| blob | 4d7433534f5cd77b7f9b240fba57ac7923df07e6 |
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 {
663 }
666 }
668 /*
669 * Add a page to the dirty page list.
670 *
671 * It is a sad fact of life that this function is called from several places
672 * deeply under spinlocking. It may not sleep.
673 *
674 * If the page has buffers, the uptodate buffers are set dirty, to preserve
675 * dirty-state coherency between the page and the buffers. It the page does
676 * not have buffers then when they are later attached they will all be set
677 * dirty.
678 *
679 * The buffers are dirtied before the page is dirtied. There's a small race
680 * window in which a writepage caller may see the page cleanness but not the
681 * buffer dirtiness. That's fine. If this code were to set the page dirty
682 * before the buffers, a concurrent writepage caller could clear the page dirty
683 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
684 * page on the dirty page list.
685 *
686 * We use private_lock to lock against try_to_free_buffers while using the
687 * page's buffer list. Also use this to protect against clean buffers being
688 * added to the page after it was set dirty.
689 *
690 * FIXME: may need to call ->reservepage here as well. That's rather up to the
691 * address_space though.
692 */
694 {
710 }
717 }
720 /*
721 * Write out and wait upon a list of buffers.
722 *
723 * We have conflicting pressures: we want to make sure that all
724 * initially dirty buffers get waited on, but that any subsequently
725 * dirtied buffers don't. After all, we don't want fsync to last
726 * forever if somebody is actively writing to the file.
727 *
728 * Do this in two main stages: first we copy dirty buffers to a
729 * temporary inode list, queueing the writes as we go. Then we clean
730 * up, waiting for those writes to complete.
731 *
732 * During this second stage, any subsequent updates to the file may end
733 * up refiling the buffer on the original inode's dirty list again, so
734 * there is a chance we will end up with a buffer queued for write but
735 * not yet completed on that list. So, as a final cleanup we go through
736 * the osync code to catch these locked, dirty buffers without requeuing
737 * any newly dirty buffers for write.
738 */
740 {
755 /* Avoid race with mark_buffer_dirty_inode() which does
756 * a lockless check and we rely on seeing the dirty bit */
764 /*
765 * Ensure any pending I/O completes so that
766 * write_dirty_buffer() actually writes the
767 * current contents - it is a noop if I/O is
768 * still in flight on potentially older
769 * contents.
770 */
773 /*
774 * Kick off IO for the previous mapping. Note
775 * that we will not run the very last mapping,
776 * wait_on_buffer() will do that for us
777 * through sync_buffer().
778 */
781 }
782 }
783 }
794 /* Avoid race with mark_buffer_dirty_inode() which does
795 * a lockless check and we rely on seeing the dirty bit */
801 }
808 }
814 else
816 }
818 /*
819 * Invalidate any and all dirty buffers on a given inode. We are
820 * probably unmounting the fs, but that doesn't mean we have already
821 * done a sync(). Just drop the buffers from the inode list.
822 *
823 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
824 * assumes that all the buffers are against the blockdev. Not true
825 * for reiserfs.
826 */
828 {
838 }
839 }
842 /*
843 * Remove any clean buffers from the inode's buffer list. This is called
844 * when we're trying to free the inode itself. Those buffers can pin it.
845 *
846 * Returns true if all buffers were removed.
847 */
849 {
863 }
865 }
867 }
869 }
871 /*
872 * Create the appropriate buffers when given a page for data area and
873 * the size of each buffer.. Use the bh->b_this_page linked list to
874 * follow the buffers created. Return NULL if unable to create more
875 * buffers.
876 *
877 * The retry flag is used to differentiate async IO (paging, swapping)
878 * which may not fail from ordinary buffer allocations.
879 */
882 {
886 try_again:
900 /* Link the buffer to its page */
902 }
904 /*
905 * In case anything failed, we just free everything we got.
906 */
907 no_grow:
914 }
916 /*
917 * Return failure for non-async IO requests. Async IO requests
918 * are not allowed to fail, so we have to wait until buffer heads
919 * become available. But we don't want tasks sleeping with
920 * partially complete buffers, so all were released above.
