIFT;
1244                 }
1245
1246                 /*
1247                  * If we had a short copy we need to release the excess delaloc
1248                  * bytes we reserved.  We need to increment outstanding_extents
1249                  * because btrfs_delalloc_release_space will decrement it, but
1250                  * we still have an outstanding extent for the chunk we actually
1251                  * managed to copy.
1252                  */
1253                 if (num_pages > dirty_pages) {
1254                         if (copied > 0) {
1255                                 spin_lock(&BTRFS_I(inode)->lock);
1256                                 BTRFS_I(inode)->outstanding_extents++;
1257                                 spin_unlock(&BTRFS_I(inode)->lock);
1258                         }
1259                         btrfs_delalloc_release_space(inode,
1260                                         (num_pages - dirty_pages) <<
1261                                         PAGE_CACHE_SHIFT);
1262                 }
1263
1264                 if (copied > 0) {
1265                         ret = btrfs_dirty_pages(root, inode, pages,
1266                                                 dirty_pages, pos, copied,
1267                                                 NULL);
1268                         if (ret) {
1269                                 btrfs_delalloc_release_space(inode,
1270                                         dirty_pages << PAGE_CACHE_SHIFT);
1271                                 btrfs_drop_pages(pages, num_pages);
1272                                 break;
1273                         }
1274                 }
1275
1276                 btrfs_drop_pages(pages, num_pages);
1277
1278                 cond_resched();
1279
1280                 balance_dirty_pages_ratelimited_nr(inode->i_mapping,
1281                                                    dirty_pages);
1282                 if (dirty_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1283                         btrfs_btree_balance_dirty(root, 1);
1284                 btrfs_throttle(root);
1285
1286                 pos += copied;
1287                 num_written += copied;
1288         }
1289
1290         kfree(pages);
1291
1292         return num_written ? num_written : ret;
1293 }
1294
1295 static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1296                                     const struct iovec *iov,
1297                                     unsigned long nr_segs, loff_t pos,
1298                                     loff_t *ppos, size_t count, size_t ocount)
1299 {
1300         struct file *file = iocb->ki_filp;
1301         struct inode *inode = fdentry(file)->d_inode;
1302         struct iov_iter i;
1303         ssize_t written;
1304         ssize_t written_buffered;
1305         loff_t endbyte;
1306         int err;
1307
1308         written = generic_file_direct_write(iocb, iov, &nr_segs, pos, ppos,
1309                                             count, ocount);
1310
1311         /*
1312          * the generic O_DIRECT will update in-memory i_size after the
1313          * DIOs are done.  But our endio handlers that update the on
1314          * disk i_size never update past the in memory i_size.  So we
1315          * need one more update here to catch any additions to the
1316          * file
1317          */
1318         if (inode->i_size != BTRFS_I(inode)->disk_i_size) {
1319                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
1320                 mark_inode_dirty(inode);
1321         }
1322
1323         if (written < 0 || written == count)
1324                 return written;
1325
1326         pos += written;
1327         count -= written;
1328         iov_iter_init(&i, iov, nr_segs, count, written);
1329         written_buffered = __btrfs_buffered_write(file, &i, pos);
1330         if (written_buffered < 0) {
1331                 err = written_buffered;
1332                 goto out;
1333         }
1334         endbyte = pos + written_buffered - 1;
1335         err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
1336         if (err)
1337                 goto out;
1338         written += written_buffered;
1339         *ppos = pos + written_buffered;
1340         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_CACHE_SHIFT,
1341                                  endbyte >> PAGE_CACHE_SHIFT);
1342 out:
1343         return written ? written : err;
1344 }
1345
1346 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
1347                                     const struct iovec *iov,
1348                                     unsigned long nr_segs, loff_t pos)
1349 {
1350         struct file *file = iocb->ki_filp;
1351         struct inode *inode = fdentry(file)->d_inode;
1352         struct btrfs_root *root = BTRFS_I(inode)->root;
1353         loff_t *ppos = &iocb->ki_pos;
1354         u64 start_pos;
1355         ssize_t num_written = 0;
1356         ssize_t err = 0;
1357         size_t count, ocount;
1358
1359         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1360
1361         mutex_lock(&inode->i_mutex);
1362
1363         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1364         if (err) {
1365                 mutex_unlock(&inode->i_mutex);
1366                 goto out;
1367         }
1368         count = ocount;
1369
1370         current->backing_dev_info = inode->i_mapping->backing_dev_info;
1371         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1372         if (err) {
1373                 mutex_unlock(&inode->i_mutex);
1374                 goto out;
1375         }
1376
1377         if (count == 0) {
1378                 mutex_unlock(&inode->i_mutex);
1379                 goto out;
1380         }
1381
1382         err = file_remove_suid(file);
1383         if (err) {
1384                 mutex_unlock(&inode->i_mutex);
1385                 goto out;
1386         }
1387
1388         /*
1389          * If BTRFS flips readonly due to some impossible error
1390          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1391          * although we have opened a file as writable, we have
1392          * to stop this write operation to ensure FS consistency.
