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>
24 #include <linux/iomap.h>
26 #include <linux/percpu.h>
27 #include <linux/slab.h>
28 #include <linux/capability.h>
29 #include <linux/blkdev.h>
30 #include <linux/file.h>
31 #include <linux/quotaops.h>
32 #include <linux/highmem.h>
33 #include <linux/export.h>
34 #include <linux/backing-dev.h>
35 #include <linux/writeback.h>
36 #include <linux/hash.h>
37 #include <linux/suspend.h>
38 #include <linux/buffer_head.h>
39 #include <linux/task_io_accounting_ops.h>
40 #include <linux/bio.h>
41 #include <linux/notifier.h>
42 #include <linux/cpu.h>
43 #include <linux/bitops.h>
44 #include <linux/mpage.h>
45 #include <linux/bit_spinlock.h>
46 #include <linux/pagevec.h>
47 #include <trace/events/block.h>
49 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
50 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
51 unsigned long bio_flags,
52 struct writeback_control *wbc);
54 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
56 void init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
58 bh->b_end_io = handler;
59 bh->b_private = private;
61 EXPORT_SYMBOL(init_buffer);
63 inline void touch_buffer(struct buffer_head *bh)
65 trace_block_touch_buffer(bh);
66 mark_page_accessed(bh->b_page);
68 EXPORT_SYMBOL(touch_buffer);
70 void __lock_buffer(struct buffer_head *bh)
72 wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
74 EXPORT_SYMBOL(__lock_buffer);
76 void unlock_buffer(struct buffer_head *bh)
78 clear_bit_unlock(BH_Lock, &bh->b_state);
79 smp_mb__after_atomic();
80 wake_up_bit(&bh->b_state, BH_Lock);
82 EXPORT_SYMBOL(unlock_buffer);
85 * Returns if the page has dirty or writeback buffers. If all the buffers
86 * are unlocked and clean then the PageDirty information is stale. If
87 * any of the pages are locked, it is assumed they are locked for IO.
89 void buffer_check_dirty_writeback(struct page *page,
90 bool *dirty, bool *writeback)
92 struct buffer_head *head, *bh;
96 BUG_ON(!PageLocked(page));
98 if (!page_has_buffers(page))
101 if (PageWriteback(page))
104 head = page_buffers(page);
107 if (buffer_locked(bh))
110 if (buffer_dirty(bh))
113 bh = bh->b_this_page;
114 } while (bh != head);
116 EXPORT_SYMBOL(buffer_check_dirty_writeback);
119 * Block until a buffer comes unlocked. This doesn't stop it
120 * from becoming locked again - you have to lock it yourself
121 * if you want to preserve its state.
123 void __wait_on_buffer(struct buffer_head * bh)
125 wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
127 EXPORT_SYMBOL(__wait_on_buffer);
130 __clear_page_buffers(struct page *page)
132 ClearPagePrivate(page);
133 set_page_private(page, 0);
137 static void buffer_io_error(struct buffer_head *bh, char *msg)
139 if (!test_bit(BH_Quiet, &bh->b_state))
140 printk_ratelimited(KERN_ERR
141 "Buffer I/O error on dev %pg, logical block %llu%s\n",
142 bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
146 * End-of-IO handler helper function which does not touch the bh after
148 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
149 * a race there is benign: unlock_buffer() only use the bh's address for
150 * hashing after unlocking the buffer, so it doesn't actually touch the bh
153 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
156 set_buffer_uptodate(bh);
158 /* This happens, due to failed read-ahead attempts. */
159 clear_buffer_uptodate(bh);
165 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
166 * unlock the buffer. This is what ll_rw_block uses too.
168 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
170 __end_buffer_read_notouch(bh, uptodate);
173 EXPORT_SYMBOL(end_buffer_read_sync);
175 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
178 set_buffer_uptodate(bh);
180 buffer_io_error(bh, ", lost sync page write");
181 set_buffer_write_io_error(bh);
182 clear_buffer_uptodate(bh);
187 EXPORT_SYMBOL(end_buffer_write_sync);
190 * Various filesystems appear to want __find_get_block to be non-blocking.
191 * But it's the page lock which protects the buffers. To get around this,
192 * we get exclusion from try_to_free_buffers with the blockdev mapping's
195 * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
196 * may be quite high. This code could TryLock the page, and if that
197 * succeeds, there is no need to take private_lock. (But if
198 * private_lock is contended then so is mapping->tree_lock).
200 static struct buffer_head *
201 __find_get_block_slow(struct block_device *bdev, sector_t block)
203 struct inode *bd_inode = bdev->bd_inode;
204 struct address_space *bd_mapping = bd_inode->i_mapping;
205 struct buffer_head *ret = NULL;
207 struct buffer_head *bh;
208 struct buffer_head *head;
212 index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
213 page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
217 spin_lock(&bd_mapping->private_lock);
218 if (!page_has_buffers(page))
220 head = page_buffers(page);
223 if (!buffer_mapped(bh))
225 else if (bh->b_blocknr == block) {
230 bh = bh->b_this_page;
231 } while (bh != head);
233 /* we might be here because some of the buffers on this page are
234 * not mapped. This is due to various races between
235 * file io on the block device and getblk. It gets dealt with
236 * elsewhere, don't buffer_error if we had some unmapped buffers
239 printk("__find_get_block_slow() failed. "
240 "block=%llu, b_blocknr=%llu\n",
241 (unsigned long long)block,
242 (unsigned long long)bh->b_blocknr);
243 printk("b_state=0x%08lx, b_size=%zu\n",
244 bh->b_state, bh->b_size);
245 printk("device %pg blocksize: %d\n", bdev,
246 1 << bd_inode->i_blkbits);
249 spin_unlock(&bd_mapping->private_lock);
256 * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
258 static void free_more_memory(void)
263 wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
266 for_each_online_node(nid) {
268 z = first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
269 gfp_zone(GFP_NOFS), NULL);
271 try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
277 * I/O completion handler for block_read_full_page() - pages
278 * which come unlocked at the end of I/O.
280 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
283 struct buffer_head *first;
284 struct buffer_head *tmp;
286 int page_uptodate = 1;
288 BUG_ON(!buffer_async_read(bh));
292 set_buffer_uptodate(bh);
294 clear_buffer_uptodate(bh);
295 buffer_io_error(bh, ", async page read");
300 * Be _very_ careful from here on. Bad things can happen if
301 * two buffer heads end IO at almost the same time and both
302 * decide that the page is now completely done.