921 */
925 /* We're _really_ low on memory. Now we just
926 * wait for old buffer heads to become free due to
927 * finishing IO. Since this is an async request and
928 * the reserve list is empty, we're sure there are
929 * async buffer heads in use.
930 */
933 }
938 {
948 }
951 {
958 }
960 }
962 /*
963 * Initialise the state of a blockdev page's buffers.
964 */
965 static sector_t
968 {
983 }
984 block++;
988 /*
989 * Caller needs to validate requested block against end of device.
990 */
992 }
994 /*
995 * Create the page-cache page that contains the requested block.
996 *
997 * This is used purely for blockdev mappings.
998 */
999 static int
1002 {
1006 sector_t end_block;
1022 }
1025 }
1027 /*
1028 * Allocate some buffers for this page
1029 */
1034 /*
1035 * Link the page to the buffers and initialise them. Take the
1036 * lock to be atomic wrt __find_get_block(), which does not
1037 * run under the page lock.
1038 */
1043 done:
1045 failed:
1049 }
1051 /*
1052 * Create buffers for the specified block device block's page. If
1053 * that page was dirty, the buffers are set dirty also.
1054 */
1055 static int
1057 {
1058 pgoff_t index;
1063 sizebits++;
1068 /*
1069 * Check for a block which wants to lie outside our maximum possible
1070 * pagecache index. (this comparison is done using sector_t types).
1071 */
1080 }
1082 /* Create a page with the proper size buffers.. */
1084 }
1088 {
1089 /* Size must be multiple of hard sectorsize */
1093 size);
1099 }
1114 }
1115 }
1117 /*
1118 * The relationship between dirty buffers and dirty pages:
1119 *
1120 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1121 * the page is tagged dirty in its radix tree.
1122 *
1123 * At all times, the dirtiness of the buffers represents the dirtiness of
1124 * subsections of the page. If the page has buffers, the page dirty bit is
1125 * merely a hint about the true dirty state.
1126 *
1127 * When a page is set dirty in its entirety, all its buffers are marked dirty
1128 * (if the page has buffers).
1129 *
1130 * When a buffer is marked dirty, its page is dirtied, but the page's other
1131 * buffers are not.
1132 *
1133 * Also. When blockdev buffers are explicitly read with bread(), they
1134 * individually become uptodate. But their backing page remains not
1135 * uptodate - even if all of its buffers are uptodate. A subsequent
1136 * block_read_full_page() against that page will discover all the uptodate
1137 * buffers, will set the page uptodate and will perform no I/O.
1138 */
1140 /**
1141 * mark_buffer_dirty - mark a buffer_head as needing writeout
1142 * @bh: the buffer_head to mark dirty
1143 *
1144 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1145 * backing page dirty, then tag the page as dirty in its address_space's radix
1146 * tree and then attach the address_space's inode to its superblock's dirty
1147 * inode list.
1148 *
1149 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1150 * mapping->tree_lock and mapping->host->i_lock.
1151 */
1153 {
1158 /*
1159 * Very *carefully* optimize the it-is-already-dirty case.
1160 *
1161 * Don't let the final "is it dirty" escape to before we
1162 * perhaps modified the buffer.
1163 */
1168 }
1176 }
1177 }
1178 }
1181 /*
1182 * Decrement a buffer_head's reference count. If all buffers against a page
1183 * have zero reference count, are clean and unlocked, and if the page is clean
1184 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1185 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1186 * a page but it ends up not being freed, and buffers may later be reattached).
1187 */
1189 {
1193 }
1195 }
1198 /*
1199 * bforget() is like brelse(), except it discards any
1200 * potentially dirty data.
1201 */
1203 {
1212 }
1214 }
1218 {
1230 }
1233 }
1235 /*
1236 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1237 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1238 * refcount elevated by one when they're in an LRU. A buffer can only appear
1239 * once in a particular CPU's LRU. A single buffer can be present in multiple
1240 * CPU's LRUs at the same time.
