| blob | c7a5602d01eed200912d3a90ca4ac6780209cb6f |
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 {
67 }
71 {
75 }
78 /*
79 * Returns if the page has dirty or writeback buffers. If all the buffers
80 * are unlocked and clean then the PageDirty information is stale. If
81 * any of the pages are locked, it is assumed they are locked for IO.
82 */
85 {
109 }
112 /*
113 * Block until a buffer comes unlocked. This doesn't stop it
114 * from becoming locked again - you have to lock it yourself
115 * if you want to preserve its state.
116 */
118 {
120 }
123 static void
125 {
129 }
132 {
140 }
142 /*
143 * End-of-IO handler helper function which does not touch the bh after
144 * unlocking it.
145 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
146 * a race there is benign: unlock_buffer() only use the bh's address for
147 * hashing after unlocking the buffer, so it doesn't actually touch the bh
148 * itself.
149 */
151 {
155 /* This happens, due to failed READA attempts. */
157 }
159 }
161 /*
162 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
163 * unlock the buffer. This is what ll_rw_block uses too.
164 */
166 {
169 }
173 {
180 }
183 }
186 /*
187 * Various filesystems appear to want __find_get_block to be non-blocking.
188 * But it's the page lock which protects the buffers. To get around this,
189 * we get exclusion from try_to_free_buffers with the blockdev mapping's
190 * private_lock.
191 *
192 * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
193 * may be quite high. This code could TryLock the page, and if that
194 * succeeds, there is no need to take private_lock. (But if
195 * private_lock is contended then so is mapping->tree_lock).
196 */
199 {
203 pgoff_t index;
226 }
230 /* we might be here because some of the buffers on this page are
231 * not mapped. This is due to various races between
232 * file io on the block device and getblk. It gets dealt with
233 * elsewhere, don't buffer_error if we had some unmapped buffers
234 */
246 }
247 out_unlock:
250 out:
252 }
254 /*
255 * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
256 */
258 {
272 }
273 }
275 /*
276 * I/O completion handler for block_read_full_page() - pages
277 * which come unlocked at the end of I/O.
278 */
280 {
296 }
298 /*
299 * Be _very_ careful from here on. Bad things can happen if
300 * two buffer heads end IO at almost the same time and both
301 * decide that the page is now completely done.
302 */
315 }
321 /*
322 * If none of the buffers had errors and they are all
323 * uptodate then we can set the page uptodate.
324 */
330 still_busy:
334 }
336 /*
337 * Completion handler for block_write_full_page() - pages which are unlocked
338 * during I/O, and which have PageWriteback cleared upon I/O completion.
339 */
341 {
358 }
371 }
373 }
379 still_busy:
383 }
386 /*
387 * If a page's buffers are under async readin (end_buffer_async_read
388 * completion) then there is a possibility that another thread of
389 * control could lock one of the buffers after it has completed
390 * but while some of the other buffers have not completed. This
391 * locked buffer would confuse end_buffer_async_read() into not unlocking
392 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
393 * that this buffer is not under async I/O.
394 *
395 * The page comes unlocked when it has no locked buffer_async buffers
396 * left.
397 *
398 * PageLocked prevents anyone starting new async I/O reads any of
399 * the buffers.
400 *
401 * PageWriteback is used to prevent simultaneous writeout of the same
402 * page.
403 *
404 * PageLocked prevents anyone from starting writeback of a page which is
405 * under read I/O (PageWriteback is only ever set against a locked page).
406 */
408 {
411 }
415 {
418 }
421 {
423 }
427 /*
428 * fs/buffer.c contains helper functions for buffer-backed address space's
429 * fsync functions. A common requirement for buffer-based filesystems is
430 * that certain data from the backing blockdev needs to be written out for
431 * a successful fsync(). For example, ext2 indirect blocks need to be
432 * written back and waited upon before fsync() returns.
433 *
434 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
435 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
436 * management of a list of dependent buffers at ->i_mapping->private_list.
437 *
438 * Locking is a little subtle: try_to_free_buffers() will remove buffers
439 * from their controlling inode's queue when they are being freed. But
440 * try_to_free_buffers() will be operating against the *blockdev* mapping
441 * at the time, not against the S_ISREG file which depends on those buffers.
442 * So the locking for private_list is via the private_lock in the address_space
443 * which backs the buffers. Which is different from the address_space
444 * against which the buffers are listed. So for a particular address_space,
445 * mapping->private_lock does *not* protect mapping->private_list! In fact,
446 * mapping->private_list will always be protected by the backing blockdev's
447 * ->private_lock.
448 *
449 * Which introduces a requirement: all buffers on an address_space's
450 * ->private_list must be from the same address_space: the blockdev's.
451 *
452 * address_spaces which do not place buffers at ->private_list via these
453 * utility functions are free to use private_lock and private_list for
454 * whatever they want. The only requirement is that list_empty(private_list)
455 * be true at clear_inode() time.
