4 * Copyright (C) 1991, 1992, 2002 Linus Torvalds
8 * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
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
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
16 * Added 32k buffer block sizes - these are required older ARM systems. - RMK
18 * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
21 #include <linux/kernel.h>
22 #include <linux/syscalls.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>
46 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
48 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
50 void init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
52 bh->b_end_io = handler;
53 bh->b_private = private;
55 EXPORT_SYMBOL(init_buffer);
57 inline void touch_buffer(struct buffer_head *bh)
59 trace_block_touch_buffer(bh);
60 mark_page_accessed(bh->b_page);
62 EXPORT_SYMBOL(touch_buffer);
64 void __lock_buffer(struct buffer_head *bh)
66 wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
68 EXPORT_SYMBOL(__lock_buffer);
70 void unlock_buffer(struct buffer_head *bh)
72 clear_bit_unlock(BH_Lock, &bh->b_state);
73 smp_mb__after_atomic();
74 wake_up_bit(&bh->b_state, BH_Lock);
76 EXPORT_SYMBOL(unlock_buffer);
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.
83 void buffer_check_dirty_writeback(struct page *page,
84 bool *dirty, bool *writeback)
86 struct buffer_head *head, *bh;
90 BUG_ON(!PageLocked(page));
92 if (!page_has_buffers(page))
95 if (PageWriteback(page))
98 head = page_buffers(page);
101 if (buffer_locked(bh))
104 if (buffer_dirty(bh))
107 bh = bh->b_this_page;
108 } while (bh != head);
110 EXPORT_SYMBOL(buffer_check_dirty_writeback);
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.
117 void __wait_on_buffer(struct buffer_head * bh)
119 wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
121 EXPORT_SYMBOL(__wait_on_buffer);
124 __clear_page_buffers(struct page *page)
126 ClearPagePrivate(page);
127 set_page_private(page, 0);
128 page_cache_release(page);
132 static int quiet_error(struct buffer_head *bh)
134 if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
140 static void buffer_io_error(struct buffer_head *bh, char *msg)
142 char b[BDEVNAME_SIZE];
143 printk(KERN_ERR "Buffer I/O error on dev %s, logical block %llu%s\n",
144 bdevname(bh->b_bdev, b),
145 (unsigned long long)bh->b_blocknr, msg);
149 * End-of-IO handler helper function which does not touch the bh after
151 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
152 * a race there is benign: unlock_buffer() only use the bh's address for
153 * hashing after unlocking the buffer, so it doesn't actually touch the bh
156 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
159 set_buffer_uptodate(bh);
161 /* This happens, due to failed READA attempts. */
162 clear_buffer_uptodate(bh);
168 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
169 * unlock the buffer. This is what ll_rw_block uses too.
171 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
173 __end_buffer_read_notouch(bh, uptodate);
176 EXPORT_SYMBOL(end_buffer_read_sync);
178 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
181 set_buffer_uptodate(bh);
183 if (!quiet_error(bh))
184 buffer_io_error(bh, ", lost sync page write");
185 set_buffer_write_io_error(bh);
186 clear_buffer_uptodate(bh);
191 EXPORT_SYMBOL(end_buffer_write_sync);
194 * Various filesystems appear to want __find_get_block to be non-blocking.
195 * But it's the page lock which protects the buffers. To get around this,
196 * we get exclusion from try_to_free_buffers with the blockdev mapping's
199 * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
200 * may be quite high. This code could TryLock the page, and if that
201 * succeeds, there is no need to take private_lock. (But if
202 * private_lock is contended then so is mapping->tree_lock).
204 static struct buffer_head *
205 __find_get_block_slow(struct block_device *bdev, sector_t block)
207 struct inode *bd_inode = bdev->bd_inode;
208 struct address_space *bd_mapping = bd_inode->i_mapping;
209 struct buffer_head *ret = NULL;
211 struct buffer_head *bh;
212 struct buffer_head *head;
216 index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
217 page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
221 spin_lock(&bd_mapping->private_lock);
222 if (!page_has_buffers(page))
224 head = page_buffers(page);
227 if (!buffer_mapped(bh))
229 else if (bh->b_blocknr == block) {
234 bh = bh->b_this_page;
235 } while (bh != head);
237 /* we might be here because some of the buffers on this page are
238 * not mapped. This is due to various races between
239 * file io on the block device and getblk. It gets dealt with
240 * elsewhere, don't buffer_error if we had some unmapped buffers
243 char b[BDEVNAME_SIZE];
245 printk("__find_get_block_slow() failed. "
246 "block=%llu, b_blocknr=%llu\n",
247 (unsigned long long)block,
248 (unsigned long long)bh->b_blocknr);
249 printk("b_state=0x%08lx, b_size=%zu\n",
250 bh->b_state, bh->b_size);
251 printk("device %s blocksize: %d\n", bdevname(bdev, b),
252 1 << bd_inode->i_blkbits);
255 spin_unlock(&bd_mapping->private_lock);
256 page_cache_release(page);
262 * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
264 static void free_more_memory(void)
269 wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
272 for_each_online_node(nid) {
273 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
274 gfp_zone(GFP_NOFS), NULL,
277 try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
283 * I/O completion handler for block_read_full_page() - pages
284 * which come unlocked at the end of I/O.
286 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
289 struct buffer_head *first;
290 struct buffer_head *tmp;
292 int page_uptodate = 1;
294 BUG_ON(!buffer_async_read(bh));
298 set_buffer_uptodate(bh);
300 clear_buffer_uptodate(bh);
301 if (!quiet_error(bh))
302 buffer_io_error(bh, ", async page read");
307 * Be _very_ careful from here on. Bad things can happen if
308 * two buffer heads end IO at almost the same time and both
309 * decide that the page is now completely done.
311 first = page_buffers(page);
312 local_irq_save(flags);
313 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
314 clear_buffer_async_read(bh);
318 if (!buffer_uptodate(tmp))
320 if (buffer_async_read(tmp)) {
321 BUG_ON(!buffer_locked(tmp));
324 tmp = tmp->b_this_page;
326 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
327 local_irq_restore(flags);
330 * If none of the buffers had errors and they are all
331 * uptodate then we can set the page uptodate.
333 if (page_uptodate && !PageError(page))
334 SetPageUptodate(page);
339 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
340 local_irq_restore(flags);
345 * Completion handler for block_write_full_page() - pages which are unlocked
346 * during I/O, and which have PageWriteback cleared upon I/O completion.
348 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
351 struct buffer_head *first;
352 struct buffer_head *tmp;
355 BUG_ON(!buffer_async_write(bh));
359 set_buffer_uptodate(bh);
361 if (!quiet_error(bh))
362 buffer_io_error(bh, ", lost async page write");
363 set_bit(AS_EIO, &page->mapping->flags);
364 set_buffer_write_io_error(bh);
365 clear_buffer_uptodate(bh);
369 first = page_buffers(page);
370 local_irq_save(flags);
371 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
373 clear_buffer_async_write(bh);
375 tmp = bh->b_this_page;
377 if (buffer_async_write(tmp)) {
378 BUG_ON(!buffer_locked(tmp));
381 tmp = tmp->b_this_page;
383 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
384 local_irq_restore(flags);
385 end_page_writeback(page);
389 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
390 local_irq_restore(flags);
393 EXPORT_SYMBOL(end_buffer_async_write);
396 * If a page's buffers are under async readin (end_buffer_async_read
397 * completion) then there is a possibility that another thread of
398 * control could lock one of the buffers after it has completed
399 * but while some of the other buffers have not completed. This
400 * locked buffer would confuse end_buffer_async_read() into not unlocking
401 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
402 * that this buffer is not under async I/O.
404 * The page comes unlocked when it has no locked buffer_async buffers
407 * PageLocked prevents anyone starting new async I/O reads any of
410 * PageWriteback is used to prevent simultaneous writeout of the same
413 * PageLocked prevents anyone from starting writeback of a page which is
414 * under read I/O (PageWriteback is only ever set against a locked page).
416 static void mark_buffer_async_read(struct buffer_head *bh)
418 bh->b_end_io = end_buffer_async_read;
419 set_buffer_async_read(bh);
422 static void mark_buffer_async_write_endio(struct buffer_head *bh,
423 bh_end_io_t *handler)
425 bh->b_end_io = handler;
426 set_buffer_async_write(bh);
429 void mark_buffer_async_write(struct buffer_head *bh)
431 mark_buffer_async_write_endio(bh, end_buffer_async_write);
433 EXPORT_SYMBOL(mark_buffer_async_write);
437 * fs/buffer.c contains helper functions for buffer-backed address space's
438 * fsync functions. A common requirement for buffer-based filesystems is
439 * that certain data from the backing blockdev needs to be written out for
440 * a successful fsync(). For example, ext2 indirect blocks need to be
441 * written back and waited upon before fsync() returns.
443 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
444 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
445 * management of a list of dependent buffers at ->i_mapping->private_list.
447 * Locking is a little subtle: try_to_free_buffers() will remove buffers
448 * from their controlling inode's queue when they are being freed. But
449 * try_to_free_buffers() will be operating against the *blockdev* mapping
450 * at the time, not against the S_ISREG file which depends on those buffers.
451 * So the locking for private_list is via the private_lock in the address_space
452 * which backs the buffers. Which is different from the address_space
453 * against which the buffers are listed. So for a particular address_space,
454 * mapping->private_lock does *not* protect mapping->private_list! In fact,
455 * mapping->private_list will always be protected by the backing blockdev's
458 * Which introduces a requirement: all buffers on an address_space's
459 * ->private_list must be from the same address_space: the blockdev's.
461 * address_spaces which do not place buffers at ->private_list via these
462 * utility functions are free to use private_lock and private_list for
463 * whatever they want. The only requirement is that list_empty(private_list)
464 * be true at clear_inode() time.
466 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
467 * filesystems should do that. invalidate_inode_buffers() should just go
468 * BUG_ON(!list_empty).
