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); /* -ENOMEM */
 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); /* Logic error */
 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); /* Logic error */
 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         int modify_tree = -1;
 571
 572         if (drop_cache)
 573                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 574
 575         path = btrfs_alloc_path();
 576         if (!path)
 577                 return -ENOMEM;
 578
 579         if (start >= BTRFS_I(inode)->disk_i_size)
 580                 modify_tree = 0;
 581
 582         while (1) {
 583                 recow = 0;
 584                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 585                                                search_start, modify_tree);
 586                 if (ret < 0)
 587                         break;
 588                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 589                         leaf = path->nodes[0];
 590                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 591                         if (key.objectid == ino &&
 592                             key.type == BTRFS_EXTENT_DATA_KEY)
 593                                 path->slots[0]--;
 594                 }
 595                 ret = 0;
 596 next_slot:
 597                 leaf = path->nodes[0];
 598                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 599                         BUG_ON(del_nr > 0);
 600                         ret = btrfs_next_leaf(root, path);
 601                         if (ret < 0)
 602                                 break;
 603                         if (ret > 0) {
 604                                 ret = 0;
 605                                 break;
 606                         }
 607                         leaf = path->nodes[0];
 608                         recow = 1;
 609                 }
 610
 611                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 612                 if (key.objectid > ino ||
 613                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 614                         break;
 615
 616                 fi = btrfs_item_ptr(leaf, path->slots[0],
 617                                     struct btrfs_file_extent_item);
 618                 extent_type = btrfs_file_extent_type(leaf, fi);
 619
 620                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 621                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 622                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 623                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 624                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 625                         extent_end = key.offset +
 626                                 btrfs_file_extent_num_bytes(leaf, fi);
 627                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 628                         extent_end = key.offset +
 629                                 btrfs_file_extent_inline_len(leaf, fi);
 630                 } else {
 631                         WARN_ON(1);
 632                         extent_end = search_start;
 633                 }
 634
 635                 if (extent_end <= search_start) {
 636                         path->slots[0]++;
 637                         goto next_slot;
 638                 }
 639
 640                 search_start = max(key.offset, start);
 641                 if (recow || !modify_tree) {
 642                         modify_tree = -1;
 643                         btrfs_release_path(path);
 644                         continue;
 645                 }
 646
 647                 /*
 648                  *     | - range to drop - |
 649                  *  | -------- extent -------- |
 650                  */
 651                 if (start > key.offset && end < extent_end) {
 652                         BUG_ON(del_nr > 0);
 653                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 654
 655                         memcpy(&new_key, &key, sizeof(new_key));
 656                         new_key.offset = start;
 657                         ret = btrfs_duplicate_item(trans, root, path,
 658                                                    &new_key);
 659                         if (ret == -EAGAIN) {
 660                                 btrfs_release_path(path);
 661                                 continue;
 662                         }
 663                         if (ret < 0)
 664                                 break;
 665
 666                         leaf = path->nodes[0];
 667                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 668                                             struct btrfs_file_extent_item);
 669                         btrfs_set_file_extent_num_bytes(leaf, fi,
 670                                                         start - key.offset);
 671
 672                         fi = btrfs_item_ptr(leaf, path->slots[0],
 673                                             struct btrfs_file_extent_item);
 674
 675                         extent_offset += start - key.offset;
 676                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 677                         btrfs_set_file_extent_num_bytes(leaf, fi,
 678                                                         extent_end - start);
 679                         btrfs_mark_buffer_dirty(leaf);
 680
 681                         if (disk_bytenr > 0) {
 682                                 ret = btrfs_inc_extent_ref(trans, root,
 683                                                 disk_bytenr, num_bytes, 0,
 684                                                 root->root_key.objectid,
 685                                                 new_key.objectid,
 686                                                 start - extent_offset, 0);
 687                                 BUG_ON(ret); /* -ENOMEM */
 688                                 *hint_byte = disk_bytenr;
 689                         }
 690                         key.offset = start;
 691                 }
 692                 /*
 693                  *  | ---- range to drop ----- |
 694                  *      | -------- extent -------- |
 695                  */
 696                 if (start <= key.offset && end < extent_end) {
 697                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 698
 699                         memcpy(&new_key, &key, sizeof(new_key));
 700                         new_key.offset = end;
 701                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 702
 703                         extent_offset += end - key.offset;
 704                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 705                         btrfs_set_file_extent_num_bytes(leaf, fi,
 706                                                         extent_end - end);
 707                         btrfs_mark_buffer_dirty(leaf);
 708                         if (disk_bytenr > 0) {
 709                                 inode_sub_bytes(inode, end - key.offset);
 710                                 *hint_byte = disk_bytenr;
 711                         }
 712                         break;
 713                 }
 714
 715                 search_start = extent_end;
 716                 /*
 717                  *       | ---- range to drop ----- |
 718                  *  | -------- extent -------- |
 719                  */
 720                 if (start > key.offset && end >= extent_end) {
 721                         BUG_ON(del_nr > 0);
 722                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 723
 724                         btrfs_set_file_extent_num_bytes(leaf, fi,