456 *
457 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
458 * filesystems should do that. invalidate_inode_buffers() should just go
459 * BUG_ON(!list_empty).
460 *
461 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
462 * take an address_space, not an inode. And it should be called
463 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
464 * queued up.
465 *
466 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
467 * list if it is already on a list. Because if the buffer is on a list,
468 * it *must* already be on the right one. If not, the filesystem is being
469 * silly. This will save a ton of locking. But first we have to ensure
470 * that buffers are taken *off* the old inode's list when they are freed
471 * (presumably in truncate). That requires careful auditing of all
472 * filesystems (do it inside bforget()). It could also be done by bringing
473 * b_inode back.
474 */
476 /*
477 * The buffer's backing address_space's private_lock must be held
478 */
480 {
486 }
489 {
491 }
493 /*
494 * osync is designed to support O_SYNC io. It waits synchronously for
495 * all already-submitted IO to complete, but does not queue any new
496 * writes to the disk.
497 *
498 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
499 * you dirty the buffers, and then use osync_inode_buffers to wait for
500 * completion. Any other dirty buffers which are not yet queued for
501 * write will not be flushed to disk by the osync.
502 */
504 {
510 repeat:
522 }
523 }
526 }
529 {
534 }
537 {
541 }
543 /**
544 * emergency_thaw_all -- forcibly thaw every frozen filesystem
545 *
546 * Used for emergency unfreeze of all filesystems via SysRq
547 */
549 {
556 }
557 }
559 /**
560 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
561 * @mapping: the mapping which wants those buffers written
562 *
563 * Starts I/O against the buffers at mapping->private_list, and waits upon
564 * that I/O.
565 *
566 * Basically, this is a convenience function for fsync().
567 * @mapping is a file or directory which needs those buffers to be written for
568 * a successful fsync().
569 */
571 {
579 }
582 /*
583 * Called when we've recently written block `bblock', and it is known that
584 * `bblock' was for a buffer_boundary() buffer. This means that the block at
585 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
586 * dirty, schedule it for IO. So that indirects merge nicely with their data.
587 */
590 {
596 }
597 }
600 {
609 }
616 }
617 }
620 /*
621 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
622 * dirty.
623 *
624 * If warn is true, then emit a warning if the page is not uptodate and has
625 * not been truncated.
626 */
629 {
638 }
641 }
643 /*
644 * Add a page to the dirty page list.
645 *
646 * It is a sad fact of life that this function is called from several places
647 * deeply under spinlocking. It may not sleep.
648 *
649 * If the page has buffers, the uptodate buffers are set dirty, to preserve
650 * dirty-state coherency between the page and the buffers. It the page does
651 * not have buffers then when they are later attached they will all be set
652 * dirty.
653 *
654 * The buffers are dirtied before the page is dirtied. There's a small race
655 * window in which a writepage caller may see the page cleanness but not the
656 * buffer dirtiness. That's fine. If this code were to set the page dirty
657 * before the buffers, a concurrent writepage caller could clear the page dirty
658 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
659 * page on the dirty page list.
660 *
661 * We use private_lock to lock against try_to_free_buffers while using the
662 * page's buffer list. Also use this to protect against clean buffers being
663 * added to the page after it was set dirty.
664 *
665 * FIXME: may need to call ->reservepage here as well. That's rather up to the
666 * address_space though.
667 */
669 {
685 }
692 }
695 /*
696 * Write out and wait upon a list of buffers.
697 *
698 * We have conflicting pressures: we want to make sure that all
699 * initially dirty buffers get waited on, but that any subsequently
700 * dirtied buffers don't. After all, we don't want fsync to last
701 * forever if somebody is actively writing to the file.
702 *
703 * Do this in two main stages: first we copy dirty buffers to a
704 * temporary inode list, queueing the writes as we go. Then we clean
705 * up, waiting for those writes to complete.
706 *
707 * During this second stage, any subsequent updates to the file may end
708 * up refiling the buffer on the original inode's dirty list again, so
709 * there is a chance we will end up with a buffer queued for write but
710 * not yet completed on that list. So, as a final cleanup we go through
711 * the osync code to catch these locked, dirty buffers without requeuing
712 * any newly dirty buffers for write.
713 */
715 {
730 /* Avoid race with mark_buffer_dirty_inode() which does
731 * a lockless check and we rely on seeing the dirty bit */
739 /*
740 * Ensure any pending I/O completes so that
741 * write_dirty_buffer() actually writes the
742 * current contents - it is a noop if I/O is
743 * still in flight on potentially older
744 * contents.
745 */
748 /*
749 * Kick off IO for the previous mapping. Note
750 * that we will not run the very last mapping,
751 * wait_on_buffer() will do that for us
752 * through sync_buffer().