470 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
471 * take an address_space, not an inode. And it should be called
472 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
475 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
476 * list if it is already on a list. Because if the buffer is on a list,
477 * it *must* already be on the right one. If not, the filesystem is being
478 * silly. This will save a ton of locking. But first we have to ensure
479 * that buffers are taken *off* the old inode's list when they are freed
480 * (presumably in truncate). That requires careful auditing of all
481 * filesystems (do it inside bforget()). It could also be done by bringing
486 * The buffer's backing address_space's private_lock must be held
488 static void __remove_assoc_queue(struct buffer_head *bh)
490 list_del_init(&bh->b_assoc_buffers);
491 WARN_ON(!bh->b_assoc_map);
492 if (buffer_write_io_error(bh))
493 set_bit(AS_EIO, &bh->b_assoc_map->flags);
494 bh->b_assoc_map = NULL;
497 int inode_has_buffers(struct inode *inode)
499 return !list_empty(&inode->i_data.private_list);
503 * osync is designed to support O_SYNC io. It waits synchronously for
504 * all already-submitted IO to complete, but does not queue any new
505 * writes to the disk.
507 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
508 * you dirty the buffers, and then use osync_inode_buffers to wait for
509 * completion. Any other dirty buffers which are not yet queued for
510 * write will not be flushed to disk by the osync.
512 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
514 struct buffer_head *bh;
520 list_for_each_prev(p, list) {
522 if (buffer_locked(bh)) {
526 if (!buffer_uptodate(bh))
537 static void do_thaw_one(struct super_block *sb, void *unused)
539 char b[BDEVNAME_SIZE];
540 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
541 printk(KERN_WARNING "Emergency Thaw on %s\n",
542 bdevname(sb->s_bdev, b));
545 static void do_thaw_all(struct work_struct *work)
547 iterate_supers(do_thaw_one, NULL);
549 printk(KERN_WARNING "Emergency Thaw complete\n");
553 * emergency_thaw_all -- forcibly thaw every frozen filesystem
555 * Used for emergency unfreeze of all filesystems via SysRq
557 void emergency_thaw_all(void)
559 struct work_struct *work;
561 work = kmalloc(sizeof(*work), GFP_ATOMIC);
563 INIT_WORK(work, do_thaw_all);
569 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
570 * @mapping: the mapping which wants those buffers written
572 * Starts I/O against the buffers at mapping->private_list, and waits upon
575 * Basically, this is a convenience function for fsync().
576 * @mapping is a file or directory which needs those buffers to be written for
577 * a successful fsync().
579 int sync_mapping_buffers(struct address_space *mapping)
581 struct address_space *buffer_mapping = mapping->private_data;
583 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
586 return fsync_buffers_list(&buffer_mapping->private_lock,
587 &mapping->private_list);
589 EXPORT_SYMBOL(sync_mapping_buffers);
592 * Called when we've recently written block `bblock', and it is known that
593 * `bblock' was for a buffer_boundary() buffer. This means that the block at
594 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
595 * dirty, schedule it for IO. So that indirects merge nicely with their data.
597 void write_boundary_block(struct block_device *bdev,
598 sector_t bblock, unsigned blocksize)
600 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
602 if (buffer_dirty(bh))
603 ll_rw_block(WRITE, 1, &bh);
608 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
610 struct address_space *mapping = inode->i_mapping;
611 struct address_space *buffer_mapping = bh->b_page->mapping;
613 mark_buffer_dirty(bh);
614 if (!mapping->private_data) {
615 mapping->private_data = buffer_mapping;
617 BUG_ON(mapping->private_data != buffer_mapping);
619 if (!bh->b_assoc_map) {
620 spin_lock(&buffer_mapping->private_lock);
621 list_move_tail(&bh->b_assoc_buffers,
622 &mapping->private_list);
623 bh->b_assoc_map = mapping;
624 spin_unlock(&buffer_mapping->private_lock);
627 EXPORT_SYMBOL(mark_buffer_dirty_inode);
630 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
633 * If warn is true, then emit a warning if the page is not uptodate and has
634 * not been truncated.
636 static void __set_page_dirty(struct page *page,
637 struct address_space *mapping, int warn)
641 spin_lock_irqsave(&mapping->tree_lock, flags);
642 if (page->mapping) { /* Race with truncate? */
643 WARN_ON_ONCE(warn && !PageUptodate(page));
644 account_page_dirtied(page, mapping);
645 radix_tree_tag_set(&mapping->page_tree,
646 page_index(page), PAGECACHE_TAG_DIRTY);
648 spin_unlock_irqrestore(&mapping->tree_lock, flags);
649 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
653 * Add a page to the dirty page list.
655 * It is a sad fact of life that this function is called from several places
656 * deeply under spinlocking. It may not sleep.
658 * If the page has buffers, the uptodate buffers are set dirty, to preserve
659 * dirty-state coherency between the page and the buffers. It the page does
660 * not have buffers then when they are later attached they will all be set
663 * The buffers are dirtied before the page is dirtied. There's a small race
664 * window in which a writepage caller may see the page cleanness but not the
665 * buffer dirtiness. That's fine. If this code were to set the page dirty
666 * before the buffers, a concurrent writepage caller could clear the page dirty
667 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
668 * page on the dirty page list.
670 * We use private_lock to lock against try_to_free_buffers while using the
671 * page's buffer list. Also use this to protect against clean buffers being
672 * added to the page after it was set dirty.
674 * FIXME: may need to call ->reservepage here as well. That's rather up to the
675 * address_space though.
677 int __set_page_dirty_buffers(struct page *page)
680 struct address_space *mapping = page_mapping(page);
682 if (unlikely(!mapping))
683 return !TestSetPageDirty(page);
685 spin_lock(&mapping->private_lock);
686 if (page_has_buffers(page)) {
687 struct buffer_head *head = page_buffers(page);
688 struct buffer_head *bh = head;
691 set_buffer_dirty(bh);
692 bh = bh->b_this_page;
693 } while (bh != head);
695 newly_dirty = !TestSetPageDirty(page);
696 spin_unlock(&mapping->private_lock);
699 __set_page_dirty(page, mapping, 1);
702 EXPORT_SYMBOL(__set_page_dirty_buffers);
705 * Write out and wait upon a list of buffers.
707 * We have conflicting pressures: we want to make sure that all
708 * initially dirty buffers get waited on, but that any subsequently
709 * dirtied buffers don't. After all, we don't want fsync to last
710 * forever if somebody is actively writing to the file.
712 * Do this in two main stages: first we copy dirty buffers to a
713 * temporary inode list, queueing the writes as we go. Then we clean
714 * up, waiting for those writes to complete.
716 * During this second stage, any subsequent updates to the file may end
717 * up refiling the buffer on the original inode's dirty list again, so
718 * there is a chance we will end up with a buffer queued for write but
719 * not yet completed on that list. So, as a final cleanup we go through
720 * the osync code to catch these locked, dirty buffers without requeuing
721 * any newly dirty buffers for write.
723 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
725 struct buffer_head *bh;
726 struct list_head tmp;
727 struct address_space *mapping;
729 struct blk_plug plug;
731 INIT_LIST_HEAD(&tmp);
732 blk_start_plug(&plug);
735 while (!list_empty(list)) {
736 bh = BH_ENTRY(list->next);
737 mapping = bh->b_assoc_map;
738 __remove_assoc_queue(bh);
739 /* Avoid race with mark_buffer_dirty_inode() which does
740 * a lockless check and we rely on seeing the dirty bit */
742 if (buffer_dirty(bh) || buffer_locked(bh)) {
743 list_add(&bh->b_assoc_buffers, &tmp);
744 bh->b_assoc_map = mapping;
745 if (buffer_dirty(bh)) {
749 * Ensure any pending I/O completes so that
750 * write_dirty_buffer() actually writes the
751 * current contents - it is a noop if I/O is
752 * still in flight on potentially older
755 write_dirty_buffer(bh, WRITE_SYNC);
758 * Kick off IO for the previous mapping. Note
759 * that we will not run the very last mapping,
760 * wait_on_buffer() will do that for us
761 * through sync_buffer().
770 blk_finish_plug(&plug);
773 while (!list_empty(&tmp)) {
774 bh = BH_ENTRY(tmp.prev);
776 mapping = bh->b_assoc_map;
777 __remove_assoc_queue(bh);
778 /* Avoid race with mark_buffer_dirty_inode() which does
779 * a lockless check and we rely on seeing the dirty bit */
781 if (buffer_dirty(bh)) {
782 list_add(&bh->b_assoc_buffers,
783 &mapping->private_list);
784 bh->b_assoc_map = mapping;
788 if (!buffer_uptodate(bh))
795 err2 = osync_buffers_list(lock, list);
803 * Invalidate any and all dirty buffers on a given inode. We are
804 * probably unmounting the fs, but that doesn't mean we have already
805 * done a sync(). Just drop the buffers from the inode list.
807 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
808 * assumes that all the buffers are against the blockdev. Not true
811 void invalidate_inode_buffers(struct inode *inode)
813 if (inode_has_buffers(inode)) {
814 struct address_space *mapping = &inode->i_data;
815 struct list_head *list = &mapping->private_list;
816 struct address_space *buffer_mapping = mapping->private_data;
818 spin_lock(&buffer_mapping->private_lock);
819 while (!list_empty(list))
820 __remove_assoc_queue(BH_ENTRY(list->next));
821 spin_unlock(&buffer_mapping->private_lock);
824 EXPORT_SYMBOL(invalidate_inode_buffers);
827 * Remove any clean buffers from the inode's buffer list. This is called
828 * when we're trying to free the inode itself. Those buffers can pin it.
830 * Returns true if all buffers were removed.
832 int remove_inode_buffers(struct inode *inode)
836 if (inode_has_buffers(inode)) {
837 struct address_space *mapping = &inode->i_data;
838 struct list_head *list = &mapping->private_list;
839 struct address_space *buffer_mapping = mapping->private_data;
841 spin_lock(&buffer_mapping->private_lock);
842 while (!list_empty(list)) {
843 struct buffer_head *bh = BH_ENTRY(list->next);
844 if (buffer_dirty(bh)) {
848 __remove_assoc_queue(bh);
850 spin_unlock(&buffer_mapping->private_lock);
856 * Create the appropriate buffers when given a page for data area and
857 * the size of each buffer.. Use the bh->b_this_page linked list to
858 * follow the buffers created. Return NULL if unable to create more
861 * The retry flag is used to differentiate async IO (paging, swapping)
862 * which may not fail from ordinary buffer allocations.