1241 *
1242 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1243 * sb_find_get_block().
1244 *
1245 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1246 * a local interrupt disable for that.
1247 */
1249 #define BH_LRU_SIZE 8
1253 };
1257 #ifdef CONFIG_SMP
1258 #define bh_lru_lock() local_irq_disable()
1259 #define bh_lru_unlock() local_irq_enable()
1260 #else
1261 #define bh_lru_lock() preempt_disable()
1262 #define bh_lru_unlock() preempt_enable()
1263 #endif
1266 {
1267 #ifdef irqs_disabled
1269 #endif
1270 }
1272 /*
1273 * The LRU management algorithm is dopey-but-simple. Sorry.
1274 */
1276 {
1300 }
1301 }
1302 }
1306 }
1311 }
1313 /*
1314 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1315 */
1318 {
1333 i--;
1334 }
1336 }
1340 }
1341 }
1344 }
1346 /*
1347 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1348 * it in the LRU and mark it as accessed. If it is not present then return
1349 * NULL
1350 */
1353 {
1360 }
1364 }
1367 /*
1368 * __getblk will locate (and, if necessary, create) the buffer_head
1369 * which corresponds to the passed block_device, block and size. The
1370 * returned buffer has its reference count incremented.
1371 *
1372 * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1373 * attempt is failing. FIXME, perhaps?
1374 */
1377 {
1384 }
1387 /*
1388 * Do async read-ahead on a buffer..
1389 */
1391 {
1396 }
1397 }
1400 /**
1401 * __bread() - reads a specified block and returns the bh
1402 * @bdev: the block_device to read from
1403 * @block: number of block
1404 * @size: size (in bytes) to read
1405 *
1406 * Reads a specified block, and returns buffer head that contains it.
1407 * It returns NULL if the block was unreadable.
1408 */
1411 {
1417 }
1420 /*
1421 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1422 * This doesn't race because it runs in each cpu either in irq
1423 * or with preempt disabled.
1424 */
1426 {
1433 }
1435 }
1438 {
1445 }
1448 }
1451 {
1453 }
1458 {
1462 /*
1463 * This catches illegal uses and preserves the offset:
1464 */
1466 else
1468 }
1471 /*
1472 * Called when truncating a buffer on a page completely.
1473 */
1475 {
1485 }
1487 /**
1488 * block_invalidatepage - invalidate part or all of a buffer-backed page
1489 *
1490 * @page: the page which is affected
1491 * @offset: start of the range to invalidate
1492 * @length: length of the range to invalidate
1493 *
1494 * block_invalidatepage() is called when all or part of the page has become
1495 * invalidated by a truncate operation.
1496 *
1497 * block_invalidatepage() does not have to release all buffers, but it must
1498 * ensure that no dirty buffer is left outside @offset and that no I/O
1499 * is underway against any of the blocks which are outside the truncation
1500 * point. Because the caller is about to free (and possibly reuse) those
1501 * blocks on-disk.
1502 */
1505 {
1514 /*
1515 * Check for overflow
1516 */
1525 /*
1526 * Are we still fully in range ?
1527 */
1531 /*
1532 * is this block fully invalidated?
1533 */
1540 /*
1541 * We release buffers only if the entire page is being invalidated.
1542 * The get_block cached value has been unconditionally invalidated,
1543 * so real IO is not possible anymore.
1544 */
1547 out:
1549 }
1553 /*
1554 * We attach and possibly dirty the buffers atomically wrt
1555 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1556 * is already excluded via the page lock.
1557 */
1560 {
1582 }
1585 }
1588 /*
1589 * We are taking a block for data and we don't want any output from any
1590 * buffer-cache aliases starting from return from that function and
1591 * until the moment when something will explicitly mark the buffer
1592 * dirty (hopefully that will not happen until we will free that block ;-)
1593 * We don't even need to mark it not-uptodate - nobody can expect
1594 * anything from a newly allocated buffer anyway. We used to used
1595 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1596 * don't want to mark the alias unmapped, for example - it would confuse
1597 * anyone who might pick it with bread() afterwards...
