| blob | cdd0320e8f84dd8328176e21e7294712f38bf494 |
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/writeback.h>
34 #include <linux/hash.h>
35 #include <linux/suspend.h>
36 #include <linux/buffer_head.h>
37 #include <linux/task_io_accounting_ops.h>
38 #include <linux/bio.h>
39 #include <linux/notifier.h>
40 #include <linux/cpu.h>
41 #include <linux/bitops.h>
42 #include <linux/mpage.h>
43 #include <linux/bit_spinlock.h>
44 #include <trace/events/block.h>
48 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
51 {
54 }
58 {
61 }
65 {
68 }
71 {
73 TASK_UNINTERRUPTIBLE);
74 }
78 {
82 }
85 /*
86 * Returns if the page has dirty or writeback buffers. If all the buffers
87 * are unlocked and clean then the PageDirty information is stale. If
88 * any of the pages are locked, it is assumed they are locked for IO.
89 */
92 {
116 }
119 /*
120 * Block until a buffer comes unlocked. This doesn't stop it
121 * from becoming locked again - you have to lock it yourself
122 * if you want to preserve its state.
123 */
125 {
127 }
130 static void
132 {
136 }
140 {
144 }
148 {
153 }
155 /*
156 * End-of-IO handler helper function which does not touch the bh after
157 * unlocking it.
158 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
159 * a race there is benign: unlock_buffer() only use the bh's address for
160 * hashing after unlocking the buffer, so it doesn't actually touch the bh
161 * itself.
162 */
164 {
168 /* This happens, due to failed READA attempts. */
170 }
172 }
174 /*
175 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
176 * unlock the buffer. This is what ll_rw_block uses too.
177 */
179 {
182 }
186 {
197 }
200 }
203 }
206 /*
207 * Various filesystems appear to want __find_get_block to be non-blocking.
208 * But it's the page lock which protects the buffers. To get around this,
209 * we get exclusion from try_to_free_buffers with the blockdev mapping's
210 * private_lock.
211 *
212 * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
213 * may be quite high. This code could TryLock the page, and if that
214 * succeeds, there is no need to take private_lock. (But if
215 * private_lock is contended then so is mapping->tree_lock).
216 */
219 {
223 pgoff_t index;
246 }
250 /* we might be here because some of the buffers on this page are
251 * not mapped. This is due to various races between
252 * file io on the block device and getblk. It gets dealt with
253 * elsewhere, don't buffer_error if we had some unmapped buffers
254 */
266 }
267 out_unlock:
270 out:
272 }
274 /*
275 * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
276 */
278 {
292 }
293 }
295 /*
296 * I/O completion handler for block_read_full_page() - pages
297 * which come unlocked at the end of I/O.
298 */
300 {
317 }
319 /*
320 * Be _very_ careful from here on. Bad things can happen if
321 * two buffer heads end IO at almost the same time and both
322 * decide that the page is now completely done.
323 */
336 }
342 /*
343 * If none of the buffers had errors and they are all
344 * uptodate then we can set the page uptodate.
345 */
351 still_busy:
355 }
357 /*
358 * Completion handler for block_write_full_page() - pages which are unlocked
359 * during I/O, and which have PageWriteback cleared upon I/O completion.
360 */
362 {
380 }
385 }
398 }
400 }
406 still_busy:
410 }
413 /*
414 * If a page's buffers are under async readin (end_buffer_async_read
415 * completion) then there is a possibility that another thread of
416 * control could lock one of the buffers after it has completed
417 * but while some of the other buffers have not completed. This
418 * locked buffer would confuse end_buffer_async_read() into not unlocking
419 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
420 * that this buffer is not under async I/O.
421 *
422 * The page comes unlocked when it has no locked buffer_async buffers
423 * left.
424 *
425 * PageLocked prevents anyone starting new async I/O reads any of
426 * the buffers.
427 *
428 * PageWriteback is used to prevent simultaneous writeout of the same
429 * page.
430 *
431 * PageLocked prevents anyone from starting writeback of a page which is
432 * under read I/O (PageWriteback is only ever set against a locked page).
433 */
435 {
438 }
442 {
445 }
448 {
450 }
454 /*
455 * fs/buffer.c contains helper functions for buffer-backed address space's
456 * fsync functions. A common requirement for buffer-based filesystems is
457 * that certain data from the backing blockdev needs to be written out for
458 * a successful fsync(). For example, ext2 indirect blocks need to be
459 * written back and waited upon before fsync() returns.
460 *
461 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
462 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
463 * management of a list of dependent buffers at ->i_mapping->private_list.
464 *
465 * Locking is a little subtle: try_to_free_buffers() will remove buffers
466 * from their controlling inode's queue when they are being freed. But
467 * try_to_free_buffers() will be operating against the *blockdev* mapping
468 * at the time, not against the S_ISREG file which depends on those buffers.