864 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
867 struct buffer_head *bh, *head;
873 while ((offset -= size) >= 0) {
874 bh = alloc_buffer_head(GFP_NOFS);
878 bh->b_this_page = head;
884 /* Link the buffer to its page */
885 set_bh_page(bh, page, offset);
889 * In case anything failed, we just free everything we got.
895 head = head->b_this_page;
896 free_buffer_head(bh);
901 * Return failure for non-async IO requests. Async IO requests
902 * are not allowed to fail, so we have to wait until buffer heads
903 * become available. But we don't want tasks sleeping with
904 * partially complete buffers, so all were released above.
909 /* We're _really_ low on memory. Now we just
910 * wait for old buffer heads to become free due to
911 * finishing IO. Since this is an async request and
912 * the reserve list is empty, we're sure there are
913 * async buffer heads in use.
918 EXPORT_SYMBOL_GPL(alloc_page_buffers);
921 link_dev_buffers(struct page *page, struct buffer_head *head)
923 struct buffer_head *bh, *tail;
928 bh = bh->b_this_page;
930 tail->b_this_page = head;
931 attach_page_buffers(page, head);
934 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
936 sector_t retval = ~((sector_t)0);
937 loff_t sz = i_size_read(bdev->bd_inode);
940 unsigned int sizebits = blksize_bits(size);
941 retval = (sz >> sizebits);
947 * Initialise the state of a blockdev page's buffers.
950 init_page_buffers(struct page *page, struct block_device *bdev,
951 sector_t block, int size)
953 struct buffer_head *head = page_buffers(page);
954 struct buffer_head *bh = head;
955 int uptodate = PageUptodate(page);
956 sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
959 if (!buffer_mapped(bh)) {
960 init_buffer(bh, NULL, NULL);
962 bh->b_blocknr = block;
964 set_buffer_uptodate(bh);
965 if (block < end_block)
966 set_buffer_mapped(bh);
969 bh = bh->b_this_page;
970 } while (bh != head);
973 * Caller needs to validate requested block against end of device.
979 * Create the page-cache page that contains the requested block.
981 * This is used purely for blockdev mappings.
984 grow_dev_page(struct block_device *bdev, sector_t block,
985 pgoff_t index, int size, int sizebits, gfp_t gfp)
987 struct inode *inode = bdev->bd_inode;
989 struct buffer_head *bh;
991 int ret = 0; /* Will call free_more_memory() */
994 gfp_mask = (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS) | gfp;
997 * XXX: __getblk_slow() can not really deal with failure and
998 * will endlessly loop on improvised global reclaim. Prefer
999 * looping in the allocator rather than here, at least that
1000 * code knows what it's doing.
1002 gfp_mask |= __GFP_NOFAIL;
1004 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
1008 BUG_ON(!PageLocked(page));
1010 if (page_has_buffers(page)) {
1011 bh = page_buffers(page);
1012 if (bh->b_size == size) {
1013 end_block = init_page_buffers(page, bdev,
1014 (sector_t)index << sizebits,
1018 if (!try_to_free_buffers(page))
1023 * Allocate some buffers for this page
1025 bh = alloc_page_buffers(page, size, 0);
1030 * Link the page to the buffers and initialise them. Take the
1031 * lock to be atomic wrt __find_get_block(), which does not
1032 * run under the page lock.
1034 spin_lock(&inode->i_mapping->private_lock);
1035 link_dev_buffers(page, bh);
1036 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1038 spin_unlock(&inode->i_mapping->private_lock);
1040 ret = (block < end_block) ? 1 : -ENXIO;
1043 page_cache_release(page);
1048 * Create buffers for the specified block device block's page. If
1049 * that page was dirty, the buffers are set dirty also.
1052 grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
1060 } while ((size << sizebits) < PAGE_SIZE);
1062 index = block >> sizebits;
1065 * Check for a block which wants to lie outside our maximum possible
1066 * pagecache index. (this comparison is done using sector_t types).
1068 if (unlikely(index != block >> sizebits)) {
1069 char b[BDEVNAME_SIZE];
1071 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1073 __func__, (unsigned long long)block,
1078 /* Create a page with the proper size buffers.. */
1079 return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1082 struct buffer_head *
1083 __getblk_slow(struct block_device *bdev, sector_t block,
1084 unsigned size, gfp_t gfp)
1086 /* Size must be multiple of hard sectorsize */
1087 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1088 (size < 512 || size > PAGE_SIZE))) {
1089 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1091 printk(KERN_ERR "logical block size: %d\n",
1092 bdev_logical_block_size(bdev));
1099 struct buffer_head *bh;
1102 bh = __find_get_block(bdev, block, size);
1106 ret = grow_buffers(bdev, block, size, gfp);
1113 EXPORT_SYMBOL(__getblk_slow);
1116 * The relationship between dirty buffers and dirty pages:
1118 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1119 * the page is tagged dirty in its radix tree.
1121 * At all times, the dirtiness of the buffers represents the dirtiness of
1122 * subsections of the page. If the page has buffers, the page dirty bit is
1123 * merely a hint about the true dirty state.
1125 * When a page is set dirty in its entirety, all its buffers are marked dirty
1126 * (if the page has buffers).
1128 * When a buffer is marked dirty, its page is dirtied, but the page's other
1131 * Also. When blockdev buffers are explicitly read with bread(), they
1132 * individually become uptodate. But their backing page remains not
1133 * uptodate - even if all of its buffers are uptodate. A subsequent
1134 * block_read_full_page() against that page will discover all the uptodate
1135 * buffers, will set the page uptodate and will perform no I/O.
1139 * mark_buffer_dirty - mark a buffer_head as needing writeout
1140 * @bh: the buffer_head to mark dirty
1142 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1143 * backing page dirty, then tag the page as dirty in its address_space's radix
1144 * tree and then attach the address_space's inode to its superblock's dirty
1147 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1148 * mapping->tree_lock and mapping->host->i_lock.
1150 void mark_buffer_dirty(struct buffer_head *bh)
1152 WARN_ON_ONCE(!buffer_uptodate(bh));
1154 trace_block_dirty_buffer(bh);
1157 * Very *carefully* optimize the it-is-already-dirty case.
1159 * Don't let the final "is it dirty" escape to before we
1160 * perhaps modified the buffer.
1162 if (buffer_dirty(bh)) {
1164 if (buffer_dirty(bh))
1168 if (!test_set_buffer_dirty(bh)) {
1169 struct page *page = bh->b_page;
1170 if (!TestSetPageDirty(page)) {
1171 struct address_space *mapping = page_mapping(page);
1173 __set_page_dirty(page, mapping, 0);
1177 EXPORT_SYMBOL(mark_buffer_dirty);
1180 * Decrement a buffer_head's reference count. If all buffers against a page
1181 * have zero reference count, are clean and unlocked, and if the page is clean
1182 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1183 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1184 * a page but it ends up not being freed, and buffers may later be reattached).
1186 void __brelse(struct buffer_head * buf)
1188 if (atomic_read(&buf->b_count)) {
1192 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1194 EXPORT_SYMBOL(__brelse);
1197 * bforget() is like brelse(), except it discards any
1198 * potentially dirty data.
1200 void __bforget(struct buffer_head *bh)
1202 clear_buffer_dirty(bh);
1203 if (bh->b_assoc_map) {
1204 struct address_space *buffer_mapping = bh->b_page->mapping;
1206 spin_lock(&buffer_mapping->private_lock);
1207 list_del_init(&bh->b_assoc_buffers);
1208 bh->b_assoc_map = NULL;
1209 spin_unlock(&buffer_mapping->private_lock);
1213 EXPORT_SYMBOL(__bforget);
1215 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1218 if (buffer_uptodate(bh)) {
1223 bh->b_end_io = end_buffer_read_sync;
1224 submit_bh(READ, bh);
1226 if (buffer_uptodate(bh))
1234 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1235 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1236 * refcount elevated by one when they're in an LRU. A buffer can only appear
1237 * once in a particular CPU's LRU. A single buffer can be present in multiple
1238 * CPU's LRUs at the same time.
1240 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1241 * sb_find_get_block().
1243 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1244 * a local interrupt disable for that.
1247 #define BH_LRU_SIZE 16
1250 struct buffer_head *bhs[BH_LRU_SIZE];
1253 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1256 #define bh_lru_lock() local_irq_disable()
1257 #define bh_lru_unlock() local_irq_enable()
1259 #define bh_lru_lock() preempt_disable()
1260 #define bh_lru_unlock() preempt_enable()
1263 static inline void check_irqs_on(void)
1265 #ifdef irqs_disabled
1266 BUG_ON(irqs_disabled());
1271 * The LRU management algorithm is dopey-but-simple. Sorry.
1273 static void bh_lru_install(struct buffer_head *bh)
1275 struct buffer_head *evictee = NULL;
1279 if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1280 struct buffer_head *bhs[BH_LRU_SIZE];
1286 for (in = 0; in < BH_LRU_SIZE; in++) {
1287 struct buffer_head *bh2 =
1288 __this_cpu_read(bh_lrus.bhs[in]);
1293 if (out >= BH_LRU_SIZE) {
1294 BUG_ON(evictee != NULL);
1301 while (out < BH_LRU_SIZE)
1303 memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1312 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1314 static struct buffer_head *
1315 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1317 struct buffer_head *ret = NULL;
1322 for (i = 0; i < BH_LRU_SIZE; i++) {
1323 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1325 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1326 bh->b_size == size) {
1329 __this_cpu_write(bh_lrus.bhs[i],
1330 __this_cpu_read(bh_lrus.bhs[i - 1]));
1333 __this_cpu_write(bh_lrus.bhs[0], bh);
1345 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1346 * it in the LRU and mark it as accessed. If it is not present then return
1349 struct buffer_head *
1350 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1352 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1355 /* __find_get_block_slow will mark the page accessed */
1356 bh = __find_get_block_slow(bdev, block);
1364 EXPORT_SYMBOL(__find_get_block);
1367 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1368 * which corresponds to the passed block_device, block and size. The
1369 * returned buffer has its reference count incremented.
