| blob | 754813a6962bc324c476b5d5035f10e761c1e28c |
1 /*
2 * linux/fs/buffer.c
3 *
4 * Copyright (C) 1991, 1992, 2002 Linus Torvalds
5 */
7 /*
8 * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9 *
10 * Removed a lot of unnecessary code and simplified things now that
11 * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12 *
13 * Speed up hash, lru, and free list operations. Use gfp() for allocating
14 * hash table, use SLAB cache for buffer heads. SMP threading. -DaveM
15 *
16 * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17 *
18 * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19 */
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
23 #include <linux/fs.h>
24 #include <linux/mm.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/capability.h>
28 #include <linux/blkdev.h>
29 #include <linux/file.h>
30 #include <linux/quotaops.h>
31 #include <linux/highmem.h>
32 #include <linux/export.h>
33 #include <linux/backing-dev.h>
34 #include <linux/writeback.h>
35 #include <linux/hash.h>
36 #include <linux/suspend.h>
37 #include <linux/buffer_head.h>
38 #include <linux/task_io_accounting_ops.h>
39 #include <linux/bio.h>
40 #include <linux/notifier.h>
41 #include <linux/cpu.h>
42 #include <linux/bitops.h>
43 #include <linux/mpage.h>
44 #include <linux/bit_spinlock.h>
45 #include <trace/events/block.h>
52 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
55 {
58 }
62 {
65 }
69 {
71 }
75 {
79 }
82 /*
83 * Returns if the page has dirty or writeback buffers. If all the buffers
84 * are unlocked and clean then the PageDirty information is stale. If
85 * any of the pages are locked, it is assumed they are locked for IO.
86 */
89 {
113 }
116 /*
117 * Block until a buffer comes unlocked. This doesn't stop it
118 * from becoming locked again - you have to lock it yourself
119 * if you want to preserve its state.
120 */
122 {
124 }
127 static void
129 {
133 }
136 {
141 }
143 /*
144 * End-of-IO handler helper function which does not touch the bh after
145 * unlocking it.
146 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
147 * a race there is benign: unlock_buffer() only use the bh's address for
148 * hashing after unlocking the buffer, so it doesn't actually touch the bh
149 * itself.
150 */
152 {
156 /* This happens, due to failed READA attempts. */
158 }
160 }
162 /*
163 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
164 * unlock the buffer. This is what ll_rw_block uses too.
165 */
167 {
170 }
174 {
181 }
184 }
187 /*
188 * Various filesystems appear to want __find_get_block to be non-blocking.
189 * But it's the page lock which protects the buffers. To get around this,
190 * we get exclusion from try_to_free_buffers with the blockdev mapping's
191 * private_lock.
192 *
193 * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
194 * may be quite high. This code could TryLock the page, and if that
195 * succeeds, there is no need to take private_lock. (But if
196 * private_lock is contended then so is mapping->tree_lock).
197 */
200 {
204 pgoff_t index;
227 }
231 /* we might be here because some of the buffers on this page are
232 * not mapped. This is due to various races between
233 * file io on the block device and getblk. It gets dealt with
234 * elsewhere, don't buffer_error if we had some unmapped buffers
235 */
245 }
246 out_unlock:
249 out:
251 }
253 /*
254 * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
255 */
257 {
271 }
272 }
274 /*
275 * I/O completion handler for block_read_full_page() - pages
276 * which come unlocked at the end of I/O.
277 */
279 {
295 }
297 /*
298 * Be _very_ careful from here on. Bad things can happen if
299 * two buffer heads end IO at almost the same time and both
300 * decide that the page is now completely done.
301 */
314 }
320 /*
321 * If none of the buffers had errors and they are all
322 * uptodate then we can set the page uptodate.
323 */
329 still_busy:
333 }
335 /*
336 * Completion handler for block_write_full_page() - pages which are unlocked
337 * during I/O, and which have PageWriteback cleared upon I/O completion.
338 */
340 {
357 }
370 }
372 }
378 still_busy:
382 }
385 /*
386 * If a page's buffers are under async readin (end_buffer_async_read
387 * completion) then there is a possibility that another thread of
388 * control could lock one of the buffers after it has completed
389 * but while some of the other buffers have not completed. This
390 * locked buffer would confuse end_buffer_async_read() into not unlocking
391 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
392 * that this buffer is not under async I/O.
393 *
394 * The page comes unlocked when it has no locked buffer_async buffers
395 * left.
396 *
397 * PageLocked prevents anyone starting new async I/O reads any of
398 * the buffers.
399 *
400 * PageWriteback is used to prevent simultaneous writeout of the same
401 * page.
402 *
403 * PageLocked prevents anyone from starting writeback of a page which is
404 * under read I/O (PageWriteback is only ever set against a locked page).
405 */
407 {
410 }
414 {
417 }
420 {
422 }
426 /*
427 * fs/buffer.c contains helper functions for buffer-backed address space's
428 * fsync functions. A common requirement for buffer-based filesystems is
429 * that certain data from the backing blockdev needs to be written out for
430 * a successful fsync(). For example, ext2 indirect blocks need to be
431 * written back and waited upon before fsync() returns.