1393          */
1394         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
1395                 mutex_unlock(&inode->i_mutex);
1396                 err = -EROFS;
1397                 goto out;
1398         }
1399
1400         err = btrfs_update_time(file);
1401         if (err) {
1402                 mutex_unlock(&inode->i_mutex);
1403                 goto out;
1404         }
1405         BTRFS_I(inode)->sequence++;
1406
1407         start_pos = round_down(pos, root->sectorsize);
1408         if (start_pos > i_size_read(inode)) {
1409                 err = btrfs_cont_expand(inode, i_size_read(inode), start_pos);
1410                 if (err) {
1411                         mutex_unlock(&inode->i_mutex);
1412                         goto out;
1413                 }
1414         }
1415
1416         if (unlikely(file->f_flags & O_DIRECT)) {
1417                 num_written = __btrfs_direct_write(iocb, iov, nr_segs,
1418                                                    pos, ppos, count, ocount);
1419         } else {
1420                 struct iov_iter i;
1421
1422                 iov_iter_init(&i, iov, nr_segs, count, num_written);
1423
1424                 num_written = __btrfs_buffered_write(file, &i, pos);
1425                 if (num_written > 0)
1426                         *ppos = pos + num_written;
1427         }
1428
1429         mutex_unlock(&inode->i_mutex);
1430
1431         /*
1432          * we want to make sure fsync finds this change
1433          * but we haven't joined a transaction running right now.
1434          *
1435          * Later on, someone is sure to update the inode and get the
1436          * real transid recorded.
1437          *
1438          * We set last_trans now to the fs_info generation + 1,
1439          * this will either be one more than the running transaction
1440          * or the generation used for the next transaction if there isn't
1441          * one running right now.
1442          */
1443         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1444         if (num_written > 0 || num_written == -EIOCBQUEUED) {
1445                 err = generic_write_sync(file, pos, num_written);
1446                 if (err < 0 && num_written > 0)
1447                         num_written = err;
1448         }
1449 out:
1450         current->backing_dev_info = NULL;
1451         return num_written ? num_written : err;
1452 }
1453
1454 int btrfs_release_file(struct inode *inode, struct file *filp)
1455 {
1456         /*
1457          * ordered_data_close is set by settattr when we are about to truncate
1458          * a file from a non-zero size to a zero size.  This tries to
1459          * flush down new bytes that may have been written if the
1460          * application were using truncate to replace a file in place.
1461          */
1462         if (BTRFS_I(inode)->ordered_data_close) {
1463                 BTRFS_I(inode)->ordered_data_close = 0;
1464                 btrfs_add_ordered_operation(NULL, BTRFS_I(inode)->root, inode);
1465                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1466                         filemap_flush(inode->i_mapping);
1467         }
1468         if (filp->private_data)
1469                 btrfs_ioctl_trans_end(filp);
1470         return 0;
1471 }
1472
1473 /*
1474  * fsync call for both files and directories.  This logs the inode into
1475  * the tree log instead of forcing full commits whenever possible.
1476  *
1477  * It needs to call filemap_fdatawait so that all ordered extent updates are
1478  * in the metadata btree are up to date for copying to the log.
1479  *
1480  * It drops the inode mutex before doing the tree log commit.  This is an
1481  * important optimization for directories because holding the mutex prevents
1482  * new operations on the dir while we write to disk.