 725                                                         start - key.offset);
 726                         btrfs_mark_buffer_dirty(leaf);
 727                         if (disk_bytenr > 0) {
 728                                 inode_sub_bytes(inode, extent_end - start);
 729                                 *hint_byte = disk_bytenr;
 730                         }
 731                         if (end == extent_end)
 732                                 break;
 733
 734                         path->slots[0]++;
 735                         goto next_slot;
 736                 }
 737
 738                 /*
 739                  *  | ---- range to drop ----- |
 740                  *    | ------ extent ------ |
 741                  */
 742                 if (start <= key.offset && end >= extent_end) {
 743                         if (del_nr == 0) {
 744                                 del_slot = path->slots[0];
 745                                 del_nr = 1;
 746                         } else {
 747                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 748                                 del_nr++;
 749                         }
 750
 751                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 752                                 inode_sub_bytes(inode,
 753                                                 extent_end - key.offset);
 754                                 extent_end = ALIGN(extent_end,
 755                                                    root->sectorsize);
 756                         } else if (disk_bytenr > 0) {
 757                                 ret = btrfs_free_extent(trans, root,
 758                                                 disk_bytenr, num_bytes, 0,
 759                                                 root->root_key.objectid,
 760                                                 key.objectid, key.offset -
 761                                                 extent_offset, 0);
 762                                 BUG_ON(ret); /* -ENOMEM */
 763                                 inode_sub_bytes(inode,
 764                                                 extent_end - key.offset);
 765                                 *hint_byte = disk_bytenr;
 766                         }
 767
 768                         if (end == extent_end)
 769                                 break;
 770
 771                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 772                                 path->slots[0]++;
 773                                 goto next_slot;
 774                         }
 775
 776                         ret = btrfs_del_items(trans, root, path, del_slot,
 777                                               del_nr);
 778                         if (ret) {
 779                                 btrfs_abort_transaction(trans, root, ret);
 780                                 goto out;
 781                         }
 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 (!ret && del_nr > 0) {
 794                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 795                 if (ret)
 796                         btrfs_abort_transaction(trans, root, ret);
 797         }
 798
 799 out:
 800         btrfs_free_path(path);
 801         return ret;
 802 }
 803
 804 static int extent_mergeable(struct extent_buffer *leaf, int slot,
 805                             u64 objectid, u64 bytenr, u64 orig_offset,
 806                             u64 *start, u64 *end)
 807 {
 808         struct btrfs_file_extent_item *fi;
 809         struct btrfs_key key;
 810         u64 extent_end;
 811
 812         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
 813                 return 0;
 814
 815         btrfs_item_key_to_cpu(leaf, &key, slot);
 816         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
 817                 return 0;
 818
 819         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
 820         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
 821             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
 822             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
 823             btrfs_file_extent_compression(leaf, fi) ||
 824             btrfs_file_extent_encryption(leaf, fi) ||
 825             btrfs_file_extent_other_encoding(leaf, fi))
 826                 return 0;
 827
 828         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 829         if ((*start && *start != key.offset) || (*end && *end != extent_end))
 830                 return 0;
 831
 832         *start = key.offset;
 833         *end = extent_end;
 834         return 1;
 835 }
 836
 837 /*
 838  * Mark extent in the range start - end as written.
 839  *
 840  * This changes extent type from 'pre-allocated' to 'regular'. If only
 841  * part of extent is marked as written, the extent will be split into
 842  * two or three.
 843  */
 844 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
 845                               struct inode *inode, u64 start, u64 end)
 846 {
 847         struct btrfs_root *root = BTRFS_I(inode)->root;
 848         struct extent_buffer *leaf;
 849         struct btrfs_path *path;
 850         struct btrfs_file_extent_item *fi;
 851         struct btrfs_key key;
 852         struct btrfs_key new_key;
 853         u64 bytenr;
 854         u64 num_bytes;
 855         u64 extent_end;
 856         u64 orig_offset;
 857         u64 other_start;
 858         u64 other_end;
 859         u64 split;
 860         int del_nr = 0;
 861         int del_slot = 0;
 862         int recow;
 863         int ret;
 864         u64 ino = btrfs_ino(inode);
 865
 866         btrfs_drop_extent_cache(inode, start, end - 1, 0);
 867
 868         path = btrfs_alloc_path();
 869         if (!path)
 870                 return -ENOMEM;
 871 again:
 872         recow = 0;
 873         split = start;
 874         key.objectid = ino;
 875         key.type = BTRFS_EXTENT_DATA_KEY;
 876         key.offset = split;
 877
 878         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 879         if (ret < 0)
 880                 goto out;
 881         if (ret > 0 && path->slots[0] > 0)
 882                 path->slots[0]--;
 883
 884         leaf = path->nodes[0];
 885         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 886         BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
 887         fi = btrfs_item_ptr(leaf, path->slots[0],
 888                             struct btrfs_file_extent_item);
 889         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
 890                BTRFS_FILE_EXTENT_PREALLOC);
 891         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 892         BUG_ON(key.offset > start || extent_end < end);
 893
 894         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 895         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 896         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
 897         memcpy(&new_key, &key, sizeof(new_key));
 898
 899         if (start == key.offset && end < extent_end) {
 900                 other_start = 0;
 901                 other_end = start;
 902                 if (extent_mergeable(leaf, path->slots[0] - 1,
 903                                      ino, bytenr, orig_offset,
 904                                      &other_start, &other_end)) {
 905                         new_key.offset = end;
 906                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 907                         fi = btrfs_item_ptr(leaf, path->slots[0],
 908                                             struct btrfs_file_extent_item);
 909                         btrfs_set_file_extent_num_bytes(leaf, fi,
 910                                                         extent_end - end);