753 */
756 }
757 }
758 }
769 /* Avoid race with mark_buffer_dirty_inode() which does
770 * a lockless check and we rely on seeing the dirty bit */
776 }
783 }
789 else
791 }
793 /*
794 * Invalidate any and all dirty buffers on a given inode. We are
795 * probably unmounting the fs, but that doesn't mean we have already
796 * done a sync(). Just drop the buffers from the inode list.
797 *
798 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
799 * assumes that all the buffers are against the blockdev. Not true
800 * for reiserfs.
801 */
803 {
813 }
814 }
817 /*
818 * Remove any clean buffers from the inode's buffer list. This is called
819 * when we're trying to free the inode itself. Those buffers can pin it.
820 *
821 * Returns true if all buffers were removed.
822 */
824 {
838 }
840 }
842 }
844 }
846 /*
847 * Create the appropriate buffers when given a page for data area and
848 * the size of each buffer.. Use the bh->b_this_page linked list to
849 * follow the buffers created. Return NULL if unable to create more
850 * buffers.
851 *
852 * The retry flag is used to differentiate async IO (paging, swapping)
853 * which may not fail from ordinary buffer allocations.
854 */
857 {
861 try_again:
875 /* Link the buffer to its page */
877 }
879 /*
880 * In case anything failed, we just free everything we got.
881 */
882 no_grow:
889 }
891 /*
892 * Return failure for non-async IO requests. Async IO requests
893 * are not allowed to fail, so we have to wait until buffer heads
894 * become available. But we don't want tasks sleeping with
895 * partially complete buffers, so all were released above.
896 */
900 /* We're _really_ low on memory. Now we just
901 * wait for old buffer heads to become free due to
902 * finishing IO. Since this is an async request and
903 * the reserve list is empty, we're sure there are
904 * async buffer heads in use.
905 */
908 }
913 {
923 }
926 {
933 }
935 }
937 /*
938 * Initialise the state of a blockdev page's buffers.
939 */
940 static sector_t
943 {
958 }
959 block++;
963 /*
964 * Caller needs to validate requested block against end of device.
965 */
967 }
969 /*
970 * Create the page-cache page that contains the requested block.
971 *
972 * This is used purely for blockdev mappings.
973 */
974 static int
977 {
981 sector_t end_block;
983 gfp_t gfp_mask;
987 /*
988 * XXX: __getblk_slow() can not really deal with failure and
989 * will endlessly loop on improvised global reclaim. Prefer
990 * looping in the allocator rather than here, at least that
991 * code knows what it's doing.
992 */
1006 size);
1008 }
1011 }
1013 /*
1014 * Allocate some buffers for this page
1015 */
1020 /*
1021 * Link the page to the buffers and initialise them. Take the
1022 * lock to be atomic wrt __find_get_block(), which does not
1023 * run under the page lock.
1024 */
1028 size);
1030 done:
1032 failed:
1036 }
1038 /*
1039 * Create buffers for the specified block device block's page. If
1040 * that page was dirty, the buffers are set dirty also.
1041 */
1042 static int
1044 {
1045 pgoff_t index;
1050 sizebits++;
1055 /*
1056 * Check for a block which wants to lie outside our maximum possible
1057 * pagecache index. (this comparison is done using sector_t types).
1058 */
1067 }
1069 /* Create a page with the proper size buffers.. */
1071 }
1076 {
1077 /* Size must be multiple of hard sectorsize */
1081 size);
1087 }
1102 }
1103 }
1106 /*
1107 * The relationship between dirty buffers and dirty pages:
1108 *
1109 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1110 * the page is tagged dirty in its radix tree.
1111 *
1112 * At all times, the dirtiness of the buffers represents the dirtiness of
1113 * subsections of the page. If the page has buffers, the page dirty bit is
1114 * merely a hint about the true dirty state.
1115 *
1116 * When a page is set dirty in its entirety, all its buffers are marked dirty
1117 * (if the page has buffers).
1118 *
1119 * When a buffer is marked dirty, its page is dirtied, but the page's other
1120 * buffers are not.
1121 *
1122 * Also. When blockdev buffers are explicitly read with bread(), they
1123 * individually become uptodate. But their backing page remains not
1124 * uptodate - even if all of its buffers are uptodate. A subsequent
1125 * block_read_full_page() against that page will discover all the uptodate
1126 * buffers, will set the page uptodate and will perform no I/O.
1127 */
1129 /**
1130 * mark_buffer_dirty - mark a buffer_head as needing writeout
1131 * @bh: the buffer_head to mark dirty
1132 *
1133 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1134 * backing page dirty, then tag the page as dirty in its address_space's radix
1135 * tree and then attach the address_space's inode to its superblock's dirty
1136 * inode list.
1137 *
1138 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1139 * mapping->tree_lock and mapping->host->i_lock.
1140 */
1142 {
1147 /*
1148 * Very *carefully* optimize the it-is-already-dirty case.
1149 *
1150 * Don't let the final "is it dirty" escape to before we
1151 * perhaps modified the buffer.
