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/swap.h>
  28 #include <linux/writeback.h>
  29 #include <linux/statfs.h>
  30 #include <linux/compat.h>
  31 #include <linux/slab.h>
  32 #include "ctree.h"
  33 #include "disk-io.h"
  34 #include "transaction.h"
  35 #include "btrfs_inode.h"
  36 #include "ioctl.h"
  37 #include "print-tree.h"
  38 #include "tree-log.h"
  39 #include "locking.h"
  40 #include "compat.h"
  41
  42
  43 /* simple helper to fault in pages and copy.  This should go away
  44  * and be replaced with calls into generic code.
  45  */
  46 static noinline int btrfs_copy_from_user(loff_t pos, int num_pages,
  47                                          int write_bytes,
  48                                          struct page **prepared_pages,
  49                                          struct iov_iter *i)
  50 {
  51         size_t copied = 0;
  52         int pg = 0;
  53         int offset = pos & (PAGE_CACHE_SIZE - 1);
  54         int total_copied = 0;
  55
  56         while (write_bytes > 0) {
  57                 size_t count = min_t(size_t,
  58                                      PAGE_CACHE_SIZE - offset, write_bytes);
  59                 struct page *page = prepared_pages[pg];
  60                 /*
  61                  * Copy data from userspace to the current page
  62                  *
  63                  * Disable pagefault to avoid recursive lock since
  64                  * the pages are already locked
  65                  */
  66                 pagefault_disable();
  67                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
  68                 pagefault_enable();
  69
  70                 /* Flush processor's dcache for this page */
  71                 flush_dcache_page(page);
  72                 iov_iter_advance(i, copied);
  73                 write_bytes -= copied;
  74                 total_copied += copied;
  75
  76                 /* Return to btrfs_file_aio_write to fault page */
  77                 if (unlikely(copied == 0)) {
  78                         break;
  79                 }
  80
  81                 if (unlikely(copied < PAGE_CACHE_SIZE - offset)) {
  82                         offset += copied;
  83                 } else {
  84                         pg++;
  85                         offset = 0;
  86                 }
  87         }
  88         return total_copied;
  89 }
  90
  91 /*
  92  * unlocks pages after btrfs_file_write is done with them
  93  */
  94 static noinline void btrfs_drop_pages(struct page **pages, size_t num_pages)
  95 {
  96         size_t i;
  97         for (i = 0; i < num_pages; i++) {
  98                 if (!pages[i])
  99                         break;
 100                 /* page checked is some magic around finding pages that
 101                  * have been modified without going through btrfs_set_page_dirty
 102                  * clear it here
 103                  */
 104                 ClearPageChecked(pages[i]);
 105                 unlock_page(pages[i]);
 106                 mark_page_accessed(pages[i]);
 107                 page_cache_release(pages[i]);
 108         }
 109 }
 110
 111 /*
 112  * after copy_from_user, pages need to be dirtied and we need to make
 113  * sure holes are created between the current EOF and the start of
 114  * any next extents (if required).
 115  *
 116  * this also makes the decision about creating an inline extent vs
 117  * doing real data extents, marking pages dirty and delalloc as required.
 118  */
 119 static noinline int dirty_and_release_pages(struct btrfs_trans_handle *trans,
 120                                    struct btrfs_root *root,
 121                                    struct file *file,
 122                                    struct page **pages,
 123                                    size_t num_pages,
 124                                    loff_t pos,
 125                                    size_t write_bytes)
 126 {
 127         int err = 0;
 128         int i;
 129         struct inode *inode = fdentry(file)->d_inode;
 130         u64 num_bytes;
 131         u64 start_pos;
 132         u64 end_of_last_block;
 133         u64 end_pos = pos + write_bytes;
 134         loff_t isize = i_size_read(inode);
 135
 136         start_pos = pos & ~((u64)root->sectorsize - 1);
 137         num_bytes = (write_bytes + pos - start_pos +
 138                     root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
 139
 140         end_of_last_block = start_pos + num_bytes - 1;
 141         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 142                                         NULL);
 143         BUG_ON(err);
 144
 145         for (i = 0; i < num_pages; i++) {
 146                 struct page *p = pages[i];
 147                 SetPageUptodate(p);
 148                 ClearPageChecked(p);
 149                 set_page_dirty(p);
 150         }
 151         if (end_pos > isize) {
 152                 i_size_write(inode, end_pos);
 153                 /* we've only changed i_size in ram, and we haven't updated
 154                  * the disk i_size.  There is no need to log the inode
 155                  * at this time.
 156                  */
 157         }
 158         return 0;
 159 }
 160
 161 /*
 162  * this drops all the extents in the cache that intersect the range
 163  * [start, end].  Existing extents are split as required.
