| blob | eba6e4f621ce0946ddccb250103656654c0844ee |
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 */
1034 }
1037 }
1039 /*
1040 * Allocate some buffers for this page
1041 */
1046 /*
1047 * Link the page to the buffers and initialise them. Take the
1048 * lock to be atomic wrt __find_get_block(), which does not
1049 * run under the page lock.
1050 */
1055 done:
1057 failed:
1061 }
1063 /*
1064 * Create buffers for the specified block device block's page. If
1065 * that page was dirty, the buffers are set dirty also.
1066 */
1067 static int
1069 {
1070 pgoff_t index;
1075 sizebits++;
1080 /*
1081 * Check for a block which wants to lie outside our maximum possible
1082 * pagecache index. (this comparison is done using sector_t types).
1083 */
1092 }
1094 /* Create a page with the proper size buffers.. */
1096 }
1100 {
1101 /* Size must be multiple of hard sectorsize */
1105 size);
1111 }
1126 }
1127 }
1129 /*
1130 * The relationship between dirty buffers and dirty pages:
1131 *
1132 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1133 * the page is tagged dirty in its radix tree.
1134 *
1135 * At all times, the dirtiness of the buffers represents the dirtiness of
1136 * subsections of the page. If the page has buffers, the page dirty bit is
1137 * merely a hint about the true dirty state.
1138 *
1139 * When a page is set dirty in its entirety, all its buffers are marked dirty
1140 * (if the page has buffers).
1141 *
1142 * When a buffer is marked dirty, its page is dirtied, but the page's other
1143 * buffers are not.
1144 *
1145 * Also. When blockdev buffers are explicitly read with bread(), they
1146 * individually become uptodate. But their backing page remains not
1147 * uptodate - even if all of its buffers are uptodate. A subsequent
1148 * block_read_full_page() against that page will discover all the uptodate
1149 * buffers, will set the page uptodate and will perform no I/O.
1150 */
1152 /**
1153 * mark_buffer_dirty - mark a buffer_head as needing writeout
1154 * @bh: the buffer_head to mark dirty
1155 *
1156 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1157 * backing page dirty, then tag the page as dirty in its address_space's radix
1158 * tree and then attach the address_space's inode to its superblock's dirty
1159 * inode list.
1160 *
1161 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1162 * mapping->tree_lock and mapping->host->i_lock.
1163 */
1165 {
1170 /*
1171 * Very *carefully* optimize the it-is-already-dirty case.
1172 *
1173 * Don't let the final "is it dirty" escape to before we
1174 * perhaps modified the buffer.
1175 */
1180 }
1188 }
1189 }
1190 }
1193 /*
1194 * Decrement a buffer_head's reference count. If all buffers against a page
1195 * have zero reference count, are clean and unlocked, and if the page is clean
1196 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1197 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1198 * a page but it ends up not being freed, and buffers may later be reattached).
1199 */
1201 {
1205 }
1207 }
1210 /*
1211 * bforget() is like brelse(), except it discards any
1212 * potentially dirty data.
1213 */
1215 {
1224 }
1226 }
1230 {
1242 }
1245 }
1247 /*
1248 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1249 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1250 * refcount elevated by one when they're in an LRU. A buffer can only appear
1251 * once in a particular CPU's LRU. A single buffer can be present in multiple
1252 * CPU's LRUs at the same time.
1253 *
1254 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1255 * sb_find_get_block().
1256 *
1257 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1258 * a local interrupt disable for that.
1259 */
1261 #define BH_LRU_SIZE 8
1265 };
1269 #ifdef CONFIG_SMP
1270 #define bh_lru_lock() local_irq_disable()
1271 #define bh_lru_unlock() local_irq_enable()
1272 #else
1273 #define bh_lru_lock() preempt_disable()
1274 #define bh_lru_unlock() preempt_enable()
1275 #endif
1278 {
1279 #ifdef irqs_disabled
1281 #endif
1282 }
1284 /*
1285 * The LRU management algorithm is dopey-but-simple. Sorry.
