| blob | 333adbc7ed5a215318db78c11aefdffab0d57cb3 |
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 }
1102 {
1103 /* Size must be multiple of hard sectorsize */
1107 size);
1113 }
1128 }
1129 }
1131 /*
1132 * The relationship between dirty buffers and dirty pages:
1133 *
1134 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1135 * the page is tagged dirty in its radix tree.
1136 *
1137 * At all times, the dirtiness of the buffers represents the dirtiness of
1138 * subsections of the page. If the page has buffers, the page dirty bit is
1139 * merely a hint about the true dirty state.
1140 *
1141 * When a page is set dirty in its entirety, all its buffers are marked dirty
1142 * (if the page has buffers).
1143 *
1144 * When a buffer is marked dirty, its page is dirtied, but the page's other
1145 * buffers are not.
1146 *
1147 * Also. When blockdev buffers are explicitly read with bread(), they
1148 * individually become uptodate. But their backing page remains not
1149 * uptodate - even if all of its buffers are uptodate. A subsequent
1150 * block_read_full_page() against that page will discover all the uptodate
1151 * buffers, will set the page uptodate and will perform no I/O.
1152 */
1154 /**
1155 * mark_buffer_dirty - mark a buffer_head as needing writeout
1156 * @bh: the buffer_head to mark dirty
1157 *
1158 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1159 * backing page dirty, then tag the page as dirty in its address_space's radix
1160 * tree and then attach the address_space's inode to its superblock's dirty
1161 * inode list.
1162 *
1163 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1164 * mapping->tree_lock and mapping->host->i_lock.
1165 */
1167 {
1172 /*
1173 * Very *carefully* optimize the it-is-already-dirty case.
1174 *
1175 * Don't let the final "is it dirty" escape to before we
1176 * perhaps modified the buffer.
1177 */
1182 }
1190 }
1191 }
1192 }
1195 /*
1196 * Decrement a buffer_head's reference count. If all buffers against a page
1197 * have zero reference count, are clean and unlocked, and if the page is clean
1198 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1199 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1200 * a page but it ends up not being freed, and buffers may later be reattached).
1201 */
1203 {
1207 }
1209 }
1212 /*
1213 * bforget() is like brelse(), except it discards any
1214 * potentially dirty data.
1215 */
1217 {
1226 }
1228 }
1232 {
1244 }
1247 }
1249 /*
1250 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1251 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1252 * refcount elevated by one when they're in an LRU. A buffer can only appear
1253 * once in a particular CPU's LRU. A single buffer can be present in multiple
1254 * CPU's LRUs at the same time.
1255 *
1256 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1257 * sb_find_get_block().
1258 *
1259 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1260 * a local interrupt disable for that.
1261 */
1263 #define BH_LRU_SIZE 8
1267 };
1271 #ifdef CONFIG_SMP
1272 #define bh_lru_lock() local_irq_disable()
1273 #define bh_lru_unlock() local_irq_enable()
1274 #else
1275 #define bh_lru_lock() preempt_disable()
1276 #define bh_lru_unlock() preempt_enable()
1277 #endif
1280 {
1281 #ifdef irqs_disabled
1283 #endif
1284 }
1286 /*
1287 * The LRU management algorithm is dopey-but-simple. Sorry.
1288 */
1290 {
1314 }
1315 }
1316 }
1320 }
1325 }
1327 /*
1328 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1329 */
1332 {
1347 i--;
1348 }
1350 }
1354 }
1355 }
1358 }
1360 /*
1361 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1362 * it in the LRU and mark it as accessed. If it is not present then return
1363 * NULL
1364 */
1367 {
1371 /* __find_get_block_slow will mark the page accessed */
1379 }
1382 /*
1383 * __getblk will locate (and, if necessary, create) the buffer_head
1384 * which corresponds to the passed block_device, block and size. The
1385 * returned buffer has its reference count incremented.
1386 *
1387 * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1388 * attempt is failing. FIXME, perhaps?
1389 */
1392 {
1399 }
1402 /*
1403 * Do async read-ahead on a buffer..
1404 */
1406 {
1411 }
1412 }
1415 /**
1416 * __bread() - reads a specified block and returns the bh
1417 * @bdev: the block_device to read from
1418 * @block: number of block
1419 * @size: size (in bytes) to read
1420 *
1421 * Reads a specified block, and returns buffer head that contains it.
1422 * It returns NULL if the block was unreadable.