1152 */
1157 }
1165 }
1166 }
1167 }
1170 /*
1171 * Decrement a buffer_head's reference count. If all buffers against a page
1172 * have zero reference count, are clean and unlocked, and if the page is clean
1173 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1174 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1175 * a page but it ends up not being freed, and buffers may later be reattached).
1176 */
1178 {
1182 }
1184 }
1187 /*
1188 * bforget() is like brelse(), except it discards any
1189 * potentially dirty data.
1190 */
1192 {
1201 }
1203 }
1207 {
1219 }
1222 }
1224 /*
1225 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1226 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1227 * refcount elevated by one when they're in an LRU. A buffer can only appear
1228 * once in a particular CPU's LRU. A single buffer can be present in multiple
1229 * CPU's LRUs at the same time.
1230 *
1231 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1232 * sb_find_get_block().
1233 *
1234 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1235 * a local interrupt disable for that.
1236 */
1238 #define BH_LRU_SIZE 16
1242 };
1246 #ifdef CONFIG_SMP
1247 #define bh_lru_lock() local_irq_disable()
1248 #define bh_lru_unlock() local_irq_enable()
1249 #else
1250 #define bh_lru_lock() preempt_disable()
1251 #define bh_lru_unlock() preempt_enable()
1252 #endif
1255 {
1256 #ifdef irqs_disabled
1258 #endif
1259 }
1261 /*
1262 * The LRU management algorithm is dopey-but-simple. Sorry.
1263 */
1265 {
1289 }
1290 }
1291 }
1295 }
1300 }
1302 /*
1303 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1304 */
1307 {
1322 i--;
1323 }
1325 }
1329 }
1330 }
1333 }
1335 /*
1336 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1337 * it in the LRU and mark it as accessed. If it is not present then return
1338 * NULL
1339 */
1342 {
1346 /* __find_get_block_slow will mark the page accessed */
1354 }
1357 /*
1358 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1359 * which corresponds to the passed block_device, block and size. The
1360 * returned buffer has its reference count incremented.
1361 *
1362 * __getblk_gfp() will lock up the machine if grow_dev_page's
1363 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1364 */
1368 {
1375 }
1378 /*
1379 * Do async read-ahead on a buffer..
1380 */
1382 {
1387 }
1388 }
1391 /**
1392 * __bread_gfp() - reads a specified block and returns the bh
1393 * @bdev: the block_device to read from
1394 * @block: number of block
1395 * @size: size (in bytes) to read
1396 * @gfp: page allocation flag
1397 *
1398 * Reads a specified block, and returns buffer head that contains it.
1399 * The page cache can be allocated from non-movable area
1400 * not to prevent page migration if you set gfp to zero.
1401 * It returns NULL if the block was unreadable.
1402 */
1406 {
1412 }
1415 /*
1416 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1417 * This doesn't race because it runs in each cpu either in irq
1418 * or with preempt disabled.
1419 */
1421 {
1428 }
1430 }
1433 {
1440 }
1443 }
1446 {
1448 }
1453 {
1457 /*
1458 * This catches illegal uses and preserves the offset:
1459 */
1461 else
1463 }
1466 /*
1467 * Called when truncating a buffer on a page completely.
1468 */
1470 /* Bits that are cleared during an invalidate */
1471 #define BUFFER_FLAGS_DISCARD \
1472 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1473 1 << BH_Delay | 1 << BH_Unwritten)
1476 {
1489 }
1491 }
1493 /**
1494 * block_invalidatepage - invalidate part or all of a buffer-backed page
1495 *
1496 * @page: the page which is affected
1497 * @offset: start of the range to invalidate
1498 * @length: length of the range to invalidate
1499 *
1500 * block_invalidatepage() is called when all or part of the page has become
1501 * invalidated by a truncate operation.
1502 *
1503 * block_invalidatepage() does not have to release all buffers, but it must
1504 * ensure that no dirty buffer is left outside @offset and that no I/O
1505 * is underway against any of the blocks which are outside the truncation
1506 * point. Because the caller is about to free (and possibly reuse) those
1507 * blocks on-disk.
1508 */
1511 {
1520 /*
1521 * Check for overflow
1522 */
1531 /*
1532 * Are we still fully in range ?
1533 */
1537 /*
1538 * is this block fully invalidated?
1539 */
1546 /*
1547 * We release buffers only if the entire page is being invalidated.
1548 * The get_block cached value has been unconditionally invalidated,
1549 * so real IO is not possible anymore.
1550 */
1553 out:
1555 }
1559 /*
1560 * We attach and possibly dirty the buffers atomically wrt
1561 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1562 * is already excluded via the page lock.
1563 */
1566 {
1588 }
1591 }
1594 /*
1595 * We are taking a block for data and we don't want any output from any
1596 * buffer-cache aliases starting from return from that function and
1597 * until the moment when something will explicitly mark the buffer
1598 * dirty (hopefully that will not happen until we will free that block ;-)
1599 * We don't even need to mark it not-uptodate - nobody can expect
1600 * anything from a newly allocated buffer anyway. We used to used
1601 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1602 * don't want to mark the alias unmapped, for example - it would confuse
1603 * anyone who might pick it with bread() afterwards...