469 * So the locking for private_list is via the private_lock in the address_space
470 * which backs the buffers. Which is different from the address_space
471 * against which the buffers are listed. So for a particular address_space,
472 * mapping->private_lock does *not* protect mapping->private_list! In fact,
473 * mapping->private_list will always be protected by the backing blockdev's
474 * ->private_lock.
475 *
476 * Which introduces a requirement: all buffers on an address_space's
477 * ->private_list must be from the same address_space: the blockdev's.
478 *
479 * address_spaces which do not place buffers at ->private_list via these
480 * utility functions are free to use private_lock and private_list for
481 * whatever they want. The only requirement is that list_empty(private_list)
482 * be true at clear_inode() time.
483 *
484 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
485 * filesystems should do that. invalidate_inode_buffers() should just go
486 * BUG_ON(!list_empty).
487 *
488 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
489 * take an address_space, not an inode. And it should be called
490 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
491 * queued up.
492 *
493 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
494 * list if it is already on a list. Because if the buffer is on a list,
495 * it *must* already be on the right one. If not, the filesystem is being
496 * silly. This will save a ton of locking. But first we have to ensure
497 * that buffers are taken *off* the old inode's list when they are freed
498 * (presumably in truncate). That requires careful auditing of all
499 * filesystems (do it inside bforget()). It could also be done by bringing
500 * b_inode back.
501 */
503 /*
504 * The buffer's backing address_space's private_lock must be held
505 */
507 {
513 }
516 {
518 }
520 /*
521 * osync is designed to support O_SYNC io. It waits synchronously for
522 * all already-submitted IO to complete, but does not queue any new
523 * writes to the disk.
524 *
525 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
526 * you dirty the buffers, and then use osync_inode_buffers to wait for
527 * completion. Any other dirty buffers which are not yet queued for
528 * write will not be flushed to disk by the osync.
529 */
531 {
537 repeat:
549 }
550 }
553 }
556 {
561 }
564 {
568 }
570 /**
571 * emergency_thaw_all -- forcibly thaw every frozen filesystem
572 *
573 * Used for emergency unfreeze of all filesystems via SysRq
574 */
576 {
583 }
584 }
586 /**
587 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
588 * @mapping: the mapping which wants those buffers written
589 *
590 * Starts I/O against the buffers at mapping->private_list, and waits upon
591 * that I/O.
592 *
593 * Basically, this is a convenience function for fsync().
594 * @mapping is a file or directory which needs those buffers to be written for
595 * a successful fsync().
596 */
598 {
606 }
609 /*
610 * Called when we've recently written block `bblock', and it is known that
611 * `bblock' was for a buffer_boundary() buffer. This means that the block at
612 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
613 * dirty, schedule it for IO. So that indirects merge nicely with their data.
614 */
617 {
623 }
624 }
627 {
636 }
643 }
644 }
647 /*
648 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
649 * dirty.
650 *
651 * If warn is true, then emit a warning if the page is not uptodate and has
652 * not been truncated.
653 */
656 {
665 }
668 }
670 /*
671 * Add a page to the dirty page list.
672 *
673 * It is a sad fact of life that this function is called from several places
674 * deeply under spinlocking. It may not sleep.
675 *
676 * If the page has buffers, the uptodate buffers are set dirty, to preserve
677 * dirty-state coherency between the page and the buffers. It the page does
678 * not have buffers then when they are later attached they will all be set
679 * dirty.
680 *
681 * The buffers are dirtied before the page is dirtied. There's a small race
682 * window in which a writepage caller may see the page cleanness but not the
683 * buffer dirtiness. That's fine. If this code were to set the page dirty
684 * before the buffers, a concurrent writepage caller could clear the page dirty
685 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
686 * page on the dirty page list.
687 *
688 * We use private_lock to lock against try_to_free_buffers while using the
689 * page's buffer list. Also use this to protect against clean buffers being
690 * added to the page after it was set dirty.
691 *
692 * FIXME: may need to call ->reservepage here as well. That's rather up to the
693 * address_space though.
694 */
696 {
712 }
719 }
722 /*
723 * Write out and wait upon a list of buffers.
724 *
725 * We have conflicting pressures: we want to make sure that all
726 * initially dirty buffers get waited on, but that any subsequently
727 * dirtied buffers don't. After all, we don't want fsync to last
728 * forever if somebody is actively writing to the file.
729 *
730 * Do this in two main stages: first we copy dirty buffers to a
731 * temporary inode list, queueing the writes as we go. Then we clean
732 * up, waiting for those writes to complete.
733 *
734 * During this second stage, any subsequent updates to the file may end
735 * up refiling the buffer on the original inode's dirty list again, so
736 * there is a chance we will end up with a buffer queued for write but
737 * not yet completed on that list. So, as a final cleanup we go through
738 * the osync code to catch these locked, dirty buffers without requeuing
739 * any newly dirty buffers for write.
740 */
742 {
757 /* Avoid race with mark_buffer_dirty_inode() which does
758 * a lockless check and we rely on seeing the dirty bit */
766 /*
767 * Ensure any pending I/O completes so that
768 * write_dirty_buffer() actually writes the
769 * current contents - it is a noop if I/O is
770 * still in flight on potentially older
771 * contents.