1423 */
1426 {
1432 }
1435 /*
1436 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1437 * This doesn't race because it runs in each cpu either in irq
1438 * or with preempt disabled.
1439 */
1441 {
1448 }
1450 }
1453 {
1460 }
1463 }
1466 {
1468 }
1473 {
1477 /*
1478 * This catches illegal uses and preserves the offset:
1479 */
1481 else
1483 }
1486 /*
1487 * Called when truncating a buffer on a page completely.
1488 */
1490 /* Bits that are cleared during an invalidate */
1491 #define BUFFER_FLAGS_DISCARD \
1492 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1493 1 << BH_Delay | 1 << BH_Unwritten)
1496 {
1509 }
1511 }
1513 /**
1514 * block_invalidatepage - invalidate part or all of a buffer-backed page
1515 *
1516 * @page: the page which is affected
1517 * @offset: start of the range to invalidate
1518 * @length: length of the range to invalidate
1519 *
1520 * block_invalidatepage() is called when all or part of the page has become
1521 * invalidated by a truncate operation.
1522 *
1523 * block_invalidatepage() does not have to release all buffers, but it must
1524 * ensure that no dirty buffer is left outside @offset and that no I/O
1525 * is underway against any of the blocks which are outside the truncation
1526 * point. Because the caller is about to free (and possibly reuse) those
1527 * blocks on-disk.
1528 */
1531 {
1540 /*
1541 * Check for overflow
1542 */
1551 /*
1552 * Are we still fully in range ?
1553 */
1557 /*
1558 * is this block fully invalidated?
1559 */
1566 /*
1567 * We release buffers only if the entire page is being invalidated.
1568 * The get_block cached value has been unconditionally invalidated,
1569 * so real IO is not possible anymore.
1570 */
1573 out:
1575 }
1579 /*
1580 * We attach and possibly dirty the buffers atomically wrt
1581 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1582 * is already excluded via the page lock.
1583 */
1586 {
1608 }
1611 }
1614 /*
1615 * We are taking a block for data and we don't want any output from any
1616 * buffer-cache aliases starting from return from that function and
1617 * until the moment when something will explicitly mark the buffer
1618 * dirty (hopefully that will not happen until we will free that block ;-)
1619 * We don't even need to mark it not-uptodate - nobody can expect
1620 * anything from a newly allocated buffer anyway. We used to used
1621 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1622 * don't want to mark the alias unmapped, for example - it would confuse
1623 * anyone who might pick it with bread() afterwards...
1624 *
1625 * Also.. Note that bforget() doesn't lock the buffer. So there can
1626 * be writeout I/O going on against recently-freed buffers. We don't
1627 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1628 * only if we really need to. That happens here.
1629 */
1631 {
1642 }
1643 }
1646 /*
1647 * Size is a power-of-two in the range 512..PAGE_SIZE,
1648 * and the case we care about most is PAGE_SIZE.
1649 *
1650 * So this *could* possibly be written with those
1651 * constraints in mind (relevant mostly if some
1652 * architecture has a slow bit-scan instruction)
1653 */
1655 {
1657 }
1659 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1660 {
1666 }
1668 /*
1669 * NOTE! All mapped/uptodate combinations are valid:
1670 *
1671 * Mapped Uptodate Meaning
1672 *
1673 * No No "unknown" - must do get_block()
1674 * No Yes "hole" - zero-filled
1675 * Yes No "allocated" - allocated on disk, not read in
1676 * Yes Yes "valid" - allocated and up-to-date in memory.
1677 *
1678 * "Dirty" is valid only with the last case (mapped+uptodate).
1679 */
1681 /*
1682 * While block_write_full_page is writing back the dirty buffers under
1683 * the page lock, whoever dirtied the buffers may decide to clean them
1684 * again at any time. We handle that by only looking at the buffer
1685 * state inside lock_buffer().
1686 *
1687 * If block_write_full_page() is called for regular writeback
1688 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1689 * locked buffer. This only can happen if someone has written the buffer
1690 * directly, with submit_bh(). At the address_space level PageWriteback
1691 * prevents this contention from occurring.
1692 *
1693 * If block_write_full_page() is called with wbc->sync_mode ==
1694 * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1695 * causes the writes to be flagged as synchronous writes.
1696 */
1700 {
1702 sector_t block;
1703 sector_t last_block;
1713 /*
1714 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1715 * here, and the (potentially unmapped) buffers may become dirty at
1716 * any time. If a buffer becomes dirty here after we've inspected it
1717 * then we just miss that fact, and the page stays dirty.