1371 * __getblk_gfp() will lock up the machine if grow_dev_page's
1372 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1374 struct buffer_head *
1375 __getblk_gfp(struct block_device *bdev, sector_t block,
1376 unsigned size, gfp_t gfp)
1378 struct buffer_head *bh = __find_get_block(bdev, block, size);
1382 bh = __getblk_slow(bdev, block, size, gfp);
1385 EXPORT_SYMBOL(__getblk_gfp);
1388 * Do async read-ahead on a buffer..
1390 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1392 struct buffer_head *bh = __getblk(bdev, block, size);
1394 ll_rw_block(READA, 1, &bh);
1398 EXPORT_SYMBOL(__breadahead);
1401 * __bread_gfp() - 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 * @gfp: page allocation flag
1407 * Reads a specified block, and returns buffer head that contains it.
1408 * The page cache can be allocated from non-movable area
1409 * not to prevent page migration if you set gfp to zero.
1410 * It returns NULL if the block was unreadable.
1412 struct buffer_head *
1413 __bread_gfp(struct block_device *bdev, sector_t block,
1414 unsigned size, gfp_t gfp)
1416 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1418 if (likely(bh) && !buffer_uptodate(bh))
1419 bh = __bread_slow(bh);
1422 EXPORT_SYMBOL(__bread_gfp);
1425 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1426 * This doesn't race because it runs in each cpu either in irq
1427 * or with preempt disabled.
1429 static void invalidate_bh_lru(void *arg)
1431 struct bh_lru *b = &get_cpu_var(bh_lrus);
1434 for (i = 0; i < BH_LRU_SIZE; i++) {
1438 put_cpu_var(bh_lrus);
1441 static bool has_bh_in_lru(int cpu, void *dummy)
1443 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1446 for (i = 0; i < BH_LRU_SIZE; i++) {
1454 void invalidate_bh_lrus(void)
1456 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1458 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1460 void set_bh_page(struct buffer_head *bh,
1461 struct page *page, unsigned long offset)
1464 BUG_ON(offset >= PAGE_SIZE);
1465 if (PageHighMem(page))
1467 * This catches illegal uses and preserves the offset:
1469 bh->b_data = (char *)(0 + offset);
1471 bh->b_data = page_address(page) + offset;
1473 EXPORT_SYMBOL(set_bh_page);
1476 * Called when truncating a buffer on a page completely.
1479 /* Bits that are cleared during an invalidate */
1480 #define BUFFER_FLAGS_DISCARD \
1481 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1482 1 << BH_Delay | 1 << BH_Unwritten)
1484 static void discard_buffer(struct buffer_head * bh)
1486 unsigned long b_state, b_state_old;
1489 clear_buffer_dirty(bh);
1491 b_state = bh->b_state;
1493 b_state_old = cmpxchg(&bh->b_state, b_state,
1494 (b_state & ~BUFFER_FLAGS_DISCARD));
1495 if (b_state_old == b_state)
1497 b_state = b_state_old;
1503 * block_invalidatepage - invalidate part or all of a buffer-backed page
1505 * @page: the page which is affected
1506 * @offset: start of the range to invalidate
1507 * @length: length of the range to invalidate
1509 * block_invalidatepage() is called when all or part of the page has become
1510 * invalidated by a truncate operation.
1512 * block_invalidatepage() does not have to release all buffers, but it must
1513 * ensure that no dirty buffer is left outside @offset and that no I/O
1514 * is underway against any of the blocks which are outside the truncation
1515 * point. Because the caller is about to free (and possibly reuse) those
1518 void block_invalidatepage(struct page *page, unsigned int offset,
1519 unsigned int length)
1521 struct buffer_head *head, *bh, *next;
1522 unsigned int curr_off = 0;
1523 unsigned int stop = length + offset;
1525 BUG_ON(!PageLocked(page));
1526 if (!page_has_buffers(page))
1530 * Check for overflow
1532 BUG_ON(stop > PAGE_CACHE_SIZE || stop < length);
1534 head = page_buffers(page);
1537 unsigned int next_off = curr_off + bh->b_size;
1538 next = bh->b_this_page;
1541 * Are we still fully in range ?
1543 if (next_off > stop)
1547 * is this block fully invalidated?
1549 if (offset <= curr_off)
1551 curr_off = next_off;
1553 } while (bh != head);
1556 * We release buffers only if the entire page is being invalidated.
1557 * The get_block cached value has been unconditionally invalidated,
1558 * so real IO is not possible anymore.
1561 try_to_release_page(page, 0);
1565 EXPORT_SYMBOL(block_invalidatepage);
1569 * We attach and possibly dirty the buffers atomically wrt
1570 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1571 * is already excluded via the page lock.
1573 void create_empty_buffers(struct page *page,
1574 unsigned long blocksize, unsigned long b_state)
1576 struct buffer_head *bh, *head, *tail;
1578 head = alloc_page_buffers(page, blocksize, 1);
1581 bh->b_state |= b_state;
1583 bh = bh->b_this_page;
1585 tail->b_this_page = head;
1587 spin_lock(&page->mapping->private_lock);
1588 if (PageUptodate(page) || PageDirty(page)) {
1591 if (PageDirty(page))
1592 set_buffer_dirty(bh);
1593 if (PageUptodate(page))
1594 set_buffer_uptodate(bh);
1595 bh = bh->b_this_page;
1596 } while (bh != head);
1598 attach_page_buffers(page, head);
1599 spin_unlock(&page->mapping->private_lock);
1601 EXPORT_SYMBOL(create_empty_buffers);
1604 * We are taking a block for data and we don't want any output from any
1605 * buffer-cache aliases starting from return from that function and
1606 * until the moment when something will explicitly mark the buffer
1607 * dirty (hopefully that will not happen until we will free that block ;-)
1608 * We don't even need to mark it not-uptodate - nobody can expect
1609 * anything from a newly allocated buffer anyway. We used to used
1610 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1611 * don't want to mark the alias unmapped, for example - it would confuse
1612 * anyone who might pick it with bread() afterwards...
1614 * Also.. Note that bforget() doesn't lock the buffer. So there can
1615 * be writeout I/O going on against recently-freed buffers. We don't
1616 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1617 * only if we really need to. That happens here.
1619 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1621 struct buffer_head *old_bh;
1625 old_bh = __find_get_block_slow(bdev, block);
1627 clear_buffer_dirty(old_bh);
1628 wait_on_buffer(old_bh);
1629 clear_buffer_req(old_bh);
1633 EXPORT_SYMBOL(unmap_underlying_metadata);
1636 * Size is a power-of-two in the range 512..PAGE_SIZE,
1637 * and the case we care about most is PAGE_SIZE.
1639 * So this *could* possibly be written with those
1640 * constraints in mind (relevant mostly if some
1641 * architecture has a slow bit-scan instruction)
1643 static inline int block_size_bits(unsigned int blocksize)
1645 return ilog2(blocksize);
1648 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1650 BUG_ON(!PageLocked(page));
1652 if (!page_has_buffers(page))
1653 create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1654 return page_buffers(page);
1658 * NOTE! All mapped/uptodate combinations are valid:
1660 * Mapped Uptodate Meaning
1662 * No No "unknown" - must do get_block()
1663 * No Yes "hole" - zero-filled
1664 * Yes No "allocated" - allocated on disk, not read in
1665 * Yes Yes "valid" - allocated and up-to-date in memory.
1667 * "Dirty" is valid only with the last case (mapped+uptodate).
1671 * While block_write_full_page is writing back the dirty buffers under
1672 * the page lock, whoever dirtied the buffers may decide to clean them
1673 * again at any time. We handle that by only looking at the buffer
1674 * state inside lock_buffer().
1676 * If block_write_full_page() is called for regular writeback
1677 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1678 * locked buffer. This only can happen if someone has written the buffer
1679 * directly, with submit_bh(). At the address_space level PageWriteback
1680 * prevents this contention from occurring.
1682 * If block_write_full_page() is called with wbc->sync_mode ==
1683 * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1684 * causes the writes to be flagged as synchronous writes.
1686 static int __block_write_full_page(struct inode *inode, struct page *page,
1687 get_block_t *get_block, struct writeback_control *wbc,
1688 bh_end_io_t *handler)
1692 sector_t last_block;
1693 struct buffer_head *bh, *head;
1694 unsigned int blocksize, bbits;
1695 int nr_underway = 0;
1696 int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1697 WRITE_SYNC : WRITE);
1699 head = create_page_buffers(page, inode,
1700 (1 << BH_Dirty)|(1 << BH_Uptodate));
1703 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1704 * here, and the (potentially unmapped) buffers may become dirty at
1705 * any time. If a buffer becomes dirty here after we've inspected it
1706 * then we just miss that fact, and the page stays dirty.
1708 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1709 * handle that here by just cleaning them.
1713 blocksize = bh->b_size;
1714 bbits = block_size_bits(blocksize);
1716 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1717 last_block = (i_size_read(inode) - 1) >> bbits;
1720 * Get all the dirty buffers mapped to disk addresses and
1721 * handle any aliases from the underlying blockdev's mapping.
1724 if (block > last_block) {
1726 * mapped buffers outside i_size will occur, because
1727 * this page can be outside i_size when there is a
1728 * truncate in progress.
1731 * The buffer was zeroed by block_write_full_page()
1733 clear_buffer_dirty(bh);
1734 set_buffer_uptodate(bh);
1735 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1737 WARN_ON(bh->b_size != blocksize);
1738 err = get_block(inode, block, bh, 1);
1741 clear_buffer_delay(bh);
1742 if (buffer_new(bh)) {
1743 /* blockdev mappings never come here */
1744 clear_buffer_new(bh);
1745 unmap_underlying_metadata(bh->b_bdev,
1749 bh = bh->b_this_page;
1751 } while (bh != head);
1754 if (!buffer_mapped(bh))
1757 * If it's a fully non-blocking write attempt and we cannot
1758 * lock the buffer then redirty the page. Note that this can
1759 * potentially cause a busy-wait loop from writeback threads
1760 * and kswapd activity, but those code paths have their own
1761 * higher-level throttling.