 911                         btrfs_set_file_extent_offset(leaf, fi,
 912                                                      end - orig_offset);
 913                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 914                                             struct btrfs_file_extent_item);
 915                         btrfs_set_file_extent_num_bytes(leaf, fi,
 916                                                         end - other_start);
 917                         btrfs_mark_buffer_dirty(leaf);
 918                         goto out;
 919                 }
 920         }
 921
 922         if (start > key.offset && end == extent_end) {
 923                 other_start = end;
 924                 other_end = 0;
 925                 if (extent_mergeable(leaf, path->slots[0] + 1,
 926                                      ino, bytenr, orig_offset,
 927                                      &other_start, &other_end)) {
 928                         fi = btrfs_item_ptr(leaf, path->slots[0],
 929                                             struct btrfs_file_extent_item);
 930                         btrfs_set_file_extent_num_bytes(leaf, fi,
 931                                                         start - key.offset);
 932                         path->slots[0]++;
 933                         new_key.offset = start;
 934                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 935
 936                         fi = btrfs_item_ptr(leaf, path->slots[0],
 937                                             struct btrfs_file_extent_item);
 938                         btrfs_set_file_extent_num_bytes(leaf, fi,
 939                                                         other_end - start);
 940                         btrfs_set_file_extent_offset(leaf, fi,
 941                                                      start - orig_offset);
 942                         btrfs_mark_buffer_dirty(leaf);
 943                         goto out;
 944                 }
 945         }
 946
 947         while (start > key.offset || end < extent_end) {
 948                 if (key.offset == start)
 949                         split = end;
 950
 951                 new_key.offset = split;
 952                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
 953                 if (ret == -EAGAIN) {
 954                         btrfs_release_path(path);
 955                         goto again;
 956                 }
 957                 if (ret < 0) {
 958                         btrfs_abort_transaction(trans, root, ret);
 959                         goto out;
 960                 }
 961
 962                 leaf = path->nodes[0];
 963                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 964                                     struct btrfs_file_extent_item);
 965                 btrfs_set_file_extent_num_bytes(leaf, fi,
 966                                                 split - key.offset);
 967
 968                 fi = btrfs_item_ptr(leaf, path->slots[0],
 969                                     struct btrfs_file_extent_item);
 970
 971                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
 972                 btrfs_set_file_extent_num_bytes(leaf, fi,
 973                                                 extent_end - split);
 974                 btrfs_mark_buffer_dirty(leaf);
 975
 976                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
 977                                            root->root_key.objectid,
 978                                            ino, orig_offset, 0);
 979                 BUG_ON(ret); /* -ENOMEM */
 980
 981                 if (split == start) {
 982                         key.offset = start;
 983                 } else {
 984                         BUG_ON(start != key.offset);
 985                         path->slots[0]--;
 986                         extent_end = end;
 987                 }
 988                 recow = 1;
 989         }
 990
 991         other_start = end;
 992         other_end = 0;
 993         if (extent_mergeable(leaf, path->slots[0] + 1,
 994                              ino, bytenr, orig_offset,
 995                              &other_start, &other_end)) {
 996                 if (recow) {
 997                         btrfs_release_path(path);
 998                         goto again;
 999                 }
1000                 extent_end = other_end;
1001                 del_slot = path->slots[0] + 1;
1002                 del_nr++;
1003                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1004                                         0, root->root_key.objectid,
1005                                         ino, orig_offset, 0);
1006                 BUG_ON(ret); /* -ENOMEM */
1007         }
1008         other_start = 0;
1009         other_end = start;
1010         if (extent_mergeable(leaf, path->slots[0] - 1,
1011                              ino, bytenr, orig_offset,
1012                              &other_start, &other_end)) {
1013                 if (recow) {
1014                         btrfs_release_path(path);
1015                         goto again;
1016                 }
1017                 key.offset = other_start;
1018                 del_slot = path->slots[0];
1019                 del_nr++;
1020                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1021                                         0, root->root_key.objectid,
1022                                         ino, orig_offset, 0);
1023                 BUG_ON(ret); /* -ENOMEM */
1024         }
1025         if (del_nr == 0) {
1026                 fi = btrfs_item_ptr(leaf, path->slots[0],
1027                            struct btrfs_file_extent_item);
1028                 btrfs_set_file_extent_type(leaf, fi,
1029                                            BTRFS_FILE_EXTENT_REG);
1030                 btrfs_mark_buffer_dirty(leaf);
1031         } else {
1032                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1033                            struct btrfs_file_extent_item);
1034                 btrfs_set_file_extent_type(leaf, fi,
1035                                            BTRFS_FILE_EXTENT_REG);
1036                 btrfs_set_file_extent_num_bytes(leaf, fi,
1037                                                 extent_end - key.offset);
1038                 btrfs_mark_buffer_dirty(leaf);
1039
1040                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1041                 if (ret < 0) {
1042                         btrfs_abort_transaction(trans, root, ret);
1043                         goto out;
1044                 }
1045         }
1046 out:
1047         btrfs_free_path(path);
1048         return 0;
1049 }
1050
1051 /*
1052  * on error we return an unlocked page and the error value
1053  * on success we return a locked page and 0
1054  */
1055 static int prepare_uptodate_page(struct page *page, u64 pos,
1056                                  bool force_uptodate)
1057 {
1058         int ret = 0;
1059
1060         if (((pos & (PAGE_CACHE_SIZE - 1)) || force_uptodate) &&
1061             !PageUptodate(page)) {
1062                 ret = btrfs_readpage(NULL, page);
1063                 if (ret)
1064                         return ret;
1065                 lock_page(page);
1066                 if (!PageUptodate(page)) {
1067                         unlock_page(page);
1068                         return -EIO;
1069                 }
1070         }
1071         return 0;
1072 }
1073
1074 /*
1075  * this gets pages into the page cache and locks them down, it also properly
1076  * waits for data=ordered extents to finish before allowing the pages to be
1077  * modified.