1598 *
1599 * Also.. Note that bforget() doesn't lock the buffer. So there can
1600 * be writeout I/O going on against recently-freed buffers. We don't
1601 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1602 * only if we really need to. That happens here.
1603 */
1605 {
1616 }
1617 }
1620 /*
1621 * Size is a power-of-two in the range 512..PAGE_SIZE,
1622 * and the case we care about most is PAGE_SIZE.
1623 *
1624 * So this *could* possibly be written with those
1625 * constraints in mind (relevant mostly if some
1626 * architecture has a slow bit-scan instruction)
1627 */
1629 {
1631 }
1633 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1634 {
1640 }
1642 /*
1643 * NOTE! All mapped/uptodate combinations are valid:
1644 *
1645 * Mapped Uptodate Meaning
1646 *
1647 * No No "unknown" - must do get_block()
1648 * No Yes "hole" - zero-filled
1649 * Yes No "allocated" - allocated on disk, not read in
1650 * Yes Yes "valid" - allocated and up-to-date in memory.
1651 *
1652 * "Dirty" is valid only with the last case (mapped+uptodate).
1653 */
1655 /*
1656 * While block_write_full_page is writing back the dirty buffers under
1657 * the page lock, whoever dirtied the buffers may decide to clean them
1658 * again at any time. We handle that by only looking at the buffer
1659 * state inside lock_buffer().
1660 *
1661 * If block_write_full_page() is called for regular writeback
1662 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1663 * locked buffer. This only can happen if someone has written the buffer
1664 * directly, with submit_bh(). At the address_space level PageWriteback
1665 * prevents this contention from occurring.
1666 *
1667 * If block_write_full_page() is called with wbc->sync_mode ==
1668 * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1669 * causes the writes to be flagged as synchronous writes.
1670 */
1674 {
1676 sector_t block;
1677 sector_t last_block;
1687 /*
1688 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1689 * here, and the (potentially unmapped) buffers may become dirty at
1690 * any time. If a buffer becomes dirty here after we've inspected it
1691 * then we just miss that fact, and the page stays dirty.
1692 *
1693 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1694 * handle that here by just cleaning them.
1695 */
1704 /*
1705 * Get all the dirty buffers mapped to disk addresses and
1706 * handle any aliases from the underlying blockdev's mapping.
1707 */
1710 /*
1711 * mapped buffers outside i_size will occur, because
1712 * this page can be outside i_size when there is a
1713 * truncate in progress.
1714 */
1715 /*
1716 * The buffer was zeroed by block_write_full_page()
1717 */
1728 /* blockdev mappings never come here */
1732 }
1733 }
1735 block++;
1741 /*
1742 * If it's a fully non-blocking write attempt and we cannot
1743 * lock the buffer then redirty the page. Note that this can
1744 * potentially cause a busy-wait loop from writeback threads
1745 * and kswapd activity, but those code paths have their own
1746 * higher-level throttling.
1747 */
1753 }
1758 }
1761 /*
1762 * The page and its buffers are protected by PageWriteback(), so we can
1763 * drop the bh refcounts early.
1764 */
1772 nr_underway++;
1773 }
1779 done:
1781 /*
1782 * The page was marked dirty, but the buffers were
1783 * clean. Someone wrote them back by hand with
1784 * ll_rw_block/submit_bh. A rare case.
1785 */
1788 /*
1789 * The page and buffer_heads can be released at any time from
1790 * here on.
1791 */
1792 }
1795 recover:
1796 /*
1797 * ENOSPC, or some other error. We may already have added some
1798 * blocks to the file, so we need to write these out to avoid
1799 * exposing stale data.
1800 * The page is currently locked and not marked for writeback
1801 */
1803 /* Recovery: lock and submit the mapped buffers */
1810 /*
1811 * The buffer may have been set dirty during
1812 * attachment to a dirty page.