772 */
775 /*
776 * Kick off IO for the previous mapping. Note
777 * that we will not run the very last mapping,
778 * wait_on_buffer() will do that for us
779 * through sync_buffer().
780 */
783 }
784 }
785 }
796 /* Avoid race with mark_buffer_dirty_inode() which does
797 * a lockless check and we rely on seeing the dirty bit */
803 }
810 }
816 else
818 }
820 /*
821 * Invalidate any and all dirty buffers on a given inode. We are
822 * probably unmounting the fs, but that doesn't mean we have already
823 * done a sync(). Just drop the buffers from the inode list.
824 *
825 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
826 * assumes that all the buffers are against the blockdev. Not true
827 * for reiserfs.
828 */
830 {
840 }
841 }
844 /*
845 * Remove any clean buffers from the inode's buffer list. This is called
846 * when we're trying to free the inode itself. Those buffers can pin it.
847 *
848 * Returns true if all buffers were removed.
849 */
851 {
865 }
867 }
869 }
871 }
873 /*
874 * Create the appropriate buffers when given a page for data area and
875 * the size of each buffer.. Use the bh->b_this_page linked list to
876 * follow the buffers created. Return NULL if unable to create more
877 * buffers.
878 *
879 * The retry flag is used to differentiate async IO (paging, swapping)
880 * which may not fail from ordinary buffer allocations.
881 */
884 {
888 try_again:
902 /* Link the buffer to its page */
904 }
906 /*
907 * In case anything failed, we just free everything we got.
908 */
909 no_grow:
916 }
918 /*
919 * Return failure for non-async IO requests. Async IO requests
920 * are not allowed to fail, so we have to wait until buffer heads
921 * become available. But we don't want tasks sleeping with
922 * partially complete buffers, so all were released above.
923 */
927 /* We're _really_ low on memory. Now we just
928 * wait for old buffer heads to become free due to
929 * finishing IO. Since this is an async request and
930 * the reserve list is empty, we're sure there are
931 * async buffer heads in use.
932 */
935 }
940 {
950 }
953 {
960 }
962 }
964 /*
965 * Initialise the state of a blockdev page's buffers.
966 */
967 static sector_t
970 {
985 }
986 block++;
990 /*
991 * Caller needs to validate requested block against end of device.
992 */
994 }
996 /*
997 * Create the page-cache page that contains the requested block.
998 *
999 * This is used purely for blockdev mappings.
1000 */
1001 static int
1004 {
1008 sector_t end_block;
1010 gfp_t gfp_mask;
1014 /*
1015 * XXX: __getblk_slow() can not really deal with failure and
1016 * will endlessly loop on improvised global reclaim. Prefer
1017 * looping in the allocator rather than here, at least that
1018 * code knows what it's doing.
1019 */
1033 size);
1035 }
1038 }
1040 /*
1041 * Allocate some buffers for this page
1042 */
1047 /*
1048 * Link the page to the buffers and initialise them. Take the
1049 * lock to be atomic wrt __find_get_block(), which does not
1050 * run under the page lock.
1051 */
1055 size);
1057 done:
1059 failed:
1063 }
1065 /*
1066 * Create buffers for the specified block device block's page. If
1067 * that page was dirty, the buffers are set dirty also.
1068 */
1069 static int
1071 {
1072 pgoff_t index;
1077 sizebits++;
1082 /*
1083 * Check for a block which wants to lie outside our maximum possible
1084 * pagecache index. (this comparison is done using sector_t types).
1085 */
1094 }
1096 /* Create a page with the proper size buffers.. */
1098 }
1103 {
1104 /* Size must be multiple of hard sectorsize */
1108 size);
1114 }
1129 }
1130 }
1133 /*
1134 * The relationship between dirty buffers and dirty pages:
1135 *
1136 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1137 * the page is tagged dirty in its radix tree.
1138 *
1139 * At all times, the dirtiness of the buffers represents the dirtiness of
1140 * subsections of the page. If the page has buffers, the page dirty bit is
1141 * merely a hint about the true dirty state.
1142 *
1143 * When a page is set dirty in its entirety, all its buffers are marked dirty
1144 * (if the page has buffers).
1145 *
1146 * When a buffer is marked dirty, its page is dirtied, but the page's other
1147 * buffers are not.
1148 *
1149 * Also. When blockdev buffers are explicitly read with bread(), they
1150 * individually become uptodate. But their backing page remains not
1151 * uptodate - even if all of its buffers are uptodate. A subsequent
1152 * block_read_full_page() against that page will discover all the uptodate
1153 * buffers, will set the page uptodate and will perform no I/O.
1154 */
1156 /**
1157 * mark_buffer_dirty - mark a buffer_head as needing writeout
1158 * @bh: the buffer_head to mark dirty
1159 *
1160 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1161 * backing page dirty, then tag the page as dirty in its address_space's radix
1162 * tree and then attach the address_space's inode to its superblock's dirty
1163 * inode list.