1078  */
1079 static noinline int prepare_pages(struct btrfs_root *root, struct file *file,
1080                          struct page **pages, size_t num_pages,
1081                          loff_t pos, unsigned long first_index,
1082                          size_t write_bytes, bool force_uptodate)
1083 {
1084         struct extent_state *cached_state = NULL;
1085         int i;
1086         unsigned long index = pos >> PAGE_CACHE_SHIFT;
1087         struct inode *inode = fdentry(file)->d_inode;
1088         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1089         int err = 0;
1090         int faili = 0;
1091         u64 start_pos;
1092         u64 last_pos;
1093
1094         start_pos = pos & ~((u64)root->sectorsize - 1);
1095         last_pos = ((u64)index + num_pages) << PAGE_CACHE_SHIFT;
1096
1097 again:
1098         for (i = 0; i < num_pages; i++) {
1099                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1100                                                mask | __GFP_WRITE);
1101                 if (!pages[i]) {
1102                         faili = i - 1;
1103                         err = -ENOMEM;
1104                         goto fail;
1105                 }
1106
1107                 if (i == 0)
1108                         err = prepare_uptodate_page(pages[i], pos,
1109                                                     force_uptodate);
1110                 if (i == num_pages - 1)
1111                         err = prepare_uptodate_page(pages[i],
1112                                                     pos + write_bytes, false);
1113                 if (err) {
1114                         page_cache_release(pages[i]);
1115                         faili = i - 1;
1116                         goto fail;
1117                 }
1118                 wait_on_page_writeback(pages[i]);
1119         }
1120         err = 0;
1121         if (start_pos < inode->i_size) {
1122                 struct btrfs_ordered_extent *ordered;
1123                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1124                                  start_pos, last_pos - 1, 0, &cached_state);
1125                 ordered = btrfs_lookup_first_ordered_extent(inode,
1126                                                             last_pos - 1);
1127                 if (ordered &&
1128                     ordered->file_offset + ordered->len > start_pos &&
1129                     ordered->file_offset < last_pos) {
1130                         btrfs_put_ordered_extent(ordered);
1131                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1132                                              start_pos, last_pos - 1,
1133                                              &cached_state, GFP_NOFS);
1134                         for (i = 0; i < num_pages; i++) {
1135                                 unlock_page(pages[i]);
1136                                 page_cache_release(pages[i]);
1137                         }
1138                         btrfs_wait_ordered_range(inode, start_pos,
1139                                                  last_pos - start_pos);
1140                         goto again;
1141                 }
1142                 if (ordered)
1143                         btrfs_put_ordered_extent(ordered);
1144
1145                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1146                                   last_pos - 1, EXTENT_DIRTY | EXTENT_DELALLOC |
1147                                   EXTENT_DO_ACCOUNTING, 0, 0, &cached_state,
1148                                   GFP_NOFS);
1149                 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1150                                      start_pos, last_pos - 1, &cached_state,
1151                                      GFP_NOFS);
1152         }
1153         for (i = 0; i < num_pages; i++) {
1154                 if (clear_page_dirty_for_io(pages[i]))
1155                         account_page_redirty(pages[i]);
1156                 set_page_extent_mapped(pages[i]);
1157                 WARN_ON(!PageLocked(pages[i]));
1158         }
1159         return 0;
1160 fail:
1161         while (faili >= 0) {
1162                 unlock_page(pages[faili]);
1163                 page_cache_release(pages[faili]);
1164                 faili--;
1165         }
1166         return err;
1167
1168 }
1169
1170 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1171                                                struct iov_iter *i,
1172                                                loff_t pos)
1173 {
1174         struct inode *inode = fdentry(file)->d_inode;
1175         struct btrfs_root *root = BTRFS_I(inode)->root;
1176         struct page **pages = NULL;
1177         unsigned long first_index;
1178         size_t num_written = 0;
1179         int nrptrs;
1180         int ret = 0;
1181         bool force_page_uptodate = false;
1182
1183         nrptrs = min((iov_iter_count(i) + PAGE_CACHE_SIZE - 1) /
1184                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1185                      (sizeof(struct page *)));
1186         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1187         nrptrs = max(nrptrs, 8);
1188         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1189         if (!pages)
1190                 return -ENOMEM;
1191
1192         first_index = pos >> PAGE_CACHE_SHIFT;
1193
1194         while (iov_iter_count(i) > 0) {
1195                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1196                 size_t write_bytes = min(iov_iter_count(i),
1197                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1198                                          offset);
1199                 size_t num_pages = (write_bytes + offset +
1200                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1201                 size_t dirty_pages;
1202                 size_t copied;
1203
1204                 WARN_ON(num_pages > nrptrs);
1205
1206                 /*
1207                  * Fault pages before locking them in prepare_pages
1208                  * to avoid recursive lock
1209                  */
1210                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1211                         ret = -EFAULT;
1212                         break;
1213                 }
1214
1215                 ret = btrfs_delalloc_reserve_space(inode,
1216                                         num_pages << PAGE_CACHE_SHIFT);
1217                 if (ret)
1218                         break;
1219
1220                 /*
1221                  * This is going to setup the pages array with the number of
1222                  * pages we want, so we don't really need to worry about the
1223                  * contents of pages from loop to loop
1224                  */
1225                 ret = prepare_pages(root, file, pages, num_pages,
1226                                     pos, first_index, write_bytes,
1227                                     force_page_uptodate);
1228                 if (ret) {
1229                         btrfs_delalloc_release_space(inode,
1230                                         num_pages << PAGE_CACHE_SHIFT);
1231                         break;
1232                 }
1233
1234                 copied = btrfs_copy_from_user(pos, num_pages,
1235                                            write_bytes, pages, i);
1236
1237                 /*
1238                  * if we have trouble faulting in the pages, fall
1239                  * back to one page at a time
1240                  */
1241                 if (copied < write_bytes)
1242                         nrptrs = 1;
1243
1244                 if (copied == 0) {
1245                         force_page_uptodate = true;
1246                         dirty_pages = 0;
1247                 } else {
1248                         force_page_uptodate = false;
1249                         dirty_pages = (copied + offset +
1250                                        PAGE_CACHE_SIZE - 1) >>
1251                                        PAGE_CACHE_SHIFT;
1252                 }
1253
1254                 /*
1255                  * If we had a short copy we need to release the excess delaloc
1256                  * bytes we reserved.  We need to increment outstanding_extents
1257                  * because btrfs_delalloc_release_space will decrement it, but
1258                  * we still have an outstanding extent for the chunk we actually
1259                  * managed to copy.