 164  */
 165 int btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
 166                             int skip_pinned)
 167 {
 168         struct extent_map *em;
 169         struct extent_map *split = NULL;
 170         struct extent_map *split2 = NULL;
 171         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 172         u64 len = end - start + 1;
 173         int ret;
 174         int testend = 1;
 175         unsigned long flags;
 176         int compressed = 0;
 177
 178         WARN_ON(end < start);
 179         if (end == (u64)-1) {
 180                 len = (u64)-1;
 181                 testend = 0;
 182         }
 183         while (1) {
 184                 if (!split)
 185                         split = alloc_extent_map(GFP_NOFS);
 186                 if (!split2)
 187                         split2 = alloc_extent_map(GFP_NOFS);
 188
 189                 write_lock(&em_tree->lock);
 190                 em = lookup_extent_mapping(em_tree, start, len);
 191                 if (!em) {
 192                         write_unlock(&em_tree->lock);
 193                         break;
 194                 }
 195                 flags = em->flags;
 196                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 197                         if (testend && em->start + em->len >= start + len) {
 198                                 free_extent_map(em);
 199                                 write_unlock(&em_tree->lock);
 200                                 break;
 201                         }
 202                         start = em->start + em->len;
 203                         if (testend)
 204                                 len = start + len - (em->start + em->len);
 205                         free_extent_map(em);
 206                         write_unlock(&em_tree->lock);
 207                         continue;
 208                 }
 209                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 210                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 211                 remove_extent_mapping(em_tree, em);
 212
 213                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
 214                     em->start < start) {
 215                         split->start = em->start;
 216                         split->len = start - em->start;
 217                         split->orig_start = em->orig_start;
 218                         split->block_start = em->block_start;
 219
 220                         if (compressed)
 221                                 split->block_len = em->block_len;
 222                         else
 223                                 split->block_len = split->len;
 224
 225                         split->bdev = em->bdev;
 226                         split->flags = flags;
 227                         split->compress_type = em->compress_type;
 228                         ret = add_extent_mapping(em_tree, split);
 229                         BUG_ON(ret);
 230                         free_extent_map(split);
 231                         split = split2;
 232                         split2 = NULL;
 233                 }
 234                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
 235                     testend && em->start + em->len > start + len) {
 236                         u64 diff = start + len - em->start;
 237
 238                         split->start = start + len;
 239                         split->len = em->start + em->len - (start + len);
 240                         split->bdev = em->bdev;
 241                         split->flags = flags;
 242                         split->compress_type = em->compress_type;
 243
 244                         if (compressed) {
 245                                 split->block_len = em->block_len;
 246                                 split->block_start = em->block_start;
 247                                 split->orig_start = em->orig_start;
 248                         } else {
 249                                 split->block_len = split->len;
 250                                 split->block_start = em->block_start + diff;
 251                                 split->orig_start = split->start;
 252                         }
 253
 254                         ret = add_extent_mapping(em_tree, split);
 255                         BUG_ON(ret);
 256                         free_extent_map(split);
 257                         split = NULL;
 258                 }
 259                 write_unlock(&em_tree->lock);
 260
 261                 /* once for us */
 262                 free_extent_map(em);
 263                 /* once for the tree*/
 264                 free_extent_map(em);
 265         }
 266         if (split)
 267                 free_extent_map(split);
 268         if (split2)
 269                 free_extent_map(split2);
 270         return 0;
 271 }
 272
 273 /*
 274  * this is very complex, but the basic idea is to drop all extents
 275  * in the range start - end.  hint_block is filled in with a block number
 276  * that would be a good hint to the block allocator for this file.
 277  *
 278  * If an extent intersects the range but is not entirely inside the range
 279  * it is either truncated or split.  Anything entirely inside the range
 280  * is deleted from the tree.
 281  */
 282 int btrfs_drop_extents(struct btrfs_trans_handle *trans, struct inode *inode,
 283                        u64 start, u64 end, u64 *hint_byte, int drop_cache)
 284 {
 285         struct btrfs_root *root = BTRFS_I(inode)->root;
 286         struct extent_buffer *leaf;
 287         struct btrfs_file_extent_item *fi;
 288         struct btrfs_path *path;
 289         struct btrfs_key key;
 290         struct btrfs_key new_key;
 291         u64 search_start = start;
 292         u64 disk_bytenr = 0;
 293         u64 num_bytes = 0;
 294         u64 extent_offset = 0;
 295         u64 extent_end = 0;
 296         int del_nr = 0;
 297         int del_slot = 0;
 298         int extent_type;
 299         int recow;
 300         int ret;
 301
 302         if (drop_cache)
 303                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 304
 305         path = btrfs_alloc_path();
 306         if (!path)
 307                 return -ENOMEM;
 308
 309         while (1) {
 310                 recow = 0;
 311                 ret = btrfs_lookup_file_extent(trans, root, path, inode->i_ino,
 312                                                search_start, -1);
 313                 if (ret < 0)
 314                         break;
 315                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 316                         leaf = path->nodes[0];
 317                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 318                         if (key.objectid == inode->i_ino &&
 319                             key.type == BTRFS_EXTENT_DATA_KEY)
 320                                 path->slots[0]--;
 321                 }
 322                 ret = 0;
 323 next_slot:
 324                 leaf = path->nodes[0];
 325                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 326                         BUG_ON(del_nr > 0);
 327                         ret = btrfs_next_leaf(root, path);
 328                         if (ret < 0)
 329                                 break;
 330                         if (ret > 0) {
 331                                 ret = 0;
 332                                 break;
 333                         }
 334                         leaf = path->nodes[0];
 335                         recow = 1;
 336                 }
 337
 338                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 339                 if (key.objectid > inode->i_ino ||
 340                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 341                         break;
 342
 343                 fi = btrfs_item_ptr(leaf, path->slots[0],
 344                                     struct btrfs_file_extent_item);
 345                 extent_type = btrfs_file_extent_type(leaf, fi);
 346
 347                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 348                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 349                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 350                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 351                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 352                         extent_end = key.offset +
 353                                 btrfs_file_extent_num_bytes(leaf, fi);
 354                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 355                         extent_end = key.offset +
 356                                 btrfs_file_extent_inline_len(leaf, fi);
 357                 } else {
 358                         WARN_ON(1);
 359                         extent_end = search_start;
 360                 }
 361
 362                 if (extent_end <= search_start) {
 363                         path->slots[0]++;