304 first = page_buffers(page);
305 local_irq_save(flags);
306 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
307 clear_buffer_async_read(bh);
311 if (!buffer_uptodate(tmp))
313 if (buffer_async_read(tmp)) {
314 BUG_ON(!buffer_locked(tmp));
317 tmp = tmp->b_this_page;
319 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
320 local_irq_restore(flags);
323 * If none of the buffers had errors and they are all
324 * uptodate then we can set the page uptodate.
326 if (page_uptodate && !PageError(page))
327 SetPageUptodate(page);
332 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
333 local_irq_restore(flags);
338 * Completion handler for block_write_full_page() - pages which are unlocked
339 * during I/O, and which have PageWriteback cleared upon I/O completion.
341 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
344 struct buffer_head *first;
345 struct buffer_head *tmp;
348 BUG_ON(!buffer_async_write(bh));
352 set_buffer_uptodate(bh);
354 buffer_io_error(bh, ", lost async page write");
355 mapping_set_error(page->mapping, -EIO);
356 set_buffer_write_io_error(bh);
357 clear_buffer_uptodate(bh);
361 first = page_buffers(page);
362 local_irq_save(flags);
363 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
365 clear_buffer_async_write(bh);
367 tmp = bh->b_this_page;
369 if (buffer_async_write(tmp)) {
370 BUG_ON(!buffer_locked(tmp));
373 tmp = tmp->b_this_page;
375 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
376 local_irq_restore(flags);
377 end_page_writeback(page);
381 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
382 local_irq_restore(flags);
385 EXPORT_SYMBOL(end_buffer_async_write);
388 * If a page's buffers are under async readin (end_buffer_async_read
389 * completion) then there is a possibility that another thread of
390 * control could lock one of the buffers after it has completed
391 * but while some of the other buffers have not completed. This
392 * locked buffer would confuse end_buffer_async_read() into not unlocking
393 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
394 * that this buffer is not under async I/O.
396 * The page comes unlocked when it has no locked buffer_async buffers
399 * PageLocked prevents anyone starting new async I/O reads any of
402 * PageWriteback is used to prevent simultaneous writeout of the same
405 * PageLocked prevents anyone from starting writeback of a page which is
406 * under read I/O (PageWriteback is only ever set against a locked page).
408 static void mark_buffer_async_read(struct buffer_head *bh)
410 bh->b_end_io = end_buffer_async_read;
411 set_buffer_async_read(bh);
414 static void mark_buffer_async_write_endio(struct buffer_head *bh,
415 bh_end_io_t *handler)
417 bh->b_end_io = handler;
418 set_buffer_async_write(bh);
421 void mark_buffer_async_write(struct buffer_head *bh)
423 mark_buffer_async_write_endio(bh, end_buffer_async_write);
425 EXPORT_SYMBOL(mark_buffer_async_write);
429 * fs/buffer.c contains helper functions for buffer-backed address space's
430 * fsync functions. A common requirement for buffer-based filesystems is
431 * that certain data from the backing blockdev needs to be written out for
432 * a successful fsync(). For example, ext2 indirect blocks need to be
433 * written back and waited upon before fsync() returns.
435 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
436 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
437 * management of a list of dependent buffers at ->i_mapping->private_list.
439 * Locking is a little subtle: try_to_free_buffers() will remove buffers
440 * from their controlling inode's queue when they are being freed. But
441 * try_to_free_buffers() will be operating against the *blockdev* mapping
442 * at the time, not against the S_ISREG file which depends on those buffers.
443 * So the locking for private_list is via the private_lock in the address_space
444 * which backs the buffers. Which is different from the address_space
445 * against which the buffers are listed. So for a particular address_space,
446 * mapping->private_lock does *not* protect mapping->private_list! In fact,
447 * mapping->private_list will always be protected by the backing blockdev's
450 * Which introduces a requirement: all buffers on an address_space's
451 * ->private_list must be from the same address_space: the blockdev's.
453 * address_spaces which do not place buffers at ->private_list via these
454 * utility functions are free to use private_lock and private_list for
455 * whatever they want. The only requirement is that list_empty(private_list)
456 * be true at clear_inode() time.
458 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
459 * filesystems should do that. invalidate_inode_buffers() should just go
460 * BUG_ON(!list_empty).
462 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
463 * take an address_space, not an inode. And it should be called
464 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
467 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
468 * list if it is already on a list. Because if the buffer is on a list,
469 * it *must* already be on the right one. If not, the filesystem is being
470 * silly. This will save a ton of locking. But first we have to ensure
471 * that buffers are taken *off* the old inode's list when they are freed
472 * (presumably in truncate). That requires careful auditing of all
473 * filesystems (do it inside bforget()). It could also be done by bringing
478 * The buffer's backing address_space's private_lock must be held
480 static void __remove_assoc_queue(struct buffer_head *bh)
482 list_del_init(&bh->b_assoc_buffers);
483 WARN_ON(!bh->b_assoc_map);
484 if (buffer_write_io_error(bh))
485 set_bit(AS_EIO, &bh->b_assoc_map->flags);
486 bh->b_assoc_map = NULL;
489 int inode_has_buffers(struct inode *inode)
491 return !list_empty(&inode->i_data.private_list);
495 * osync is designed to support O_SYNC io. It waits synchronously for
496 * all already-submitted IO to complete, but does not queue any new
497 * writes to the disk.
499 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
500 * you dirty the buffers, and then use osync_inode_buffers to wait for
501 * completion. Any other dirty buffers which are not yet queued for
502 * write will not be flushed to disk by the osync.
504 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
506 struct buffer_head *bh;
512 list_for_each_prev(p, list) {
514 if (buffer_locked(bh)) {
518 if (!buffer_uptodate(bh))
529 static void do_thaw_one(struct super_block *sb, void *unused)
531 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
532 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
535 static void do_thaw_all(struct work_struct *work)
537 iterate_supers(do_thaw_one, NULL);
539 printk(KERN_WARNING "Emergency Thaw complete\n");
543 * emergency_thaw_all -- forcibly thaw every frozen filesystem
545 * Used for emergency unfreeze of all filesystems via SysRq
547 void emergency_thaw_all(void)
549 struct work_struct *work;
551 work = kmalloc(sizeof(*work), GFP_ATOMIC);
553 INIT_WORK(work, do_thaw_all);
559 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
560 * @mapping: the mapping which wants those buffers written
562 * Starts I/O against the buffers at mapping->private_list, and waits upon
565 * Basically, this is a convenience function for fsync().
566 * @mapping is a file or directory which needs those buffers to be written for
567 * a successful fsync().