1604 *
1605 * Also.. Note that bforget() doesn't lock the buffer. So there can
1606 * be writeout I/O going on against recently-freed buffers. We don't
1607 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1608 * only if we really need to. That happens here.
1609 */
1611 {
1622 }
1623 }
1626 /*
1627 * Size is a power-of-two in the range 512..PAGE_SIZE,
1628 * and the case we care about most is PAGE_SIZE.
1629 *
1630 * So this *could* possibly be written with those
1631 * constraints in mind (relevant mostly if some
1632 * architecture has a slow bit-scan instruction)
1633 */
1635 {
1637 }
1639 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1640 {
1646 }
1648 /*
1649 * NOTE! All mapped/uptodate combinations are valid:
1650 *
1651 * Mapped Uptodate Meaning
1652 *
1653 * No No "unknown" - must do get_block()
1654 * No Yes "hole" - zero-filled
1655 * Yes No "allocated" - allocated on disk, not read in
1656 * Yes Yes "valid" - allocated and up-to-date in memory.
1657 *
1658 * "Dirty" is valid only with the last case (mapped+uptodate).
1659 */
1661 /*
1662 * While block_write_full_page is writing back the dirty buffers under
1663 * the page lock, whoever dirtied the buffers may decide to clean them
1664 * again at any time. We handle that by only looking at the buffer
1665 * state inside lock_buffer().
1666 *
1667 * If block_write_full_page() is called for regular writeback
1668 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1669 * locked buffer. This only can happen if someone has written the buffer
1670 * directly, with submit_bh(). At the address_space level PageWriteback
1671 * prevents this contention from occurring.
1672 *
1673 * If block_write_full_page() is called with wbc->sync_mode ==
1674 * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1675 * causes the writes to be flagged as synchronous writes.
1676 */
1680 {
1682 sector_t block;
1683 sector_t last_block;
1693 /*
1694 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1695 * here, and the (potentially unmapped) buffers may become dirty at
1696 * any time. If a buffer becomes dirty here after we've inspected it
1697 * then we just miss that fact, and the page stays dirty.
1698 *
1699 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1700 * handle that here by just cleaning them.
1701 */
1710 /*
1711 * Get all the dirty buffers mapped to disk addresses and
1712 * handle any aliases from the underlying blockdev's mapping.
1713 */
1716 /*
1717 * mapped buffers outside i_size will occur, because
1718 * this page can be outside i_size when there is a
1719 * truncate in progress.
1720 */
1721 /*
1722 * The buffer was zeroed by block_write_full_page()
1723 */
1734 /* blockdev mappings never come here */
1738 }
1739 }
1741 block++;
1747 /*
1748 * If it's a fully non-blocking write attempt and we cannot
1749 * lock the buffer then redirty the page. Note that this can
1750 * potentially cause a busy-wait loop from writeback threads
1751 * and kswapd activity, but those code paths have their own
1752 * higher-level throttling.
1753 */
1759 }
1764 }
1767 /*
1768 * The page and its buffers are protected by PageWriteback(), so we can
1769 * drop the bh refcounts early.
1770 */
1778 nr_underway++;
1779 }
1785 done:
1787 /*
1788 * The page was marked dirty, but the buffers were
1789 * clean. Someone wrote them back by hand with
1790 * ll_rw_block/submit_bh. A rare case.
1791 */
1794 /*
1795 * The page and buffer_heads can be released at any time from
1796 * here on.
1797 */
1798 }
1801 recover:
1802 /*
1803 * ENOSPC, or some other error. We may already have added some
1804 * blocks to the file, so we need to write these out to avoid
1805 * exposing stale data.
1806 * The page is currently locked and not marked for writeback
1807 */
1809 /* Recovery: lock and submit the mapped buffers */
1816 /*
1817 * The buffer may have been set dirty during
1818 * attachment to a dirty page.
1819 */
1821 }
1832 nr_underway++;
1833 }
1838 }
1840 /*
1841 * If a page has any new buffers, zero them out here, and mark them uptodate
1842 * and dirty so they'll be written out (in order to prevent uninitialised
1843 * block data from leaking). And clear the new bit.
1844 */
1846 {
1869 }
1873 }
1874 }
1879 }
1884 {
1889 sector_t block;
1912 }
1914 }
1930 }
1936 }
1937 }
1942 }
1948 }
1949 }
1950 /*
1951 * If we issued read requests - let them complete.
1952 */
1957 }
1961 }
1966 {
1984 }
1991 /*
1992 * If this is a partial write which happened to make all buffers
1993 * uptodate then we can optimize away a bogus readpage() for
1994 * the next read(). Here we 'discover' whether the page went
1995 * uptodate as a result of this (potentially partial) write.