432 *
433 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
434 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
435 * management of a list of dependent buffers at ->i_mapping->private_list.
436 *
437 * Locking is a little subtle: try_to_free_buffers() will remove buffers
438 * from their controlling inode's queue when they are being freed. But
439 * try_to_free_buffers() will be operating against the *blockdev* mapping
440 * at the time, not against the S_ISREG file which depends on those buffers.
441 * So the locking for private_list is via the private_lock in the address_space
442 * which backs the buffers. Which is different from the address_space
443 * against which the buffers are listed. So for a particular address_space,
444 * mapping->private_lock does *not* protect mapping->private_list! In fact,
445 * mapping->private_list will always be protected by the backing blockdev's
446 * ->private_lock.
447 *
448 * Which introduces a requirement: all buffers on an address_space's
449 * ->private_list must be from the same address_space: the blockdev's.
450 *
451 * address_spaces which do not place buffers at ->private_list via these
452 * utility functions are free to use private_lock and private_list for
453 * whatever they want. The only requirement is that list_empty(private_list)
454 * be true at clear_inode() time.
455 *
456 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
457 * filesystems should do that. invalidate_inode_buffers() should just go
458 * BUG_ON(!list_empty).
459 *
460 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
461 * take an address_space, not an inode. And it should be called
462 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
463 * queued up.
464 *
465 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
466 * list if it is already on a list. Because if the buffer is on a list,
467 * it *must* already be on the right one. If not, the filesystem is being
468 * silly. This will save a ton of locking. But first we have to ensure
469 * that buffers are taken *off* the old inode's list when they are freed
470 * (presumably in truncate). That requires careful auditing of all
471 * filesystems (do it inside bforget()). It could also be done by bringing
472 * b_inode back.
473 */
475 /*
476 * The buffer's backing address_space's private_lock must be held
477 */
479 {
485 }
488 {
490 }
492 /*
493 * osync is designed to support O_SYNC io. It waits synchronously for
494 * all already-submitted IO to complete, but does not queue any new
495 * writes to the disk.
496 *
497 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
498 * you dirty the buffers, and then use osync_inode_buffers to wait for
499 * completion. Any other dirty buffers which are not yet queued for
500 * write will not be flushed to disk by the osync.
501 */
503 {
509 repeat:
521 }
522 }
525 }
528 {
531 }
534 {
538 }
540 /**
541 * emergency_thaw_all -- forcibly thaw every frozen filesystem
542 *
543 * Used for emergency unfreeze of all filesystems via SysRq
544 */
546 {
553 }
554 }
556 /**
557 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
558 * @mapping: the mapping which wants those buffers written
559 *
560 * Starts I/O against the buffers at mapping->private_list, and waits upon
561 * that I/O.
562 *
563 * Basically, this is a convenience function for fsync().
564 * @mapping is a file or directory which needs those buffers to be written for
565 * a successful fsync().
566 */
568 {
576 }
579 /*
580 * Called when we've recently written block `bblock', and it is known that
581 * `bblock' was for a buffer_boundary() buffer. This means that the block at
582 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
583 * dirty, schedule it for IO. So that indirects merge nicely with their data.
584 */
587 {
593 }
594 }
597 {
606 }
613 }
614 }
617 /*
618 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
619 * dirty.
620 *
621 * If warn is true, then emit a warning if the page is not uptodate and has
622 * not been truncated.
623 *
624 * The caller must hold lock_page_memcg().
625 */
628 {
637 }
639 }
641 /*
642 * Add a page to the dirty page list.
643 *
644 * It is a sad fact of life that this function is called from several places
645 * deeply under spinlocking. It may not sleep.
646 *
647 * If the page has buffers, the uptodate buffers are set dirty, to preserve
648 * dirty-state coherency between the page and the buffers. It the page does
649 * not have buffers then when they are later attached they will all be set
650 * dirty.
651 *
652 * The buffers are dirtied before the page is dirtied. There's a small race
653 * window in which a writepage caller may see the page cleanness but not the
654 * buffer dirtiness. That's fine. If this code were to set the page dirty
655 * before the buffers, a concurrent writepage caller could clear the page dirty
656 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
657 * page on the dirty page list.
658 *
659 * We use private_lock to lock against try_to_free_buffers while using the
660 * page's buffer list. Also use this to protect against clean buffers being
661 * added to the page after it was set dirty.
662 *
663 * FIXME: may need to call ->reservepage here as well. That's rather up to the
664 * address_space though.
665 */
667 {
683 }
684 /*
685 * Lock out page->mem_cgroup migration to keep PageDirty
686 * synchronized with per-memcg dirty page counters.
687 */
701 }
704 /*
705 * Write out and wait upon a list of buffers.
706 *
707 * We have conflicting pressures: we want to make sure that all
708 * initially dirty buffers get waited on, but that any subsequently
709 * dirtied buffers don't. After all, we don't want fsync to last
710 * forever if somebody is actively writing to the file.
711 *
712 * Do this in two main stages: first we copy dirty buffers to a
713 * temporary inode list, queueing the writes as we go. Then we clean
714 * up, waiting for those writes to complete.