1164 *
1165 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1166 * mapping->tree_lock and mapping->host->i_lock.
1167 */
1169 {
1174 /*
1175 * Very *carefully* optimize the it-is-already-dirty case.
1176 *
1177 * Don't let the final "is it dirty" escape to before we
1178 * perhaps modified the buffer.
1179 */
1184 }
1192 }
1193 }
1194 }
1197 /*
1198 * Decrement a buffer_head's reference count. If all buffers against a page
1199 * have zero reference count, are clean and unlocked, and if the page is clean
1200 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1201 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1202 * a page but it ends up not being freed, and buffers may later be reattached).
1203 */
1205 {
1209 }
1211 }
1214 /*
1215 * bforget() is like brelse(), except it discards any
1216 * potentially dirty data.
1217 */
1219 {
1228 }
1230 }
1234 {
1246 }
1249 }
1251 /*
1252 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1253 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1254 * refcount elevated by one when they're in an LRU. A buffer can only appear
1255 * once in a particular CPU's LRU. A single buffer can be present in multiple
1256 * CPU's LRUs at the same time.
1257 *
1258 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1259 * sb_find_get_block().
1260 *
1261 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1262 * a local interrupt disable for that.
1263 */
1265 #define BH_LRU_SIZE 8
1269 };
1273 #ifdef CONFIG_SMP
1274 #define bh_lru_lock() local_irq_disable()
1275 #define bh_lru_unlock() local_irq_enable()
1276 #else
1277 #define bh_lru_lock() preempt_disable()
1278 #define bh_lru_unlock() preempt_enable()
1279 #endif
1282 {
1283 #ifdef irqs_disabled
1285 #endif
1286 }
1288 /*
1289 * The LRU management algorithm is dopey-but-simple. Sorry.
1290 */
1292 {
1316 }
1317 }
1318 }
1322 }
1327 }
1329 /*
1330 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1331 */
1334 {
1349 i--;
1350 }
1352 }
1356 }
1357 }
1360 }
1362 /*
1363 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1364 * it in the LRU and mark it as accessed. If it is not present then return
1365 * NULL
1366 */
1369 {
1373 /* __find_get_block_slow will mark the page accessed */
1381 }
1384 /*
1385 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1386 * which corresponds to the passed block_device, block and size. The
1387 * returned buffer has its reference count incremented.
1388 *
1389 * __getblk_gfp() will lock up the machine if grow_dev_page's
1390 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1391 */
1395 {
1402 }
1405 /*
1406 * Do async read-ahead on a buffer..
1407 */
1409 {
1414 }
1415 }
1418 /**
1419 * __bread_gfp() - reads a specified block and returns the bh
1420 * @bdev: the block_device to read from
1421 * @block: number of block
1422 * @size: size (in bytes) to read
1423 * @gfp: page allocation flag
1424 *
1425 * Reads a specified block, and returns buffer head that contains it.
1426 * The page cache can be allocated from non-movable area
1427 * not to prevent page migration if you set gfp to zero.
1428 * It returns NULL if the block was unreadable.
1429 */
1433 {
1439 }
1442 /*
1443 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1444 * This doesn't race because it runs in each cpu either in irq
1445 * or with preempt disabled.
1446 */
1448 {
1455 }
1457 }
1460 {
1467 }
1470 }
1473 {
1475 }
1480 {
1484 /*
1485 * This catches illegal uses and preserves the offset:
1486 */
1488 else
1490 }
1493 /*
1494 * Called when truncating a buffer on a page completely.
1495 */
1497 /* Bits that are cleared during an invalidate */
1498 #define BUFFER_FLAGS_DISCARD \
1499 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1500 1 << BH_Delay | 1 << BH_Unwritten)
1503 {
1516 }
1518 }
1520 /**
1521 * block_invalidatepage - invalidate part or all of a buffer-backed page
1522 *
1523 * @page: the page which is affected
1524 * @offset: start of the range to invalidate
1525 * @length: length of the range to invalidate
1526 *
1527 * block_invalidatepage() is called when all or part of the page has become
1528 * invalidated by a truncate operation.
1529 *
1530 * block_invalidatepage() does not have to release all buffers, but it must
1531 * ensure that no dirty buffer is left outside @offset and that no I/O
1532 * is underway against any of the blocks which are outside the truncation
1533 * point. Because the caller is about to free (and possibly reuse) those
1534 * blocks on-disk.
1535 */
1538 {
1547 /*
1548 * Check for overflow
1549 */
1558 /*
1559 * Are we still fully in range ?
1560 */
1564 /*
1565 * is this block fully invalidated?
1566 */
1573 /*
1574 * We release buffers only if the entire page is being invalidated.
1575 * The get_block cached value has been unconditionally invalidated,
1576 * so real IO is not possible anymore.
1577 */
1580 out:
1582 }
1586 /*
1587 * We attach and possibly dirty the buffers atomically wrt
1588 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1589 * is already excluded via the page lock.