1718 *
1719 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1720 * handle that here by just cleaning them.
1721 */
1730 /*
1731 * Get all the dirty buffers mapped to disk addresses and
1732 * handle any aliases from the underlying blockdev's mapping.
1733 */
1736 /*
1737 * mapped buffers outside i_size will occur, because
1738 * this page can be outside i_size when there is a
1739 * truncate in progress.
1740 */
1741 /*
1742 * The buffer was zeroed by block_write_full_page()
1743 */
1754 /* blockdev mappings never come here */
1758 }
1759 }
1761 block++;
1767 /*
1768 * If it's a fully non-blocking write attempt and we cannot
1769 * lock the buffer then redirty the page. Note that this can
1770 * potentially cause a busy-wait loop from writeback threads
1771 * and kswapd activity, but those code paths have their own
1772 * higher-level throttling.
1773 */
1779 }
1784 }
1787 /*
1788 * The page and its buffers are protected by PageWriteback(), so we can
1789 * drop the bh refcounts early.
1790 */
1798 nr_underway++;
1799 }
1805 done:
1807 /*
1808 * The page was marked dirty, but the buffers were
1809 * clean. Someone wrote them back by hand with
1810 * ll_rw_block/submit_bh. A rare case.
1811 */
1814 /*
1815 * The page and buffer_heads can be released at any time from
1816 * here on.
1817 */
1818 }
1821 recover:
1822 /*
1823 * ENOSPC, or some other error. We may already have added some
1824 * blocks to the file, so we need to write these out to avoid
1825 * exposing stale data.
1826 * The page is currently locked and not marked for writeback
1827 */
1829 /* Recovery: lock and submit the mapped buffers */
1836 /*
1837 * The buffer may have been set dirty during
1838 * attachment to a dirty page.
1839 */
1841 }
1852 nr_underway++;
1853 }
1858 }
1860 /*
1861 * If a page has any new buffers, zero them out here, and mark them uptodate
1862 * and dirty so they'll be written out (in order to prevent uninitialised
1863 * block data from leaking). And clear the new bit.
1864 */
1866 {
1889 }
1893 }
1894 }
1899 }
1904 {
1909 sector_t block;
1932 }
1934 }
1950 }
1956 }
1957 }
1962 }
1968 }
1969 }
1970 /*
1971 * If we issued read requests - let them complete.
1972 */
1977 }
1981 }
1986 {
2004 }
2011 /*
2012 * If this is a partial write which happened to make all buffers
2013 * uptodate then we can optimize away a bogus readpage() for
2014 * the next read(). Here we 'discover' whether the page went
2015 * uptodate as a result of this (potentially partial) write.
2016 */
2020 }
2022 /*
2023 * block_write_begin takes care of the basic task of block allocation and
2024 * bringing partial write blocks uptodate first.
2025 *
2026 * The filesystem needs to handle block truncation upon failure.
2027 */
2030 {
2044 }
2048 }
2054 {
2061 /*
2062 * The buffers that were written will now be uptodate, so we
2063 * don't have to worry about a readpage reading them and
2064 * overwriting a partial write. However if we have encountered
2065 * a short write and only partially written into a buffer, it
2066 * will not be marked uptodate, so a readpage might come in and
2067 * destroy our partial write.
2068 *
2069 * Do the simplest thing, and just treat any short write to a
2070 * non uptodate page as a zero-length write, and force the
2071 * caller to redo the whole thing.
2072 */
2077 }
2080 /* This could be a short (even 0-length) commit */
2084 }
2090 {
2097 /*
2098 * No need to use i_size_read() here, the i_size
2099 * cannot change under us because we hold i_mutex.
2100 *
2101 * But it's important to update i_size while still holding page lock:
2102 * page writeout could otherwise come in and zero beyond i_size.
2103 */
2107 }
2114 /*
2115 * Don't mark the inode dirty under page lock. First, it unnecessarily
2116 * makes the holding time of page lock longer. Second, it forces lock
2117 * ordering of page lock and transaction start for journaling
2118 * filesystems.
2119 */
2124 }
2127 /*
2128 * block_is_partially_uptodate checks whether buffers within a page are
2129 * uptodate or not.
2130 *
2131 * Returns true if all buffers which correspond to a file portion
2132 * we want to read are uptodate.