715 *
716 * During this second stage, any subsequent updates to the file may end
717 * up refiling the buffer on the original inode's dirty list again, so
718 * there is a chance we will end up with a buffer queued for write but
719 * not yet completed on that list. So, as a final cleanup we go through
720 * the osync code to catch these locked, dirty buffers without requeuing
721 * any newly dirty buffers for write.
722 */
724 {
739 /* Avoid race with mark_buffer_dirty_inode() which does
740 * a lockless check and we rely on seeing the dirty bit */
748 /*
749 * Ensure any pending I/O completes so that
750 * write_dirty_buffer() actually writes the
751 * current contents - it is a noop if I/O is
752 * still in flight on potentially older
753 * contents.
754 */
757 /*
758 * Kick off IO for the previous mapping. Note
759 * that we will not run the very last mapping,
760 * wait_on_buffer() will do that for us
761 * through sync_buffer().
762 */
765 }
766 }
767 }
778 /* Avoid race with mark_buffer_dirty_inode() which does
779 * a lockless check and we rely on seeing the dirty bit */
785 }
792 }
798 else
800 }
802 /*
803 * Invalidate any and all dirty buffers on a given inode. We are
804 * probably unmounting the fs, but that doesn't mean we have already
805 * done a sync(). Just drop the buffers from the inode list.
806 *
807 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
808 * assumes that all the buffers are against the blockdev. Not true
809 * for reiserfs.
810 */
812 {
822 }
823 }
826 /*
827 * Remove any clean buffers from the inode's buffer list. This is called
828 * when we're trying to free the inode itself. Those buffers can pin it.
829 *
830 * Returns true if all buffers were removed.
831 */
833 {
847 }
849 }
851 }
853 }
855 /*
856 * Create the appropriate buffers when given a page for data area and
857 * the size of each buffer.. Use the bh->b_this_page linked list to
858 * follow the buffers created. Return NULL if unable to create more
859 * buffers.
860 *
861 * The retry flag is used to differentiate async IO (paging, swapping)
862 * which may not fail from ordinary buffer allocations.
863 */
866 {
870 try_again:
884 /* Link the buffer to its page */
886 }
888 /*
889 * In case anything failed, we just free everything we got.
890 */
891 no_grow:
898 }
900 /*
901 * Return failure for non-async IO requests. Async IO requests
902 * are not allowed to fail, so we have to wait until buffer heads
903 * become available. But we don't want tasks sleeping with
904 * partially complete buffers, so all were released above.
905 */
909 /* We're _really_ low on memory. Now we just
910 * wait for old buffer heads to become free due to
911 * finishing IO. Since this is an async request and
912 * the reserve list is empty, we're sure there are
913 * async buffer heads in use.
914 */
917 }
922 {
932 }
935 {
942 }
944 }
946 /*
947 * Initialise the state of a blockdev page's buffers.
948 */
949 static sector_t
952 {
967 }
968 block++;
972 /*
973 * Caller needs to validate requested block against end of device.
974 */
976 }
978 /*
979 * Create the page-cache page that contains the requested block.
980 *
981 * This is used purely for blockdev mappings.
982 */
983 static int
986 {
990 sector_t end_block;
992 gfp_t gfp_mask;
996 /*
997 * XXX: __getblk_slow() can not really deal with failure and
998 * will endlessly loop on improvised global reclaim. Prefer
999 * looping in the allocator rather than here, at least that
1000 * code knows what it's doing.
1001 */
1015 size);
1017 }
1020 }
1022 /*
1023 * Allocate some buffers for this page
1024 */
1029 /*
1030 * Link the page to the buffers and initialise them. Take the
1031 * lock to be atomic wrt __find_get_block(), which does not
1032 * run under the page lock.
1033 */
1037 size);
1039 done:
1041 failed:
1045 }
1047 /*
1048 * Create buffers for the specified block device block's page. If
1049 * that page was dirty, the buffers are set dirty also.
1050 */
1051 static int
1053 {
1054 pgoff_t index;
1059 sizebits++;
1064 /*
1065 * Check for a block which wants to lie outside our maximum possible
1066 * pagecache index. (this comparison is done using sector_t types).
1067 */
1072 bdev);
1074 }
1076 /* Create a page with the proper size buffers.. */
1078 }
1083 {
1084 /* Size must be multiple of hard sectorsize */
1088 size);
1094 }
1109 }
1110 }
1113 /*
1114 * The relationship between dirty buffers and dirty pages:
1115 *
1116 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1117 * the page is tagged dirty in its radix tree.
1118 *
1119 * At all times, the dirtiness of the buffers represents the dirtiness of
1120 * subsections of the page. If the page has buffers, the page dirty bit is
1121 * merely a hint about the true dirty state.
1122 *
1123 * When a page is set dirty in its entirety, all its buffers are marked dirty
1124 * (if the page has buffers).
1125 *
1126 * When a buffer is marked dirty, its page is dirtied, but the page's other
1127 * buffers are not.