1286 */
1288 {
1312 }
1313 }
1314 }
1318 }
1323 }
1325 /*
1326 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1327 */
1330 {
1345 i--;
1346 }
1348 }
1352 }
1353 }
1356 }
1358 /*
1359 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1360 * it in the LRU and mark it as accessed. If it is not present then return
1361 * NULL
1362 */
1365 {
1369 /* __find_get_block_slow will mark the page accessed */
1377 }
1380 /*
1381 * __getblk will locate (and, if necessary, create) the buffer_head
1382 * which corresponds to the passed block_device, block and size. The
1383 * returned buffer has its reference count incremented.
1384 *
1385 * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1386 * attempt is failing. FIXME, perhaps?
1387 */
1390 {
1397 }
1400 /*
1401 * Do async read-ahead on a buffer..
1402 */
1404 {
1409 }
1410 }
1413 /**
1414 * __bread() - reads a specified block and returns the bh
1415 * @bdev: the block_device to read from
1416 * @block: number of block
1417 * @size: size (in bytes) to read
1418 *
1419 * Reads a specified block, and returns buffer head that contains it.
1420 * It returns NULL if the block was unreadable.
1421 */
1424 {
1430 }
1433 /*
1434 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1435 * This doesn't race because it runs in each cpu either in irq
1436 * or with preempt disabled.
1437 */
1439 {
1446 }
1448 }
1451 {
1458 }
1461 }
1464 {
1466 }
1471 {
1475 /*
1476 * This catches illegal uses and preserves the offset:
1477 */
1479 else
1481 }
1484 /*
1485 * Called when truncating a buffer on a page completely.
1486 */
1488 /* Bits that are cleared during an invalidate */
1489 #define BUFFER_FLAGS_DISCARD \
1490 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1491 1 << BH_Delay | 1 << BH_Unwritten)
1494 {
1507 }
1509 }
1511 /**
1512 * block_invalidatepage - invalidate part or all of a buffer-backed page
1513 *
1514 * @page: the page which is affected
1515 * @offset: start of the range to invalidate
1516 * @length: length of the range to invalidate
1517 *
1518 * block_invalidatepage() is called when all or part of the page has become
1519 * invalidated by a truncate operation.
1520 *
1521 * block_invalidatepage() does not have to release all buffers, but it must
1522 * ensure that no dirty buffer is left outside @offset and that no I/O
1523 * is underway against any of the blocks which are outside the truncation
1524 * point. Because the caller is about to free (and possibly reuse) those
1525 * blocks on-disk.
1526 */
1529 {
1538 /*
1539 * Check for overflow
1540 */
1549 /*
1550 * Are we still fully in range ?
1551 */
1555 /*
1556 * is this block fully invalidated?
1557 */
1564 /*
1565 * We release buffers only if the entire page is being invalidated.
1566 * The get_block cached value has been unconditionally invalidated,
1567 * so real IO is not possible anymore.
1568 */
1571 out:
1573 }
1577 /*
1578 * We attach and possibly dirty the buffers atomically wrt
1579 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1580 * is already excluded via the page lock.
1581 */
1584 {
1606 }
1609 }
1612 /*
1613 * We are taking a block for data and we don't want any output from any
1614 * buffer-cache aliases starting from return from that function and
1615 * until the moment when something will explicitly mark the buffer
1616 * dirty (hopefully that will not happen until we will free that block ;-)
1617 * We don't even need to mark it not-uptodate - nobody can expect
1618 * anything from a newly allocated buffer anyway. We used to used
1619 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1620 * don't want to mark the alias unmapped, for example - it would confuse
1621 * anyone who might pick it with bread() afterwards...
1622 *
1623 * Also.. Note that bforget() doesn't lock the buffer. So there can
1624 * be writeout I/O going on against recently-freed buffers. We don't
1625 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1626 * only if we really need to. That happens here.