1813 */
1815 }
1826 nr_underway++;
1827 }
1832 }
1834 /*
1835 * If a page has any new buffers, zero them out here, and mark them uptodate
1836 * and dirty so they'll be written out (in order to prevent uninitialised
1837 * block data from leaking). And clear the new bit.
1838 */
1840 {
1863 }
1867 }
1868 }
1873 }
1878 {
1883 sector_t block;
1906 }
1908 }
1924 }
1930 }
1931 }
1936 }
1942 }
1943 }
1944 /*
1945 * If we issued read requests - let them complete.
1946 */
1951 }
1955 }
1960 {
1978 }
1985 /*
1986 * If this is a partial write which happened to make all buffers
1987 * uptodate then we can optimize away a bogus readpage() for
1988 * the next read(). Here we 'discover' whether the page went
1989 * uptodate as a result of this (potentially partial) write.
1990 */
1994 }
1996 /*
1997 * block_write_begin takes care of the basic task of block allocation and
1998 * bringing partial write blocks uptodate first.
1999 *
2000 * The filesystem needs to handle block truncation upon failure.
2001 */
2004 {
2018 }
2022 }
2028 {
2035 /*
2036 * The buffers that were written will now be uptodate, so we
2037 * don't have to worry about a readpage reading them and
2038 * overwriting a partial write. However if we have encountered
2039 * a short write and only partially written into a buffer, it
2040 * will not be marked uptodate, so a readpage might come in and
2041 * destroy our partial write.
2042 *
2043 * Do the simplest thing, and just treat any short write to a
2044 * non uptodate page as a zero-length write, and force the
2045 * caller to redo the whole thing.
2046 */
2051 }
2054 /* This could be a short (even 0-length) commit */
2058 }
2064 {
2070 /*
2071 * No need to use i_size_read() here, the i_size
2072 * cannot change under us because we hold i_mutex.
2073 *
2074 * But it's important to update i_size while still holding page lock:
2075 * page writeout could otherwise come in and zero beyond i_size.
2076 */
2080 }
2085 /*
2086 * Don't mark the inode dirty under page lock. First, it unnecessarily
2087 * makes the holding time of page lock longer. Second, it forces lock
2088 * ordering of page lock and transaction start for journaling
2089 * filesystems.
2090 */
2095 }
2098 /*
2099 * block_is_partially_uptodate checks whether buffers within a page are
2100 * uptodate or not.
2101 *
2102 * Returns true if all buffers which correspond to a file portion
2103 * we want to read are uptodate.
2104 */
2107 {
2131 }
2134 }
2140 }
2143 /*
2144 * Generic "read page" function for block devices that have the normal
2145 * get_block functionality. This is most of the block device filesystems.
2146 * Reads the page asynchronously --- the unlock_buffer() and
2147 * set/clear_buffer_uptodate() functions propagate buffer state into the
2148 * page struct once IO has completed.
2149 */
2151 {
2182 }
2188 }
2189 /*
2190 * get_block() might have updated the buffer
2191 * synchronously
2192 */
2195 }
2203 /*
2204 * All buffers are uptodate - we can set the page uptodate
2205 * as well. But not if get_block() returned an error.
2206 */
2211 }
2213 /* Stage two: lock the buffers */
2218 }
2220 /*
2221 * Stage 3: start the IO. Check for uptodateness
2222 * inside the buffer lock in case another process reading
2223 * the underlying blockdev brought it uptodate (the sct fix).
2224 */
2229 else
2231 }
2233 }
2236 /* utility function for filesystems that need to do work on expanding
2237 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2238 * deal with the hole.
2239 */
2241 {
2260 out:
2262 }
2267 {
2273 loff_t curpos;
2285 }
2289 AOP_FLAG_UNINTERRUPTIBLE,
2302 }
2304 /* page covers the boundary, find the boundary offset */
2307 /* if we will expand the thing last block will be filled */
2310 }
2314 }
2318 AOP_FLAG_UNINTERRUPTIBLE,
2329 }
2330 out:
2332 }
2334 /*
2335 * For moronic filesystems that do not allow holes in file.