1128 *
1129 * Also. When blockdev buffers are explicitly read with bread(), they
1130 * individually become uptodate. But their backing page remains not
1131 * uptodate - even if all of its buffers are uptodate. A subsequent
1132 * block_read_full_page() against that page will discover all the uptodate
1133 * buffers, will set the page uptodate and will perform no I/O.
1134 */
1136 /**
1137 * mark_buffer_dirty - mark a buffer_head as needing writeout
1138 * @bh: the buffer_head to mark dirty
1139 *
1140 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1141 * backing page dirty, then tag the page as dirty in its address_space's radix
1142 * tree and then attach the address_space's inode to its superblock's dirty
1143 * inode list.
1144 *
1145 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1146 * mapping->tree_lock and mapping->host->i_lock.
1147 */
1149 {
1154 /*
1155 * Very *carefully* optimize the it-is-already-dirty case.
1156 *
1157 * Don't let the final "is it dirty" escape to before we
1158 * perhaps modified the buffer.
1159 */
1164 }
1175 }
1179 }
1180 }
1183 /*
1184 * Decrement a buffer_head's reference count. If all buffers against a page
1185 * have zero reference count, are clean and unlocked, and if the page is clean
1186 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1187 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1188 * a page but it ends up not being freed, and buffers may later be reattached).
1189 */
1191 {
1195 }
1197 }
1200 /*
1201 * bforget() is like brelse(), except it discards any
1202 * potentially dirty data.
1203 */
1205 {
1214 }
1216 }
1220 {
1232 }
1235 }
1237 /*
1238 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1239 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1240 * refcount elevated by one when they're in an LRU. A buffer can only appear
1241 * once in a particular CPU's LRU. A single buffer can be present in multiple
1242 * CPU's LRUs at the same time.
1243 *
1244 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1245 * sb_find_get_block().
1246 *
1247 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1248 * a local interrupt disable for that.
1249 */
1251 #define BH_LRU_SIZE 16
1255 };
1259 #ifdef CONFIG_SMP
1260 #define bh_lru_lock() local_irq_disable()
1261 #define bh_lru_unlock() local_irq_enable()
1262 #else
1263 #define bh_lru_lock() preempt_disable()
1264 #define bh_lru_unlock() preempt_enable()
1265 #endif
1268 {
1269 #ifdef irqs_disabled
1271 #endif
1272 }
1274 /*
1275 * The LRU management algorithm is dopey-but-simple. Sorry.
1276 */
1278 {
1302 }
1303 }
1304 }
1308 }
1313 }
1315 /*
1316 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1317 */
1320 {
1335 i--;
1336 }
1338 }
1342 }
1343 }
1346 }
1348 /*
1349 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1350 * it in the LRU and mark it as accessed. If it is not present then return
1351 * NULL
1352 */
1355 {
1359 /* __find_get_block_slow will mark the page accessed */
1367 }
1370 /*
1371 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1372 * which corresponds to the passed block_device, block and size. The
1373 * returned buffer has its reference count incremented.
1374 *
1375 * __getblk_gfp() will lock up the machine if grow_dev_page's
1376 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1377 */
1381 {
1388 }
1391 /*
1392 * Do async read-ahead on a buffer..
1393 */
1395 {
1400 }
1401 }
1404 /**
1405 * __bread_gfp() - reads a specified block and returns the bh
1406 * @bdev: the block_device to read from
1407 * @block: number of block
1408 * @size: size (in bytes) to read
1409 * @gfp: page allocation flag
1410 *
1411 * Reads a specified block, and returns buffer head that contains it.
1412 * The page cache can be allocated from non-movable area
1413 * not to prevent page migration if you set gfp to zero.
1414 * It returns NULL if the block was unreadable.
1415 */
1419 {
1425 }
1428 /*
1429 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1430 * This doesn't race because it runs in each cpu either in irq
1431 * or with preempt disabled.
1432 */
1434 {
1441 }
1443 }
1446 {
1453 }
1456 }
1459 {
1461 }
1466 {
1470 /*
1471 * This catches illegal uses and preserves the offset:
1472 */
1474 else
1476 }
1479 /*
1480 * Called when truncating a buffer on a page completely.
1481 */
1483 /* Bits that are cleared during an invalidate */
1484 #define BUFFER_FLAGS_DISCARD \
1485 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1486 1 << BH_Delay | 1 << BH_Unwritten)
1489 {
1502 }
1504 }
1506 /**
1507 * block_invalidatepage - invalidate part or all of a buffer-backed page
1508 *
1509 * @page: the page which is affected
1510 * @offset: start of the range to invalidate
1511 * @length: length of the range to invalidate
1512 *
1513 * block_invalidatepage() is called when all or part of the page has become
1514 * invalidated by a truncate operation.
1515 *
1516 * block_invalidatepage() does not have to release all buffers, but it must
1517 * ensure that no dirty buffer is left outside @offset and that no I/O
1518 * is underway against any of the blocks which are outside the truncation
1519 * point. Because the caller is about to free (and possibly reuse) those
1520 * blocks on-disk.
1521 */
1524 {
1533 /*
1534 * Check for overflow
1535 */
1544 /*
1545 * Are we still fully in range ?