2133 */
2136 {
2160 }
2163 }
2169 }
2172 /*
2173 * Generic "read page" function for block devices that have the normal
2174 * get_block functionality. This is most of the block device filesystems.
2175 * Reads the page asynchronously --- the unlock_buffer() and
2176 * set/clear_buffer_uptodate() functions propagate buffer state into the
2177 * page struct once IO has completed.
2178 */
2180 {
2211 }
2217 }
2218 /*
2219 * get_block() might have updated the buffer
2220 * synchronously
2221 */
2224 }
2232 /*
2233 * All buffers are uptodate - we can set the page uptodate
2234 * as well. But not if get_block() returned an error.
2235 */
2240 }
2242 /* Stage two: lock the buffers */
2247 }
2249 /*
2250 * Stage 3: start the IO. Check for uptodateness
2251 * inside the buffer lock in case another process reading
2252 * the underlying blockdev brought it uptodate (the sct fix).
2253 */
2258 else
2260 }
2262 }
2265 /* utility function for filesystems that need to do work on expanding
2266 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2267 * deal with the hole.
2268 */
2270 {
2289 out:
2291 }
2296 {
2302 loff_t curpos;
2314 }
2318 AOP_FLAG_UNINTERRUPTIBLE,
2335 }
2336 }
2338 /* page covers the boundary, find the boundary offset */
2341 /* if we will expand the thing last block will be filled */
2344 }
2348 }
2352 AOP_FLAG_UNINTERRUPTIBLE,
2363 }
2364 out:
2366 }
2368 /*
2369 * For moronic filesystems that do not allow holes in file.
2370 * We may have to extend the file.
2371 */
2376 {
2390 }
2393 }
2397 {
2401 }
2404 /*
2405 * block_page_mkwrite() is not allowed to change the file size as it gets
2406 * called from a page fault handler when a page is first dirtied. Hence we must
2407 * be careful to check for EOF conditions here. We set the page up correctly
2408 * for a written page which means we get ENOSPC checking when writing into
2409 * holes and correct delalloc and unwritten extent mapping on filesystems that
2410 * support these features.
2411 *
2412 * We are not allowed to take the i_mutex here so we have to play games to
2413 * protect against truncate races as the page could now be beyond EOF. Because
2414 * truncate writes the inode size before removing pages, once we have the
2415 * page lock we can determine safely if the page is beyond EOF. If it is not
2416 * beyond EOF, then the page is guaranteed safe against truncation until we
2417 * unlock the page.
2418 *
2419 * Direct callers of this function should protect against filesystem freezing
2420 * using sb_start_write() - sb_end_write() functions.
2421 */
2423 get_block_t get_block)
2424 {
2428 loff_t size;
2435 /* We overload EFAULT to mean page got truncated */
2438 }
2440 /* page is wholly or partially inside EOF */
2443 else
2455 out_unlock:
2458 }
2462 get_block_t get_block)
2463 {
2469 /*
2470 * Update file times before taking page lock. We may end up failing the
2471 * fault so this update may be superfluous but who really cares...
2472 */
2478 }
2481 /*
2482 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2483 * immediately, while under the page lock. So it needs a special end_io
2484 * handler which does not touch the bh after unlocking it.
2485 */
2487 {
2489 }
2491 /*
2492 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2493 * the page (converting it to circular linked list and taking care of page
2494 * dirty races).
2495 */
2497 {
2513 }
2515 /*
2516 * On entry, the page is fully not uptodate.
2517 * On exit the page is fully uptodate in the areas outside (from,to)
2518 * The filesystem needs to handle block truncation upon failure.
2519 */
2524 {
2530 pgoff_t index;
2534 sector_t block_in_file;
2554 }
2559 /*
2560 * Allocate buffers so that we can keep track of state, and potentially
2561 * attach them to the page if an error occurs. In the common case of
2562 * no error, they will just be freed again without ever being attached
2563 * to the page (which is all OK, because we're under the page lock).
2564 *
2565 * Be careful: the buffer linked list is a NULL terminated one, rather
2566 * than the circular one we're used to.
2567 */
2572 }
2576 /*
2577 * We loop across all blocks in the page, whether or not they are
2578 * part of the affected region. This is so we can discover if the
2579 * page is fully mapped-to-disk.
2580 */
2602 }
2607 }
2614 nr_reads++;
2615 }
2616 }
2619 /*
2620 * The page is locked, so these buffers are protected from
2621 * any VM or truncate activity. Hence we don't need to care
2622 * for the buffer_head refcounts.