1763 if (wbc->sync_mode != WB_SYNC_NONE) {
1765 } else if (!trylock_buffer(bh)) {
1766 redirty_page_for_writepage(wbc, page);
1769 if (test_clear_buffer_dirty(bh)) {
1770 mark_buffer_async_write_endio(bh, handler);
1774 } while ((bh = bh->b_this_page) != head);
1777 * The page and its buffers are protected by PageWriteback(), so we can
1778 * drop the bh refcounts early.
1780 BUG_ON(PageWriteback(page));
1781 set_page_writeback(page);
1784 struct buffer_head *next = bh->b_this_page;
1785 if (buffer_async_write(bh)) {
1786 submit_bh(write_op, bh);
1790 } while (bh != head);
1795 if (nr_underway == 0) {
1797 * The page was marked dirty, but the buffers were
1798 * clean. Someone wrote them back by hand with
1799 * ll_rw_block/submit_bh. A rare case.
1801 end_page_writeback(page);
1804 * The page and buffer_heads can be released at any time from
1812 * ENOSPC, or some other error. We may already have added some
1813 * blocks to the file, so we need to write these out to avoid
1814 * exposing stale data.
1815 * The page is currently locked and not marked for writeback
1818 /* Recovery: lock and submit the mapped buffers */
1820 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1821 !buffer_delay(bh)) {
1823 mark_buffer_async_write_endio(bh, handler);
1826 * The buffer may have been set dirty during
1827 * attachment to a dirty page.
1829 clear_buffer_dirty(bh);
1831 } while ((bh = bh->b_this_page) != head);
1833 BUG_ON(PageWriteback(page));
1834 mapping_set_error(page->mapping, err);
1835 set_page_writeback(page);
1837 struct buffer_head *next = bh->b_this_page;
1838 if (buffer_async_write(bh)) {
1839 clear_buffer_dirty(bh);
1840 submit_bh(write_op, bh);
1844 } while (bh != head);
1850 * If a page has any new buffers, zero them out here, and mark them uptodate
1851 * and dirty so they'll be written out (in order to prevent uninitialised
1852 * block data from leaking). And clear the new bit.
1854 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1856 unsigned int block_start, block_end;
1857 struct buffer_head *head, *bh;
1859 BUG_ON(!PageLocked(page));
1860 if (!page_has_buffers(page))
1863 bh = head = page_buffers(page);
1866 block_end = block_start + bh->b_size;
1868 if (buffer_new(bh)) {
1869 if (block_end > from && block_start < to) {
1870 if (!PageUptodate(page)) {
1871 unsigned start, size;
1873 start = max(from, block_start);
1874 size = min(to, block_end) - start;
1876 zero_user(page, start, size);
1877 set_buffer_uptodate(bh);
1880 clear_buffer_new(bh);
1881 mark_buffer_dirty(bh);
1885 block_start = block_end;
1886 bh = bh->b_this_page;
1887 } while (bh != head);
1889 EXPORT_SYMBOL(page_zero_new_buffers);
1891 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1892 get_block_t *get_block)
1894 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1895 unsigned to = from + len;
1896 struct inode *inode = page->mapping->host;
1897 unsigned block_start, block_end;
1900 unsigned blocksize, bbits;
1901 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1903 BUG_ON(!PageLocked(page));
1904 BUG_ON(from > PAGE_CACHE_SIZE);
1905 BUG_ON(to > PAGE_CACHE_SIZE);
1908 head = create_page_buffers(page, inode, 0);
1909 blocksize = head->b_size;
1910 bbits = block_size_bits(blocksize);
1912 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1914 for(bh = head, block_start = 0; bh != head || !block_start;
1915 block++, block_start=block_end, bh = bh->b_this_page) {
1916 block_end = block_start + blocksize;
1917 if (block_end <= from || block_start >= to) {
1918 if (PageUptodate(page)) {
1919 if (!buffer_uptodate(bh))
1920 set_buffer_uptodate(bh);
1925 clear_buffer_new(bh);
1926 if (!buffer_mapped(bh)) {
1927 WARN_ON(bh->b_size != blocksize);
1928 err = get_block(inode, block, bh, 1);
1931 if (buffer_new(bh)) {
1932 unmap_underlying_metadata(bh->b_bdev,
1934 if (PageUptodate(page)) {
1935 clear_buffer_new(bh);
1936 set_buffer_uptodate(bh);
1937 mark_buffer_dirty(bh);
1940 if (block_end > to || block_start < from)
1941 zero_user_segments(page,
1947 if (PageUptodate(page)) {
1948 if (!buffer_uptodate(bh))
1949 set_buffer_uptodate(bh);
1952 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1953 !buffer_unwritten(bh) &&
1954 (block_start < from || block_end > to)) {
1955 ll_rw_block(READ, 1, &bh);
1960 * If we issued read requests - let them complete.
1962 while(wait_bh > wait) {
1963 wait_on_buffer(*--wait_bh);
1964 if (!buffer_uptodate(*wait_bh))
1968 page_zero_new_buffers(page, from, to);
1971 EXPORT_SYMBOL(__block_write_begin);
1973 static int __block_commit_write(struct inode *inode, struct page *page,
1974 unsigned from, unsigned to)
1976 unsigned block_start, block_end;
1979 struct buffer_head *bh, *head;
1981 bh = head = page_buffers(page);
1982 blocksize = bh->b_size;
1986 block_end = block_start + blocksize;
1987 if (block_end <= from || block_start >= to) {
1988 if (!buffer_uptodate(bh))
1991 set_buffer_uptodate(bh);
1992 mark_buffer_dirty(bh);
1994 clear_buffer_new(bh);
1996 block_start = block_end;
1997 bh = bh->b_this_page;
1998 } while (bh != head);
2001 * If this is a partial write which happened to make all buffers
2002 * uptodate then we can optimize away a bogus readpage() for
2003 * the next read(). Here we 'discover' whether the page went
2004 * uptodate as a result of this (potentially partial) write.
2007 SetPageUptodate(page);
2012 * block_write_begin takes care of the basic task of block allocation and
2013 * bringing partial write blocks uptodate first.
2015 * The filesystem needs to handle block truncation upon failure.
2017 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2018 unsigned flags, struct page **pagep, get_block_t *get_block)
2020 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
2024 page = grab_cache_page_write_begin(mapping, index, flags);
2028 status = __block_write_begin(page, pos, len, get_block);
2029 if (unlikely(status)) {
2031 page_cache_release(page);
2038 EXPORT_SYMBOL(block_write_begin);
2040 int block_write_end(struct file *file, struct address_space *mapping,
2041 loff_t pos, unsigned len, unsigned copied,
2042 struct page *page, void *fsdata)
2044 struct inode *inode = mapping->host;
2047 start = pos & (PAGE_CACHE_SIZE - 1);
2049 if (unlikely(copied < len)) {
2051 * The buffers that were written will now be uptodate, so we
2052 * don't have to worry about a readpage reading them and
2053 * overwriting a partial write. However if we have encountered
2054 * a short write and only partially written into a buffer, it
2055 * will not be marked uptodate, so a readpage might come in and
2056 * destroy our partial write.
2058 * Do the simplest thing, and just treat any short write to a
2059 * non uptodate page as a zero-length write, and force the
2060 * caller to redo the whole thing.
2062 if (!PageUptodate(page))
2065 page_zero_new_buffers(page, start+copied, start+len);
2067 flush_dcache_page(page);
2069 /* This could be a short (even 0-length) commit */
2070 __block_commit_write(inode, page, start, start+copied);
2074 EXPORT_SYMBOL(block_write_end);
2076 int generic_write_end(struct file *file, struct address_space *mapping,
2077 loff_t pos, unsigned len, unsigned copied,
2078 struct page *page, void *fsdata)
2080 struct inode *inode = mapping->host;
2081 loff_t old_size = inode->i_size;
2082 int i_size_changed = 0;
2084 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2087 * No need to use i_size_read() here, the i_size
2088 * cannot change under us because we hold i_mutex.
2090 * But it's important to update i_size while still holding page lock:
2091 * page writeout could otherwise come in and zero beyond i_size.
2093 if (pos+copied > inode->i_size) {
2094 i_size_write(inode, pos+copied);
2099 page_cache_release(page);
2102 pagecache_isize_extended(inode, old_size, pos);
2104 * Don't mark the inode dirty under page lock. First, it unnecessarily
2105 * makes the holding time of page lock longer. Second, it forces lock
2106 * ordering of page lock and transaction start for journaling
2110 mark_inode_dirty(inode);
2114 EXPORT_SYMBOL(generic_write_end);
2117 * block_is_partially_uptodate checks whether buffers within a page are
2120 * Returns true if all buffers which correspond to a file portion
2121 * we want to read are uptodate.
2123 int block_is_partially_uptodate(struct page *page, unsigned long from,
2124 unsigned long count)
2126 unsigned block_start, block_end, blocksize;
2128 struct buffer_head *bh, *head;
2131 if (!page_has_buffers(page))
2134 head = page_buffers(page);
2135 blocksize = head->b_size;
2136 to = min_t(unsigned, PAGE_CACHE_SIZE - from, count);
2138 if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2144 block_end = block_start + blocksize;
2145 if (block_end > from && block_start < to) {
2146 if (!buffer_uptodate(bh)) {
2150 if (block_end >= to)
2153 block_start = block_end;
2154 bh = bh->b_this_page;
2155 } while (bh != head);
2159 EXPORT_SYMBOL(block_is_partially_uptodate);
2162 * Generic "read page" function for block devices that have the normal
2163 * get_block functionality. This is most of the block device filesystems.