1546 */
1550 /*
1551 * is this block fully invalidated?
1552 */
1559 /*
1560 * We release buffers only if the entire page is being invalidated.
1561 * The get_block cached value has been unconditionally invalidated,
1562 * so real IO is not possible anymore.
1563 */
1566 out:
1568 }
1572 /*
1573 * We attach and possibly dirty the buffers atomically wrt
1574 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1575 * is already excluded via the page lock.
1576 */
1579 {
1601 }
1604 }
1607 /*
1608 * We are taking a block for data and we don't want any output from any
1609 * buffer-cache aliases starting from return from that function and
1610 * until the moment when something will explicitly mark the buffer
1611 * dirty (hopefully that will not happen until we will free that block ;-)
1612 * We don't even need to mark it not-uptodate - nobody can expect
1613 * anything from a newly allocated buffer anyway. We used to used
1614 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1615 * don't want to mark the alias unmapped, for example - it would confuse
1616 * anyone who might pick it with bread() afterwards...
1617 *
1618 * Also.. Note that bforget() doesn't lock the buffer. So there can
1619 * be writeout I/O going on against recently-freed buffers. We don't
1620 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1621 * only if we really need to. That happens here.
1622 */
1624 {
1635 }
1636 }
1639 /*
1640 * Size is a power-of-two in the range 512..PAGE_SIZE,
1641 * and the case we care about most is PAGE_SIZE.
1642 *
1643 * So this *could* possibly be written with those
1644 * constraints in mind (relevant mostly if some
1645 * architecture has a slow bit-scan instruction)
1646 */
1648 {
1650 }
1652 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1653 {
1659 }
1661 /*
1662 * NOTE! All mapped/uptodate combinations are valid:
1663 *
1664 * Mapped Uptodate Meaning
1665 *
1666 * No No "unknown" - must do get_block()
1667 * No Yes "hole" - zero-filled
1668 * Yes No "allocated" - allocated on disk, not read in
1669 * Yes Yes "valid" - allocated and up-to-date in memory.
1670 *
1671 * "Dirty" is valid only with the last case (mapped+uptodate).
1672 */
1674 /*
1675 * While block_write_full_page is writing back the dirty buffers under
1676 * the page lock, whoever dirtied the buffers may decide to clean them
1677 * again at any time. We handle that by only looking at the buffer
1678 * state inside lock_buffer().
1679 *
1680 * If block_write_full_page() is called for regular writeback
1681 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1682 * locked buffer. This only can happen if someone has written the buffer
1683 * directly, with submit_bh(). At the address_space level PageWriteback
1684 * prevents this contention from occurring.
1685 *
1686 * If block_write_full_page() is called with wbc->sync_mode ==
1687 * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1688 * causes the writes to be flagged as synchronous writes.
1689 */
1693 {
1695 sector_t block;
1696 sector_t last_block;
1705 /*
1706 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1707 * here, and the (potentially unmapped) buffers may become dirty at
1708 * any time. If a buffer becomes dirty here after we've inspected it
1709 * then we just miss that fact, and the page stays dirty.
1710 *
1711 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1712 * handle that here by just cleaning them.
1713 */
1722 /*
1723 * Get all the dirty buffers mapped to disk addresses and
1724 * handle any aliases from the underlying blockdev's mapping.
1725 */
1728 /*
1729 * mapped buffers outside i_size will occur, because
1730 * this page can be outside i_size when there is a
1731 * truncate in progress.
1732 */
1733 /*
1734 * The buffer was zeroed by block_write_full_page()
1735 */
1746 /* blockdev mappings never come here */
1750 }
1751 }
1753 block++;
1759 /*
1760 * If it's a fully non-blocking write attempt and we cannot
1761 * lock the buffer then redirty the page. Note that this can
1762 * potentially cause a busy-wait loop from writeback threads
1763 * and kswapd activity, but those code paths have their own
1764 * higher-level throttling.
1765 */
1771 }
1776 }
1779 /*
1780 * The page and its buffers are protected by PageWriteback(), so we can
1781 * drop the bh refcounts early.
1782 */
1790 nr_underway++;
1791 }
1797 done:
1799 /*
1800 * The page was marked dirty, but the buffers were
1801 * clean. Someone wrote them back by hand with
1802 * ll_rw_block/submit_bh. A rare case.
1803 */
1806 /*
1807 * The page and buffer_heads can be released at any time from
1808 * here on.
1809 */
1810 }
1813 recover:
1814 /*
1815 * ENOSPC, or some other error. We may already have added some
1816 * blocks to the file, so we need to write these out to avoid
1817 * exposing stale data.
1818 * The page is currently locked and not marked for writeback
1819 */
1821 /* Recovery: lock and submit the mapped buffers */
1828 /*
1829 * The buffer may have been set dirty during
1830 * attachment to a dirty page.
1831 */
1833 }
1844 nr_underway++;
1845 }
1850 }
1852 /*
1853 * If a page has any new buffers, zero them out here, and mark them uptodate
1854 * and dirty so they'll be written out (in order to prevent uninitialised
1855 * block data from leaking). And clear the new bit.