2336 * We may have to extend the file.
2337 */
2342 {
2356 }
2359 }
2363 {
2367 }
2370 /*
2371 * block_page_mkwrite() is not allowed to change the file size as it gets
2372 * called from a page fault handler when a page is first dirtied. Hence we must
2373 * be careful to check for EOF conditions here. We set the page up correctly
2374 * for a written page which means we get ENOSPC checking when writing into
2375 * holes and correct delalloc and unwritten extent mapping on filesystems that
2376 * support these features.
2377 *
2378 * We are not allowed to take the i_mutex here so we have to play games to
2379 * protect against truncate races as the page could now be beyond EOF. Because
2380 * truncate writes the inode size before removing pages, once we have the
2381 * page lock we can determine safely if the page is beyond EOF. If it is not
2382 * beyond EOF, then the page is guaranteed safe against truncation until we
2383 * unlock the page.
2384 *
2385 * Direct callers of this function should protect against filesystem freezing
2386 * using sb_start_write() - sb_end_write() functions.
2387 */
2389 get_block_t get_block)
2390 {
2394 loff_t size;
2401 /* We overload EFAULT to mean page got truncated */
2404 }
2406 /* page is wholly or partially inside EOF */
2409 else
2421 out_unlock:
2424 }
2428 get_block_t get_block)
2429 {
2435 /*
2436 * Update file times before taking page lock. We may end up failing the
2437 * fault so this update may be superfluous but who really cares...
2438 */
2444 }
2447 /*
2448 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2449 * immediately, while under the page lock. So it needs a special end_io
2450 * handler which does not touch the bh after unlocking it.
2451 */
2453 {
2455 }
2457 /*
2458 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2459 * the page (converting it to circular linked list and taking care of page
2460 * dirty races).
2461 */
2463 {
2479 }
2481 /*
2482 * On entry, the page is fully not uptodate.
2483 * On exit the page is fully uptodate in the areas outside (from,to)
2484 * The filesystem needs to handle block truncation upon failure.
2485 */
2490 {
2496 pgoff_t index;
2500 sector_t block_in_file;
2520 }
2525 /*
2526 * Allocate buffers so that we can keep track of state, and potentially
2527 * attach them to the page if an error occurs. In the common case of
2528 * no error, they will just be freed again without ever being attached
2529 * to the page (which is all OK, because we're under the page lock).
2530 *
2531 * Be careful: the buffer linked list is a NULL terminated one, rather
2532 * than the circular one we're used to.
2533 */
2538 }
2542 /*
2543 * We loop across all blocks in the page, whether or not they are
2544 * part of the affected region. This is so we can discover if the
2545 * page is fully mapped-to-disk.
2546 */
2568 }
2573 }
2580 nr_reads++;
2581 }
2582 }
2585 /*
2586 * The page is locked, so these buffers are protected from
2587 * any VM or truncate activity. Hence we don't need to care
2588 * for the buffer_head refcounts.
2589 */
2594 }
2597 }
2606 failed:
2608 /*
2609 * Error recovery is a bit difficult. We need to zero out blocks that
2610 * were newly allocated, and dirty them to ensure they get written out.
2611 * Buffers need to be attached to the page at this point, otherwise
2612 * the handling of potential IO errors during writeout would be hard
2613 * (could try doing synchronous writeout, but what if that fails too?)
2614 */
2618 out_release:
2624 }
2630 {
2647 }
2656 }
2659 }
2662 /*
2663 * nobh_writepage() - based on block_full_write_page() except
2664 * that it tries to operate without attaching bufferheads to
2665 * the page.
2666 */
2669 {
2676 /* Is the page fully inside i_size? */
2680 /* Is the page fully outside i_size? (truncate in progress) */
2683 /*
2684 * The page may have dirty, unmapped buffers. For example,
2685 * they may have been added in ext3_writepage(). Make them
2686 * freeable here, so the page does not leak.