 364                         goto next_slot;
 365                 }
 366
 367                 search_start = max(key.offset, start);
 368                 if (recow) {
 369                         btrfs_release_path(root, path);
 370                         continue;
 371                 }
 372
 373                 /*
 374                  *     | - range to drop - |
 375                  *  | -------- extent -------- |
 376                  */
 377                 if (start > key.offset && end < extent_end) {
 378                         BUG_ON(del_nr > 0);
 379                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 380
 381                         memcpy(&new_key, &key, sizeof(new_key));
 382                         new_key.offset = start;
 383                         ret = btrfs_duplicate_item(trans, root, path,
 384                                                    &new_key);
 385                         if (ret == -EAGAIN) {
 386                                 btrfs_release_path(root, path);
 387                                 continue;
 388                         }
 389                         if (ret < 0)
 390                                 break;
 391
 392                         leaf = path->nodes[0];
 393                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 394                                             struct btrfs_file_extent_item);
 395                         btrfs_set_file_extent_num_bytes(leaf, fi,
 396                                                         start - key.offset);
 397
 398                         fi = btrfs_item_ptr(leaf, path->slots[0],
 399                                             struct btrfs_file_extent_item);
 400
 401                         extent_offset += start - key.offset;
 402                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 403                         btrfs_set_file_extent_num_bytes(leaf, fi,
 404                                                         extent_end - start);
 405                         btrfs_mark_buffer_dirty(leaf);
 406
 407                         if (disk_bytenr > 0) {
 408                                 ret = btrfs_inc_extent_ref(trans, root,
 409                                                 disk_bytenr, num_bytes, 0,
 410                                                 root->root_key.objectid,
 411                                                 new_key.objectid,
 412                                                 start - extent_offset);
 413                                 BUG_ON(ret);
 414                                 *hint_byte = disk_bytenr;
 415                         }
 416                         key.offset = start;
 417                 }
 418                 /*
 419                  *  | ---- range to drop ----- |
 420                  *      | -------- extent -------- |
 421                  */
 422                 if (start <= key.offset && end < extent_end) {
 423                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 424
 425                         memcpy(&new_key, &key, sizeof(new_key));
 426                         new_key.offset = end;
 427                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 428
 429                         extent_offset += end - key.offset;
 430                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 431                         btrfs_set_file_extent_num_bytes(leaf, fi,
 432                                                         extent_end - end);
 433                         btrfs_mark_buffer_dirty(leaf);
 434                         if (disk_bytenr > 0) {
 435                                 inode_sub_bytes(inode, end - key.offset);
 436                                 *hint_byte = disk_bytenr;
 437                         }
 438                         break;
 439                 }
 440
 441                 search_start = extent_end;
 442                 /*
 443                  *       | ---- range to drop ----- |
 444                  *  | -------- extent -------- |
 445                  */
 446                 if (start > key.offset && end >= extent_end) {
 447                         BUG_ON(del_nr > 0);
 448                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 449
 450                         btrfs_set_file_extent_num_bytes(leaf, fi,
 451                                                         start - key.offset);
 452                         btrfs_mark_buffer_dirty(leaf);
 453                         if (disk_bytenr > 0) {
 454                                 inode_sub_bytes(inode, extent_end - start);
 455                                 *hint_byte = disk_bytenr;
 456                         }
 457                         if (end == extent_end)
 458                                 break;
 459
 460                         path->slots[0]++;
 461                         goto next_slot;
 462                 }
 463
 464                 /*
 465                  *  | ---- range to drop ----- |
 466                  *    | ------ extent ------ |
 467                  */
 468                 if (start <= key.offset && end >= extent_end) {
 469                         if (del_nr == 0) {
 470                                 del_slot = path->slots[0];
 471                                 del_nr = 1;
 472                         } else {
 473                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 474                                 del_nr++;
 475                         }
 476
 477                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 478                                 inode_sub_bytes(inode,
 479                                                 extent_end - key.offset);
 480                                 extent_end = ALIGN(extent_end,
 481                                                    root->sectorsize);
 482                         } else if (disk_bytenr > 0) {
 483                                 ret = btrfs_free_extent(trans, root,
 484                                                 disk_bytenr, num_bytes, 0,
 485                                                 root->root_key.objectid,
 486                                                 key.objectid, key.offset -
 487                                                 extent_offset);
 488                                 BUG_ON(ret);
 489                                 inode_sub_bytes(inode,
 490                                                 extent_end - key.offset);
 491                                 *hint_byte = disk_bytenr;
 492                         }
 493
 494                         if (end == extent_end)
 495                                 break;
 496
 497                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 498                                 path->slots[0]++;
 499                                 goto next_slot;
 500                         }
 501
 502                         ret = btrfs_del_items(trans, root, path, del_slot,
 503                                               del_nr);
 504                         BUG_ON(ret);
 505
 506                         del_nr = 0;
 507                         del_slot = 0;
 508
 509                         btrfs_release_path(root, path);
 510                         continue;
 511                 }
 512
 513                 BUG_ON(1);
 514         }
 515
 516         if (del_nr > 0) {
 517                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 518                 BUG_ON(ret);
 519         }
 520
 521         btrfs_free_path(path);
 522         return ret;
 523 }
 524
 525 static int extent_mergeable(struct extent_buffer *leaf, int slot,
 526                             u64 objectid, u64 bytenr, u64 orig_offset,
 527                             u64 *start, u64 *end)
 528 {
 529         struct btrfs_file_extent_item *fi;
 530         struct btrfs_key key;
 531         u64 extent_end;
 532
 533         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
 534                 return 0;
 535
 536         btrfs_item_key_to_cpu(leaf, &key, slot);
 537         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
 538                 return 0;
 539
 540         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
 541         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
 542             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
 543             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
 544             btrfs_file_extent_compression(leaf, fi) ||
 545             btrfs_file_extent_encryption(leaf, fi) ||
 546             btrfs_file_extent_other_encoding(leaf, fi))
 547                 return 0;
 548
 549         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 550         if ((*start && *start != key.offset) || (*end && *end != extent_end))
 551                 return 0;
 552
 553         *start = key.offset;
 554         *end = extent_end;
 555         return 1;
 556 }
 557
 558 /*
 559  * Mark extent in the range start - end as written.