1856 */
1858 {
1881 }
1885 }
1886 }
1891 }
1896 {
1901 sector_t block;
1924 }
1926 }
1942 }
1948 }
1949 }
1954 }
1960 }
1961 }
1962 /*
1963 * If we issued read requests - let them complete.
1964 */
1969 }
1973 }
1978 {
1996 }
2003 /*
2004 * If this is a partial write which happened to make all buffers
2005 * uptodate then we can optimize away a bogus readpage() for
2006 * the next read(). Here we 'discover' whether the page went
2007 * uptodate as a result of this (potentially partial) write.
2008 */
2012 }
2014 /*
2015 * block_write_begin takes care of the basic task of block allocation and
2016 * bringing partial write blocks uptodate first.
2017 *
2018 * The filesystem needs to handle block truncation upon failure.
2019 */
2022 {
2036 }
2040 }
2046 {
2053 /*
2054 * The buffers that were written will now be uptodate, so we
2055 * don't have to worry about a readpage reading them and
2056 * overwriting a partial write. However if we have encountered
2057 * a short write and only partially written into a buffer, it
2058 * will not be marked uptodate, so a readpage might come in and
2059 * destroy our partial write.
2060 *
2061 * Do the simplest thing, and just treat any short write to a
2062 * non uptodate page as a zero-length write, and force the
2063 * caller to redo the whole thing.
2064 */
2069 }
2072 /* This could be a short (even 0-length) commit */
2076 }
2082 {
2089 /*
2090 * No need to use i_size_read() here, the i_size
2091 * cannot change under us because we hold i_mutex.
2092 *
2093 * But it's important to update i_size while still holding page lock:
2094 * page writeout could otherwise come in and zero beyond i_size.
2095 */
2099 }
2106 /*
2107 * Don't mark the inode dirty under page lock. First, it unnecessarily
2108 * makes the holding time of page lock longer. Second, it forces lock
2109 * ordering of page lock and transaction start for journaling
2110 * filesystems.
2111 */
2116 }
2119 /*
2120 * block_is_partially_uptodate checks whether buffers within a page are
2121 * uptodate or not.
2122 *
2123 * Returns true if all buffers which correspond to a file portion
2124 * we want to read are uptodate.
2125 */
2128 {
2152 }
2155 }
2161 }
2164 /*
2165 * Generic "read page" function for block devices that have the normal
2166 * get_block functionality. This is most of the block device filesystems.
2167 * Reads the page asynchronously --- the unlock_buffer() and
2168 * set/clear_buffer_uptodate() functions propagate buffer state into the
2169 * page struct once IO has completed.
2170 */
2172 {
2203 }
2209 }
2210 /*
2211 * get_block() might have updated the buffer
2212 * synchronously
2213 */
2216 }
2224 /*
2225 * All buffers are uptodate - we can set the page uptodate
2226 * as well. But not if get_block() returned an error.
2227 */
2232 }
2234 /* Stage two: lock the buffers */
2239 }
2241 /*
2242 * Stage 3: start the IO. Check for uptodateness
2243 * inside the buffer lock in case another process reading
2244 * the underlying blockdev brought it uptodate (the sct fix).
2245 */
2250 else
2252 }
2254 }
2257 /* utility function for filesystems that need to do work on expanding
2258 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2259 * deal with the hole.
2260 */
2262 {
2281 out:
2283 }
2288 {
2294 loff_t curpos;
2306 }
2310 AOP_FLAG_UNINTERRUPTIBLE,
2327 }
2328 }
2330 /* page covers the boundary, find the boundary offset */
2333 /* if we will expand the thing last block will be filled */
2336 }
2340 }
2344 AOP_FLAG_UNINTERRUPTIBLE,
2355 }
2356 out:
2358 }
2360 /*
2361 * For moronic filesystems that do not allow holes in file.
2362 * We may have to extend the file.
2363 */
2368 {
2382 }
2385 }
2389 {
2393 }
2396 /*
2397 * block_page_mkwrite() is not allowed to change the file size as it gets
2398 * called from a page fault handler when a page is first dirtied. Hence we must
2399 * be careful to check for EOF conditions here. We set the page up correctly
2400 * for a written page which means we get ENOSPC checking when writing into
2401 * holes and correct delalloc and unwritten extent mapping on filesystems that
2402 * support these features.
2403 *
2404 * We are not allowed to take the i_mutex here so we have to play games to
2405 * protect against truncate races as the page could now be beyond EOF. Because
2406 * truncate writes the inode size before removing pages, once we have the
2407 * page lock we can determine safely if the page is beyond EOF. If it is not
2408 * beyond EOF, then the page is guaranteed safe against truncation until we
2409 * unlock the page.
2410 *
2411 * Direct callers of this function should protect against filesystem freezing
2412 * using sb_start_pagefault() - sb_end_pagefault() functions.
2413 */
2415 get_block_t get_block)
2416 {
2420 loff_t size;
2427 /* We overload EFAULT to mean page got truncated */
2430 }
2432 /* page is wholly or partially inside EOF */
2435 else
2447 out_unlock:
2450 }
2453 /*
2454 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2455 * immediately, while under the page lock. So it needs a special end_io
2456 * handler which does not touch the bh after unlocking it.