569 int sync_mapping_buffers(struct address_space *mapping)
571 struct address_space *buffer_mapping = mapping->private_data;
573 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
576 return fsync_buffers_list(&buffer_mapping->private_lock,
577 &mapping->private_list);
579 EXPORT_SYMBOL(sync_mapping_buffers);
582 * Called when we've recently written block `bblock', and it is known that
583 * `bblock' was for a buffer_boundary() buffer. This means that the block at
584 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
585 * dirty, schedule it for IO. So that indirects merge nicely with their data.
587 void write_boundary_block(struct block_device *bdev,
588 sector_t bblock, unsigned blocksize)
590 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
592 if (buffer_dirty(bh))
593 ll_rw_block(REQ_OP_WRITE, 0, 1, &bh);
598 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
600 struct address_space *mapping = inode->i_mapping;
601 struct address_space *buffer_mapping = bh->b_page->mapping;
603 mark_buffer_dirty(bh);
604 if (!mapping->private_data) {
605 mapping->private_data = buffer_mapping;
607 BUG_ON(mapping->private_data != buffer_mapping);
609 if (!bh->b_assoc_map) {
610 spin_lock(&buffer_mapping->private_lock);
611 list_move_tail(&bh->b_assoc_buffers,
612 &mapping->private_list);
613 bh->b_assoc_map = mapping;
614 spin_unlock(&buffer_mapping->private_lock);
617 EXPORT_SYMBOL(mark_buffer_dirty_inode);
620 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
623 * If warn is true, then emit a warning if the page is not uptodate and has
624 * not been truncated.
626 * The caller must hold lock_page_memcg().
628 static void __set_page_dirty(struct page *page, struct address_space *mapping,
633 spin_lock_irqsave(&mapping->tree_lock, flags);
634 if (page->mapping) { /* Race with truncate? */
635 WARN_ON_ONCE(warn && !PageUptodate(page));
636 account_page_dirtied(page, mapping);
637 radix_tree_tag_set(&mapping->page_tree,
638 page_index(page), PAGECACHE_TAG_DIRTY);
640 spin_unlock_irqrestore(&mapping->tree_lock, flags);
644 * Add a page to the dirty page list.
646 * It is a sad fact of life that this function is called from several places
647 * deeply under spinlocking. It may not sleep.
649 * If the page has buffers, the uptodate buffers are set dirty, to preserve
650 * dirty-state coherency between the page and the buffers. It the page does
651 * not have buffers then when they are later attached they will all be set
654 * The buffers are dirtied before the page is dirtied. There's a small race
655 * window in which a writepage caller may see the page cleanness but not the
656 * buffer dirtiness. That's fine. If this code were to set the page dirty
657 * before the buffers, a concurrent writepage caller could clear the page dirty
658 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
659 * page on the dirty page list.
661 * We use private_lock to lock against try_to_free_buffers while using the
662 * page's buffer list. Also use this to protect against clean buffers being
663 * added to the page after it was set dirty.
665 * FIXME: may need to call ->reservepage here as well. That's rather up to the
666 * address_space though.
668 int __set_page_dirty_buffers(struct page *page)
671 struct address_space *mapping = page_mapping(page);
673 if (unlikely(!mapping))
674 return !TestSetPageDirty(page);
676 spin_lock(&mapping->private_lock);
677 if (page_has_buffers(page)) {
678 struct buffer_head *head = page_buffers(page);
679 struct buffer_head *bh = head;
682 set_buffer_dirty(bh);
683 bh = bh->b_this_page;
684 } while (bh != head);
687 * Lock out page->mem_cgroup migration to keep PageDirty
688 * synchronized with per-memcg dirty page counters.
690 lock_page_memcg(page);
691 newly_dirty = !TestSetPageDirty(page);
692 spin_unlock(&mapping->private_lock);
695 __set_page_dirty(page, mapping, 1);
697 unlock_page_memcg(page);
700 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
704 EXPORT_SYMBOL(__set_page_dirty_buffers);
707 * Write out and wait upon a list of buffers.
709 * We have conflicting pressures: we want to make sure that all
710 * initially dirty buffers get waited on, but that any subsequently
711 * dirtied buffers don't. After all, we don't want fsync to last
712 * forever if somebody is actively writing to the file.
714 * Do this in two main stages: first we copy dirty buffers to a
715 * temporary inode list, queueing the writes as we go. Then we clean
716 * up, waiting for those writes to complete.
718 * During this second stage, any subsequent updates to the file may end
719 * up refiling the buffer on the original inode's dirty list again, so
720 * there is a chance we will end up with a buffer queued for write but
721 * not yet completed on that list. So, as a final cleanup we go through
722 * the osync code to catch these locked, dirty buffers without requeuing
723 * any newly dirty buffers for write.
725 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
727 struct buffer_head *bh;
728 struct list_head tmp;
729 struct address_space *mapping;
731 struct blk_plug plug;
733 INIT_LIST_HEAD(&tmp);
734 blk_start_plug(&plug);
737 while (!list_empty(list)) {
738 bh = BH_ENTRY(list->next);
739 mapping = bh->b_assoc_map;
740 __remove_assoc_queue(bh);
741 /* Avoid race with mark_buffer_dirty_inode() which does
742 * a lockless check and we rely on seeing the dirty bit */
744 if (buffer_dirty(bh) || buffer_locked(bh)) {
745 list_add(&bh->b_assoc_buffers, &tmp);
746 bh->b_assoc_map = mapping;
747 if (buffer_dirty(bh)) {
751 * Ensure any pending I/O completes so that
752 * write_dirty_buffer() actually writes the
753 * current contents - it is a noop if I/O is
754 * still in flight on potentially older
757 write_dirty_buffer(bh, REQ_SYNC);
760 * Kick off IO for the previous mapping. Note
761 * that we will not run the very last mapping,
762 * wait_on_buffer() will do that for us
763 * through sync_buffer().
772 blk_finish_plug(&plug);
775 while (!list_empty(&tmp)) {
776 bh = BH_ENTRY(tmp.prev);
778 mapping = bh->b_assoc_map;
779 __remove_assoc_queue(bh);
780 /* Avoid race with mark_buffer_dirty_inode() which does
781 * a lockless check and we rely on seeing the dirty bit */
783 if (buffer_dirty(bh)) {
784 list_add(&bh->b_assoc_buffers,
785 &mapping->private_list);
786 bh->b_assoc_map = mapping;
790 if (!buffer_uptodate(bh))
797 err2 = osync_buffers_list(lock, list);
805 * Invalidate any and all dirty buffers on a given inode. We are
806 * probably unmounting the fs, but that doesn't mean we have already
807 * done a sync(). Just drop the buffers from the inode list.