2623 */
2628 }
2631 }
2640 failed:
2642 /*
2643 * Error recovery is a bit difficult. We need to zero out blocks that
2644 * were newly allocated, and dirty them to ensure they get written out.
2645 * Buffers need to be attached to the page at this point, otherwise
2646 * the handling of potential IO errors during writeout would be hard
2647 * (could try doing synchronous writeout, but what if that fails too?)
2648 */
2652 out_release:
2658 }
2664 {
2681 }
2690 }
2693 }
2696 /*
2697 * nobh_writepage() - based on block_full_write_page() except
2698 * that it tries to operate without attaching bufferheads to
2699 * the page.
2700 */
2703 {
2710 /* Is the page fully inside i_size? */
2714 /* Is the page fully outside i_size? (truncate in progress) */
2717 /*
2718 * The page may have dirty, unmapped buffers. For example,
2719 * they may have been added in ext3_writepage(). Make them
2720 * freeable here, so the page does not leak.
2721 */
2722 #if 0
2723 /* Not really sure about this - do we need this ? */
2726 #endif
2729 }
2731 /*
2732 * The page straddles i_size. It must be zeroed out on each and every
2733 * writepage invocation because it may be mmapped. "A file is mapped
2734 * in multiples of the page size. For a file that is not a multiple of
2735 * the page size, the remaining memory is zeroed when mapped, and
2736 * writes to that region are not written out to the file."
2737 */
2739 out:
2743 end_buffer_async_write);
2745 }
2750 {
2754 sector_t iblock;
2764 /* Block boundary? Nothing to do */
2777 has_buffers:
2781 }
2783 /* Find the buffer that contains "offset" */
2786 iblock++;
2788 }
2795 /* unmapped? It's a hole - nothing to do */
2799 /* Ok, it's mapped. Make sure it's up-to-date */
2805 }
2810 }
2813 }
2818 unlock:
2821 out:
2823 }
2828 {
2832 sector_t iblock;
2842 /* Block boundary? Nothing to do */
2857 /* Find the buffer that contains "offset" */
2862 iblock++;
2864 }
2872 /* unmapped? It's a hole - nothing to do */
2875 }
2877 /* Ok, it's mapped. Make sure it's up-to-date */
2885 /* Uhhuh. Read error. Complain and punt. */
2888 }
2894 unlock:
2897 out:
2899 }
2902 /*
2903 * The generic ->writepage function for buffer-backed address_spaces
2904 * this form passes in the end_io handler used to finish the IO.
2905 */
2908 {
2914 /* Is the page fully inside i_size? */
2917 handler);
2919 /* Is the page fully outside i_size? (truncate in progress) */
2922 /*
2923 * The page may have dirty, unmapped buffers. For example,
2924 * they may have been added in ext3_writepage(). Make them
2925 * freeable here, so the page does not leak.
2926 */
2930 }
2932 /*
2933 * The page straddles i_size. It must be zeroed out on each and every
2934 * writepage invocation because it may be mmapped. "A file is mapped
2935 * in multiples of the page size. For a file that is not a multiple of
2936 * the page size, the remaining memory is zeroed when mapped, and
2937 * writes to that region are not written out to the file."
2938 */
2941 }
2944 /*
2945 * The generic ->writepage function for buffer-backed address_spaces
2946 */
2949 {
2951 end_buffer_async_write);
2952 }
2957 {
2965 }
2969 {
2974 }
2981 }
2983 /*
2984 * This allows us to do IO even on the odd last sectors
2985 * of a device, even if the bh block size is some multiple
2986 * of the physical sector size.
2987 *
2988 * We'll just truncate the bio to the size of the device,
2989 * and clear the end of the buffer head manually.
2990 *
2991 * Truly out-of-range accesses will turn into actual IO
2992 * errors, this only handles the "we need to be able to
2993 * do IO at the final sector" case.
2994 */
2996 {
2997 sector_t maxsector;
3004 /*
3005 * If the *whole* IO is past the end of the device,
3006 * let it through, and the IO layer will turn it into
3007 * an EIO.