2164 * Reads the page asynchronously --- the unlock_buffer() and
2165 * set/clear_buffer_uptodate() functions propagate buffer state into the
2166 * page struct once IO has completed.
2168 int block_read_full_page(struct page *page, get_block_t *get_block)
2170 struct inode *inode = page->mapping->host;
2171 sector_t iblock, lblock;
2172 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2173 unsigned int blocksize, bbits;
2175 int fully_mapped = 1;
2177 head = create_page_buffers(page, inode, 0);
2178 blocksize = head->b_size;
2179 bbits = block_size_bits(blocksize);
2181 iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
2182 lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2188 if (buffer_uptodate(bh))
2191 if (!buffer_mapped(bh)) {
2195 if (iblock < lblock) {
2196 WARN_ON(bh->b_size != blocksize);
2197 err = get_block(inode, iblock, bh, 0);
2201 if (!buffer_mapped(bh)) {
2202 zero_user(page, i * blocksize, blocksize);
2204 set_buffer_uptodate(bh);
2208 * get_block() might have updated the buffer
2211 if (buffer_uptodate(bh))
2215 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2218 SetPageMappedToDisk(page);
2222 * All buffers are uptodate - we can set the page uptodate
2223 * as well. But not if get_block() returned an error.
2225 if (!PageError(page))
2226 SetPageUptodate(page);
2231 /* Stage two: lock the buffers */
2232 for (i = 0; i < nr; i++) {
2235 mark_buffer_async_read(bh);
2239 * Stage 3: start the IO. Check for uptodateness
2240 * inside the buffer lock in case another process reading
2241 * the underlying blockdev brought it uptodate (the sct fix).
2243 for (i = 0; i < nr; i++) {
2245 if (buffer_uptodate(bh))
2246 end_buffer_async_read(bh, 1);
2248 submit_bh(READ, bh);
2252 EXPORT_SYMBOL(block_read_full_page);
2254 /* utility function for filesystems that need to do work on expanding
2255 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2256 * deal with the hole.
2258 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2260 struct address_space *mapping = inode->i_mapping;
2265 err = inode_newsize_ok(inode, size);
2269 err = pagecache_write_begin(NULL, mapping, size, 0,
2270 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2275 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2281 EXPORT_SYMBOL(generic_cont_expand_simple);
2283 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2284 loff_t pos, loff_t *bytes)
2286 struct inode *inode = mapping->host;
2287 unsigned blocksize = 1 << inode->i_blkbits;
2290 pgoff_t index, curidx;
2292 unsigned zerofrom, offset, len;
2295 index = pos >> PAGE_CACHE_SHIFT;
2296 offset = pos & ~PAGE_CACHE_MASK;
2298 while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2299 zerofrom = curpos & ~PAGE_CACHE_MASK;
2300 if (zerofrom & (blocksize-1)) {
2301 *bytes |= (blocksize-1);
2304 len = PAGE_CACHE_SIZE - zerofrom;
2306 err = pagecache_write_begin(file, mapping, curpos, len,
2307 AOP_FLAG_UNINTERRUPTIBLE,
2311 zero_user(page, zerofrom, len);
2312 err = pagecache_write_end(file, mapping, curpos, len, len,
2319 balance_dirty_pages_ratelimited(mapping);
2321 if (unlikely(fatal_signal_pending(current))) {
2327 /* page covers the boundary, find the boundary offset */
2328 if (index == curidx) {
2329 zerofrom = curpos & ~PAGE_CACHE_MASK;
2330 /* if we will expand the thing last block will be filled */
2331 if (offset <= zerofrom) {
2334 if (zerofrom & (blocksize-1)) {
2335 *bytes |= (blocksize-1);
2338 len = offset - zerofrom;
2340 err = pagecache_write_begin(file, mapping, curpos, len,
2341 AOP_FLAG_UNINTERRUPTIBLE,
2345 zero_user(page, zerofrom, len);
2346 err = pagecache_write_end(file, mapping, curpos, len, len,
2358 * For moronic filesystems that do not allow holes in file.
2359 * We may have to extend the file.
2361 int cont_write_begin(struct file *file, struct address_space *mapping,
2362 loff_t pos, unsigned len, unsigned flags,
2363 struct page **pagep, void **fsdata,
2364 get_block_t *get_block, loff_t *bytes)
2366 struct inode *inode = mapping->host;
2367 unsigned blocksize = 1 << inode->i_blkbits;
2371 err = cont_expand_zero(file, mapping, pos, bytes);
2375 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2376 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2377 *bytes |= (blocksize-1);
2381 return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2383 EXPORT_SYMBOL(cont_write_begin);
2385 int block_commit_write(struct page *page, unsigned from, unsigned to)
2387 struct inode *inode = page->mapping->host;
2388 __block_commit_write(inode,page,from,to);
2391 EXPORT_SYMBOL(block_commit_write);
2394 * block_page_mkwrite() is not allowed to change the file size as it gets
2395 * called from a page fault handler when a page is first dirtied. Hence we must
2396 * be careful to check for EOF conditions here. We set the page up correctly
2397 * for a written page which means we get ENOSPC checking when writing into
2398 * holes and correct delalloc and unwritten extent mapping on filesystems that
2399 * support these features.
2401 * We are not allowed to take the i_mutex here so we have to play games to
2402 * protect against truncate races as the page could now be beyond EOF. Because
2403 * truncate writes the inode size before removing pages, once we have the
2404 * page lock we can determine safely if the page is beyond EOF. If it is not
2405 * beyond EOF, then the page is guaranteed safe against truncation until we
2408 * Direct callers of this function should protect against filesystem freezing
2409 * using sb_start_write() - sb_end_write() functions.
2411 int __block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2412 get_block_t get_block)
2414 struct page *page = vmf->page;
2415 struct inode *inode = file_inode(vma->vm_file);
2421 size = i_size_read(inode);
2422 if ((page->mapping != inode->i_mapping) ||
2423 (page_offset(page) > size)) {
2424 /* We overload EFAULT to mean page got truncated */
2429 /* page is wholly or partially inside EOF */
2430 if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2431 end = size & ~PAGE_CACHE_MASK;
2433 end = PAGE_CACHE_SIZE;
2435 ret = __block_write_begin(page, 0, end, get_block);
2437 ret = block_commit_write(page, 0, end);
2439 if (unlikely(ret < 0))
2441 set_page_dirty(page);
2442 wait_for_stable_page(page);
2448 EXPORT_SYMBOL(__block_page_mkwrite);
2450 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2451 get_block_t get_block)
2454 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
2456 sb_start_pagefault(sb);
2459 * Update file times before taking page lock. We may end up failing the
2460 * fault so this update may be superfluous but who really cares...
2462 file_update_time(vma->vm_file);
2464 ret = __block_page_mkwrite(vma, vmf, get_block);
2465 sb_end_pagefault(sb);
2466 return block_page_mkwrite_return(ret);
2468 EXPORT_SYMBOL(block_page_mkwrite);
2471 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2472 * immediately, while under the page lock. So it needs a special end_io
2473 * handler which does not touch the bh after unlocking it.
2475 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2477 __end_buffer_read_notouch(bh, uptodate);
2481 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2482 * the page (converting it to circular linked list and taking care of page
2485 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2487 struct buffer_head *bh;
2489 BUG_ON(!PageLocked(page));
2491 spin_lock(&page->mapping->private_lock);
2494 if (PageDirty(page))
2495 set_buffer_dirty(bh);
2496 if (!bh->b_this_page)
2497 bh->b_this_page = head;
2498 bh = bh->b_this_page;
2499 } while (bh != head);
2500 attach_page_buffers(page, head);
2501 spin_unlock(&page->mapping->private_lock);
2505 * On entry, the page is fully not uptodate.
2506 * On exit the page is fully uptodate in the areas outside (from,to)
2507 * The filesystem needs to handle block truncation upon failure.
2509 int nobh_write_begin(struct address_space *mapping,
2510 loff_t pos, unsigned len, unsigned flags,
2511 struct page **pagep, void **fsdata,
2512 get_block_t *get_block)
2514 struct inode *inode = mapping->host;
2515 const unsigned blkbits = inode->i_blkbits;
2516 const unsigned blocksize = 1 << blkbits;
2517 struct buffer_head *head, *bh;
2521 unsigned block_in_page;
2522 unsigned block_start, block_end;
2523 sector_t block_in_file;
2526 int is_mapped_to_disk = 1;
2528 index = pos >> PAGE_CACHE_SHIFT;
2529 from = pos & (PAGE_CACHE_SIZE - 1);
2532 page = grab_cache_page_write_begin(mapping, index, flags);
2538 if (page_has_buffers(page)) {
2539 ret = __block_write_begin(page, pos, len, get_block);
2545 if (PageMappedToDisk(page))
2549 * Allocate buffers so that we can keep track of state, and potentially
2550 * attach them to the page if an error occurs. In the common case of
2551 * no error, they will just be freed again without ever being attached
2552 * to the page (which is all OK, because we're under the page lock).
2554 * Be careful: the buffer linked list is a NULL terminated one, rather
2555 * than the circular one we're used to.
2557 head = alloc_page_buffers(page, blocksize, 0);
2563 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2566 * We loop across all blocks in the page, whether or not they are
2567 * part of the affected region. This is so we can discover if the
2568 * page is fully mapped-to-disk.
2570 for (block_start = 0, block_in_page = 0, bh = head;
2571 block_start < PAGE_CACHE_SIZE;
2572 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2575 block_end = block_start + blocksize;
2578 if (block_start >= to)
2580 ret = get_block(inode, block_in_file + block_in_page,
2584 if (!buffer_mapped(bh))
2585 is_mapped_to_disk = 0;
2587 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2588 if (PageUptodate(page)) {
2589 set_buffer_uptodate(bh);
2592 if (buffer_new(bh) || !buffer_mapped(bh)) {
2593 zero_user_segments(page, block_start, from,
2597 if (buffer_uptodate(bh))
2598 continue; /* reiserfs does this */
2599 if (block_start < from || block_end > to) {
2601 bh->b_end_io = end_buffer_read_nobh;
2602 submit_bh(READ, bh);
2609 * The page is locked, so these buffers are protected from
2610 * any VM or truncate activity. Hence we don't need to care
2611 * for the buffer_head refcounts.