1627 */
1629 {
1640 }
1641 }
1644 /*
1645 * Size is a power-of-two in the range 512..PAGE_SIZE,
1646 * and the case we care about most is PAGE_SIZE.
1647 *
1648 * So this *could* possibly be written with those
1649 * constraints in mind (relevant mostly if some
1650 * architecture has a slow bit-scan instruction)
1651 */
1653 {
1655 }
1657 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1658 {
1664 }
1666 /*
1667 * NOTE! All mapped/uptodate combinations are valid:
1668 *
1669 * Mapped Uptodate Meaning
1670 *
1671 * No No "unknown" - must do get_block()
1672 * No Yes "hole" - zero-filled
1673 * Yes No "allocated" - allocated on disk, not read in
1674 * Yes Yes "valid" - allocated and up-to-date in memory.
1675 *
1676 * "Dirty" is valid only with the last case (mapped+uptodate).
1677 */
1679 /*
1680 * While block_write_full_page is writing back the dirty buffers under
1681 * the page lock, whoever dirtied the buffers may decide to clean them
1682 * again at any time. We handle that by only looking at the buffer
1683 * state inside lock_buffer().
1684 *
1685 * If block_write_full_page() is called for regular writeback
1686 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1687 * locked buffer. This only can happen if someone has written the buffer
1688 * directly, with submit_bh(). At the address_space level PageWriteback
1689 * prevents this contention from occurring.
1690 *
1691 * If block_write_full_page() is called with wbc->sync_mode ==
1692 * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1693 * causes the writes to be flagged as synchronous writes.
1694 */
1698 {
1700 sector_t block;
1701 sector_t last_block;
1711 /*
1712 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1713 * here, and the (potentially unmapped) buffers may become dirty at
1714 * any time. If a buffer becomes dirty here after we've inspected it
1715 * then we just miss that fact, and the page stays dirty.
1716 *
1717 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1718 * handle that here by just cleaning them.
1719 */
1728 /*
1729 * Get all the dirty buffers mapped to disk addresses and
1730 * handle any aliases from the underlying blockdev's mapping.
1731 */
1734 /*
1735 * mapped buffers outside i_size will occur, because
1736 * this page can be outside i_size when there is a
1737 * truncate in progress.
1738 */
1739 /*
1740 * The buffer was zeroed by block_write_full_page()
1741 */
1752 /* blockdev mappings never come here */
1756 }
1757 }
1759 block++;
1765 /*
1766 * If it's a fully non-blocking write attempt and we cannot
1767 * lock the buffer then redirty the page. Note that this can
1768 * potentially cause a busy-wait loop from writeback threads
1769 * and kswapd activity, but those code paths have their own
1770 * higher-level throttling.
1771 */
1777 }
1782 }
1785 /*
1786 * The page and its buffers are protected by PageWriteback(), so we can
1787 * drop the bh refcounts early.
1788 */
1796 nr_underway++;
1797 }
1803 done:
1805 /*
1806 * The page was marked dirty, but the buffers were
1807 * clean. Someone wrote them back by hand with
1808 * ll_rw_block/submit_bh. A rare case.
1809 */
1812 /*
1813 * The page and buffer_heads can be released at any time from
1814 * here on.
1815 */
1816 }
1819 recover:
1820 /*
1821 * ENOSPC, or some other error. We may already have added some
1822 * blocks to the file, so we need to write these out to avoid
1823 * exposing stale data.
1824 * The page is currently locked and not marked for writeback
1825 */
1827 /* Recovery: lock and submit the mapped buffers */
1834 /*
1835 * The buffer may have been set dirty during
1836 * attachment to a dirty page.
1837 */
1839 }
1850 nr_underway++;
1851 }
1856 }
1858 /*
1859 * If a page has any new buffers, zero them out here, and mark them uptodate
1860 * and dirty so they'll be written out (in order to prevent uninitialised
1861 * block data from leaking). And clear the new bit.