1590 */
1593 {
1615 }
1618 }
1621 /*
1622 * We are taking a block for data and we don't want any output from any
1623 * buffer-cache aliases starting from return from that function and
1624 * until the moment when something will explicitly mark the buffer
1625 * dirty (hopefully that will not happen until we will free that block ;-)
1626 * We don't even need to mark it not-uptodate - nobody can expect
1627 * anything from a newly allocated buffer anyway. We used to used
1628 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1629 * don't want to mark the alias unmapped, for example - it would confuse
1630 * anyone who might pick it with bread() afterwards...
1631 *
1632 * Also.. Note that bforget() doesn't lock the buffer. So there can
1633 * be writeout I/O going on against recently-freed buffers. We don't
1634 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1635 * only if we really need to. That happens here.
1636 */
1638 {
1649 }
1650 }
1653 /*
1654 * Size is a power-of-two in the range 512..PAGE_SIZE,
1655 * and the case we care about most is PAGE_SIZE.
1656 *
1657 * So this *could* possibly be written with those
1658 * constraints in mind (relevant mostly if some
1659 * architecture has a slow bit-scan instruction)
1660 */
1662 {
1664 }
1666 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1667 {
1673 }
1675 /*
1676 * NOTE! All mapped/uptodate combinations are valid:
1677 *
1678 * Mapped Uptodate Meaning
1679 *
1680 * No No "unknown" - must do get_block()
1681 * No Yes "hole" - zero-filled
1682 * Yes No "allocated" - allocated on disk, not read in
1683 * Yes Yes "valid" - allocated and up-to-date in memory.
1684 *
1685 * "Dirty" is valid only with the last case (mapped+uptodate).
1686 */
1688 /*
1689 * While block_write_full_page is writing back the dirty buffers under
1690 * the page lock, whoever dirtied the buffers may decide to clean them
1691 * again at any time. We handle that by only looking at the buffer
1692 * state inside lock_buffer().
1693 *
1694 * If block_write_full_page() is called for regular writeback
1695 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1696 * locked buffer. This only can happen if someone has written the buffer
1697 * directly, with submit_bh(). At the address_space level PageWriteback
1698 * prevents this contention from occurring.
1699 *
1700 * If block_write_full_page() is called with wbc->sync_mode ==
1701 * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1702 * causes the writes to be flagged as synchronous writes.
1703 */
1707 {
1709 sector_t block;
1710 sector_t last_block;
1720 /*
1721 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1722 * here, and the (potentially unmapped) buffers may become dirty at
1723 * any time. If a buffer becomes dirty here after we've inspected it
1724 * then we just miss that fact, and the page stays dirty.
1725 *
1726 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1727 * handle that here by just cleaning them.
1728 */
1737 /*
1738 * Get all the dirty buffers mapped to disk addresses and
1739 * handle any aliases from the underlying blockdev's mapping.
1740 */
1743 /*
1744 * mapped buffers outside i_size will occur, because
1745 * this page can be outside i_size when there is a
1746 * truncate in progress.
1747 */
1748 /*
1749 * The buffer was zeroed by block_write_full_page()
1750 */
1761 /* blockdev mappings never come here */
1765 }
1766 }
1768 block++;
1774 /*
1775 * If it's a fully non-blocking write attempt and we cannot
1776 * lock the buffer then redirty the page. Note that this can
1777 * potentially cause a busy-wait loop from writeback threads
1778 * and kswapd activity, but those code paths have their own
1779 * higher-level throttling.
1780 */
1786 }
1791 }
1794 /*
1795 * The page and its buffers are protected by PageWriteback(), so we can
1796 * drop the bh refcounts early.
1797 */
1805 nr_underway++;
1806 }
1812 done:
1814 /*
1815 * The page was marked dirty, but the buffers were
1816 * clean. Someone wrote them back by hand with
1817 * ll_rw_block/submit_bh. A rare case.
1818 */
1821 /*
1822 * The page and buffer_heads can be released at any time from
1823 * here on.
1824 */
1825 }
1828 recover:
1829 /*
1830 * ENOSPC, or some other error. We may already have added some
1831 * blocks to the file, so we need to write these out to avoid
1832 * exposing stale data.
1833 * The page is currently locked and not marked for writeback
1834 */
1836 /* Recovery: lock and submit the mapped buffers */
1843 /*
1844 * The buffer may have been set dirty during
1845 * attachment to a dirty page.
1846 */
1848 }
1859 nr_underway++;
1860 }
1865 }
1867 /*
1868 * If a page has any new buffers, zero them out here, and mark them uptodate
1869 * and dirty so they'll be written out (in order to prevent uninitialised
1870 * block data from leaking). And clear the new bit.
1871 */
1873 {
1896 }
1900 }
1901 }
1906 }
1911 {
1916 sector_t block;
1939 }
1941 }
1957 }
1963 }
1964 }
1969 }
1975 }
1976 }
1977 /*
1978 * If we issued read requests - let them complete.