1260                  */
1261                 if (num_pages > dirty_pages) {
1262                         if (copied > 0) {
1263                                 spin_lock(&BTRFS_I(inode)->lock);
1264                                 BTRFS_I(inode)->outstanding_extents++;
1265                                 spin_unlock(&BTRFS_I(inode)->lock);
1266                         }
1267                         btrfs_delalloc_release_space(inode,
1268                                         (num_pages - dirty_pages) <<
1269                                         PAGE_CACHE_SHIFT);
1270                 }
1271
1272                 if (copied > 0) {
1273                         ret = btrfs_dirty_pages(root, inode, pages,
1274                                                 dirty_pages, pos, copied,
1275                                                 NULL);
1276                         if (ret) {
1277                                 btrfs_delalloc_release_space(inode,
1278                                         dirty_pages << PAGE_CACHE_SHIFT);
1279                                 btrfs_drop_pages(pages, num_pages);
1280                                 break;
1281                         }
1282                 }
1283
1284                 btrfs_drop_pages(pages, num_pages);
1285
1286                 cond_resched();
1287
1288                 balance_dirty_pages_ratelimited_nr(inode->i_mapping,
1289                                                    dirty_pages);
1290                 if (dirty_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1291                         btrfs_btree_balance_dirty(root, 1);
1292
1293                 pos += copied;
1294                 num_written += copied;
1295         }
1296
1297         kfree(pages);
1298
1299         return num_written ? num_written : ret;
1300 }
1301
1302 static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1303                                     const struct iovec *iov,
1304                                     unsigned long nr_segs, loff_t pos,
1305                                     loff_t *ppos, size_t count, size_t ocount)
1306 {
1307         struct file *file = iocb->ki_filp;
1308         struct inode *inode = fdentry(file)->d_inode;
1309         struct iov_iter i;
1310         ssize_t written;
1311         ssize_t written_buffered;
1312         loff_t endbyte;
1313         int err;
1314
1315         written = generic_file_direct_write(iocb, iov, &nr_segs, pos, ppos,
1316                                             count, ocount);
1317
1318         /*
1319          * the generic O_DIRECT will update in-memory i_size after the
1320          * DIOs are done.  But our endio handlers that update the on
1321          * disk i_size never update past the in memory i_size.  So we
1322          * need one more update here to catch any additions to the
1323          * file
1324          */
1325         if (inode->i_size != BTRFS_I(inode)->disk_i_size) {
1326                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
1327                 mark_inode_dirty(inode);
1328         }
1329
1330         if (written < 0 || written == count)
1331                 return written;
1332
1333         pos += written;
1334         count -= written;
1335         iov_iter_init(&i, iov, nr_segs, count, written);
1336         written_buffered = __btrfs_buffered_write(file, &i, pos);
1337         if (written_buffered < 0) {
1338                 err = written_buffered;
1339                 goto out;
1340         }
1341         endbyte = pos + written_buffered - 1;
1342         err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
1343         if (err)
1344                 goto out;
1345         written += written_buffered;
1346         *ppos = pos + written_buffered;
1347         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_CACHE_SHIFT,
1348                                  endbyte >> PAGE_CACHE_SHIFT);
1349 out:
1350         return written ? written : err;
1351 }
1352
1353 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
1354                                     const struct iovec *iov,
1355                                     unsigned long nr_segs, loff_t pos)
1356 {
1357         struct file *file = iocb->ki_filp;
1358         struct inode *inode = fdentry(file)->d_inode;
1359         struct btrfs_root *root = BTRFS_I(inode)->root;
1360         loff_t *ppos = &iocb->ki_pos;
1361         u64 start_pos;
1362         ssize_t num_written = 0;
1363         ssize_t err = 0;
1364         size_t count, ocount;
1365
1366         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1367
1368         mutex_lock(&inode->i_mutex);
1369
1370         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1371         if (err) {
1372                 mutex_unlock(&inode->i_mutex);
1373                 goto out;
1374         }
1375         count = ocount;
1376
1377         current->backing_dev_info = inode->i_mapping->backing_dev_info;
1378         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1379         if (err) {
1380                 mutex_unlock(&inode->i_mutex);
1381                 goto out;
1382         }
1383
1384         if (count == 0) {
1385                 mutex_unlock(&inode->i_mutex);
1386                 goto out;
1387         }
1388
1389         err = file_remove_suid(file);
1390         if (err) {
1391                 mutex_unlock(&inode->i_mutex);
1392                 goto out;
1393         }
1394
1395         /*
1396          * If BTRFS flips readonly due to some impossible error
1397          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1398          * although we have opened a file as writable, we have
1399          * to stop this write operation to ensure FS consistency.