2613 for (bh = head; bh; bh = bh->b_this_page) {
2615 if (!buffer_uptodate(bh))
2622 if (is_mapped_to_disk)
2623 SetPageMappedToDisk(page);
2625 *fsdata = head; /* to be released by nobh_write_end */
2632 * Error recovery is a bit difficult. We need to zero out blocks that
2633 * were newly allocated, and dirty them to ensure they get written out.
2634 * Buffers need to be attached to the page at this point, otherwise
2635 * the handling of potential IO errors during writeout would be hard
2636 * (could try doing synchronous writeout, but what if that fails too?)
2638 attach_nobh_buffers(page, head);
2639 page_zero_new_buffers(page, from, to);
2643 page_cache_release(page);
2648 EXPORT_SYMBOL(nobh_write_begin);
2650 int nobh_write_end(struct file *file, struct address_space *mapping,
2651 loff_t pos, unsigned len, unsigned copied,
2652 struct page *page, void *fsdata)
2654 struct inode *inode = page->mapping->host;
2655 struct buffer_head *head = fsdata;
2656 struct buffer_head *bh;
2657 BUG_ON(fsdata != NULL && page_has_buffers(page));
2659 if (unlikely(copied < len) && head)
2660 attach_nobh_buffers(page, head);
2661 if (page_has_buffers(page))
2662 return generic_write_end(file, mapping, pos, len,
2663 copied, page, fsdata);
2665 SetPageUptodate(page);
2666 set_page_dirty(page);
2667 if (pos+copied > inode->i_size) {
2668 i_size_write(inode, pos+copied);
2669 mark_inode_dirty(inode);
2673 page_cache_release(page);
2677 head = head->b_this_page;
2678 free_buffer_head(bh);
2683 EXPORT_SYMBOL(nobh_write_end);
2686 * nobh_writepage() - based on block_full_write_page() except
2687 * that it tries to operate without attaching bufferheads to
2690 int nobh_writepage(struct page *page, get_block_t *get_block,
2691 struct writeback_control *wbc)
2693 struct inode * const inode = page->mapping->host;
2694 loff_t i_size = i_size_read(inode);
2695 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2699 /* Is the page fully inside i_size? */
2700 if (page->index < end_index)
2703 /* Is the page fully outside i_size? (truncate in progress) */
2704 offset = i_size & (PAGE_CACHE_SIZE-1);
2705 if (page->index >= end_index+1 || !offset) {
2707 * The page may have dirty, unmapped buffers. For example,
2708 * they may have been added in ext3_writepage(). Make them
2709 * freeable here, so the page does not leak.
2712 /* Not really sure about this - do we need this ? */
2713 if (page->mapping->a_ops->invalidatepage)
2714 page->mapping->a_ops->invalidatepage(page, offset);
2717 return 0; /* don't care */
2721 * The page straddles i_size. It must be zeroed out on each and every
2722 * writepage invocation because it may be mmapped. "A file is mapped
2723 * in multiples of the page size. For a file that is not a multiple of
2724 * the page size, the remaining memory is zeroed when mapped, and
2725 * writes to that region are not written out to the file."
2727 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2729 ret = mpage_writepage(page, get_block, wbc);
2731 ret = __block_write_full_page(inode, page, get_block, wbc,
2732 end_buffer_async_write);
2735 EXPORT_SYMBOL(nobh_writepage);
2737 int nobh_truncate_page(struct address_space *mapping,
2738 loff_t from, get_block_t *get_block)
2740 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2741 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2744 unsigned length, pos;
2745 struct inode *inode = mapping->host;
2747 struct buffer_head map_bh;
2750 blocksize = 1 << inode->i_blkbits;
2751 length = offset & (blocksize - 1);
2753 /* Block boundary? Nothing to do */
2757 length = blocksize - length;
2758 iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2760 page = grab_cache_page(mapping, index);
2765 if (page_has_buffers(page)) {
2768 page_cache_release(page);
2769 return block_truncate_page(mapping, from, get_block);
2772 /* Find the buffer that contains "offset" */
2774 while (offset >= pos) {
2779 map_bh.b_size = blocksize;
2781 err = get_block(inode, iblock, &map_bh, 0);
2784 /* unmapped? It's a hole - nothing to do */
2785 if (!buffer_mapped(&map_bh))
2788 /* Ok, it's mapped. Make sure it's up-to-date */
2789 if (!PageUptodate(page)) {
2790 err = mapping->a_ops->readpage(NULL, page);
2792 page_cache_release(page);
2796 if (!PageUptodate(page)) {
2800 if (page_has_buffers(page))
2803 zero_user(page, offset, length);
2804 set_page_dirty(page);
2809 page_cache_release(page);
2813 EXPORT_SYMBOL(nobh_truncate_page);
2815 int block_truncate_page(struct address_space *mapping,
2816 loff_t from, get_block_t *get_block)
2818 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2819 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2822 unsigned length, pos;
2823 struct inode *inode = mapping->host;
2825 struct buffer_head *bh;
2828 blocksize = 1 << inode->i_blkbits;
2829 length = offset & (blocksize - 1);
2831 /* Block boundary? Nothing to do */
2835 length = blocksize - length;
2836 iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2838 page = grab_cache_page(mapping, index);
2843 if (!page_has_buffers(page))
2844 create_empty_buffers(page, blocksize, 0);
2846 /* Find the buffer that contains "offset" */
2847 bh = page_buffers(page);
2849 while (offset >= pos) {
2850 bh = bh->b_this_page;
2856 if (!buffer_mapped(bh)) {
2857 WARN_ON(bh->b_size != blocksize);
2858 err = get_block(inode, iblock, bh, 0);
2861 /* unmapped? It's a hole - nothing to do */
2862 if (!buffer_mapped(bh))
2866 /* Ok, it's mapped. Make sure it's up-to-date */
2867 if (PageUptodate(page))
2868 set_buffer_uptodate(bh);
2870 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2872 ll_rw_block(READ, 1, &bh);
2874 /* Uhhuh. Read error. Complain and punt. */
2875 if (!buffer_uptodate(bh))
2879 zero_user(page, offset, length);
2880 mark_buffer_dirty(bh);
2885 page_cache_release(page);
2889 EXPORT_SYMBOL(block_truncate_page);
2892 * The generic ->writepage function for buffer-backed address_spaces
2894 int block_write_full_page(struct page *page, get_block_t *get_block,
2895 struct writeback_control *wbc)
2897 struct inode * const inode = page->mapping->host;
2898 loff_t i_size = i_size_read(inode);
2899 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2902 /* Is the page fully inside i_size? */
2903 if (page->index < end_index)
2904 return __block_write_full_page(inode, page, get_block, wbc,
2905 end_buffer_async_write);
2907 /* Is the page fully outside i_size? (truncate in progress) */
2908 offset = i_size & (PAGE_CACHE_SIZE-1);
2909 if (page->index >= end_index+1 || !offset) {
2911 * The page may have dirty, unmapped buffers. For example,
2912 * they may have been added in ext3_writepage(). Make them
2913 * freeable here, so the page does not leak.
2915 do_invalidatepage(page, 0, PAGE_CACHE_SIZE);
2917 return 0; /* don't care */
2921 * The page straddles i_size. It must be zeroed out on each and every
2922 * writepage invocation because it may be mmapped. "A file is mapped
2923 * in multiples of the page size. For a file that is not a multiple of
2924 * the page size, the remaining memory is zeroed when mapped, and
2925 * writes to that region are not written out to the file."
2927 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2928 return __block_write_full_page(inode, page, get_block, wbc,
2929 end_buffer_async_write);
2931 EXPORT_SYMBOL(block_write_full_page);
2933 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2934 get_block_t *get_block)
2936 struct buffer_head tmp;
2937 struct inode *inode = mapping->host;
2940 tmp.b_size = 1 << inode->i_blkbits;
2941 get_block(inode, block, &tmp, 0);
2942 return tmp.b_blocknr;
2944 EXPORT_SYMBOL(generic_block_bmap);
2946 static void end_bio_bh_io_sync(struct bio *bio, int err)
2948 struct buffer_head *bh = bio->bi_private;
2950 if (err == -EOPNOTSUPP) {
2951 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2954 if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2955 set_bit(BH_Quiet, &bh->b_state);
2957 bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2962 * This allows us to do IO even on the odd last sectors
2963 * of a device, even if the block size is some multiple
2964 * of the physical sector size.
2966 * We'll just truncate the bio to the size of the device,
2967 * and clear the end of the buffer head manually.
2969 * Truly out-of-range accesses will turn into actual IO
2970 * errors, this only handles the "we need to be able to
2971 * do IO at the final sector" case.