2457 */
2459 {
2461 }
2463 /*
2464 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2465 * the page (converting it to circular linked list and taking care of page
2466 * dirty races).
2467 */
2469 {
2485 }
2487 /*
2488 * On entry, the page is fully not uptodate.
2489 * On exit the page is fully uptodate in the areas outside (from,to)
2490 * The filesystem needs to handle block truncation upon failure.
2491 */
2496 {
2502 pgoff_t index;
2506 sector_t block_in_file;
2526 }
2531 /*
2532 * Allocate buffers so that we can keep track of state, and potentially
2533 * attach them to the page if an error occurs. In the common case of
2534 * no error, they will just be freed again without ever being attached
2535 * to the page (which is all OK, because we're under the page lock).
2536 *
2537 * Be careful: the buffer linked list is a NULL terminated one, rather
2538 * than the circular one we're used to.
2539 */
2544 }
2548 /*
2549 * We loop across all blocks in the page, whether or not they are
2550 * part of the affected region. This is so we can discover if the
2551 * page is fully mapped-to-disk.
2552 */
2574 }
2579 }
2586 nr_reads++;
2587 }
2588 }
2591 /*
2592 * The page is locked, so these buffers are protected from
2593 * any VM or truncate activity. Hence we don't need to care
2594 * for the buffer_head refcounts.
2595 */
2600 }
2603 }
2612 failed:
2614 /*
2615 * Error recovery is a bit difficult. We need to zero out blocks that
2616 * were newly allocated, and dirty them to ensure they get written out.
2617 * Buffers need to be attached to the page at this point, otherwise
2618 * the handling of potential IO errors during writeout would be hard
2619 * (could try doing synchronous writeout, but what if that fails too?)
2620 */
2624 out_release:
2630 }
2636 {
2653 }
2662 }
2665 }
2668 /*
2669 * nobh_writepage() - based on block_full_write_page() except
2670 * that it tries to operate without attaching bufferheads to
2671 * the page.
2672 */
2675 {
2682 /* Is the page fully inside i_size? */
2686 /* Is the page fully outside i_size? (truncate in progress) */
2689 /*
2690 * The page may have dirty, unmapped buffers. For example,
2691 * they may have been added in ext3_writepage(). Make them
2692 * freeable here, so the page does not leak.
2693 */
2694 #if 0
2695 /* Not really sure about this - do we need this ? */
2698 #endif
2701 }
2703 /*
2704 * The page straddles i_size. It must be zeroed out on each and every
2705 * writepage invocation because it may be mmapped. "A file is mapped
2706 * in multiples of the page size. For a file that is not a multiple of
2707 * the page size, the remaining memory is zeroed when mapped, and
2708 * writes to that region are not written out to the file."
2709 */
2711 out:
2715 end_buffer_async_write);
2717 }
2722 {
2726 sector_t iblock;
2736 /* Block boundary? Nothing to do */
2749 has_buffers:
2753 }
2755 /* Find the buffer that contains "offset" */
2758 iblock++;
2760 }
2767 /* unmapped? It's a hole - nothing to do */
2771 /* Ok, it's mapped. Make sure it's up-to-date */
2777 }
2782 }
2785 }
2790 unlock:
2793 out:
2795 }
2800 {
2804 sector_t iblock;
2814 /* Block boundary? Nothing to do */
2829 /* Find the buffer that contains "offset" */
2834 iblock++;
2836 }
2844 /* unmapped? It's a hole - nothing to do */
2847 }
2849 /* Ok, it's mapped. Make sure it's up-to-date */
2857 /* Uhhuh. Read error. Complain and punt. */
2860 }
2866 unlock:
2869 out:
2871 }
2874 /*
2875 * The generic ->writepage function for buffer-backed address_spaces
2876 */
2879 {
2885 /* Is the page fully inside i_size? */
2888 end_buffer_async_write);
2890 /* Is the page fully outside i_size? (truncate in progress) */
2893 /*
2894 * The page may have dirty, unmapped buffers. For example,
2895 * they may have been added in ext3_writepage(). Make them
2896 * freeable here, so the page does not leak.
2897 */
2901 }
2903 /*
2904 * The page straddles i_size. It must be zeroed out on each and every
2905 * writepage invocation because it may be mmapped. "A file is mapped
2906 * in multiples of the page size. For a file that is not a multiple of
2907 * the page size, the remaining memory is zeroed when mapped, and
2908 * writes to that region are not written out to the file."
2909 */
2912 end_buffer_async_write);
2913 }
2918 {
2926 }
2930 {
2938 }
2940 /*
2941 * This allows us to do IO even on the odd last sectors
2942 * of a device, even if the block size is some multiple
2943 * of the physical sector size.
2944 *
2945 * We'll just truncate the bio to the size of the device,
2946 * and clear the end of the buffer head manually.
2947 *
2948 * Truly out-of-range accesses will turn into actual IO
2949 * errors, this only handles the "we need to be able to
2950 * do IO at the final sector" case.