1979 */
1984 }
1988 }
1993 {
2011 }
2018 /*
2019 * If this is a partial write which happened to make all buffers
2020 * uptodate then we can optimize away a bogus readpage() for
2021 * the next read(). Here we 'discover' whether the page went
2022 * uptodate as a result of this (potentially partial) write.
2023 */
2027 }
2029 /*
2030 * block_write_begin takes care of the basic task of block allocation and
2031 * bringing partial write blocks uptodate first.
2032 *
2033 * The filesystem needs to handle block truncation upon failure.
2034 */
2037 {
2051 }
2055 }
2061 {
2068 /*
2069 * The buffers that were written will now be uptodate, so we
2070 * don't have to worry about a readpage reading them and
2071 * overwriting a partial write. However if we have encountered
2072 * a short write and only partially written into a buffer, it
2073 * will not be marked uptodate, so a readpage might come in and
2074 * destroy our partial write.
2075 *
2076 * Do the simplest thing, and just treat any short write to a
2077 * non uptodate page as a zero-length write, and force the
2078 * caller to redo the whole thing.
2079 */
2084 }
2087 /* This could be a short (even 0-length) commit */
2091 }
2097 {
2104 /*
2105 * No need to use i_size_read() here, the i_size
2106 * cannot change under us because we hold i_mutex.
2107 *
2108 * But it's important to update i_size while still holding page lock:
2109 * page writeout could otherwise come in and zero beyond i_size.
2110 */
2114 }
2121 /*
2122 * Don't mark the inode dirty under page lock. First, it unnecessarily
2123 * makes the holding time of page lock longer. Second, it forces lock
2124 * ordering of page lock and transaction start for journaling
2125 * filesystems.
2126 */
2131 }
2134 /*
2135 * block_is_partially_uptodate checks whether buffers within a page are
2136 * uptodate or not.
2137 *
2138 * Returns true if all buffers which correspond to a file portion
2139 * we want to read are uptodate.
2140 */
2143 {
2167 }
2170 }
2176 }
2179 /*
2180 * Generic "read page" function for block devices that have the normal
2181 * get_block functionality. This is most of the block device filesystems.
2182 * Reads the page asynchronously --- the unlock_buffer() and
2183 * set/clear_buffer_uptodate() functions propagate buffer state into the
2184 * page struct once IO has completed.
2185 */
2187 {
2218 }
2224 }
2225 /*
2226 * get_block() might have updated the buffer
2227 * synchronously
2228 */
2231 }
2239 /*
2240 * All buffers are uptodate - we can set the page uptodate
2241 * as well. But not if get_block() returned an error.
2242 */
2247 }
2249 /* Stage two: lock the buffers */
2254 }
2256 /*
2257 * Stage 3: start the IO. Check for uptodateness
2258 * inside the buffer lock in case another process reading
2259 * the underlying blockdev brought it uptodate (the sct fix).
2260 */
2265 else
2267 }
2269 }
2272 /* utility function for filesystems that need to do work on expanding
2273 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2274 * deal with the hole.
2275 */
2277 {
2296 out:
2298 }
2303 {
2309 loff_t curpos;
2321 }
2325 AOP_FLAG_UNINTERRUPTIBLE,
2342 }
2343 }
2345 /* page covers the boundary, find the boundary offset */
2348 /* if we will expand the thing last block will be filled */
2351 }
2355 }
2359 AOP_FLAG_UNINTERRUPTIBLE,
2370 }
2371 out:
2373 }
2375 /*
2376 * For moronic filesystems that do not allow holes in file.
2377 * We may have to extend the file.
2378 */
2383 {
2397 }
2400 }
2404 {
2408 }
2411 /*
2412 * block_page_mkwrite() is not allowed to change the file size as it gets
2413 * called from a page fault handler when a page is first dirtied. Hence we must
2414 * be careful to check for EOF conditions here. We set the page up correctly
2415 * for a written page which means we get ENOSPC checking when writing into
2416 * holes and correct delalloc and unwritten extent mapping on filesystems that
2417 * support these features.
2418 *
2419 * We are not allowed to take the i_mutex here so we have to play games to
2420 * protect against truncate races as the page could now be beyond EOF. Because
2421 * truncate writes the inode size before removing pages, once we have the
2422 * page lock we can determine safely if the page is beyond EOF. If it is not
2423 * beyond EOF, then the page is guaranteed safe against truncation until we
2424 * unlock the page.
2425 *
2426 * Direct callers of this function should protect against filesystem freezing
2427 * using sb_start_write() - sb_end_write() functions.
2428 */
2430 get_block_t get_block)
2431 {
2435 loff_t size;
2442 /* We overload EFAULT to mean page got truncated */
2445 }
2447 /* page is wholly or partially inside EOF */
2450 else
2462 out_unlock:
2465 }
2469 get_block_t get_block)
2470 {
2476 /*
2477 * Update file times before taking page lock. We may end up failing the
2478 * fault so this update may be superfluous but who really cares...
2479 */
2485 }
2488 /*
2489 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2490 * immediately, while under the page lock. So it needs a special end_io
2491 * handler which does not touch the bh after unlocking it.