1400          */
1401         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
1402                 mutex_unlock(&inode->i_mutex);
1403                 err = -EROFS;
1404                 goto out;
1405         }
1406
1407         err = btrfs_update_time(file);
1408         if (err) {
1409                 mutex_unlock(&inode->i_mutex);
1410                 goto out;
1411         }
1412         BTRFS_I(inode)->sequence++;
1413
1414         start_pos = round_down(pos, root->sectorsize);
1415         if (start_pos > i_size_read(inode)) {
1416                 err = btrfs_cont_expand(inode, i_size_read(inode), start_pos);
1417                 if (err) {
1418                         mutex_unlock(&inode->i_mutex);
1419                         goto out;
1420                 }
1421         }
1422
1423         if (unlikely(file->f_flags & O_DIRECT)) {
1424                 num_written = __btrfs_direct_write(iocb, iov, nr_segs,
1425                                                    pos, ppos, count, ocount);
1426         } else {
1427                 struct iov_iter i;
1428
1429                 iov_iter_init(&i, iov, nr_segs, count, num_written);
1430
1431                 num_written = __btrfs_buffered_write(file, &i, pos);
1432                 if (num_written > 0)
1433                         *ppos = pos + num_written;
1434         }
1435
1436         mutex_unlock(&inode->i_mutex);
1437
1438         /*
1439          * we want to make sure fsync finds this change
1440          * but we haven't joined a transaction running right now.
1441          *
1442          * Later on, someone is sure to update the inode and get the
1443          * real transid recorded.
1444          *
1445          * We set last_trans now to the fs_info generation + 1,
1446          * this will either be one more than the running transaction
1447          * or the generation used for the next transaction if there isn't
1448          * one running right now.
1449          */
1450         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1451         if (num_written > 0 || num_written == -EIOCBQUEUED) {
1452                 err = generic_write_sync(file, pos, num_written);
1453                 if (err < 0 && num_written > 0)
1454                         num_written = err;
1455         }
1456 out:
1457         current->backing_dev_info = NULL;
1458         return num_written ? num_written : err;
1459 }
1460
1461 int btrfs_release_file(struct inode *inode, struct file *filp)
1462 {
1463         /*
1464          * ordered_data_close is set by settattr when we are about to truncate
1465          * a file from a non-zero size to a zero size.  This tries to
1466          * flush down new bytes that may have been written if the
1467          * application were using truncate to replace a file in place.
1468          */
1469         if (BTRFS_I(inode)->ordered_data_close) {
1470                 BTRFS_I(inode)->ordered_data_close = 0;
1471                 btrfs_add_ordered_operation(NULL, BTRFS_I(inode)->root, inode);
1472                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1473                         filemap_flush(inode->i_mapping);
1474         }
1475         if (filp->private_data)
1476                 btrfs_ioctl_trans_end(filp);
1477         return 0;
1478 }
1479
1480 /*
1481  * fsync call for both files and directories.  This logs the inode into
1482  * the tree log instead of forcing full commits whenever possible.
1483  *
1484  * It needs to call filemap_fdatawait so that all ordered extent updates are
1485  * in the metadata btree are up to date for copying to the log.
1486  *
1487  * It drops the inode mutex before doing the tree log commit.  This is an
1488  * important optimization for directories because holding the mutex prevents
1489  * new operations on the dir while we write to disk.
1490  */
1491 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1492 {
1493         struct dentry *dentry = file->f_path.dentry;
1494         struct inode *inode = dentry->d_inode;
1495         struct btrfs_root *root = BTRFS_I(inode)->root;
1496         int ret = 0;
1497         struct btrfs_trans_handle *trans;
1498
1499         trace_btrfs_sync_file(file, datasync);
1500
1501         ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1502         if (ret)
1503                 return ret;
1504         mutex_lock(&inode->i_mutex);
1505
1506         /* we wait first, since the writeback may change the inode */
1507         root->log_batch++;
1508         btrfs_wait_ordered_range(inode, 0, (u64)-1);
1509         root->log_batch++;
1510
1511         /*
1512          * check the transaction that last modified this inode
1513          * and see if its already been committed
1514          */
1515         if (!BTRFS_I(inode)->last_trans) {
1516                 mutex_unlock(&inode->i_mutex);
1517                 goto out;
1518         }
1519
1520         /*
1521          * if the last transaction that changed this file was before
1522          * the current transaction, we can bail out now without any
1523          * syncing
1524          */
1525         smp_mb();
1526         if (BTRFS_I(inode)->last_trans <=
1527             root->fs_info->last_trans_committed) {
1528                 BTRFS_I(inode)->last_trans = 0;
1529                 mutex_unlock(&inode->i_mutex);
1530                 goto out;
1531         }
1532
1533         /*
1534          * ok we haven't committed the transaction yet, lets do a commit
1535          */
1536         if (file->private_data)
1537                 btrfs_ioctl_trans_end(file);
1538
1539         trans = btrfs_start_transaction(root, 0);
1540         if (IS_ERR(trans)) {
1541                 ret = PTR_ERR(trans);
1542                 mutex_unlock(&inode->i_mutex);
1543                 goto out;
1544         }
1545
1546         ret = btrfs_log_dentry_safe(trans, root, dentry);
1547         if (ret < 0) {
1548                 mutex_unlock(&inode->i_mutex);
1549                 goto out;
1550         }
1551
1552         /* we've logged all the items and now have a consistent
1553          * version of the file in the log.  It is possible that
1554          * someone will come in and modify the file, but that's
1555          * fine because the log is consistent on disk, and we
1556          * have references to all of the file's extents
1557          *
1558          * It is possible that someone will come in and log the
1559          * file again, but that will end up using the synchronization
1560          * inside btrfs_sync_log to keep things safe.