2951 */
2953 {
2954 sector_t maxsector;
2962 /*
2963 * If the *whole* IO is past the end of the device,
2964 * let it through, and the IO layer will turn it into
2965 * an EIO.
2966 */
2974 /* Uhhuh. We've got a bio that straddles the device size! */
2977 /* Truncate the bio.. */
2981 /* ..and clear the end of the buffer for reads */
2984 truncated_bytes);
2985 }
2986 }
2990 {
2999 /*
3000 * Only clear out a write error when rewriting
3001 */
3005 /*
3006 * from here on down, it's all bio -- do the initial mapping,
3007 * submit_bio -> generic_make_request may further map this bio around
3008 */
3014 }
3026 /* Take care of bh's that straddle the end of the device */
3036 }
3039 {
3041 }
3045 {
3047 }
3050 /**
3051 * ll_rw_block: low-level access to block devices (DEPRECATED)
3052 * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
3053 * @nr: number of &struct buffer_heads in the array
3054 * @bhs: array of pointers to &struct buffer_head
3055 *
3056 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3057 * requests an I/O operation on them, either a %READ or a %WRITE. The third
3058 * %READA option is described in the documentation for generic_make_request()
3059 * which ll_rw_block() calls.
3060 *
3061 * This function drops any buffer that it cannot get a lock on (with the
3062 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3063 * request, and any buffer that appears to be up-to-date when doing read
3064 * request. Further it marks as clean buffers that are processed for
3065 * writing (the buffer cache won't assume that they are actually clean
3066 * until the buffer gets unlocked).
3067 *
3068 * ll_rw_block sets b_end_io to simple completion handler that marks
3069 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3070 * any waiters.
3071 *
3072 * All of the buffers must be for the same device, and must also be a
3073 * multiple of the current approved size for the device.
3074 */
3076 {
3090 }
3097 }
3098 }
3100 }
3101 }
3105 {
3110 }
3114 }
3117 /*
3118 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3119 * and then start new I/O and then wait upon it. The caller must have a ref on
3120 * the buffer_head.
3121 */
3123 {
3137 }
3139 }
3143 {
3145 }
3148 /*
3149 * try_to_free_buffers() checks if all the buffers on this particular page
3150 * are unused, and releases them if so.
3151 *
3152 * Exclusion against try_to_free_buffers may be obtained by either
3153 * locking the page or by holding its mapping's private_lock.
3154 *
3155 * If the page is dirty but all the buffers are clean then we need to
3156 * be sure to mark the page clean as well. This is because the page
3157 * may be against a block device, and a later reattachment of buffers
3158 * to a dirty page will set *all* buffers dirty. Which would corrupt
3159 * filesystem data on the same device.
3160 *
3161 * The same applies to regular filesystem pages: if all the buffers are
3162 * clean then we set the page clean and proceed. To do that, we require
3163 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3164 * private_lock.
3165 *
3166 * try_to_free_buffers() is non-blocking.
3167 */
3169 {
3172 }
3174 static int
3176 {
3199 failed:
3201 }
3204 {
3216 }
3221 /*
3222 * If the filesystem writes its buffers by hand (eg ext3)
3223 * then we can have clean buffers against a dirty page. We
3224 * clean the page here; otherwise the VM will never notice
3225 * that the filesystem did any IO at all.
3226 *
3227 * Also, during truncate, discard_buffer will have marked all
3228 * the page's buffers clean. We discover that here and clean
3229 * the page also.
3230 *
3231 * private_lock must be held over this entire operation in order
3232 * to synchronise against __set_page_dirty_buffers and prevent the
3233 * dirty bit from being lost.
3234 */
3238 out:
3247 }
3249 }
3252 /*
3253 * There are no bdflush tunables left. But distributions are
3254 * still running obsolete flush daemons, so we terminate them here.
3255 *
3256 * Use of bdflush() is deprecated and will be removed in a future kernel.
3257 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3258 */
3260 {
3267 msg_count++;
3269 "warning: process `%s' used the obsolete bdflush"
3272 }
3277 }
3279 /*
3280 * Buffer-head allocation
3281 */
3284 /*
3285 * Once the number of bh's in the machine exceeds this level, we start
3286 * stripping them in writeback.
3287 */
3295 };
3300 {
3310 }
3313 {
3321 }
3323 }
3327 {
3334 }
3338 {
3345 }
3348 }
3352 {
3356 }
3358 /**
3359 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3360 * @bh: struct buffer_head
3361 *
3362 * Return true if the buffer is up-to-date and false,
3363 * with the buffer locked, if not.
3364 */
3366 {
3372 }
3374 }
3377 /**
3378 * bh_submit_read - Submit a locked buffer for reading
3379 * @bh: struct buffer_head
3380 *
3381 * Returns zero on success and -EIO on error.
3382 */
3384 {
3390 }
3399 }
3403 {
3409 SLAB_MEM_SPREAD),
3410 NULL);
3412 /*
3413 * Limit the bh occupancy to 10% of ZONE_NORMAL
3414 */
3418 }