2687 */
2688 #if 0
2689 /* Not really sure about this - do we need this ? */
2692 #endif
2695 }
2697 /*
2698 * The page straddles i_size. It must be zeroed out on each and every
2699 * writepage invocation because it may be mmapped. "A file is mapped
2700 * in multiples of the page size. For a file that is not a multiple of
2701 * the page size, the remaining memory is zeroed when mapped, and
2702 * writes to that region are not written out to the file."
2703 */
2705 out:
2709 end_buffer_async_write);
2711 }
2716 {
2720 sector_t iblock;
2730 /* Block boundary? Nothing to do */
2743 has_buffers:
2747 }
2749 /* Find the buffer that contains "offset" */
2752 iblock++;
2754 }
2761 /* unmapped? It's a hole - nothing to do */
2765 /* Ok, it's mapped. Make sure it's up-to-date */
2771 }
2776 }
2779 }
2784 unlock:
2787 out:
2789 }
2794 {
2798 sector_t iblock;
2808 /* Block boundary? Nothing to do */
2823 /* Find the buffer that contains "offset" */
2828 iblock++;
2830 }
2838 /* unmapped? It's a hole - nothing to do */
2841 }
2843 /* Ok, it's mapped. Make sure it's up-to-date */
2851 /* Uhhuh. Read error. Complain and punt. */
2854 }
2860 unlock:
2863 out:
2865 }
2868 /*
2869 * The generic ->writepage function for buffer-backed address_spaces
2870 * this form passes in the end_io handler used to finish the IO.
2871 */
2874 {
2880 /* Is the page fully inside i_size? */
2883 handler);
2885 /* Is the page fully outside i_size? (truncate in progress) */
2888 /*
2889 * The page may have dirty, unmapped buffers. For example,
2890 * they may have been added in ext3_writepage(). Make them
2891 * freeable here, so the page does not leak.
2892 */
2896 }
2898 /*
2899 * The page straddles i_size. It must be zeroed out on each and every
2900 * writepage invocation because it may be mmapped. "A file is mapped
2901 * in multiples of the page size. For a file that is not a multiple of
2902 * the page size, the remaining memory is zeroed when mapped, and
2903 * writes to that region are not written out to the file."
2904 */
2907 }
2910 /*
2911 * The generic ->writepage function for buffer-backed address_spaces
2912 */
2915 {
2917 end_buffer_async_write);
2918 }
2923 {
2931 }
2935 {
2940 }
2947 }
2949 /*
2950 * This allows us to do IO even on the odd last sectors
2951 * of a device, even if the bh block size is some multiple
2952 * of the physical sector size.
2953 *
2954 * We'll just truncate the bio to the size of the device,
2955 * and clear the end of the buffer head manually.
2956 *
2957 * Truly out-of-range accesses will turn into actual IO
2958 * errors, this only handles the "we need to be able to
2959 * do IO at the final sector" case.
2960 */
2962 {
2963 sector_t maxsector;
2970 /*
2971 * If the *whole* IO is past the end of the device,
2972 * let it through, and the IO layer will turn it into
2973 * an EIO.
2974 */
2983 /* Uhhuh. We've got a bh that straddles the device size! */
2986 /* Truncate the bio.. */
2990 /* ..and clear the end of the buffer for reads */
2996 }
2997 }
3000 {
3010 /*
3011 * Only clear out a write error when rewriting
3012 */
3016 /*
3017 * from here on down, it's all bio -- do the initial mapping,
3018 * submit_bio -> generic_make_request may further map this bio around
3019 */
3035 /* Take care of bh's that straddle the end of the device */
3051 }
3055 {
3057 }
3060 /**
3061 * ll_rw_block: low-level access to block devices (DEPRECATED)
3062 * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
3063 * @nr: number of &struct buffer_heads in the array
3064 * @bhs: array of pointers to &struct buffer_head
3065 *
3066 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3067 * requests an I/O operation on them, either a %READ or a %WRITE. The third
3068 * %READA option is described in the documentation for generic_make_request()
3069 * which ll_rw_block() calls.