 560  *
 561  * This changes extent type from 'pre-allocated' to 'regular'. If only
 562  * part of extent is marked as written, the extent will be split into
 563  * two or three.
 564  */
 565 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
 566                               struct inode *inode, u64 start, u64 end)
 567 {
 568         struct btrfs_root *root = BTRFS_I(inode)->root;
 569         struct extent_buffer *leaf;
 570         struct btrfs_path *path;
 571         struct btrfs_file_extent_item *fi;
 572         struct btrfs_key key;
 573         struct btrfs_key new_key;
 574         u64 bytenr;
 575         u64 num_bytes;
 576         u64 extent_end;
 577         u64 orig_offset;
 578         u64 other_start;
 579         u64 other_end;
 580         u64 split;
 581         int del_nr = 0;
 582         int del_slot = 0;
 583         int recow;
 584         int ret;
 585
 586         btrfs_drop_extent_cache(inode, start, end - 1, 0);
 587
 588         path = btrfs_alloc_path();
 589         BUG_ON(!path);
 590 again:
 591         recow = 0;
 592         split = start;
 593         key.objectid = inode->i_ino;
 594         key.type = BTRFS_EXTENT_DATA_KEY;
 595         key.offset = split;
 596
 597         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 598         if (ret > 0 && path->slots[0] > 0)
 599                 path->slots[0]--;
 600
 601         leaf = path->nodes[0];
 602         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 603         BUG_ON(key.objectid != inode->i_ino ||
 604                key.type != BTRFS_EXTENT_DATA_KEY);
 605         fi = btrfs_item_ptr(leaf, path->slots[0],
 606                             struct btrfs_file_extent_item);
 607         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
 608                BTRFS_FILE_EXTENT_PREALLOC);
 609         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 610         BUG_ON(key.offset > start || extent_end < end);
 611
 612         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 613         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 614         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
 615         memcpy(&new_key, &key, sizeof(new_key));
 616
 617         if (start == key.offset && end < extent_end) {
 618                 other_start = 0;
 619                 other_end = start;
 620                 if (extent_mergeable(leaf, path->slots[0] - 1,
 621                                      inode->i_ino, bytenr, orig_offset,
 622                                      &other_start, &other_end)) {
 623                         new_key.offset = end;
 624                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 625                         fi = btrfs_item_ptr(leaf, path->slots[0],
 626                                             struct btrfs_file_extent_item);
 627                         btrfs_set_file_extent_num_bytes(leaf, fi,
 628                                                         extent_end - end);
 629                         btrfs_set_file_extent_offset(leaf, fi,
 630                                                      end - orig_offset);
 631                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 632                                             struct btrfs_file_extent_item);
 633                         btrfs_set_file_extent_num_bytes(leaf, fi,
 634                                                         end - other_start);
 635                         btrfs_mark_buffer_dirty(leaf);
 636                         goto out;
 637                 }
 638         }
 639
 640         if (start > key.offset && end == extent_end) {
 641                 other_start = end;
 642                 other_end = 0;
 643                 if (extent_mergeable(leaf, path->slots[0] + 1,
 644                                      inode->i_ino, bytenr, orig_offset,
 645                                      &other_start, &other_end)) {
 646                         fi = btrfs_item_ptr(leaf, path->slots[0],
 647                                             struct btrfs_file_extent_item);
 648                         btrfs_set_file_extent_num_bytes(leaf, fi,
 649                                                         start - key.offset);
 650                         path->slots[0]++;
 651                         new_key.offset = start;
 652                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 653
 654                         fi = btrfs_item_ptr(leaf, path->slots[0],
 655                                             struct btrfs_file_extent_item);
 656                         btrfs_set_file_extent_num_bytes(leaf, fi,
 657                                                         other_end - start);
 658                         btrfs_set_file_extent_offset(leaf, fi,
 659                                                      start - orig_offset);
 660                         btrfs_mark_buffer_dirty(leaf);
 661                         goto out;
 662                 }
 663         }
 664
 665         while (start > key.offset || end < extent_end) {
 666                 if (key.offset == start)
 667                         split = end;
 668
 669                 new_key.offset = split;
 670                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
 671                 if (ret == -EAGAIN) {
 672                         btrfs_release_path(root, path);
 673                         goto again;
 674                 }
 675                 BUG_ON(ret < 0);
 676
 677                 leaf = path->nodes[0];
 678                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 679                                     struct btrfs_file_extent_item);
 680                 btrfs_set_file_extent_num_bytes(leaf, fi,
 681                                                 split - key.offset);
 682
 683                 fi = btrfs_item_ptr(leaf, path->slots[0],
 684                                     struct btrfs_file_extent_item);
 685
 686                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
 687                 btrfs_set_file_extent_num_bytes(leaf, fi,
 688                                                 extent_end - split);
 689                 btrfs_mark_buffer_dirty(leaf);
 690
 691                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
 692                                            root->root_key.objectid,
 693                                            inode->i_ino, orig_offset);
 694                 BUG_ON(ret);
 695