1862 */
1864 {
1887 }
1891 }
1892 }
1897 }
1902 {
1907 sector_t block;
1930 }
1932 }
1948 }
1954 }
1955 }
1960 }
1966 }
1967 }
1968 /*
1969 * If we issued read requests - let them complete.
1970 */
1975 }
1979 }
1984 {
2002 }
2009 /*
2010 * If this is a partial write which happened to make all buffers
2011 * uptodate then we can optimize away a bogus readpage() for
2012 * the next read(). Here we 'discover' whether the page went
2013 * uptodate as a result of this (potentially partial) write.
2014 */
2018 }
2020 /*
2021 * block_write_begin takes care of the basic task of block allocation and
2022 * bringing partial write blocks uptodate first.
2023 *
2024 * The filesystem needs to handle block truncation upon failure.
2025 */
2028 {
2042 }
2046 }
2052 {
2059 /*
2060 * The buffers that were written will now be uptodate, so we
2061 * don't have to worry about a readpage reading them and
2062 * overwriting a partial write. However if we have encountered
2063 * a short write and only partially written into a buffer, it
2064 * will not be marked uptodate, so a readpage might come in and
2065 * destroy our partial write.
2066 *
2067 * Do the simplest thing, and just treat any short write to a
2068 * non uptodate page as a zero-length write, and force the
2069 * caller to redo the whole thing.
2070 */
2075 }
2078 /* This could be a short (even 0-length) commit */
2082 }
2088 {
2094 /*
2095 * No need to use i_size_read() here, the i_size
2096 * cannot change under us because we hold i_mutex.
2097 *
2098 * But it's important to update i_size while still holding page lock:
2099 * page writeout could otherwise come in and zero beyond i_size.
2100 */
2104 }
2109 /*
2110 * Don't mark the inode dirty under page lock. First, it unnecessarily
2111 * makes the holding time of page lock longer. Second, it forces lock
2112 * ordering of page lock and transaction start for journaling
2113 * filesystems.
2114 */
2119 }
2122 /*
2123 * block_is_partially_uptodate checks whether buffers within a page are
2124 * uptodate or not.
2125 *
2126 * Returns true if all buffers which correspond to a file portion
2127 * we want to read are uptodate.
2128 */
2131 {
2155 }
2158 }
2164 }
2167 /*
2168 * Generic "read page" function for block devices that have the normal
2169 * get_block functionality. This is most of the block device filesystems.
2170 * Reads the page asynchronously --- the unlock_buffer() and
2171 * set/clear_buffer_uptodate() functions propagate buffer state into the
2172 * page struct once IO has completed.
2173 */
2175 {
2206 }
2212 }
2213 /*
2214 * get_block() might have updated the buffer
2215 * synchronously
2216 */
2219 }
2227 /*
2228 * All buffers are uptodate - we can set the page uptodate
2229 * as well. But not if get_block() returned an error.
2230 */
2235 }
2237 /* Stage two: lock the buffers */
2242 }
2244 /*
2245 * Stage 3: start the IO. Check for uptodateness
2246 * inside the buffer lock in case another process reading
2247 * the underlying blockdev brought it uptodate (the sct fix).
2248 */
2253 else
2255 }
2257 }
2260 /* utility function for filesystems that need to do work on expanding
2261 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2262 * deal with the hole.
2263 */
2265 {
2284 out:
2286 }
2291 {
2297 loff_t curpos;
2309 }
2313 AOP_FLAG_UNINTERRUPTIBLE,
2326 }
2328 /* page covers the boundary, find the boundary offset */
2331 /* if we will expand the thing last block will be filled */
2334 }
2338 }
2342 AOP_FLAG_UNINTERRUPTIBLE,
2353 }
2354 out:
2356 }
2358 /*
2359 * For moronic filesystems that do not allow holes in file.