2973 void guard_bio_eod(int rw, struct bio *bio)
2976 struct bio_vec *bvec = &bio->bi_io_vec[bio->bi_vcnt - 1];
2977 unsigned truncated_bytes;
2979 maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
2984 * If the *whole* IO is past the end of the device,
2985 * let it through, and the IO layer will turn it into
2988 if (unlikely(bio->bi_iter.bi_sector >= maxsector))
2991 maxsector -= bio->bi_iter.bi_sector;
2992 if (likely((bio->bi_iter.bi_size >> 9) <= maxsector))
2995 /* Uhhuh. We've got a bio that straddles the device size! */
2996 truncated_bytes = bio->bi_iter.bi_size - (maxsector << 9);
2998 /* Truncate the bio.. */
2999 bio->bi_iter.bi_size -= truncated_bytes;
3000 bvec->bv_len -= truncated_bytes;
3002 /* ..and clear the end of the buffer for reads */
3003 if ((rw & RW_MASK) == READ) {
3004 zero_user(bvec->bv_page, bvec->bv_offset + bvec->bv_len,
3009 int _submit_bh(int rw, struct buffer_head *bh, unsigned long bio_flags)
3014 BUG_ON(!buffer_locked(bh));
3015 BUG_ON(!buffer_mapped(bh));
3016 BUG_ON(!bh->b_end_io);
3017 BUG_ON(buffer_delay(bh));
3018 BUG_ON(buffer_unwritten(bh));
3021 * Only clear out a write error when rewriting
3023 if (test_set_buffer_req(bh) && (rw & WRITE))
3024 clear_buffer_write_io_error(bh);
3027 * from here on down, it's all bio -- do the initial mapping,
3028 * submit_bio -> generic_make_request may further map this bio around
3030 bio = bio_alloc(GFP_NOIO, 1);
3032 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3033 bio->bi_bdev = bh->b_bdev;
3034 bio->bi_io_vec[0].bv_page = bh->b_page;
3035 bio->bi_io_vec[0].bv_len = bh->b_size;
3036 bio->bi_io_vec[0].bv_offset = bh_offset(bh);
3039 bio->bi_iter.bi_size = bh->b_size;
3041 bio->bi_end_io = end_bio_bh_io_sync;
3042 bio->bi_private = bh;
3043 bio->bi_flags |= bio_flags;
3045 /* Take care of bh's that straddle the end of the device */
3046 guard_bio_eod(rw, bio);
3048 if (buffer_meta(bh))
3050 if (buffer_prio(bh))
3054 submit_bio(rw, bio);
3056 if (bio_flagged(bio, BIO_EOPNOTSUPP))
3062 EXPORT_SYMBOL_GPL(_submit_bh);
3064 int submit_bh(int rw, struct buffer_head *bh)
3066 return _submit_bh(rw, bh, 0);
3068 EXPORT_SYMBOL(submit_bh);
3071 * ll_rw_block: low-level access to block devices (DEPRECATED)
3072 * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
3073 * @nr: number of &struct buffer_heads in the array
3074 * @bhs: array of pointers to &struct buffer_head
3076 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3077 * requests an I/O operation on them, either a %READ or a %WRITE. The third
3078 * %READA option is described in the documentation for generic_make_request()
3079 * which ll_rw_block() calls.
3081 * This function drops any buffer that it cannot get a lock on (with the
3082 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3083 * request, and any buffer that appears to be up-to-date when doing read
3084 * request. Further it marks as clean buffers that are processed for
3085 * writing (the buffer cache won't assume that they are actually clean
3086 * until the buffer gets unlocked).
3088 * ll_rw_block sets b_end_io to simple completion handler that marks
3089 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3092 * All of the buffers must be for the same device, and must also be a
3093 * multiple of the current approved size for the device.
3095 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
3099 for (i = 0; i < nr; i++) {
3100 struct buffer_head *bh = bhs[i];
3102 if (!trylock_buffer(bh))
3105 if (test_clear_buffer_dirty(bh)) {
3106 bh->b_end_io = end_buffer_write_sync;
3108 submit_bh(WRITE, bh);
3112 if (!buffer_uptodate(bh)) {
3113 bh->b_end_io = end_buffer_read_sync;
3122 EXPORT_SYMBOL(ll_rw_block);
3124 void write_dirty_buffer(struct buffer_head *bh, int rw)
3127 if (!test_clear_buffer_dirty(bh)) {
3131 bh->b_end_io = end_buffer_write_sync;
3135 EXPORT_SYMBOL(write_dirty_buffer);
3138 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3139 * and then start new I/O and then wait upon it. The caller must have a ref on
3142 int __sync_dirty_buffer(struct buffer_head *bh, int rw)
3146 WARN_ON(atomic_read(&bh->b_count) < 1);
3148 if (test_clear_buffer_dirty(bh)) {
3150 bh->b_end_io = end_buffer_write_sync;
3151 ret = submit_bh(rw, bh);
3153 if (!ret && !buffer_uptodate(bh))
3160 EXPORT_SYMBOL(__sync_dirty_buffer);
3162 int sync_dirty_buffer(struct buffer_head *bh)
3164 return __sync_dirty_buffer(bh, WRITE_SYNC);
3166 EXPORT_SYMBOL(sync_dirty_buffer);
3169 * try_to_free_buffers() checks if all the buffers on this particular page
3170 * are unused, and releases them if so.
3172 * Exclusion against try_to_free_buffers may be obtained by either
3173 * locking the page or by holding its mapping's private_lock.
3175 * If the page is dirty but all the buffers are clean then we need to
3176 * be sure to mark the page clean as well. This is because the page
3177 * may be against a block device, and a later reattachment of buffers
3178 * to a dirty page will set *all* buffers dirty. Which would corrupt
3179 * filesystem data on the same device.
3181 * The same applies to regular filesystem pages: if all the buffers are
3182 * clean then we set the page clean and proceed. To do that, we require
3183 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3186 * try_to_free_buffers() is non-blocking.
3188 static inline int buffer_busy(struct buffer_head *bh)
3190 return atomic_read(&bh->b_count) |
3191 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3195 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3197 struct buffer_head *head = page_buffers(page);
3198 struct buffer_head *bh;
3202 if (buffer_write_io_error(bh) && page->mapping)
3203 set_bit(AS_EIO, &page->mapping->flags);
3204 if (buffer_busy(bh))
3206 bh = bh->b_this_page;
3207 } while (bh != head);
3210 struct buffer_head *next = bh->b_this_page;
3212 if (bh->b_assoc_map)
3213 __remove_assoc_queue(bh);
3215 } while (bh != head);
3216 *buffers_to_free = head;
3217 __clear_page_buffers(page);
3223 int try_to_free_buffers(struct page *page)
3225 struct address_space * const mapping = page->mapping;
3226 struct buffer_head *buffers_to_free = NULL;
3229 BUG_ON(!PageLocked(page));
3230 if (PageWriteback(page))
3233 if (mapping == NULL) { /* can this still happen? */
3234 ret = drop_buffers(page, &buffers_to_free);
3238 spin_lock(&mapping->private_lock);
3239 ret = drop_buffers(page, &buffers_to_free);
3242 * If the filesystem writes its buffers by hand (eg ext3)
3243 * then we can have clean buffers against a dirty page. We
3244 * clean the page here; otherwise the VM will never notice
3245 * that the filesystem did any IO at all.
3247 * Also, during truncate, discard_buffer will have marked all
3248 * the page's buffers clean. We discover that here and clean
3251 * private_lock must be held over this entire operation in order
3252 * to synchronise against __set_page_dirty_buffers and prevent the
3253 * dirty bit from being lost.
3256 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3257 spin_unlock(&mapping->private_lock);
3259 if (buffers_to_free) {
3260 struct buffer_head *bh = buffers_to_free;
3263 struct buffer_head *next = bh->b_this_page;
3264 free_buffer_head(bh);
3266 } while (bh != buffers_to_free);
3270 EXPORT_SYMBOL(try_to_free_buffers);
3273 * There are no bdflush tunables left. But distributions are
3274 * still running obsolete flush daemons, so we terminate them here.
3276 * Use of bdflush() is deprecated and will be removed in a future kernel.
3277 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3279 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3281 static int msg_count;
3283 if (!capable(CAP_SYS_ADMIN))
3286 if (msg_count < 5) {
3289 "warning: process `%s' used the obsolete bdflush"
3290 " system call\n", current->comm);
3291 printk(KERN_INFO "Fix your initscripts?\n");
3300 * Buffer-head allocation
3302 static struct kmem_cache *bh_cachep __read_mostly;
3305 * Once the number of bh's in the machine exceeds this level, we start
3306 * stripping them in writeback.
3308 static unsigned long max_buffer_heads;
3310 int buffer_heads_over_limit;
3312 struct bh_accounting {
3313 int nr; /* Number of live bh's */
3314 int ratelimit; /* Limit cacheline bouncing */
3317 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3319 static void recalc_bh_state(void)
3324 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3326 __this_cpu_write(bh_accounting.ratelimit, 0);
3327 for_each_online_cpu(i)
3328 tot += per_cpu(bh_accounting, i).nr;
3329 buffer_heads_over_limit = (tot > max_buffer_heads);
3332 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3334 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3336 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3338 __this_cpu_inc(bh_accounting.nr);
3344 EXPORT_SYMBOL(alloc_buffer_head);
3346 void free_buffer_head(struct buffer_head *bh)
3348 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3349 kmem_cache_free(bh_cachep, bh);
3351 __this_cpu_dec(bh_accounting.nr);
3355 EXPORT_SYMBOL(free_buffer_head);
3357 static void buffer_exit_cpu(int cpu)
3360 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3362 for (i = 0; i < BH_LRU_SIZE; i++) {
3366 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3367 per_cpu(bh_accounting, cpu).nr = 0;
3370 static int buffer_cpu_notify(struct notifier_block *self,
3371 unsigned long action, void *hcpu)
3373 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3374 buffer_exit_cpu((unsigned long)hcpu);
3379 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3380 * @bh: struct buffer_head
3382 * Return true if the buffer is up-to-date and false,
3383 * with the buffer locked, if not.
3385 int bh_uptodate_or_lock(struct buffer_head *bh)
3387 if (!buffer_uptodate(bh)) {
3389 if (!buffer_uptodate(bh))
3395 EXPORT_SYMBOL(bh_uptodate_or_lock);
3398 * bh_submit_read - Submit a locked buffer for reading
3399 * @bh: struct buffer_head
3401 * Returns zero on success and -EIO on error.
3403 int bh_submit_read(struct buffer_head *bh)
3405 BUG_ON(!buffer_locked(bh));
3407 if (buffer_uptodate(bh)) {
3413 bh->b_end_io = end_buffer_read_sync;
3414 submit_bh(READ, bh);
3416 if (buffer_uptodate(bh))
3420 EXPORT_SYMBOL(bh_submit_read);
3422 void __init buffer_init(void)
3424 unsigned long nrpages;
3426 bh_cachep = kmem_cache_create("buffer_head",
3427 sizeof(struct buffer_head), 0,
3428 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3433 * Limit the bh occupancy to 10% of ZONE_NORMAL
3435 nrpages = (nr_free_buffer_pages() * 10) / 100;
3436 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3437 hotcpu_notifier(buffer_cpu_notify, 0);