1996 */
2000 }
2002 /*
2003 * block_write_begin takes care of the basic task of block allocation and
2004 * bringing partial write blocks uptodate first.
2005 *
2006 * The filesystem needs to handle block truncation upon failure.
2007 */
2010 {
2024 }
2028 }
2034 {
2041 /*
2042 * The buffers that were written will now be uptodate, so we
2043 * don't have to worry about a readpage reading them and
2044 * overwriting a partial write. However if we have encountered
2045 * a short write and only partially written into a buffer, it
2046 * will not be marked uptodate, so a readpage might come in and
2047 * destroy our partial write.
2048 *
2049 * Do the simplest thing, and just treat any short write to a
2050 * non uptodate page as a zero-length write, and force the
2051 * caller to redo the whole thing.
2052 */
2057 }
2060 /* This could be a short (even 0-length) commit */
2064 }
2070 {
2077 /*
2078 * No need to use i_size_read() here, the i_size
2079 * cannot change under us because we hold i_mutex.
2080 *
2081 * But it's important to update i_size while still holding page lock:
2082 * page writeout could otherwise come in and zero beyond i_size.
2083 */
2087 }
2094 /*
2095 * Don't mark the inode dirty under page lock. First, it unnecessarily
2096 * makes the holding time of page lock longer. Second, it forces lock
2097 * ordering of page lock and transaction start for journaling
2098 * filesystems.
2099 */
2104 }
2107 /*
2108 * block_is_partially_uptodate checks whether buffers within a page are
2109 * uptodate or not.
2110 *
2111 * Returns true if all buffers which correspond to a file portion
2112 * we want to read are uptodate.
2113 */
2116 {
2140 }
2143 }
2149 }
2152 /*
2153 * Generic "read page" function for block devices that have the normal
2154 * get_block functionality. This is most of the block device filesystems.
2155 * Reads the page asynchronously --- the unlock_buffer() and
2156 * set/clear_buffer_uptodate() functions propagate buffer state into the
2157 * page struct once IO has completed.
2158 */
2160 {
2191 }
2197 }
2198 /*
2199 * get_block() might have updated the buffer
2200 * synchronously
2201 */
2204 }
2212 /*
2213 * All buffers are uptodate - we can set the page uptodate
2214 * as well. But not if get_block() returned an error.
2215 */
2220 }
2222 /* Stage two: lock the buffers */
2227 }
2229 /*
2230 * Stage 3: start the IO. Check for uptodateness
2231 * inside the buffer lock in case another process reading
2232 * the underlying blockdev brought it uptodate (the sct fix).
2233 */
2238 else
2240 }
2242 }
2245 /* utility function for filesystems that need to do work on expanding
2246 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2247 * deal with the hole.
2248 */
2250 {
2269 out:
2271 }
2276 {
2282 loff_t curpos;
2294 }
2298 AOP_FLAG_UNINTERRUPTIBLE,
2315 }
2316 }
2318 /* page covers the boundary, find the boundary offset */
2321 /* if we will expand the thing last block will be filled */
2324 }
2328 }
2332 AOP_FLAG_UNINTERRUPTIBLE,
2343 }
2344 out:
2346 }
2348 /*
2349 * For moronic filesystems that do not allow holes in file.
2350 * We may have to extend the file.
2351 */
2356 {
2370 }
2373 }
2377 {
2381 }
2384 /*
2385 * block_page_mkwrite() is not allowed to change the file size as it gets
2386 * called from a page fault handler when a page is first dirtied. Hence we must
2387 * be careful to check for EOF conditions here. We set the page up correctly
2388 * for a written page which means we get ENOSPC checking when writing into
2389 * holes and correct delalloc and unwritten extent mapping on filesystems that
2390 * support these features.
2391 *
2392 * We are not allowed to take the i_mutex here so we have to play games to
2393 * protect against truncate races as the page could now be beyond EOF. Because
2394 * truncate writes the inode size before removing pages, once we have the
2395 * page lock we can determine safely if the page is beyond EOF. If it is not
2396 * beyond EOF, then the page is guaranteed safe against truncation until we
2397 * unlock the page.
2398 *
2399 * Direct callers of this function should protect against filesystem freezing
2400 * using sb_start_write() - sb_end_write() functions.
2401 */
2403 get_block_t get_block)
2404 {
2408 loff_t size;
2415 /* We overload EFAULT to mean page got truncated */
2418 }
2420 /* page is wholly or partially inside EOF */
2423 else
2435 out_unlock:
2438 }
2442 get_block_t get_block)
2443 {
2449 /*
2450 * Update file times before taking page lock. We may end up failing the
2451 * fault so this update may be superfluous but who really cares...
2452 */
2458 }
2461 /*
2462 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2463 * immediately, while under the page lock. So it needs a special end_io
2464 * handler which does not touch the bh after unlocking it.
2465 */
2467 {
2469 }
2471 /*
2472 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2473 * the page (converting it to circular linked list and taking care of page
2474 * dirty races).