809 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
810 * assumes that all the buffers are against the blockdev. Not true
813 void invalidate_inode_buffers(struct inode *inode)
815 if (inode_has_buffers(inode)) {
816 struct address_space *mapping = &inode->i_data;
817 struct list_head *list = &mapping->private_list;
818 struct address_space *buffer_mapping = mapping->private_data;
820 spin_lock(&buffer_mapping->private_lock);
821 while (!list_empty(list))
822 __remove_assoc_queue(BH_ENTRY(list->next));
823 spin_unlock(&buffer_mapping->private_lock);
826 EXPORT_SYMBOL(invalidate_inode_buffers);
829 * Remove any clean buffers from the inode's buffer list. This is called
830 * when we're trying to free the inode itself. Those buffers can pin it.
832 * Returns true if all buffers were removed.
834 int remove_inode_buffers(struct inode *inode)
838 if (inode_has_buffers(inode)) {
839 struct address_space *mapping = &inode->i_data;
840 struct list_head *list = &mapping->private_list;
841 struct address_space *buffer_mapping = mapping->private_data;
843 spin_lock(&buffer_mapping->private_lock);
844 while (!list_empty(list)) {
845 struct buffer_head *bh = BH_ENTRY(list->next);
846 if (buffer_dirty(bh)) {
850 __remove_assoc_queue(bh);
852 spin_unlock(&buffer_mapping->private_lock);
858 * Create the appropriate buffers when given a page for data area and
859 * the size of each buffer.. Use the bh->b_this_page linked list to
860 * follow the buffers created. Return NULL if unable to create more
863 * The retry flag is used to differentiate async IO (paging, swapping)
864 * which may not fail from ordinary buffer allocations.
866 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
869 struct buffer_head *bh, *head;
875 while ((offset -= size) >= 0) {
876 bh = alloc_buffer_head(GFP_NOFS);
880 bh->b_this_page = head;
886 /* Link the buffer to its page */
887 set_bh_page(bh, page, offset);
891 * In case anything failed, we just free everything we got.
897 head = head->b_this_page;
898 free_buffer_head(bh);
903 * Return failure for non-async IO requests. Async IO requests
904 * are not allowed to fail, so we have to wait until buffer heads
905 * become available. But we don't want tasks sleeping with
906 * partially complete buffers, so all were released above.
911 /* We're _really_ low on memory. Now we just
912 * wait for old buffer heads to become free due to
913 * finishing IO. Since this is an async request and
914 * the reserve list is empty, we're sure there are
915 * async buffer heads in use.
920 EXPORT_SYMBOL_GPL(alloc_page_buffers);
923 link_dev_buffers(struct page *page, struct buffer_head *head)
925 struct buffer_head *bh, *tail;
930 bh = bh->b_this_page;
932 tail->b_this_page = head;
933 attach_page_buffers(page, head);
936 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
938 sector_t retval = ~((sector_t)0);
939 loff_t sz = i_size_read(bdev->bd_inode);
942 unsigned int sizebits = blksize_bits(size);
943 retval = (sz >> sizebits);
949 * Initialise the state of a blockdev page's buffers.
952 init_page_buffers(struct page *page, struct block_device *bdev,
953 sector_t block, int size)
955 struct buffer_head *head = page_buffers(page);
956 struct buffer_head *bh = head;
957 int uptodate = PageUptodate(page);
958 sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
961 if (!buffer_mapped(bh)) {
962 init_buffer(bh, NULL, NULL);
964 bh->b_blocknr = block;
966 set_buffer_uptodate(bh);
967 if (block < end_block)
968 set_buffer_mapped(bh);
971 bh = bh->b_this_page;
972 } while (bh != head);
975 * Caller needs to validate requested block against end of device.
981 * Create the page-cache page that contains the requested block.
983 * This is used purely for blockdev mappings.
986 grow_dev_page(struct block_device *bdev, sector_t block,
987 pgoff_t index, int size, int sizebits, gfp_t gfp)
989 struct inode *inode = bdev->bd_inode;
991 struct buffer_head *bh;
993 int ret = 0; /* Will call free_more_memory() */
996 gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
999 * XXX: __getblk_slow() can not really deal with failure and
1000 * will endlessly loop on improvised global reclaim. Prefer
1001 * looping in the allocator rather than here, at least that
1002 * code knows what it's doing.
1004 gfp_mask |= __GFP_NOFAIL;
1006 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
1010 BUG_ON(!PageLocked(page));
1012 if (page_has_buffers(page)) {
1013 bh = page_buffers(page);
1014 if (bh->b_size == size) {
1015 end_block = init_page_buffers(page, bdev,
1016 (sector_t)index << sizebits,
1020 if (!try_to_free_buffers(page))
1025 * Allocate some buffers for this page
1027 bh = alloc_page_buffers(page, size, 0);
1032 * Link the page to the buffers and initialise them. Take the
1033 * lock to be atomic wrt __find_get_block(), which does not
1034 * run under the page lock.
1036 spin_lock(&inode->i_mapping->private_lock);
1037 link_dev_buffers(page, bh);
1038 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1040 spin_unlock(&inode->i_mapping->private_lock);
1042 ret = (block < end_block) ? 1 : -ENXIO;
1050 * Create buffers for the specified block device block's page. If
1051 * that page was dirty, the buffers are set dirty also.
1054 grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
1062 } while ((size << sizebits) < PAGE_SIZE);
1064 index = block >> sizebits;
1067 * Check for a block which wants to lie outside our maximum possible
1068 * pagecache index. (this comparison is done using sector_t types).
1070 if (unlikely(index != block >> sizebits)) {
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 static 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);
1115 * The relationship between dirty buffers and dirty pages:
1117 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1118 * the page is tagged dirty in its radix tree.
1120 * At all times, the dirtiness of the buffers represents the dirtiness of
1121 * subsections of the page. If the page has buffers, the page dirty bit is
1122 * merely a hint about the true dirty state.
1124 * When a page is set dirty in its entirety, all its buffers are marked dirty
1125 * (if the page has buffers).
1127 * When a buffer is marked dirty, its page is dirtied, but the page's other
1130 * Also. When blockdev buffers are explicitly read with bread(), they
1131 * individually become uptodate. But their backing page remains not
1132 * uptodate - even if all of its buffers are uptodate. A subsequent
1133 * block_read_full_page() against that page will discover all the uptodate
1134 * buffers, will set the page uptodate and will perform no I/O.