1483  */
1484 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1485 {
1486         struct dentry *dentry = file->f_path.dentry;
1487         struct inode *inode = dentry->d_inode;
1488         struct btrfs_root *root = BTRFS_I(inode)->root;
1489         int ret = 0;
1490         struct btrfs_trans_handle *trans;
1491
1492         trace_btrfs_sync_file(file, datasync);
1493
1494         ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1495         if (ret)
1496                 return ret;
1497         mutex_lock(&inode->i_mutex);
1498
1499         /* we wait first, since the writeback may change the inode */
1500         root->log_batch++;
1501         btrfs_wait_ordered_range(inode, 0, (u64)-1);
1502         root->log_batch++;
1503
1504         /*
1505          * check the transaction that last modified this inode
1506          * and see if its already been committed
1507          */
1508         if (!BTRFS_I(inode)->last_trans) {
1509                 mutex_unlock(&inode->i_mutex);
1510                 goto out;
1511         }
1512
1513         /*
1514          * if the last transaction that changed this file was before
1515          * the current transaction, we can bail out now without any
1516          * syncing
1517          */
1518         smp_mb();
1519         if (BTRFS_I(inode)->last_trans <=
1520             root->fs_info->last_trans_committed) {
1521                 BTRFS_I(inode)->last_trans = 0;
1522                 mutex_unlock(&inode->i_mutex);
1523                 goto out;
1524         }
1525
1526         /*
1527          * ok we haven't committed the transaction yet, lets do a commit
1528          */
1529         if (file->private_data)
1530                 btrfs_ioctl_trans_end(file);
1531
1532         trans = btrfs_start_transaction(root, 0);
1533         if (IS_ERR(trans)) {
1534                 ret = PTR_ERR(trans);
1535                 mutex_unlock(&inode->i_mutex);
1536                 goto out;
1537         }
1538
1539         ret = btrfs_log_dentry_safe(trans, root, dentry);
1540         if (ret < 0) {
1541                 mutex_unlock(&inode->i_mutex);
1542                 goto out;
1543         }
1544
1545         /* we've logged all the items and now have a consistent
1546          * version of the file in the log.  It is possible that
1547          * someone will come in and modify the file, but that's
1548          * fine because the log is consistent on disk, and we
1549          * have references to all of the file's extents
1550          *
1551          * It is possible that someone will come in and log the
1552          * file again, but that will end up using the synchronization
1553          * inside btrfs_sync_log to keep things safe.
1554          */
1555         mutex_unlock(&inode->i_mutex);
1556
1557         if (ret != BTRFS_NO_LOG_SYNC) {
1558                 if (ret > 0) {
1559                         ret = btrfs_commit_transaction(trans, root);
1560                 } else {
1561                         ret = btrfs_sync_log(trans, root);
1562                         if (ret == 0)
1563                                 ret = btrfs_end_transaction(trans, root);
1564                         else
1565                                 ret = btrfs_commit_transaction(trans, root);
1566                 }
1567         } else {
1568                 ret = btrfs_end_transaction(trans, root);
1569         }
1570 out:
1571         return ret > 0 ? -EIO : ret;
1572 }
1573
1574 static const struct vm_operations_struct btrfs_file_vm_ops = {
1575         .fault          = filemap_fault,
1576         .page_mkwrite   = btrfs_page_mkwrite,
1577 };
1578
1579 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
1580 {
1581         struct address_space *mapping = filp->f_mapping;
1582
1583         if (!mapping->a_ops->readpage)
1584                 return -ENOEXEC;
1585
1586         file_accessed(filp);
1587         vma->vm_ops = &btrfs_file_vm_ops;
1588         vma->vm_flags |= VM_CAN_NONLINEAR;
1589
1590         return 0;
1591 }
1592
1593 static long btrfs_fallocate(struct file *file, int mode,
1594                             loff_t offset, loff_t len)
1595 {
1596         struct inode *inode = file->f_path.dentry->d_inode;
1597         struct extent_state *cached_state = NULL;
1598         u64 cur_offset;
1599         u64 last_byte;
1600         u64 alloc_start;
1601         u64 alloc_end;
1602         u64 alloc_hint = 0;
1603         u64 locked_end;
1604         u64 mask = BTRFS_I(inode)->root->sectorsize - 1;
1605         struct extent_map *em;
1606         int ret;
1607
1608         alloc_start = offset & ~mask;
1609         alloc_end =  (offset + len + mask) & ~mask;
1610
1611         /* We only support the FALLOC_FL_KEEP_SIZE mode */
1612         if (mode & ~FALLOC_FL_KEEP_SIZE)
1613                 return -EOPNOTSUPP;
1614
1615         /*
1616          * wait for ordered IO before we have any locks.  We'll loop again
1617          * below with the locks held.