 696                 if (split == start) {
 697                         key.offset = start;
 698                 } else {
 699                         BUG_ON(start != key.offset);
 700                         path->slots[0]--;
 701                         extent_end = end;
 702                 }
 703                 recow = 1;
 704         }
 705
 706         other_start = end;
 707         other_end = 0;
 708         if (extent_mergeable(leaf, path->slots[0] + 1,
 709                              inode->i_ino, bytenr, orig_offset,
 710                              &other_start, &other_end)) {
 711                 if (recow) {
 712                         btrfs_release_path(root, path);
 713                         goto again;
 714                 }
 715                 extent_end = other_end;
 716                 del_slot = path->slots[0] + 1;
 717                 del_nr++;
 718                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
 719                                         0, root->root_key.objectid,
 720                                         inode->i_ino, orig_offset);
 721                 BUG_ON(ret);
 722         }
 723         other_start = 0;
 724         other_end = start;
 725         if (extent_mergeable(leaf, path->slots[0] - 1,
 726                              inode->i_ino, bytenr, orig_offset,
 727                              &other_start, &other_end)) {
 728                 if (recow) {
 729                         btrfs_release_path(root, path);
 730                         goto again;
 731                 }
 732                 key.offset = other_start;
 733                 del_slot = path->slots[0];
 734                 del_nr++;
 735                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
 736                                         0, root->root_key.objectid,
 737                                         inode->i_ino, orig_offset);
 738                 BUG_ON(ret);
 739         }
 740         if (del_nr == 0) {
 741                 fi = btrfs_item_ptr(leaf, path->slots[0],
 742                            struct btrfs_file_extent_item);
 743                 btrfs_set_file_extent_type(leaf, fi,
 744                                            BTRFS_FILE_EXTENT_REG);
 745                 btrfs_mark_buffer_dirty(leaf);
 746         } else {
 747                 fi = btrfs_item_ptr(leaf, del_slot - 1,
 748                            struct btrfs_file_extent_item);
 749                 btrfs_set_file_extent_type(leaf, fi,
 750                                            BTRFS_FILE_EXTENT_REG);
 751                 btrfs_set_file_extent_num_bytes(leaf, fi,
 752                                                 extent_end - key.offset);
 753                 btrfs_mark_buffer_dirty(leaf);
 754
 755                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 756                 BUG_ON(ret);
 757         }
 758 out:
 759         btrfs_free_path(path);
 760         return 0;
 761 }
 762
 763 /*
 764  * this gets pages into the page cache and locks them down, it also properly
 765  * waits for data=ordered extents to finish before allowing the pages to be
 766  * modified.
 767  */
 768 static noinline int prepare_pages(struct btrfs_root *root, struct file *file,
 769                          struct page **pages, size_t num_pages,
 770                          loff_t pos, unsigned long first_index,
 771                          unsigned long last_index, size_t write_bytes)
 772 {
 773         struct extent_state *cached_state = NULL;
 774         int i;
 775         unsigned long index = pos >> PAGE_CACHE_SHIFT;
 776         struct inode *inode = fdentry(file)->d_inode;
 777         int err = 0;
 778         u64 start_pos;
 779         u64 last_pos;
 780
 781         start_pos = pos & ~((u64)root->sectorsize - 1);
 782         last_pos = ((u64)index + num_pages) << PAGE_CACHE_SHIFT;
 783
 784         if (start_pos > inode->i_size) {
 785                 err = btrfs_cont_expand(inode, start_pos);
 786                 if (err)
 787                         return err;
 788         }
 789
 790         memset(pages, 0, num_pages * sizeof(struct page *));
 791 again:
 792         for (i = 0; i < num_pages; i++) {
 793                 pages[i] = grab_cache_page(inode->i_mapping, index + i);
 794                 if (!pages[i]) {
 795                         err = -ENOMEM;
 796                         BUG_ON(1);
 797                 }
 798                 wait_on_page_writeback(pages[i]);
 799         }
 800         if (start_pos < inode->i_size) {
 801                 struct btrfs_ordered_extent *ordered;
 802                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
 803                                  start_pos, last_pos - 1, 0, &cached_state,
 804                                  GFP_NOFS);
 805                 ordered = btrfs_lookup_first_ordered_extent(inode,
 806                                                             last_pos - 1);
 807                 if (ordered &&
 808                     ordered->file_offset + ordered->len > start_pos &&
 809                     ordered->file_offset < last_pos) {
 810                         btrfs_put_ordered_extent(ordered);
 811                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
 812                                              start_pos, last_pos - 1,
 813                                              &cached_state, GFP_NOFS);
 814                         for (i = 0; i < num_pages; i++) {
 815                                 unlock_page(pages[i]);
 816                                 page_cache_release(pages[i]);
 817                         }
 818                         btrfs_wait_ordered_range(inode, start_pos,
 819                                                  last_pos - start_pos);
 820                         goto again;
 821                 }
 822                 if (ordered)
 823                         btrfs_put_ordered_extent(ordered);
 824
 825                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
 826                                   last_pos - 1, EXTENT_DIRTY | EXTENT_DELALLOC |
 827                                   EXTENT_DO_ACCOUNTING, 0, 0, &cached_state,
 828                                   GFP_NOFS);
 829                 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
 830                                      start_pos, last_pos - 1, &cached_state,
 831                                      GFP_NOFS);
 832         }
 833         for (i = 0; i < num_pages; i++) {