1138 * mark_buffer_dirty - mark a buffer_head as needing writeout
1139 * @bh: the buffer_head to mark dirty
1141 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1142 * backing page dirty, then tag the page as dirty in its address_space's radix
1143 * tree and then attach the address_space's inode to its superblock's dirty
1146 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1147 * mapping->tree_lock and mapping->host->i_lock.
1149 void mark_buffer_dirty(struct buffer_head *bh)
1151 WARN_ON_ONCE(!buffer_uptodate(bh));
1153 trace_block_dirty_buffer(bh);
1156 * Very *carefully* optimize the it-is-already-dirty case.
1158 * Don't let the final "is it dirty" escape to before we
1159 * perhaps modified the buffer.
1161 if (buffer_dirty(bh)) {
1163 if (buffer_dirty(bh))
1167 if (!test_set_buffer_dirty(bh)) {
1168 struct page *page = bh->b_page;
1169 struct address_space *mapping = NULL;
1171 lock_page_memcg(page);
1172 if (!TestSetPageDirty(page)) {
1173 mapping = page_mapping(page);
1175 __set_page_dirty(page, mapping, 0);
1177 unlock_page_memcg(page);
1179 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1182 EXPORT_SYMBOL(mark_buffer_dirty);
1185 * Decrement a buffer_head's reference count. If all buffers against a page
1186 * have zero reference count, are clean and unlocked, and if the page is clean
1187 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1188 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1189 * a page but it ends up not being freed, and buffers may later be reattached).
1191 void __brelse(struct buffer_head * buf)
1193 if (atomic_read(&buf->b_count)) {
1197 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1199 EXPORT_SYMBOL(__brelse);
1202 * bforget() is like brelse(), except it discards any
1203 * potentially dirty data.
1205 void __bforget(struct buffer_head *bh)
1207 clear_buffer_dirty(bh);
1208 if (bh->b_assoc_map) {
1209 struct address_space *buffer_mapping = bh->b_page->mapping;
1211 spin_lock(&buffer_mapping->private_lock);
1212 list_del_init(&bh->b_assoc_buffers);
1213 bh->b_assoc_map = NULL;
1214 spin_unlock(&buffer_mapping->private_lock);
1218 EXPORT_SYMBOL(__bforget);
1220 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1223 if (buffer_uptodate(bh)) {
1228 bh->b_end_io = end_buffer_read_sync;
1229 submit_bh(REQ_OP_READ, 0, bh);
1231 if (buffer_uptodate(bh))
1239 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1240 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1241 * refcount elevated by one when they're in an LRU. A buffer can only appear
1242 * once in a particular CPU's LRU. A single buffer can be present in multiple
1243 * CPU's LRUs at the same time.
1245 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1246 * sb_find_get_block().
1248 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1249 * a local interrupt disable for that.
1252 #define BH_LRU_SIZE 16
1255 struct buffer_head *bhs[BH_LRU_SIZE];
1258 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1261 #define bh_lru_lock() local_irq_disable()
1262 #define bh_lru_unlock() local_irq_enable()
1264 #define bh_lru_lock() preempt_disable()
1265 #define bh_lru_unlock() preempt_enable()
1268 static inline void check_irqs_on(void)
1270 #ifdef irqs_disabled
1271 BUG_ON(irqs_disabled());
1276 * The LRU management algorithm is dopey-but-simple. Sorry.
1278 static void bh_lru_install(struct buffer_head *bh)
1280 struct buffer_head *evictee = NULL;
1284 if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1285 struct buffer_head *bhs[BH_LRU_SIZE];
1291 for (in = 0; in < BH_LRU_SIZE; in++) {
1292 struct buffer_head *bh2 =
1293 __this_cpu_read(bh_lrus.bhs[in]);
1298 if (out >= BH_LRU_SIZE) {
1299 BUG_ON(evictee != NULL);
1306 while (out < BH_LRU_SIZE)
1308 memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1317 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1319 static struct buffer_head *
1320 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1322 struct buffer_head *ret = NULL;
1327 for (i = 0; i < BH_LRU_SIZE; i++) {
1328 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1330 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1331 bh->b_size == size) {
1334 __this_cpu_write(bh_lrus.bhs[i],
1335 __this_cpu_read(bh_lrus.bhs[i - 1]));
1338 __this_cpu_write(bh_lrus.bhs[0], bh);
1350 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1351 * it in the LRU and mark it as accessed. If it is not present then return
1354 struct buffer_head *
1355 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1357 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1360 /* __find_get_block_slow will mark the page accessed */
1361 bh = __find_get_block_slow(bdev, block);
1369 EXPORT_SYMBOL(__find_get_block);
1372 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1373 * which corresponds to the passed block_device, block and size. The
1374 * returned buffer has its reference count incremented.
1376 * __getblk_gfp() will lock up the machine if grow_dev_page's
1377 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1379 struct buffer_head *
1380 __getblk_gfp(struct block_device *bdev, sector_t block,
1381 unsigned size, gfp_t gfp)
1383 struct buffer_head *bh = __find_get_block(bdev, block, size);
1387 bh = __getblk_slow(bdev, block, size, gfp);
1390 EXPORT_SYMBOL(__getblk_gfp);
1393 * Do async read-ahead on a buffer..
1395 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1397 struct buffer_head *bh = __getblk(bdev, block, size);
1399 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1403 EXPORT_SYMBOL(__breadahead);
1406 * __bread_gfp() - reads a specified block and returns the bh
1407 * @bdev: the block_device to read from
1408 * @block: number of block
1409 * @size: size (in bytes) to read
1410 * @gfp: page allocation flag
1412 * Reads a specified block, and returns buffer head that contains it.
1413 * The page cache can be allocated from non-movable area
1414 * not to prevent page migration if you set gfp to zero.
1415 * It returns NULL if the block was unreadable.
1417 struct buffer_head *
1418 __bread_gfp(struct block_device *bdev, sector_t block,
1419 unsigned size, gfp_t gfp)
1421 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1423 if (likely(bh) && !buffer_uptodate(bh))
1424 bh = __bread_slow(bh);
1427 EXPORT_SYMBOL(__bread_gfp);
1430 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1431 * This doesn't race because it runs in each cpu either in irq
1432 * or with preempt disabled.