2360 * We may have to extend the file.
2361 */
2366 {
2380 }
2383 }
2387 {
2391 }
2394 /*
2395 * block_page_mkwrite() is not allowed to change the file size as it gets
2396 * called from a page fault handler when a page is first dirtied. Hence we must
2397 * be careful to check for EOF conditions here. We set the page up correctly
2398 * for a written page which means we get ENOSPC checking when writing into
2399 * holes and correct delalloc and unwritten extent mapping on filesystems that
2400 * support these features.
2401 *
2402 * We are not allowed to take the i_mutex here so we have to play games to
2403 * protect against truncate races as the page could now be beyond EOF. Because
2404 * truncate writes the inode size before removing pages, once we have the
2405 * page lock we can determine safely if the page is beyond EOF. If it is not
2406 * beyond EOF, then the page is guaranteed safe against truncation until we
2407 * unlock the page.
2408 *
2409 * Direct callers of this function should protect against filesystem freezing
2410 * using sb_start_write() - sb_end_write() functions.
2411 */
2413 get_block_t get_block)
2414 {
2418 loff_t size;
2425 /* We overload EFAULT to mean page got truncated */
2428 }
2430 /* page is wholly or partially inside EOF */
2433 else
2445 out_unlock:
2448 }
2452 get_block_t get_block)
2453 {
2459 /*
2460 * Update file times before taking page lock. We may end up failing the
2461 * fault so this update may be superfluous but who really cares...
2462 */
2468 }
2471 /*
2472 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2473 * immediately, while under the page lock. So it needs a special end_io
2474 * handler which does not touch the bh after unlocking it.
2475 */
2477 {
2479 }
2481 /*
2482 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2483 * the page (converting it to circular linked list and taking care of page
2484 * dirty races).
2485 */
2487 {
2503 }
2505 /*
2506 * On entry, the page is fully not uptodate.
2507 * On exit the page is fully uptodate in the areas outside (from,to)
2508 * The filesystem needs to handle block truncation upon failure.
2509 */
2514 {
2520 pgoff_t index;
2524 sector_t block_in_file;
2544 }
2549 /*
2550 * Allocate buffers so that we can keep track of state, and potentially
2551 * attach them to the page if an error occurs. In the common case of
2552 * no error, they will just be freed again without ever being attached
2553 * to the page (which is all OK, because we're under the page lock).
2554 *
2555 * Be careful: the buffer linked list is a NULL terminated one, rather
2556 * than the circular one we're used to.
2557 */
2562 }
2566 /*
2567 * We loop across all blocks in the page, whether or not they are
2568 * part of the affected region. This is so we can discover if the
2569 * page is fully mapped-to-disk.
2570 */
2592 }
2597 }
2604 nr_reads++;
2605 }
2606 }
2609 /*
2610 * The page is locked, so these buffers are protected from
2611 * any VM or truncate activity. Hence we don't need to care
2612 * for the buffer_head refcounts.
2613 */
2618 }
2621 }
2630 failed:
2632 /*
2633 * Error recovery is a bit difficult. We need to zero out blocks that
2634 * were newly allocated, and dirty them to ensure they get written out.
2635 * Buffers need to be attached to the page at this point, otherwise
2636 * the handling of potential IO errors during writeout would be hard
2637 * (could try doing synchronous writeout, but what if that fails too?)
2638 */
2642 out_release:
2648 }
2654 {
2671 }
2680 }
2683 }
2686 /*
2687 * nobh_writepage() - based on block_full_write_page() except
2688 * that it tries to operate without attaching bufferheads to
2689 * the page.
2690 */
2693 {
2700 /* Is the page fully inside i_size? */
2704 /* Is the page fully outside i_size? (truncate in progress) */
2707 /*
2708 * The page may have dirty, unmapped buffers. For example,
2709 * they may have been added in ext3_writepage(). Make them
2710 * freeable here, so the page does not leak.