3070 *
3071 * This function drops any buffer that it cannot get a lock on (with the
3072 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3073 * request, and any buffer that appears to be up-to-date when doing read
3074 * request. Further it marks as clean buffers that are processed for
3075 * writing (the buffer cache won't assume that they are actually clean
3076 * until the buffer gets unlocked).
3077 *
3078 * ll_rw_block sets b_end_io to simple completion handler that marks
3079 * the buffer up-to-date (if approriate), unlocks the buffer and wakes
3080 * any waiters.
3081 *
3082 * All of the buffers must be for the same device, and must also be a
3083 * multiple of the current approved size for the device.
3084 */
3086 {
3100 }
3107 }
3108 }
3110 }
3111 }
3115 {
3120 }
3124 }
3127 /*
3128 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3129 * and then start new I/O and then wait upon it. The caller must have a ref on
3130 * the buffer_head.
3131 */
3133 {
3147 }
3149 }
3153 {
3155 }
3158 /*
3159 * try_to_free_buffers() checks if all the buffers on this particular page
3160 * are unused, and releases them if so.
3161 *
3162 * Exclusion against try_to_free_buffers may be obtained by either
3163 * locking the page or by holding its mapping's private_lock.
3164 *
3165 * If the page is dirty but all the buffers are clean then we need to
3166 * be sure to mark the page clean as well. This is because the page
3167 * may be against a block device, and a later reattachment of buffers
3168 * to a dirty page will set *all* buffers dirty. Which would corrupt
3169 * filesystem data on the same device.
3170 *
3171 * The same applies to regular filesystem pages: if all the buffers are
3172 * clean then we set the page clean and proceed. To do that, we require
3173 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3174 * private_lock.
3175 *
3176 * try_to_free_buffers() is non-blocking.
3177 */
3179 {
3182 }
3184 static int
3186 {
3209 failed:
3211 }
3214 {
3226 }
3231 /*
3232 * If the filesystem writes its buffers by hand (eg ext3)
3233 * then we can have clean buffers against a dirty page. We
3234 * clean the page here; otherwise the VM will never notice
3235 * that the filesystem did any IO at all.
3236 *
3237 * Also, during truncate, discard_buffer will have marked all
3238 * the page's buffers clean. We discover that here and clean
3239 * the page also.
3240 *
3241 * private_lock must be held over this entire operation in order
3242 * to synchronise against __set_page_dirty_buffers and prevent the
3243 * dirty bit from being lost.
3244 */
3248 out:
3257 }
3259 }
3262 /*
3263 * There are no bdflush tunables left. But distributions are
3264 * still running obsolete flush daemons, so we terminate them here.
3265 *
3266 * Use of bdflush() is deprecated and will be removed in a future kernel.
3267 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3268 */
3270 {
3277 msg_count++;
3279 "warning: process `%s' used the obsolete bdflush"
3282 }
3287 }
3289 /*
3290 * Buffer-head allocation
3291 */
3294 /*
3295 * Once the number of bh's in the machine exceeds this level, we start
3296 * stripping them in writeback.
3297 */
3305 };
3310 {
3320 }
3323 {
3331 }
3333 }
3337 {
3344 }
3348 {
3355 }
3358 }
3362 {
3366 }
3368 /**
3369 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3370 * @bh: struct buffer_head
3371 *
3372 * Return true if the buffer is up-to-date and false,
3373 * with the buffer locked, if not.
3374 */
3376 {
3382 }
3384 }
3387 /**
3388 * bh_submit_read - Submit a locked buffer for reading
3389 * @bh: struct buffer_head
3390 *
3391 * Returns zero on success and -EIO on error.
3392 */
3394 {
3400 }
3409 }
3413 {
3419 SLAB_MEM_SPREAD),
3420 NULL);
3422 /*
3423 * Limit the bh occupancy to 10% of ZONE_NORMAL
3424 */
3428 }