 834                 clear_page_dirty_for_io(pages[i]);
 835                 set_page_extent_mapped(pages[i]);
 836                 WARN_ON(!PageLocked(pages[i]));
 837         }
 838         return 0;
 839 }
 840
 841 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
 842                                     const struct iovec *iov,
 843                                     unsigned long nr_segs, loff_t pos)
 844 {
 845         struct file *file = iocb->ki_filp;
 846         struct inode *inode = fdentry(file)->d_inode;
 847         struct btrfs_root *root = BTRFS_I(inode)->root;
 848         struct page *pinned[2];
 849         struct page **pages = NULL;
 850         struct iov_iter i;
 851         loff_t *ppos = &iocb->ki_pos;
 852         loff_t start_pos;
 853         ssize_t num_written = 0;
 854         ssize_t err = 0;
 855         size_t count;
 856         size_t ocount;
 857         int ret = 0;
 858         int nrptrs;
 859         unsigned long first_index;
 860         unsigned long last_index;
 861         int will_write;
 862         int buffered = 0;
 863         int copied = 0;
 864         int dirty_pages = 0;
 865
 866         will_write = ((file->f_flags & O_DSYNC) || IS_SYNC(inode) ||
 867                       (file->f_flags & O_DIRECT));
 868
 869         pinned[0] = NULL;
 870         pinned[1] = NULL;
 871
 872         start_pos = pos;
 873
 874         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
 875
 876         mutex_lock(&inode->i_mutex);
 877
 878         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
 879         if (err)
 880                 goto out;
 881         count = ocount;
 882
 883         current->backing_dev_info = inode->i_mapping->backing_dev_info;
 884         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
 885         if (err)
 886                 goto out;
 887
 888         if (count == 0)
 889                 goto out;
 890
 891         err = file_remove_suid(file);
 892         if (err)
 893                 goto out;
 894
 895         /*
 896          * If BTRFS flips readonly due to some impossible error
 897          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
 898          * although we have opened a file as writable, we have
 899          * to stop this write operation to ensure FS consistency.
 900          */
 901         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
 902                 err = -EROFS;
 903                 goto out;
 904         }
 905
 906         file_update_time(file);
 907         BTRFS_I(inode)->sequence++;
 908
 909         if (unlikely(file->f_flags & O_DIRECT)) {
 910                 num_written = generic_file_direct_write(iocb, iov, &nr_segs,
 911                                                         pos, ppos, count,
 912                                                         ocount);
 913                 /*
 914                  * the generic O_DIRECT will update in-memory i_size after the
 915                  * DIOs are done.  But our endio handlers that update the on
 916                  * disk i_size never update past the in memory i_size.  So we
 917                  * need one more update here to catch any additions to the
 918                  * file
 919                  */
 920                 if (inode->i_size != BTRFS_I(inode)->disk_i_size) {
 921                         btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
 922                         mark_inode_dirty(inode);
 923                 }
 924
 925                 if (num_written < 0) {
 926                         ret = num_written;
 927                         num_written = 0;
 928                         goto out;
 929                 } else if (num_written == count) {
 930                         /* pick up pos changes done by the generic code */
 931                         pos = *ppos;
 932                         goto out;
 933                 }
 934                 /*
 935                  * We are going to do buffered for the rest of the range, so we
 936                  * need to make sure to invalidate the buffered pages when we're
 937                  * done.
 938                  */
 939                 buffered = 1;
 940                 pos += num_written;
 941         }
 942
 943         iov_iter_init(&i, iov, nr_segs, count, num_written);
 944         nrptrs = min((iov_iter_count(&i) + PAGE_CACHE_SIZE - 1) /
 945                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
 946                      (sizeof(struct page *)));
 947         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
 948
 949         /* generic_write_checks can change our pos */
 950         start_pos = pos;
 951
 952         first_index = pos >> PAGE_CACHE_SHIFT;
 953         last_index = (pos + iov_iter_count(&i)) >> PAGE_CACHE_SHIFT;
 954
 955         /*
 956          * there are lots of better ways to do this, but this code
 957          * makes sure the first and last page in the file range are
 958          * up to date and ready for cow
 959          */
 960         if ((pos & (PAGE_CACHE_SIZE - 1))) {
 961                 pinned[0] = grab_cache_page(inode->i_mapping, first_index);
 962                 if (!PageUptodate(pinned[0])) {
 963                         ret = btrfs_readpage(NULL, pinned[0]);
 964                         BUG_ON(ret);
 965                         wait_on_page_locked(pinned[0]);
 966                 } else {
 967                         unlock_page(pinned[0]);
 968                 }
 969         }
 970         if ((pos + iov_iter_count(&i)) & (PAGE_CACHE_SIZE - 1)) {
 971                 pinned[1] = grab_cache_page(inode->i_mapping, last_index);
 972                 if (!PageUptodate(pinned[1])) {
 973                         ret = btrfs_readpage(NULL, pinned[1]);
 974                         BUG_ON(ret);
 975                         wait_on_page_locked(pinned[1]);
 976                 } else {
 977                         unlock_page(pinned[1]);
 978                 }
 979         }
 980
 981         while (iov_iter_count(&i) > 0) {
 982                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
 983                 size_t write_bytes = min(iov_iter_count(&i),
 984                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
 985                                          offset);
 986                 size_t num_pages = (write_bytes + PAGE_CACHE_SIZE - 1) >>
 987                                         PAGE_CACHE_SHIFT;
 988