1618          */
1619         btrfs_wait_ordered_range(inode, alloc_start, alloc_end - alloc_start);
1620
1621         mutex_lock(&inode->i_mutex);
1622         ret = inode_newsize_ok(inode, alloc_end);
1623         if (ret)
1624                 goto out;
1625
1626         if (alloc_start > inode->i_size) {
1627                 ret = btrfs_cont_expand(inode, i_size_read(inode),
1628                                         alloc_start);
1629                 if (ret)
1630                         goto out;
1631         }
1632
1633         locked_end = alloc_end - 1;
1634         while (1) {
1635                 struct btrfs_ordered_extent *ordered;
1636
1637                 /* the extent lock is ordered inside the running
1638                  * transaction
1639                  */
1640                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
1641                                  locked_end, 0, &cached_state, GFP_NOFS);
1642                 ordered = btrfs_lookup_first_ordered_extent(inode,
1643                                                             alloc_end - 1);
1644                 if (ordered &&
1645                     ordered->file_offset + ordered->len > alloc_start &&
1646                     ordered->file_offset < alloc_end) {
1647                         btrfs_put_ordered_extent(ordered);
1648                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1649                                              alloc_start, locked_end,
1650                                              &cached_state, GFP_NOFS);
1651                         /*
1652                          * we can't wait on the range with the transaction
1653                          * running or with the extent lock held
1654                          */
1655                         btrfs_wait_ordered_range(inode, alloc_start,
1656                                                  alloc_end - alloc_start);
1657                 } else {
1658                         if (ordered)
1659                                 btrfs_put_ordered_extent(ordered);
1660                         break;
1661                 }
1662         }
1663
1664         cur_offset = alloc_start;
1665         while (1) {
1666                 u64 actual_end;
1667
1668                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
1669                                       alloc_end - cur_offset, 0);
1670                 BUG_ON(IS_ERR_OR_NULL(em));
1671                 last_byte = min(extent_map_end(em), alloc_end);
1672                 actual_end = min_t(u64, extent_map_end(em), offset + len);
1673                 last_byte = (last_byte + mask) & ~mask;
1674
1675                 if (em->block_start == EXTENT_MAP_HOLE ||
1676                     (cur_offset >= inode->i_size &&
1677                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
1678
1679                         /*
1680                          * Make sure we have enough space before we do the
1681                          * allocation.
1682                          */
1683                         ret = btrfs_check_data_free_space(inode, last_byte -
1684                                                           cur_offset);
1685                         if (ret) {
1686                                 free_extent_map(em);
1687                                 break;
1688                         }
1689
1690                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
1691                                                         last_byte - cur_offset,
1692                                                         1 << inode->i_blkbits,
1693                                                         offset + len,
1694                                                         &alloc_hint);
1695
1696                         /* Let go of our reservation. */
1697                         btrfs_free_reserved_data_space(inode, last_byte -
1698                                                        cur_offset);
1699                         if (ret < 0) {
1700                                 free_extent_map(em);
1701                                 break;
1702                         }
1703                 } else if (actual_end > inode->i_size &&
1704                            !(mode & FALLOC_FL_KEEP_SIZE)) {
1705                         /*
1706                          * We didn't need to allocate any more space, but we
1707                          * still extended the size of the file so we need to
1708                          * update i_size.
1709                          */
1710                         inode->i_ctime = CURRENT_TIME;
1711                         i_size_write(inode, actual_end);
1712                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
1713                 }
1714                 free_extent_map(em);
1715
1716                 cur_offset = last_byte;
1717                 if (cur_offset >= alloc_end) {
1718                         ret = 0;
1719                         break;
1720                 }
1721         }
1722         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
1723                              &cached_state, GFP_NOFS);
1724 out:
1725         mutex_unlock(&inode->i_mutex);
1726         return ret;
1727 }
1728
1729 static int find_desired_extent(struct inode *inode, loff_t *offset, int origin)
1730 {
1731         struct btrfs_root *root = BTRFS_I(inode)->root;
1732         struct extent_map *em;
1733         struct extent_state *cached_state = NULL;
1734         u64 lockstart = *offset;
1735         u64 lockend = i_size_read(inode);
1736         u64 start = *offset;
1737         u64 orig_start = *offset;
1738         u64 len = i_size_read(inode);
1739         u64 last_end = 0;
1740         int ret = 0;
1741
1742         lockend = max_t(u64, root->sectorsize, lockend);
1743         if (lockend <= lockstart)
1744                 lockend = lockstart + root->sectorsize;
1745
1746         len = lockend - lockstart + 1;
1747
1748         len = max_t(u64, len, root->sectorsize);
1749         if (inode->i_size == 0)
1750                 return -ENXIO;
1751
1752         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend, 0,
1753                          &cached_state, GFP_NOFS);
1754
1755         /*
1756          * Delalloc is such a pain.  If we have a hole and we have pending
1757          * delalloc for a portion of the hole we will get back a hole that
1758          * exists for the entire range since it hasn't been actually written
1759          * yet.  So to take care of this case we need to look for an extent just
1760          * before the position we want in case there is outstanding delalloc
1761          * going on here.