3008 */
3017 /* Uhhuh. We've got a bh that straddles the device size! */
3020 /* Truncate the bio.. */
3024 /* ..and clear the end of the buffer for reads */
3030 }
3031 }
3034 {
3044 /*
3045 * Only clear out a write error when rewriting
3046 */
3050 /*
3051 * from here on down, it's all bio -- do the initial mapping,
3052 * submit_bio -> generic_make_request may further map this bio around
3053 */
3069 /* Take care of bh's that straddle the end of the device */
3085 }
3089 {
3091 }
3094 /**
3095 * ll_rw_block: low-level access to block devices (DEPRECATED)
3096 * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
3097 * @nr: number of &struct buffer_heads in the array
3098 * @bhs: array of pointers to &struct buffer_head
3099 *
3100 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3101 * requests an I/O operation on them, either a %READ or a %WRITE. The third
3102 * %READA option is described in the documentation for generic_make_request()
3103 * which ll_rw_block() calls.
3104 *
3105 * This function drops any buffer that it cannot get a lock on (with the
3106 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3107 * request, and any buffer that appears to be up-to-date when doing read
3108 * request. Further it marks as clean buffers that are processed for
3109 * writing (the buffer cache won't assume that they are actually clean
3110 * until the buffer gets unlocked).
3111 *
3112 * ll_rw_block sets b_end_io to simple completion handler that marks
3113 * the buffer up-to-date (if approriate), unlocks the buffer and wakes
3114 * any waiters.
3115 *
3116 * All of the buffers must be for the same device, and must also be a
3117 * multiple of the current approved size for the device.
3118 */
3120 {
3134 }
3141 }
3142 }
3144 }
3145 }
3149 {
3154 }
3158 }
3161 /*
3162 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3163 * and then start new I/O and then wait upon it. The caller must have a ref on
3164 * the buffer_head.
3165 */
3167 {
3181 }
3183 }
3187 {
3189 }
3192 /*
3193 * try_to_free_buffers() checks if all the buffers on this particular page
3194 * are unused, and releases them if so.
3195 *
3196 * Exclusion against try_to_free_buffers may be obtained by either
3197 * locking the page or by holding its mapping's private_lock.
3198 *
3199 * If the page is dirty but all the buffers are clean then we need to
3200 * be sure to mark the page clean as well. This is because the page
3201 * may be against a block device, and a later reattachment of buffers
3202 * to a dirty page will set *all* buffers dirty. Which would corrupt
3203 * filesystem data on the same device.
3204 *
3205 * The same applies to regular filesystem pages: if all the buffers are
3206 * clean then we set the page clean and proceed. To do that, we require
3207 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3208 * private_lock.
3209 *
3210 * try_to_free_buffers() is non-blocking.
3211 */
3213 {
3216 }
3218 static int
3220 {
3243 failed:
3245 }
3248 {
3260 }
3265 /*
3266 * If the filesystem writes its buffers by hand (eg ext3)
3267 * then we can have clean buffers against a dirty page. We
3268 * clean the page here; otherwise the VM will never notice
3269 * that the filesystem did any IO at all.
3270 *
3271 * Also, during truncate, discard_buffer will have marked all
3272 * the page's buffers clean. We discover that here and clean
3273 * the page also.
3274 *
3275 * private_lock must be held over this entire operation in order
3276 * to synchronise against __set_page_dirty_buffers and prevent the
3277 * dirty bit from being lost.
3278 */
3282 out:
3291 }
3293 }
3296 /*
3297 * There are no bdflush tunables left. But distributions are
3298 * still running obsolete flush daemons, so we terminate them here.
3299 *
3300 * Use of bdflush() is deprecated and will be removed in a future kernel.
3301 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3302 */
3304 {
3311 msg_count++;
3313 "warning: process `%s' used the obsolete bdflush"
3316 }
3321 }
3323 /*
3324 * Buffer-head allocation
3325 */
3328 /*
3329 * Once the number of bh's in the machine exceeds this level, we start
3330 * stripping them in writeback.
3331 */
3339 };
3344 {
3354 }
3357 {
3365 }
3367 }
3371 {
3378 }
3382 {
3389 }
3392 }
3396 {
3400 }
3402 /**
3403 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3404 * @bh: struct buffer_head
3405 *
3406 * Return true if the buffer is up-to-date and false,
3407 * with the buffer locked, if not.
3408 */
3410 {
3416 }
3418 }
3421 /**
3422 * bh_submit_read - Submit a locked buffer for reading
3423 * @bh: struct buffer_head
3424 *
3425 * Returns zero on success and -EIO on error.
3426 */
3428 {
3434 }
3443 }
3447 {
3453 SLAB_MEM_SPREAD),
3454 NULL);
3456 /*
3457 * Limit the bh occupancy to 10% of ZONE_NORMAL
3458 */
3462 }