2492 */
2494 {
2496 }
2498 /*
2499 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2500 * the page (converting it to circular linked list and taking care of page
2501 * dirty races).
2502 */
2504 {
2520 }
2522 /*
2523 * On entry, the page is fully not uptodate.
2524 * On exit the page is fully uptodate in the areas outside (from,to)
2525 * The filesystem needs to handle block truncation upon failure.
2526 */
2531 {
2537 pgoff_t index;
2541 sector_t block_in_file;
2561 }
2566 /*
2567 * Allocate buffers so that we can keep track of state, and potentially
2568 * attach them to the page if an error occurs. In the common case of
2569 * no error, they will just be freed again without ever being attached
2570 * to the page (which is all OK, because we're under the page lock).
2571 *
2572 * Be careful: the buffer linked list is a NULL terminated one, rather
2573 * than the circular one we're used to.
2574 */
2579 }
2583 /*
2584 * We loop across all blocks in the page, whether or not they are
2585 * part of the affected region. This is so we can discover if the
2586 * page is fully mapped-to-disk.
2587 */
2609 }
2614 }
2621 nr_reads++;
2622 }
2623 }
2626 /*
2627 * The page is locked, so these buffers are protected from
2628 * any VM or truncate activity. Hence we don't need to care
2629 * for the buffer_head refcounts.
2630 */
2635 }
2638 }
2647 failed:
2649 /*
2650 * Error recovery is a bit difficult. We need to zero out blocks that
2651 * were newly allocated, and dirty them to ensure they get written out.
2652 * Buffers need to be attached to the page at this point, otherwise
2653 * the handling of potential IO errors during writeout would be hard
2654 * (could try doing synchronous writeout, but what if that fails too?)
2655 */
2659 out_release:
2665 }
2671 {
2688 }
2697 }
2700 }
2703 /*
2704 * nobh_writepage() - based on block_full_write_page() except
2705 * that it tries to operate without attaching bufferheads to
2706 * the page.
2707 */
2710 {
2717 /* Is the page fully inside i_size? */
2721 /* Is the page fully outside i_size? (truncate in progress) */
2724 /*
2725 * The page may have dirty, unmapped buffers. For example,
2726 * they may have been added in ext3_writepage(). Make them
2727 * freeable here, so the page does not leak.
2728 */
2729 #if 0
2730 /* Not really sure about this - do we need this ? */
2733 #endif
2736 }
2738 /*
2739 * The page straddles i_size. It must be zeroed out on each and every
2740 * writepage invocation because it may be mmapped. "A file is mapped
2741 * in multiples of the page size. For a file that is not a multiple of
2742 * the page size, the remaining memory is zeroed when mapped, and
2743 * writes to that region are not written out to the file."
2744 */
2746 out:
2750 end_buffer_async_write);
2752 }
2757 {
2761 sector_t iblock;
2771 /* Block boundary? Nothing to do */
2784 has_buffers:
2788 }
2790 /* Find the buffer that contains "offset" */
2793 iblock++;
2795 }
2802 /* unmapped? It's a hole - nothing to do */
2806 /* Ok, it's mapped. Make sure it's up-to-date */
2812 }
2817 }
2820 }
2825 unlock:
2828 out:
2830 }
2835 {
2839 sector_t iblock;
2849 /* Block boundary? Nothing to do */
2864 /* Find the buffer that contains "offset" */
2869 iblock++;
2871 }
2879 /* unmapped? It's a hole - nothing to do */
2882 }
2884 /* Ok, it's mapped. Make sure it's up-to-date */
2892 /* Uhhuh. Read error. Complain and punt. */
2895 }
2901 unlock:
2904 out:
2906 }
2909 /*
2910 * The generic ->writepage function for buffer-backed address_spaces
2911 */
2914 {
2920 /* Is the page fully inside i_size? */
2923 end_buffer_async_write);
2925 /* Is the page fully outside i_size? (truncate in progress) */
2928 /*
2929 * The page may have dirty, unmapped buffers. For example,
2930 * they may have been added in ext3_writepage(). Make them
2931 * freeable here, so the page does not leak.
2932 */
2936 }
2938 /*
2939 * The page straddles i_size. It must be zeroed out on each and every
2940 * writepage invocation because it may be mmapped. "A file is mapped
2941 * in multiples of the page size. For a file that is not a multiple of
2942 * the page size, the remaining memory is zeroed when mapped, and
2943 * writes to that region are not written out to the file."
2944 */
2947 end_buffer_async_write);
2948 }
2953 {
2961 }
2965 {
2970 }
2977 }
2979 /*
2980 * This allows us to do IO even on the odd last sectors
2981 * of a device, even if the bh block size is some multiple
2982 * of the physical sector size.
2983 *
2984 * We'll just truncate the bio to the size of the device,
2985 * and clear the end of the buffer head manually.
2986 *
2987 * Truly out-of-range accesses will turn into actual IO
2988 * errors, this only handles the "we need to be able to
2989 * do IO at the final sector" case.