2475 */
2477 {
2493 }
2495 /*
2496 * On entry, the page is fully not uptodate.
2497 * On exit the page is fully uptodate in the areas outside (from,to)
2498 * The filesystem needs to handle block truncation upon failure.
2499 */
2504 {
2510 pgoff_t index;
2514 sector_t block_in_file;
2534 }
2539 /*
2540 * Allocate buffers so that we can keep track of state, and potentially
2541 * attach them to the page if an error occurs. In the common case of
2542 * no error, they will just be freed again without ever being attached
2543 * to the page (which is all OK, because we're under the page lock).
2544 *
2545 * Be careful: the buffer linked list is a NULL terminated one, rather
2546 * than the circular one we're used to.
2547 */
2552 }
2556 /*
2557 * We loop across all blocks in the page, whether or not they are
2558 * part of the affected region. This is so we can discover if the
2559 * page is fully mapped-to-disk.
2560 */
2582 }
2587 }
2594 nr_reads++;
2595 }
2596 }
2599 /*
2600 * The page is locked, so these buffers are protected from
2601 * any VM or truncate activity. Hence we don't need to care
2602 * for the buffer_head refcounts.
2603 */
2608 }
2611 }
2620 failed:
2622 /*
2623 * Error recovery is a bit difficult. We need to zero out blocks that
2624 * were newly allocated, and dirty them to ensure they get written out.
2625 * Buffers need to be attached to the page at this point, otherwise
2626 * the handling of potential IO errors during writeout would be hard
2627 * (could try doing synchronous writeout, but what if that fails too?)
2628 */
2632 out_release:
2638 }
2644 {
2661 }
2670 }
2673 }
2676 /*
2677 * nobh_writepage() - based on block_full_write_page() except
2678 * that it tries to operate without attaching bufferheads to
2679 * the page.
2680 */
2683 {
2690 /* Is the page fully inside i_size? */
2694 /* Is the page fully outside i_size? (truncate in progress) */
2697 /*
2698 * The page may have dirty, unmapped buffers. For example,
2699 * they may have been added in ext3_writepage(). Make them
2700 * freeable here, so the page does not leak.
2701 */
2702 #if 0
2703 /* Not really sure about this - do we need this ? */
2706 #endif
2709 }
2711 /*
2712 * The page straddles i_size. It must be zeroed out on each and every
2713 * writepage invocation because it may be mmapped. "A file is mapped
2714 * in multiples of the page size. For a file that is not a multiple of
2715 * the page size, the remaining memory is zeroed when mapped, and
2716 * writes to that region are not written out to the file."
2717 */
2719 out:
2723 end_buffer_async_write);
2725 }
2730 {
2734 sector_t iblock;
2744 /* Block boundary? Nothing to do */
2757 has_buffers:
2761 }
2763 /* Find the buffer that contains "offset" */
2766 iblock++;
2768 }
2775 /* unmapped? It's a hole - nothing to do */
2779 /* Ok, it's mapped. Make sure it's up-to-date */
2785 }
2790 }
2793 }
2798 unlock:
2801 out:
2803 }
2808 {
2812 sector_t iblock;
2822 /* Block boundary? Nothing to do */
2837 /* Find the buffer that contains "offset" */
2842 iblock++;
2844 }
2852 /* unmapped? It's a hole - nothing to do */
2855 }
2857 /* Ok, it's mapped. Make sure it's up-to-date */
2865 /* Uhhuh. Read error. Complain and punt. */
2868 }
2874 unlock:
2877 out:
2879 }
2882 /*
2883 * The generic ->writepage function for buffer-backed address_spaces
2884 */
2887 {
2893 /* Is the page fully inside i_size? */
2896 end_buffer_async_write);
2898 /* Is the page fully outside i_size? (truncate in progress) */
2901 /*
2902 * The page may have dirty, unmapped buffers. For example,
2903 * they may have been added in ext3_writepage(). Make them
2904 * freeable here, so the page does not leak.
2905 */
2909 }
2911 /*
2912 * The page straddles i_size. It must be zeroed out on each and every
2913 * writepage invocation because it may be mmapped. "A file is mapped
2914 * in multiples of the page size. For a file that is not a multiple of
2915 * the page size, the remaining memory is zeroed when mapped, and
2916 * writes to that region are not written out to the file."
2917 */
2920 end_buffer_async_write);
2921 }
2926 {
2934 }
2938 {
2943 }
2950 }
2952 /*
2953 * This allows us to do IO even on the odd last sectors
2954 * of a device, even if the block size is some multiple
2955 * of the physical sector size.
2956 *
2957 * We'll just truncate the bio to the size of the device,
2958 * and clear the end of the buffer head manually.
2959 *
2960 * Truly out-of-range accesses will turn into actual IO
2961 * errors, this only handles the "we need to be able to
2962 * do IO at the final sector" case.