1434 static void invalidate_bh_lru(void *arg)
1436 struct bh_lru *b = &get_cpu_var(bh_lrus);
1439 for (i = 0; i < BH_LRU_SIZE; i++) {
1443 put_cpu_var(bh_lrus);
1446 static bool has_bh_in_lru(int cpu, void *dummy)
1448 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1451 for (i = 0; i < BH_LRU_SIZE; i++) {
1459 void invalidate_bh_lrus(void)
1461 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1463 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1465 void set_bh_page(struct buffer_head *bh,
1466 struct page *page, unsigned long offset)
1469 BUG_ON(offset >= PAGE_SIZE);
1470 if (PageHighMem(page))
1472 * This catches illegal uses and preserves the offset:
1474 bh->b_data = (char *)(0 + offset);
1476 bh->b_data = page_address(page) + offset;
1478 EXPORT_SYMBOL(set_bh_page);
1481 * Called when truncating a buffer on a page completely.
1484 /* Bits that are cleared during an invalidate */
1485 #define BUFFER_FLAGS_DISCARD \
1486 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1487 1 << BH_Delay | 1 << BH_Unwritten)
1489 static void discard_buffer(struct buffer_head * bh)
1491 unsigned long b_state, b_state_old;
1494 clear_buffer_dirty(bh);
1496 b_state = bh->b_state;
1498 b_state_old = cmpxchg(&bh->b_state, b_state,
1499 (b_state & ~BUFFER_FLAGS_DISCARD));
1500 if (b_state_old == b_state)
1502 b_state = b_state_old;
1508 * block_invalidatepage - invalidate part or all of a buffer-backed page
1510 * @page: the page which is affected
1511 * @offset: start of the range to invalidate
1512 * @length: length of the range to invalidate
1514 * block_invalidatepage() is called when all or part of the page has become
1515 * invalidated by a truncate operation.
1517 * block_invalidatepage() does not have to release all buffers, but it must
1518 * ensure that no dirty buffer is left outside @offset and that no I/O
1519 * is underway against any of the blocks which are outside the truncation
1520 * point. Because the caller is about to free (and possibly reuse) those
1523 void block_invalidatepage(struct page *page, unsigned int offset,
1524 unsigned int length)
1526 struct buffer_head *head, *bh, *next;
1527 unsigned int curr_off = 0;
1528 unsigned int stop = length + offset;
1530 BUG_ON(!PageLocked(page));
1531 if (!page_has_buffers(page))
1535 * Check for overflow
1537 BUG_ON(stop > PAGE_SIZE || stop < length);
1539 head = page_buffers(page);
1542 unsigned int next_off = curr_off + bh->b_size;
1543 next = bh->b_this_page;
1546 * Are we still fully in range ?
1548 if (next_off > stop)
1552 * is this block fully invalidated?
1554 if (offset <= curr_off)
1556 curr_off = next_off;
1558 } while (bh != head);
1561 * We release buffers only if the entire page is being invalidated.
1562 * The get_block cached value has been unconditionally invalidated,
1563 * so real IO is not possible anymore.
1566 try_to_release_page(page, 0);
1570 EXPORT_SYMBOL(block_invalidatepage);
1574 * We attach and possibly dirty the buffers atomically wrt
1575 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1576 * is already excluded via the page lock.
1578 void create_empty_buffers(struct page *page,
1579 unsigned long blocksize, unsigned long b_state)
1581 struct buffer_head *bh, *head, *tail;
1583 head = alloc_page_buffers(page, blocksize, 1);
1586 bh->b_state |= b_state;
1588 bh = bh->b_this_page;
1590 tail->b_this_page = head;
1592 spin_lock(&page->mapping->private_lock);
1593 if (PageUptodate(page) || PageDirty(page)) {
1596 if (PageDirty(page))
1597 set_buffer_dirty(bh);
1598 if (PageUptodate(page))
1599 set_buffer_uptodate(bh);
1600 bh = bh->b_this_page;
1601 } while (bh != head);
1603 attach_page_buffers(page, head);
1604 spin_unlock(&page->mapping->private_lock);
1606 EXPORT_SYMBOL(create_empty_buffers);
1609 * We are taking a block for data and we don't want any output from any
1610 * buffer-cache aliases starting from return from that function and
1611 * until the moment when something will explicitly mark the buffer
1612 * dirty (hopefully that will not happen until we will free that block ;-)
1613 * We don't even need to mark it not-uptodate - nobody can expect
1614 * anything from a newly allocated buffer anyway. We used to used
1615 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1616 * don't want to mark the alias unmapped, for example - it would confuse
1617 * anyone who might pick it with bread() afterwards...
1619 * Also.. Note that bforget() doesn't lock the buffer. So there can
1620 * be writeout I/O going on against recently-freed buffers. We don't
1621 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1622 * only if we really need to. That happens here.
1624 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1626 struct buffer_head *old_bh;
1630 old_bh = __find_get_block_slow(bdev, block);
1632 clear_buffer_dirty(old_bh);
1633 wait_on_buffer(old_bh);
1634 clear_buffer_req(old_bh);
1638 EXPORT_SYMBOL(unmap_underlying_metadata);
1641 * clean_bdev_aliases: clean a range of buffers in block device
1642 * @bdev: Block device to clean buffers in
1643 * @block: Start of a range of blocks to clean
1644 * @len: Number of blocks to clean
1646 * We are taking a range of blocks for data and we don't want writeback of any
1647 * buffer-cache aliases starting from return from this function and until the
1648 * moment when something will explicitly mark the buffer dirty (hopefully that
1649 * will not happen until we will free that block ;-) We don't even need to mark
1650 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1651 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1652 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1653 * would confuse anyone who might pick it with bread() afterwards...
1655 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1656 * writeout I/O going on against recently-freed buffers. We don't wait on that
1657 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1658 * need to. That happens here.
1660 void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1662 struct inode *bd_inode = bdev->bd_inode;
1663 struct address_space *bd_mapping = bd_inode->i_mapping;
1664 struct pagevec pvec;
1665 pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1668 struct buffer_head *bh;
1669 struct buffer_head *head;
1671 end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1672 pagevec_init(&pvec, 0);
1673 while (index <= end && pagevec_lookup(&pvec, bd_mapping, index,
1674 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
1675 for (i = 0; i < pagevec_count(&pvec); i++) {
1676 struct page *page = pvec.pages[i];
1678 index = page->index;
1681 if (!page_has_buffers(page))
1684 * We use page lock instead of bd_mapping->private_lock
1685 * to pin buffers here since we can afford to sleep and
1686 * it scales better than a global spinlock lock.