2990 */
2992 {
2993 sector_t maxsector;
3000 /*
3001 * If the *whole* IO is past the end of the device,
3002 * let it through, and the IO layer will turn it into
3003 * an EIO.
3004 */
3013 /* Uhhuh. We've got a bh that straddles the device size! */
3016 /* Truncate the bio.. */
3020 /* ..and clear the end of the buffer for reads */
3026 }
3027 }
3030 {
3040 /*
3041 * Only clear out a write error when rewriting
3042 */
3046 /*
3047 * from here on down, it's all bio -- do the initial mapping,
3048 * submit_bio -> generic_make_request may further map this bio around
3049 */
3065 /* Take care of bh's that straddle the end of the device */
3081 }
3085 {
3087 }
3090 /**
3091 * ll_rw_block: low-level access to block devices (DEPRECATED)
3092 * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
3093 * @nr: number of &struct buffer_heads in the array
3094 * @bhs: array of pointers to &struct buffer_head
3095 *
3096 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3097 * requests an I/O operation on them, either a %READ or a %WRITE. The third
3098 * %READA option is described in the documentation for generic_make_request()
3099 * which ll_rw_block() calls.
3100 *
3101 * This function drops any buffer that it cannot get a lock on (with the
3102 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3103 * request, and any buffer that appears to be up-to-date when doing read
3104 * request. Further it marks as clean buffers that are processed for
3105 * writing (the buffer cache won't assume that they are actually clean
3106 * until the buffer gets unlocked).
3107 *
3108 * ll_rw_block sets b_end_io to simple completion handler that marks
3109 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3110 * any waiters.
3111 *
3112 * All of the buffers must be for the same device, and must also be a
3113 * multiple of the current approved size for the device.
3114 */
3116 {
3130 }
3137 }
3138 }
3140 }
3141 }
3145 {
3150 }
3154 }
3157 /*
3158 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3159 * and then start new I/O and then wait upon it. The caller must have a ref on
3160 * the buffer_head.
3161 */
3163 {
3177 }
3179 }
3183 {
3185 }
3188 /*
3189 * try_to_free_buffers() checks if all the buffers on this particular page
3190 * are unused, and releases them if so.
3191 *
3192 * Exclusion against try_to_free_buffers may be obtained by either
3193 * locking the page or by holding its mapping's private_lock.
3194 *
3195 * If the page is dirty but all the buffers are clean then we need to
3196 * be sure to mark the page clean as well. This is because the page
3197 * may be against a block device, and a later reattachment of buffers
3198 * to a dirty page will set *all* buffers dirty. Which would corrupt
3199 * filesystem data on the same device.
3200 *
3201 * The same applies to regular filesystem pages: if all the buffers are
3202 * clean then we set the page clean and proceed. To do that, we require
3203 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3204 * private_lock.
3205 *
3206 * try_to_free_buffers() is non-blocking.
3207 */
3209 {
3212 }
3214 static int
3216 {
3239 failed:
3241 }
3244 {
3256 }
3261 /*
3262 * If the filesystem writes its buffers by hand (eg ext3)
3263 * then we can have clean buffers against a dirty page. We
3264 * clean the page here; otherwise the VM will never notice
3265 * that the filesystem did any IO at all.
3266 *
3267 * Also, during truncate, discard_buffer will have marked all
3268 * the page's buffers clean. We discover that here and clean
3269 * the page also.
3270 *
3271 * private_lock must be held over this entire operation in order
3272 * to synchronise against __set_page_dirty_buffers and prevent the
3273 * dirty bit from being lost.
3274 */
3278 out:
3287 }
3289 }
3292 /*
3293 * There are no bdflush tunables left. But distributions are
3294 * still running obsolete flush daemons, so we terminate them here.
3295 *
3296 * Use of bdflush() is deprecated and will be removed in a future kernel.
3297 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3298 */
3300 {
3307 msg_count++;
3309 "warning: process `%s' used the obsolete bdflush"
3312 }
3317 }
3319 /*
3320 * Buffer-head allocation
3321 */
3324 /*
3325 * Once the number of bh's in the machine exceeds this level, we start
3326 * stripping them in writeback.
3327 */
3335 };
3340 {
3350 }
3353 {
3361 }
3363 }
3367 {
3374 }
3378 {
3385 }
3388 }
3392 {
3396 }
3398 /**
3399 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3400 * @bh: struct buffer_head
3401 *
3402 * Return true if the buffer is up-to-date and false,
3403 * with the buffer locked, if not.
3404 */
3406 {
3412 }
3414 }
3417 /**
3418 * bh_submit_read - Submit a locked buffer for reading
3419 * @bh: struct buffer_head
3420 *
3421 * Returns zero on success and -EIO on error.
3422 */
3424 {
3430 }
3439 }
3443 {
3449 SLAB_MEM_SPREAD),
3450 NULL);
3452 /*
3453 * Limit the bh occupancy to 10% of ZONE_NORMAL
3454 */
3458 }