1561          */
1562         mutex_unlock(&inode->i_mutex);
1563
1564         if (ret != BTRFS_NO_LOG_SYNC) {
1565                 if (ret > 0) {
1566                         ret = btrfs_commit_transaction(trans, root);
1567                 } else {
1568                         ret = btrfs_sync_log(trans, root);
1569                         if (ret == 0)
1570                                 ret = btrfs_end_transaction(trans, root);
1571                         else
1572                                 ret = btrfs_commit_transaction(trans, root);
1573                 }
1574         } else {
1575                 ret = btrfs_end_transaction(trans, root);
1576         }
1577 out:
1578         return ret > 0 ? -EIO : ret;
1579 }
1580
1581 static const struct vm_operations_struct btrfs_file_vm_ops = {
1582         .fault          = filemap_fault,
1583         .page_mkwrite   = btrfs_page_mkwrite,
1584 };
1585
1586 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
1587 {
1588         struct address_space *mapping = filp->f_mapping;
1589
1590         if (!mapping->a_ops->readpage)
1591                 return -ENOEXEC;
1592
1593         file_accessed(filp);
1594         vma->vm_ops = &btrfs_file_vm_ops;
1595         vma->vm_flags |= VM_CAN_NONLINEAR;
1596
1597         return 0;
1598 }
1599
1600 static long btrfs_fallocate(struct file *file, int mode,
1601                             loff_t offset, loff_t len)
1602 {
1603         struct inode *inode = file->f_path.dentry->d_inode;
1604         struct extent_state *cached_state = NULL;
1605         u64 cur_offset;
1606         u64 last_byte;
1607         u64 alloc_start;
1608         u64 alloc_end;
1609         u64 alloc_hint = 0;
1610         u64 locked_end;
1611         u64 mask = BTRFS_I(inode)->root->sectorsize - 1;
1612         struct extent_map *em;
1613         int ret;
1614
1615         alloc_start = offset & ~mask;
1616         alloc_end =  (offset + len + mask) & ~mask;
1617
1618         /* We only support the FALLOC_FL_KEEP_SIZE mode */
1619         if (mode & ~FALLOC_FL_KEEP_SIZE)
1620                 return -EOPNOTSUPP;
1621
1622         /*
1623          * Make sure we have enough space before we do the
1624          * allocation.
1625          */
1626         ret = btrfs_check_data_free_space(inode, len);
1627         if (ret)
1628                 return ret;
1629
1630         /*
1631          * wait for ordered IO before we have any locks.  We'll loop again
1632          * below with the locks held.
1633          */
1634         btrfs_wait_ordered_range(inode, alloc_start, alloc_end - alloc_start);
1635
1636         mutex_lock(&inode->i_mutex);
1637         ret = inode_newsize_ok(inode, alloc_end);
1638         if (ret)
1639                 goto out;
1640
1641         if (alloc_start > inode->i_size) {
1642                 ret = btrfs_cont_expand(inode, i_size_read(inode),
1643                                         alloc_start);
1644                 if (ret)
1645                         goto out;
1646         }
1647
1648         locked_end = alloc_end - 1;
1649         while (1) {
1650                 struct btrfs_ordered_extent *ordered;
1651
1652                 /* the extent lock is ordered inside the running
1653                  * transaction
1654                  */
1655                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
1656                                  locked_end, 0, &cached_state);
1657                 ordered = btrfs_lookup_first_ordered_extent(inode,
1658                                                             alloc_end - 1);
1659                 if (ordered &&
1660                     ordered->file_offset + ordered->len > alloc_start &&
1661                     ordered->file_offset < alloc_end) {
1662                         btrfs_put_ordered_extent(ordered);
1663                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1664                                              alloc_start, locked_end,
1665                                              &cached_state, GFP_NOFS);
1666                         /*
1667                          * we can't wait on the range with the transaction
1668                          * running or with the extent lock held
1669                          */
1670                         btrfs_wait_ordered_range(inode, alloc_start,
1671                                                  alloc_end - alloc_start);
1672                 } else {
1673                         if (ordered)
1674                                 btrfs_put_ordered_extent(ordered);
1675                         break;
1676                 }
1677         }
1678
1679         cur_offset = alloc_start;
1680         while (1) {
1681                 u64 actual_end;
1682
1683                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
1684                                       alloc_end - cur_offset, 0);
1685                 if (IS_ERR_OR_NULL(em)) {
1686                         if (!em)
1687                                 ret = -ENOMEM;
1688                         else
1689                                 ret = PTR_ERR(em);
1690                         break;
1691                 }
1692                 last_byte = min(extent_map_end(em), alloc_end);
1693                 actual_end = min_t(u64, extent_map_end(em), offset + len);
1694                 last_byte = (last_byte + mask) & ~mask;
1695
1696                 if (em->block_start == EXTENT_MAP_HOLE ||
1697                     (cur_offset >= inode->i_size &&
1698                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
1699                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
1700                                                         last_byte - cur_offset,
1701                                                         1 << inode->i_blkbits,
1702                                                         offset + len,
1703                                                         &alloc_hint);
1704
1705                         if (ret < 0) {
1706                                 free_extent_map(em);
1707                                 break;
1708                         }
1709                 } else if (actual_end > inode->i_size &&
1710                            !(mode & FALLOC_FL_KEEP_SIZE)) {
1711                         /*
1712                          * We didn't need to allocate any more space, but we
1713                          * still extended the size of the file so we need to
1714                          * update i_size.