 989                 WARN_ON(num_pages > nrptrs);
 990                 memset(pages, 0, sizeof(struct page *) * nrptrs);
 991
 992                 /*
 993                  * Fault pages before locking them in prepare_pages
 994                  * to avoid recursive lock
 995                  */
 996                 if (unlikely(iov_iter_fault_in_readable(&i, write_bytes))) {
 997                         ret = -EFAULT;
 998                         goto out;
 999                 }
1000
1001                 ret = btrfs_delalloc_reserve_space(inode,
1002                                         num_pages << PAGE_CACHE_SHIFT);
1003                 if (ret)
1004                         goto out;
1005
1006                 ret = prepare_pages(root, file, pages, num_pages,
1007                                     pos, first_index, last_index,
1008                                     write_bytes);
1009                 if (ret) {
1010                         btrfs_delalloc_release_space(inode,
1011                                         num_pages << PAGE_CACHE_SHIFT);
1012                         goto out;
1013                 }
1014
1015                 copied = btrfs_copy_from_user(pos, num_pages,
1016                                            write_bytes, pages, &i);
1017                 dirty_pages = (copied + PAGE_CACHE_SIZE - 1) >>
1018                                         PAGE_CACHE_SHIFT;
1019
1020                 if (num_pages > dirty_pages) {
1021                         if (copied > 0)
1022                                 atomic_inc(
1023                                         &BTRFS_I(inode)->outstanding_extents);
1024                         btrfs_delalloc_release_space(inode,
1025                                         (num_pages - dirty_pages) <<
1026                                         PAGE_CACHE_SHIFT);
1027                 }
1028
1029                 if (copied > 0) {
1030                         dirty_and_release_pages(NULL, root, file, pages,
1031                                                 dirty_pages, pos, copied);
1032                 }
1033
1034                 btrfs_drop_pages(pages, num_pages);
1035
1036                 if (copied > 0) {
1037                         if (will_write) {
1038                                 filemap_fdatawrite_range(inode->i_mapping, pos,
1039                                                          pos + copied - 1);
1040                         } else {
1041                                 balance_dirty_pages_ratelimited_nr(
1042                                                         inode->i_mapping,
1043                                                         dirty_pages);
1044                                 if (dirty_pages <
1045                                 (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1046                                         btrfs_btree_balance_dirty(root, 1);
1047                                 btrfs_throttle(root);
1048                         }
1049                 }
1050
1051                 pos += copied;
1052                 num_written += copied;
1053
1054                 cond_resched();
1055         }
1056 out:
1057         mutex_unlock(&inode->i_mutex);
1058         if (ret)
1059                 err = ret;
1060
1061         kfree(pages);
1062         if (pinned[0])
1063                 page_cache_release(pinned[0]);
1064         if (pinned[1])
1065                 page_cache_release(pinned[1]);
1066         *ppos = pos;
1067
1068         /*
1069          * we want to make sure fsync finds this change
1070          * but we haven't joined a transaction running right now.
1071          *
1072          * Later on, someone is sure to update the inode and get the
1073          * real transid recorded.
1074          *
1075          * We set last_trans now to the fs_info generation + 1,
1076          * this will either be one more than the running transaction
1077          * or the generation used for the next transaction if there isn't
1078          * one running right now.
1079          */
1080         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1081
1082         if (num_written > 0 && will_write) {
1083                 struct btrfs_trans_handle *trans;
1084
1085                 err = btrfs_wait_ordered_range(inode, start_pos, num_written);
1086                 if (err)
1087                         num_written = err;
1088
1089                 if ((file->f_flags & O_DSYNC) || IS_SYNC(inode)) {
1090                         trans = btrfs_start_transaction(root, 0);
1091                         if (IS_ERR(trans)) {
1092                                 num_written = PTR_ERR(trans);
1093                                 goto done;
1094                         }
1095                         mutex_lock(&inode->i_mutex);
1096                         ret = btrfs_log_dentry_safe(trans, root,
1097                                                     file->f_dentry);
1098                         mutex_unlock(&inode->i_mutex);
1099                         if (ret == 0) {
1100                                 ret = btrfs_sync_log(trans, root);
1101                                 if (ret == 0)
1102                                         btrfs_end_transaction(trans, root);
1103                                 else
1104                                         btrfs_commit_transaction(trans, root);
1105                         } else if (ret != BTRFS_NO_LOG_SYNC) {
1106                                 btrfs_commit_transaction(trans, root);
1107                         } else {
1108                                 btrfs_end_transaction(trans, root);
1109                         }
1110                 }
1111                 if (file->f_flags & O_DIRECT && buffered) {
1112                         invalidate_mapping_pages(inode->i_mapping,
1113                               start_pos >> PAGE_CACHE_SHIFT,
1114                              (start_pos + num_written - 1) >> PAGE_CACHE_SHIFT);
1115                 }
1116         }
1117 done:
1118         current->backing_dev_info = NULL;
1119         return num_written ? num_written : err;
1120 }
1121
1122 int btrfs_release_file(struct inode *inode, struct file *filp)
1123 {
1124         /*
1125          * ordered_data_close is set by settattr when we are about to truncate
1126          * a file from a non-zero size to a zero size.  This tries to
1127          * flush down new bytes that may have been written if the
1128          * application were using truncate to replace a file in place.