1762          */
1763         if (origin == SEEK_HOLE && start != 0) {
1764                 if (start <= root->sectorsize)
1765                         em = btrfs_get_extent_fiemap(inode, NULL, 0, 0,
1766                                                      root->sectorsize, 0);
1767                 else
1768                         em = btrfs_get_extent_fiemap(inode, NULL, 0,
1769                                                      start - root->sectorsize,
1770                                                      root->sectorsize, 0);
1771                 if (IS_ERR(em)) {
1772                         ret = -ENXIO;
1773                         goto out;
1774                 }
1775                 last_end = em->start + em->len;
1776                 if (em->block_start == EXTENT_MAP_DELALLOC)
1777                         last_end = min_t(u64, last_end, inode->i_size);
1778                 free_extent_map(em);
1779         }
1780
1781         while (1) {
1782                 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
1783                 if (IS_ERR(em)) {
1784                         ret = -ENXIO;
1785                         break;
1786                 }
1787
1788                 if (em->block_start == EXTENT_MAP_HOLE) {
1789                         if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
1790                                 if (last_end <= orig_start) {
1791                                         free_extent_map(em);
1792                                         ret = -ENXIO;
1793                                         break;
1794                                 }
1795                         }
1796
1797                         if (origin == SEEK_HOLE) {
1798                                 *offset = start;
1799                                 free_extent_map(em);
1800                                 break;
1801                         }
1802                 } else {
1803                         if (origin == SEEK_DATA) {
1804                                 if (em->block_start == EXTENT_MAP_DELALLOC) {
1805                                         if (start >= inode->i_size) {
1806                                                 free_extent_map(em);
1807                                                 ret = -ENXIO;
1808                                                 break;
1809                                         }
1810                                 }
1811
1812                                 *offset = start;
1813                                 free_extent_map(em);
1814                                 break;
1815                         }
1816                 }
1817
1818                 start = em->start + em->len;
1819                 last_end = em->start + em->len;
1820
1821                 if (em->block_start == EXTENT_MAP_DELALLOC)
1822                         last_end = min_t(u64, last_end, inode->i_size);
1823
1824                 if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
1825                         free_extent_map(em);
1826                         ret = -ENXIO;
1827                         break;
1828                 }
1829                 free_extent_map(em);
1830                 cond_resched();
1831         }
1832         if (!ret)
1833                 *offset = min(*offset, inode->i_size);
1834 out:
1835         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
1836                              &cached_state, GFP_NOFS);
1837         return ret;
1838 }
1839
1840 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int origin)
1841 {
1842         struct inode *inode = file->f_mapping->host;
1843         int ret;
1844
1845         mutex_lock(&inode->i_mutex);
1846         switch (origin) {
1847         case SEEK_END:
1848         case SEEK_CUR:
1849                 offset = generic_file_llseek(file, offset, origin);
1850                 goto out;
1851         case SEEK_DATA:
1852         case SEEK_HOLE:
1853                 if (offset >= i_size_read(inode)) {
1854                         mutex_unlock(&inode->i_mutex);
1855                         return -ENXIO;
1856                 }
1857
1858                 ret = find_desired_extent(inode, &offset, origin);
1859                 if (ret) {
1860                         mutex_unlock(&inode->i_mutex);
1861                         return ret;
1862                 }
1863         }
1864
1865         if (offset < 0 && !(file->f_mode & FMODE_UNSIGNED_OFFSET)) {
1866                 offset = -EINVAL;
1867                 goto out;
1868         }
1869         if (offset > inode->i_sb->s_maxbytes) {
1870                 offset = -EINVAL;
1871                 goto out;
1872         }
1873
1874         /* Special lock needed here? */
1875         if (offset != file->f_pos) {
1876                 file->f_pos = offset;
1877                 file->f_version = 0;
1878         }
1879 out:
1880         mutex_unlock(&inode->i_mutex);
1881         return offset;
1882 }
1883
1884 const struct file_operations btrfs_file_operations = {
1885         .llseek         = btrfs_file_llseek,
1886         .read           = do_sync_read,
1887         .write          = do_sync_write,
1888         .aio_read       = generic_file_aio_read,
1889         .splice_read    = generic_file_splice_read,
1890         .aio_write      = btrfs_file_aio_write,
1891         .mmap           = btrfs_file_mmap,
1892         .open           = generic_file_open,
1893         .release        = btrfs_release_file,
1894         .fsync          = btrfs_sync_file,
1895         .fallocate      = btrfs_fallocate,
1896         .unlocked_ioctl = btrfs_ioctl,
1897 #ifdef CONFIG_COMPAT
1898         .compat_ioctl   = btrfs_ioctl,
1899 #endif
1900 };