2711 */
2712 #if 0
2713 /* Not really sure about this - do we need this ? */
2716 #endif
2719 }
2721 /*
2722 * The page straddles i_size. It must be zeroed out on each and every
2723 * writepage invocation because it may be mmapped. "A file is mapped
2724 * in multiples of the page size. For a file that is not a multiple of
2725 * the page size, the remaining memory is zeroed when mapped, and
2726 * writes to that region are not written out to the file."
2727 */
2729 out:
2733 end_buffer_async_write);
2735 }
2740 {
2744 sector_t iblock;
2754 /* Block boundary? Nothing to do */
2767 has_buffers:
2771 }
2773 /* Find the buffer that contains "offset" */
2776 iblock++;
2778 }
2785 /* unmapped? It's a hole - nothing to do */
2789 /* Ok, it's mapped. Make sure it's up-to-date */
2795 }
2800 }
2803 }
2808 unlock:
2811 out:
2813 }
2818 {
2822 sector_t iblock;
2832 /* Block boundary? Nothing to do */
2847 /* Find the buffer that contains "offset" */
2852 iblock++;
2854 }
2862 /* unmapped? It's a hole - nothing to do */
2865 }
2867 /* Ok, it's mapped. Make sure it's up-to-date */
2875 /* Uhhuh. Read error. Complain and punt. */
2878 }
2884 unlock:
2887 out:
2889 }
2892 /*
2893 * The generic ->writepage function for buffer-backed address_spaces
2894 */
2897 {
2903 /* Is the page fully inside i_size? */
2906 end_buffer_async_write);
2908 /* Is the page fully outside i_size? (truncate in progress) */
2911 /*
2912 * The page may have dirty, unmapped buffers. For example,
2913 * they may have been added in ext3_writepage(). Make them
2914 * freeable here, so the page does not leak.
2915 */
2919 }
2921 /*
2922 * The page straddles i_size. It must be zeroed out on each and every
2923 * writepage invocation because it may be mmapped. "A file is mapped
2924 * in multiples of the page size. For a file that is not a multiple of
2925 * the page size, the remaining memory is zeroed when mapped, and
2926 * writes to that region are not written out to the file."
2927 */
2930 end_buffer_async_write);
2931 }
2936 {
2944 }
2948 {
2953 }
2960 }
2962 /*
2963 * This allows us to do IO even on the odd last sectors
2964 * of a device, even if the bh block size is some multiple
2965 * of the physical sector size.
2966 *
2967 * We'll just truncate the bio to the size of the device,
2968 * and clear the end of the buffer head manually.
2969 *
2970 * Truly out-of-range accesses will turn into actual IO
2971 * errors, this only handles the "we need to be able to
2972 * do IO at the final sector" case.
2973 */
2975 {
2976 sector_t maxsector;
2983 /*
2984 * If the *whole* IO is past the end of the device,
2985 * let it through, and the IO layer will turn it into
2986 * an EIO.
2987 */
2996 /* Uhhuh. We've got a bh that straddles the device size! */
2999 /* Truncate the bio.. */
3003 /* ..and clear the end of the buffer for reads */
3009 }
3010 }
3013 {
3023 /*
3024 * Only clear out a write error when rewriting
3025 */
3029 /*
3030 * from here on down, it's all bio -- do the initial mapping,
3031 * submit_bio -> generic_make_request may further map this bio around
3032 */
3048 /* Take care of bh's that straddle the end of the device */
3064 }
3068 {
3070 }
3073 /**
3074 * ll_rw_block: low-level access to block devices (DEPRECATED)
3075 * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
3076 * @nr: number of &struct buffer_heads in the array
3077 * @bhs: array of pointers to &struct buffer_head
3078 *
3079 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3080 * requests an I/O operation on them, either a %READ or a %WRITE. The third
3081 * %READA option is described in the documentation for generic_make_request()
3082 * which ll_rw_block() calls.