1129          */
1130         if (BTRFS_I(inode)->ordered_data_close) {
1131                 BTRFS_I(inode)->ordered_data_close = 0;
1132                 btrfs_add_ordered_operation(NULL, BTRFS_I(inode)->root, inode);
1133                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1134                         filemap_flush(inode->i_mapping);
1135         }
1136         if (filp->private_data)
1137                 btrfs_ioctl_trans_end(filp);
1138         return 0;
1139 }
1140
1141 /*
1142  * fsync call for both files and directories.  This logs the inode into
1143  * the tree log instead of forcing full commits whenever possible.
1144  *
1145  * It needs to call filemap_fdatawait so that all ordered extent updates are
1146  * in the metadata btree are up to date for copying to the log.
1147  *
1148  * It drops the inode mutex before doing the tree log commit.  This is an
1149  * important optimization for directories because holding the mutex prevents
1150  * new operations on the dir while we write to disk.
1151  */
1152 int btrfs_sync_file(struct file *file, int datasync)
1153 {
1154         struct dentry *dentry = file->f_path.dentry;
1155         struct inode *inode = dentry->d_inode;
1156         struct btrfs_root *root = BTRFS_I(inode)->root;
1157         int ret = 0;
1158         struct btrfs_trans_handle *trans;
1159
1160
1161         /* we wait first, since the writeback may change the inode */
1162         root->log_batch++;
1163         /* the VFS called filemap_fdatawrite for us */
1164         btrfs_wait_ordered_range(inode, 0, (u64)-1);
1165         root->log_batch++;
1166
1167         /*
1168          * check the transaction that last modified this inode
1169          * and see if its already been committed
1170          */
1171         if (!BTRFS_I(inode)->last_trans)
1172                 goto out;
1173
1174         /*
1175          * if the last transaction that changed this file was before
1176          * the current transaction, we can bail out now without any
1177          * syncing
1178          */
1179         mutex_lock(&root->fs_info->trans_mutex);
1180         if (BTRFS_I(inode)->last_trans <=
1181             root->fs_info->last_trans_committed) {
1182                 BTRFS_I(inode)->last_trans = 0;
1183                 mutex_unlock(&root->fs_info->trans_mutex);
1184                 goto out;
1185         }
1186         mutex_unlock(&root->fs_info->trans_mutex);
1187
1188         /*
1189          * ok we haven't committed the transaction yet, lets do a commit
1190          */
1191         if (file->private_data)
1192                 btrfs_ioctl_trans_end(file);
1193
1194         trans = btrfs_start_transaction(root, 0);
1195         if (IS_ERR(trans)) {
1196                 ret = PTR_ERR(trans);
1197                 goto out;
1198         }
1199
1200         ret = btrfs_log_dentry_safe(trans, root, dentry);
1201         if (ret < 0)
1202                 goto out;
1203
1204         /* we've logged all the items and now have a consistent
1205          * version of the file in the log.  It is possible that
1206          * someone will come in and modify the file, but that's
1207          * fine because the log is consistent on disk, and we
1208          * have references to all of the file's extents
1209          *
1210          * It is possible that someone will come in and log the
1211          * file again, but that will end up using the synchronization
1212          * inside btrfs_sync_log to keep things safe.
1213          */
1214         mutex_unlock(&dentry->d_inode->i_mutex);
1215
1216         if (ret != BTRFS_NO_LOG_SYNC) {
1217                 if (ret > 0) {
1218                         ret = btrfs_commit_transaction(trans, root);
1219                 } else {
1220                         ret = btrfs_sync_log(trans, root);
1221                         if (ret == 0)
1222                                 ret = btrfs_end_transaction(trans, root);
1223                         else
1224                                 ret = btrfs_commit_transaction(trans, root);
1225                 }
1226         } else {
1227                 ret = btrfs_end_transaction(trans, root);
1228         }
1229         mutex_lock(&dentry->d_inode->i_mutex);
1230 out:
1231         return ret > 0 ? -EIO : ret;
1232 }
1233
1234 static const struct vm_operations_struct btrfs_file_vm_ops = {
1235         .fault          = filemap_fault,
1236         .page_mkwrite   = btrfs_page_mkwrite,
1237 };
1238
1239 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
1240 {
1241         struct address_space *mapping = filp->f_mapping;
1242
1243         if (!mapping->a_ops->readpage)
1244                 return -ENOEXEC;
1245
1246         file_accessed(filp);
1247         vma->vm_ops = &btrfs_file_vm_ops;
1248         vma->vm_flags |= VM_CAN_NONLINEAR;
1249
1250         return 0;
1251 }
1252
1253 const struct file_operations btrfs_file_operations = {
1254         .llseek         = generic_file_llseek,
1255         .read           = do_sync_read,
1256         .write          = do_sync_write,
1257         .aio_read       = generic_file_aio_read,
1258         .splice_read    = generic_file_splice_read,
1259         .aio_write      = btrfs_file_aio_write,
1260         .mmap           = btrfs_file_mmap,
1261         .open           = generic_file_open,
1262         .release        = btrfs_release_file,
1263         .fsync          = btrfs_sync_file,
1264         .unlocked_ioctl = btrfs_ioctl,
1265 #ifdef CONFIG_COMPAT
1266         .compat_ioctl   = btrfs_ioctl,
1267 #endif
1268 };