1715                          */
1716                         inode->i_ctime = CURRENT_TIME;
1717                         i_size_write(inode, actual_end);
1718                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
1719                 }
1720                 free_extent_map(em);
1721
1722                 cur_offset = last_byte;
1723                 if (cur_offset >= alloc_end) {
1724                         ret = 0;
1725                         break;
1726                 }
1727         }
1728         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
1729                              &cached_state, GFP_NOFS);
1730 out:
1731         mutex_unlock(&inode->i_mutex);
1732         /* Let go of our reservation. */
1733         btrfs_free_reserved_data_space(inode, len);
1734         return ret;
1735 }
1736
1737 static int find_desired_extent(struct inode *inode, loff_t *offset, int origin)
1738 {
1739         struct btrfs_root *root = BTRFS_I(inode)->root;
1740         struct extent_map *em;
1741         struct extent_state *cached_state = NULL;
1742         u64 lockstart = *offset;
1743         u64 lockend = i_size_read(inode);
1744         u64 start = *offset;
1745         u64 orig_start = *offset;
1746         u64 len = i_size_read(inode);
1747         u64 last_end = 0;
1748         int ret = 0;
1749
1750         lockend = max_t(u64, root->sectorsize, lockend);
1751         if (lockend <= lockstart)
1752                 lockend = lockstart + root->sectorsize;
1753
1754         len = lockend - lockstart + 1;
1755
1756         len = max_t(u64, len, root->sectorsize);
1757         if (inode->i_size == 0)
1758                 return -ENXIO;
1759
1760         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend, 0,
1761                          &cached_state);
1762
1763         /*
1764          * Delalloc is such a pain.  If we have a hole and we have pending
1765          * delalloc for a portion of the hole we will get back a hole that
1766          * exists for the entire range since it hasn't been actually written
1767          * yet.  So to take care of this case we need to look for an extent just
1768          * before the position we want in case there is outstanding delalloc
1769          * going on here.
1770          */
1771         if (origin == SEEK_HOLE && start != 0) {
1772                 if (start <= root->sectorsize)
1773                         em = btrfs_get_extent_fiemap(inode, NULL, 0, 0,
1774                                                      root->sectorsize, 0);
1775                 else
1776                         em = btrfs_get_extent_fiemap(inode, NULL, 0,
1777                                                      start - root->sectorsize,
1778                                                      root->sectorsize, 0);
1779                 if (IS_ERR(em)) {
1780                         ret = PTR_ERR(em);
1781                         goto out;
1782                 }
1783                 last_end = em->start + em->len;
1784                 if (em->block_start == EXTENT_MAP_DELALLOC)
1785                         last_end = min_t(u64, last_end, inode->i_size);
1786                 free_extent_map(em);
1787         }
1788
1789         while (1) {
1790                 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
1791                 if (IS_ERR(em)) {
1792                         ret = PTR_ERR(em);
1793                         break;
1794                 }
1795
1796                 if (em->block_start == EXTENT_MAP_HOLE) {
1797                         if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
1798                                 if (last_end <= orig_start) {
1799                                         free_extent_map(em);
1800                                         ret = -ENXIO;
1801                                         break;
1802                                 }
1803                         }
1804
1805                         if (origin == SEEK_HOLE) {
1806                                 *offset = start;
1807                                 free_extent_map(em);
1808                                 break;
1809                         }
1810                 } else {
1811                         if (origin == SEEK_DATA) {
1812                                 if (em->block_start == EXTENT_MAP_DELALLOC) {
1813                                         if (start >= inode->i_size) {
1814                                                 free_extent_map(em);
1815                                                 ret = -ENXIO;
1816                                                 break;
1817                                         }
1818                                 }
1819
1820                                 *offset = start;
1821                                 free_extent_map(em);
1822                                 break;
1823                         }
1824                 }
1825
1826                 start = em->start + em->len;
1827                 last_end = em->start + em->len;
1828
1829                 if (em->block_start == EXTENT_MAP_DELALLOC)
1830                         last_end = min_t(u64, last_end, inode->i_size);
1831
1832                 if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
1833                         free_extent_map(em);
1834                         ret = -ENXIO;
1835                         break;
1836                 }
1837                 free_extent_map(em);
1838                 cond_resched();
1839         }
1840         if (!ret)
1841                 *offset = min(*offset, inode->i_size);
1842 out:
1843         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
1844                              &cached_state, GFP_NOFS);
1845         return ret;
1846 }
1847
1848 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int origin)
1849 {
1850         struct inode *inode = file->f_mapping->host;
1851         int ret;
1852
1853         mutex_lock(&inode->i_mutex);
1854         switch (origin) {
1855         case SEEK_END:
1856         case SEEK_CUR:
1857                 offset = generic_file_llseek(file, offset, origin);
1858                 goto out;
1859         case SEEK_DATA:
1860         case SEEK_HOLE:
1861                 if (offset >= i_size_read(inode)) {
1862                         mutex_unlock(&inode->i_mutex);
1863                         return -ENXIO;
1864                 }
1865
1866                 ret = find_desired_extent(inode, &offset, origin);
1867                 if (ret) {
1868                         mutex_unlock(&inode->i_mutex);
1869                         return ret;
1870                 }
1871         }
1872
1873         if (offset < 0 && !(file->f_mode & FMODE_UNSIGNED_OFFSET)) {
1874                 offset = -EINVAL;
1875                 goto out;
1876         }
1877         if (offset > inode->i_sb->s_maxbytes) {
1878                 offset = -EINVAL;
1879                 goto out;
1880         }
1881
1882         /* Special lock needed here? */
1883         if (offset != file->f_pos) {
1884                 file->f_pos = offset;
1885                 file->f_version = 0;
1886         }
1887 out:
1888         mutex_unlock(&inode->i_mutex);
1889         return offset;
1890 }
1891
1892 const struct file_operations btrfs_file_operations = {
1893         .llseek         = btrfs_file_llseek,
1894         .read           = do_sync_read,
1895         .write          = do_sync_write,
1896         .aio_read       = generic_file_aio_read,
1897         .splice_read    = generic_file_splice_read,
1898         .aio_write      = btrfs_file_aio_write,
1899         .mmap           = btrfs_file_mmap,
1900         .open           = generic_file_open,
1901         .release        = btrfs_release_file,
1902         .fsync          = btrfs_sync_file,
1903         .fallocate      = btrfs_fallocate,
1904         .unlocked_ioctl = btrfs_ioctl,
1905 #ifdef CONFIG_COMPAT
1906         .compat_ioctl   = btrfs_ioctl,
1907 #endif
1908 };