1689 /* Recheck when the page is locked which pins bhs */
1690 if (!page_has_buffers(page))
1692 head = page_buffers(page);
1695 if (!buffer_mapped(bh))
1697 if (bh->b_blocknr >= block + len)
1699 clear_buffer_dirty(bh);
1701 clear_buffer_req(bh);
1703 bh = bh->b_this_page;
1704 } while (bh != head);
1708 pagevec_release(&pvec);
1713 EXPORT_SYMBOL(clean_bdev_aliases);
1716 * Size is a power-of-two in the range 512..PAGE_SIZE,
1717 * and the case we care about most is PAGE_SIZE.
1719 * So this *could* possibly be written with those
1720 * constraints in mind (relevant mostly if some
1721 * architecture has a slow bit-scan instruction)
1723 static inline int block_size_bits(unsigned int blocksize)
1725 return ilog2(blocksize);
1728 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1730 BUG_ON(!PageLocked(page));
1732 if (!page_has_buffers(page))
1733 create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1734 return page_buffers(page);
1738 * NOTE! All mapped/uptodate combinations are valid:
1740 * Mapped Uptodate Meaning
1742 * No No "unknown" - must do get_block()
1743 * No Yes "hole" - zero-filled
1744 * Yes No "allocated" - allocated on disk, not read in
1745 * Yes Yes "valid" - allocated and up-to-date in memory.
1747 * "Dirty" is valid only with the last case (mapped+uptodate).
1751 * While block_write_full_page is writing back the dirty buffers under
1752 * the page lock, whoever dirtied the buffers may decide to clean them
1753 * again at any time. We handle that by only looking at the buffer
1754 * state inside lock_buffer().
1756 * If block_write_full_page() is called for regular writeback
1757 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1758 * locked buffer. This only can happen if someone has written the buffer
1759 * directly, with submit_bh(). At the address_space level PageWriteback
1760 * prevents this contention from occurring.
1762 * If block_write_full_page() is called with wbc->sync_mode ==
1763 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1764 * causes the writes to be flagged as synchronous writes.
1766 int __block_write_full_page(struct inode *inode, struct page *page,
1767 get_block_t *get_block, struct writeback_control *wbc,
1768 bh_end_io_t *handler)
1772 sector_t last_block;
1773 struct buffer_head *bh, *head;
1774 unsigned int blocksize, bbits;
1775 int nr_underway = 0;
1776 int write_flags = wbc_to_write_flags(wbc);
1778 head = create_page_buffers(page, inode,
1779 (1 << BH_Dirty)|(1 << BH_Uptodate));
1782 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1783 * here, and the (potentially unmapped) buffers may become dirty at
1784 * any time. If a buffer becomes dirty here after we've inspected it
1785 * then we just miss that fact, and the page stays dirty.
1787 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1788 * handle that here by just cleaning them.
1792 blocksize = bh->b_size;
1793 bbits = block_size_bits(blocksize);
1795 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1796 last_block = (i_size_read(inode) - 1) >> bbits;
1799 * Get all the dirty buffers mapped to disk addresses and
1800 * handle any aliases from the underlying blockdev's mapping.
1803 if (block > last_block) {
1805 * mapped buffers outside i_size will occur, because
1806 * this page can be outside i_size when there is a
1807 * truncate in progress.
1810 * The buffer was zeroed by block_write_full_page()
1812 clear_buffer_dirty(bh);
1813 set_buffer_uptodate(bh);
1814 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1816 WARN_ON(bh->b_size != blocksize);
1817 err = get_block(inode, block, bh, 1);
1820 clear_buffer_delay(bh);
1821 if (buffer_new(bh)) {
1822 /* blockdev mappings never come here */
1823 clear_buffer_new(bh);
1824 unmap_underlying_metadata(bh->b_bdev,
1828 bh = bh->b_this_page;
1830 } while (bh != head);
1833 if (!buffer_mapped(bh))
1836 * If it's a fully non-blocking write attempt and we cannot
1837 * lock the buffer then redirty the page. Note that this can
1838 * potentially cause a busy-wait loop from writeback threads
1839 * and kswapd activity, but those code paths have their own
1840 * higher-level throttling.
1842 if (wbc->sync_mode != WB_SYNC_NONE) {
1844 } else if (!trylock_buffer(bh)) {
1845 redirty_page_for_writepage(wbc, page);
1848 if (test_clear_buffer_dirty(bh)) {
1849 mark_buffer_async_write_endio(bh, handler);
1853 } while ((bh = bh->b_this_page) != head);
1856 * The page and its buffers are protected by PageWriteback(), so we can
1857 * drop the bh refcounts early.
1859 BUG_ON(PageWriteback(page));
1860 set_page_writeback(page);
1863 struct buffer_head *next = bh->b_this_page;
1864 if (buffer_async_write(bh)) {
1865 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh, 0, wbc);
1869 } while (bh != head);
1874 if (nr_underway == 0) {
1876 * The page was marked dirty, but the buffers were
1877 * clean. Someone wrote them back by hand with
1878 * ll_rw_block/submit_bh. A rare case.
1880 end_page_writeback(page);
1883 * The page and buffer_heads can be released at any time from
1891 * ENOSPC, or some other error. We may already have added some
1892 * blocks to the file, so we need to write these out to avoid
1893 * exposing stale data.
1894 * The page is currently locked and not marked for writeback
1897 /* Recovery: lock and submit the mapped buffers */
1899 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1900 !buffer_delay(bh)) {
1902 mark_buffer_async_write_endio(bh, handler);
1905 * The buffer may have been set dirty during
1906 * attachment to a dirty page.
1908 clear_buffer_dirty(bh);
1910 } while ((bh = bh->b_this_page) != head);
1912 BUG_ON(PageWriteback(page));
1913 mapping_set_error(page->mapping, err);
1914 set_page_writeback(page);
1916 struct buffer_head *next = bh->b_this_page;
1917 if (buffer_async_write(bh)) {
1918 clear_buffer_dirty(bh);
1919 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh, 0, wbc);
1923 } while (bh != head);
1927 EXPORT_SYMBOL(__block_write_full_page);
1930 * If a page has any new buffers, zero them out here, and mark them uptodate
1931 * and dirty so they'll be written out (in order to prevent uninitialised
1932 * block data from leaking). And clear the new bit.