2963 */
2965 {
2966 sector_t maxsector;
2974 /*
2975 * If the *whole* IO is past the end of the device,
2976 * let it through, and the IO layer will turn it into
2977 * an EIO.
2978 */
2986 /* Uhhuh. We've got a bio that straddles the device size! */
2989 /* Truncate the bio.. */
2993 /* ..and clear the end of the buffer for reads */
2996 truncated_bytes);
2997 }
2998 }
3001 {
3011 /*
3012 * Only clear out a write error when rewriting
3013 */
3017 /*
3018 * from here on down, it's all bio -- do the initial mapping,
3019 * submit_bio -> generic_make_request may further map this bio around
3020 */
3036 /* Take care of bh's that straddle the end of the device */
3052 }
3056 {
3058 }
3061 /**
3062 * ll_rw_block: low-level access to block devices (DEPRECATED)
3063 * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
3064 * @nr: number of &struct buffer_heads in the array
3065 * @bhs: array of pointers to &struct buffer_head
3066 *
3067 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3068 * requests an I/O operation on them, either a %READ or a %WRITE. The third
3069 * %READA option is described in the documentation for generic_make_request()
3070 * which ll_rw_block() calls.
3071 *
3072 * This function drops any buffer that it cannot get a lock on (with the
3073 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3074 * request, and any buffer that appears to be up-to-date when doing read
3075 * request. Further it marks as clean buffers that are processed for
3076 * writing (the buffer cache won't assume that they are actually clean
3077 * until the buffer gets unlocked).
3078 *
3079 * ll_rw_block sets b_end_io to simple completion handler that marks
3080 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3081 * any waiters.
3082 *
3083 * All of the buffers must be for the same device, and must also be a
3084 * multiple of the current approved size for the device.
3085 */
3087 {
3101 }
3108 }
3109 }
3111 }
3112 }
3116 {
3121 }
3125 }
3128 /*
3129 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3130 * and then start new I/O and then wait upon it. The caller must have a ref on
3131 * the buffer_head.
3132 */
3134 {
3148 }
3150 }
3154 {
3156 }
3159 /*
3160 * try_to_free_buffers() checks if all the buffers on this particular page
3161 * are unused, and releases them if so.
3162 *
3163 * Exclusion against try_to_free_buffers may be obtained by either
3164 * locking the page or by holding its mapping's private_lock.
3165 *
3166 * If the page is dirty but all the buffers are clean then we need to
3167 * be sure to mark the page clean as well. This is because the page
3168 * may be against a block device, and a later reattachment of buffers
3169 * to a dirty page will set *all* buffers dirty. Which would corrupt
3170 * filesystem data on the same device.
3171 *
3172 * The same applies to regular filesystem pages: if all the buffers are
3173 * clean then we set the page clean and proceed. To do that, we require
3174 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3175 * private_lock.
3176 *
3177 * try_to_free_buffers() is non-blocking.
3178 */
3180 {
3183 }
3185 static int
3187 {
3210 failed:
3212 }
3215 {
3227 }
3232 /*
3233 * If the filesystem writes its buffers by hand (eg ext3)
3234 * then we can have clean buffers against a dirty page. We
3235 * clean the page here; otherwise the VM will never notice
3236 * that the filesystem did any IO at all.
3237 *
3238 * Also, during truncate, discard_buffer will have marked all
3239 * the page's buffers clean. We discover that here and clean
3240 * the page also.
3241 *
3242 * private_lock must be held over this entire operation in order
3243 * to synchronise against __set_page_dirty_buffers and prevent the
3244 * dirty bit from being lost.
3245 */
3249 out:
3258 }
3260 }
3263 /*
3264 * There are no bdflush tunables left. But distributions are
3265 * still running obsolete flush daemons, so we terminate them here.
3266 *
3267 * Use of bdflush() is deprecated and will be removed in a future kernel.
3268 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3269 */
3271 {
3278 msg_count++;
3280 "warning: process `%s' used the obsolete bdflush"
3283 }
3288 }
3290 /*
3291 * Buffer-head allocation
3292 */
3295 /*
3296 * Once the number of bh's in the machine exceeds this level, we start
3297 * stripping them in writeback.
3298 */
3306 };
3311 {
3321 }
3324 {
3332 }
3334 }
3338 {
3345 }
3349 {
3356 }
3359 }
3363 {
3367 }
3369 /**
3370 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3371 * @bh: struct buffer_head
3372 *
3373 * Return true if the buffer is up-to-date and false,
3374 * with the buffer locked, if not.
3375 */
3377 {
3383 }
3385 }
3388 /**
3389 * bh_submit_read - Submit a locked buffer for reading
3390 * @bh: struct buffer_head
3391 *
3392 * Returns zero on success and -EIO on error.
3393 */
3395 {
3401 }
3410 }
3414 {
3420 SLAB_MEM_SPREAD),
3421 NULL);
3423 /*
3424 * Limit the bh occupancy to 10% of ZONE_NORMAL
3425 */
3429 }