3083 *
3084 * This function drops any buffer that it cannot get a lock on (with the
3085 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3086 * request, and any buffer that appears to be up-to-date when doing read
3087 * request. Further it marks as clean buffers that are processed for
3088 * writing (the buffer cache won't assume that they are actually clean
3089 * until the buffer gets unlocked).
3090 *
3091 * ll_rw_block sets b_end_io to simple completion handler that marks
3092 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3093 * any waiters.
3094 *
3095 * All of the buffers must be for the same device, and must also be a
3096 * multiple of the current approved size for the device.
3097 */
3099 {
3113 }
3120 }
3121 }
3123 }
3124 }
3128 {
3133 }
3137 }
3140 /*
3141 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3142 * and then start new I/O and then wait upon it. The caller must have a ref on
3143 * the buffer_head.
3144 */
3146 {
3160 }
3162 }
3166 {
3168 }
3171 /*
3172 * try_to_free_buffers() checks if all the buffers on this particular page
3173 * are unused, and releases them if so.
3174 *
3175 * Exclusion against try_to_free_buffers may be obtained by either
3176 * locking the page or by holding its mapping's private_lock.
3177 *
3178 * If the page is dirty but all the buffers are clean then we need to
3179 * be sure to mark the page clean as well. This is because the page
3180 * may be against a block device, and a later reattachment of buffers
3181 * to a dirty page will set *all* buffers dirty. Which would corrupt
3182 * filesystem data on the same device.
3183 *
3184 * The same applies to regular filesystem pages: if all the buffers are
3185 * clean then we set the page clean and proceed. To do that, we require
3186 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3187 * private_lock.
3188 *
3189 * try_to_free_buffers() is non-blocking.
3190 */
3192 {
3195 }
3197 static int
3199 {
3222 failed:
3224 }
3227 {
3239 }
3244 /*
3245 * If the filesystem writes its buffers by hand (eg ext3)
3246 * then we can have clean buffers against a dirty page. We
3247 * clean the page here; otherwise the VM will never notice
3248 * that the filesystem did any IO at all.
3249 *
3250 * Also, during truncate, discard_buffer will have marked all
3251 * the page's buffers clean. We discover that here and clean
3252 * the page also.
3253 *
3254 * private_lock must be held over this entire operation in order
3255 * to synchronise against __set_page_dirty_buffers and prevent the
3256 * dirty bit from being lost.
3257 */
3261 out:
3270 }
3272 }
3275 /*
3276 * There are no bdflush tunables left. But distributions are
3277 * still running obsolete flush daemons, so we terminate them here.
3278 *
3279 * Use of bdflush() is deprecated and will be removed in a future kernel.
3280 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3281 */
3283 {
3290 msg_count++;
3292 "warning: process `%s' used the obsolete bdflush"
3295 }
3300 }
3302 /*
3303 * Buffer-head allocation
3304 */
3307 /*
3308 * Once the number of bh's in the machine exceeds this level, we start
3309 * stripping them in writeback.
3310 */
3318 };
3323 {
3333 }
3336 {
3344 }
3346 }
3350 {
3357 }
3361 {
3368 }
3371 }
3375 {
3379 }
3381 /**
3382 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3383 * @bh: struct buffer_head
3384 *
3385 * Return true if the buffer is up-to-date and false,
3386 * with the buffer locked, if not.
3387 */
3389 {
3395 }
3397 }
3400 /**
3401 * bh_submit_read - Submit a locked buffer for reading
3402 * @bh: struct buffer_head
3403 *
3404 * Returns zero on success and -EIO on error.
3405 */
3407 {
3413 }
3422 }
3426 {
3432 SLAB_MEM_SPREAD),
3433 NULL);
3435 /*
3436 * Limit the bh occupancy to 10% of ZONE_NORMAL
3437 */
3441 }