1934 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1936 unsigned int block_start, block_end;
1937 struct buffer_head *head, *bh;
1939 BUG_ON(!PageLocked(page));
1940 if (!page_has_buffers(page))
1943 bh = head = page_buffers(page);
1946 block_end = block_start + bh->b_size;
1948 if (buffer_new(bh)) {
1949 if (block_end > from && block_start < to) {
1950 if (!PageUptodate(page)) {
1951 unsigned start, size;
1953 start = max(from, block_start);
1954 size = min(to, block_end) - start;
1956 zero_user(page, start, size);
1957 set_buffer_uptodate(bh);
1960 clear_buffer_new(bh);
1961 mark_buffer_dirty(bh);
1965 block_start = block_end;
1966 bh = bh->b_this_page;
1967 } while (bh != head);
1969 EXPORT_SYMBOL(page_zero_new_buffers);
1972 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1973 struct iomap *iomap)
1975 loff_t offset = block << inode->i_blkbits;
1977 bh->b_bdev = iomap->bdev;
1980 * Block points to offset in file we need to map, iomap contains
1981 * the offset at which the map starts. If the map ends before the
1982 * current block, then do not map the buffer and let the caller
1985 BUG_ON(offset >= iomap->offset + iomap->length);
1987 switch (iomap->type) {
1990 * If the buffer is not up to date or beyond the current EOF,
1991 * we need to mark it as new to ensure sub-block zeroing is
1992 * executed if necessary.
1994 if (!buffer_uptodate(bh) ||
1995 (offset >= i_size_read(inode)))
1998 case IOMAP_DELALLOC:
1999 if (!buffer_uptodate(bh) ||
2000 (offset >= i_size_read(inode)))
2002 set_buffer_uptodate(bh);
2003 set_buffer_mapped(bh);
2004 set_buffer_delay(bh);
2006 case IOMAP_UNWRITTEN:
2008 * For unwritten regions, we always need to ensure that
2009 * sub-block writes cause the regions in the block we are not
2010 * writing to are zeroed. Set the buffer as new to ensure this.
2013 set_buffer_unwritten(bh);
2016 if (offset >= i_size_read(inode))
2018 bh->b_blocknr = (iomap->blkno >> (inode->i_blkbits - 9)) +
2019 ((offset - iomap->offset) >> inode->i_blkbits);
2020 set_buffer_mapped(bh);
2025 int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
2026 get_block_t *get_block, struct iomap *iomap)
2028 unsigned from = pos & (PAGE_SIZE - 1);
2029 unsigned to = from + len;
2030 struct inode *inode = page->mapping->host;
2031 unsigned block_start, block_end;
2034 unsigned blocksize, bbits;
2035 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
2037 BUG_ON(!PageLocked(page));
2038 BUG_ON(from > PAGE_SIZE);
2039 BUG_ON(to > PAGE_SIZE);
2042 head = create_page_buffers(page, inode, 0);
2043 blocksize = head->b_size;
2044 bbits = block_size_bits(blocksize);
2046 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
2048 for(bh = head, block_start = 0; bh != head || !block_start;
2049 block++, block_start=block_end, bh = bh->b_this_page) {
2050 block_end = block_start + blocksize;
2051 if (block_end <= from || block_start >= to) {
2052 if (PageUptodate(page)) {
2053 if (!buffer_uptodate(bh))
2054 set_buffer_uptodate(bh);
2059 clear_buffer_new(bh);
2060 if (!buffer_mapped(bh)) {
2061 WARN_ON(bh->b_size != blocksize);
2063 err = get_block(inode, block, bh, 1);
2067 iomap_to_bh(inode, block, bh, iomap);
2070 if (buffer_new(bh)) {
2071 unmap_underlying_metadata(bh->b_bdev,
2073 if (PageUptodate(page)) {
2074 clear_buffer_new(bh);
2075 set_buffer_uptodate(bh);
2076 mark_buffer_dirty(bh);
2079 if (block_end > to || block_start < from)
2080 zero_user_segments(page,
2086 if (PageUptodate(page)) {
2087 if (!buffer_uptodate(bh))
2088 set_buffer_uptodate(bh);
2091 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2092 !buffer_unwritten(bh) &&
2093 (block_start < from || block_end > to)) {
2094 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2099 * If we issued read requests - let them complete.
2101 while(wait_bh > wait) {
2102 wait_on_buffer(*--wait_bh);
2103 if (!buffer_uptodate(*wait_bh))
2107 page_zero_new_buffers(page, from, to);
2111 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2112 get_block_t *get_block)
2114 return __block_write_begin_int(page, pos, len, get_block, NULL);
2116 EXPORT_SYMBOL(__block_write_begin);
2118 static int __block_commit_write(struct inode *inode, struct page *page,
2119 unsigned from, unsigned to)
2121 unsigned block_start, block_end;
2124 struct buffer_head *bh, *head;
2126 bh = head = page_buffers(page);
2127 blocksize = bh->b_size;
2131 block_end = block_start + blocksize;
2132 if (block_end <= from || block_start >= to) {
2133 if (!buffer_uptodate(bh))
2136 set_buffer_uptodate(bh);
2137 mark_buffer_dirty(bh);
2139 clear_buffer_new(bh);
2141 block_start = block_end;
2142 bh = bh->b_this_page;
2143 } while (bh != head);
2146 * If this is a partial write which happened to make all buffers
2147 * uptodate then we can optimize away a bogus readpage() for
2148 * the next read(). Here we 'discover' whether the page went
2149 * uptodate as a result of this (potentially partial) write.
2152 SetPageUptodate(page);
2157 * block_write_begin takes care of the basic task of block allocation and
2158 * bringing partial write blocks uptodate first.
2160 * The filesystem needs to handle block truncation upon failure.
2162 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2163 unsigned flags, struct page **pagep, get_block_t *get_block)
2165 pgoff_t index = pos >> PAGE_SHIFT;
2169 page = grab_cache_page_write_begin(mapping, index, flags);
2173 status = __block_write_begin(page, pos, len, get_block);
2174 if (unlikely(status)) {
2183 EXPORT_SYMBOL(block_write_begin);
2185 int block_write_end(struct file *file, struct address_space *mapping,
2186 loff_t pos, unsigned len, unsigned copied,
2187 struct page *page, void *fsdata)
2189 struct inode *inode = mapping->host;
2192 start = pos & (PAGE_SIZE - 1);
2194 if (unlikely(copied < len)) {
2196 * The buffers that were written will now be uptodate, so we
2197 * don't have to worry about a readpage reading them and
2198 * overwriting a partial write. However if we have encountered
2199 * a short write and only partially written into a buffer, it
2200 * will not be marked uptodate, so a readpage might come in and
2201 * destroy our partial write.
2203 * Do the simplest thing, and just treat any short write to a
2204 * non uptodate page as a zero-length write, and force the
2205 * caller to redo the whole thing.
2207 if (!PageUptodate(page))
2210 page_zero_new_buffers(page, start+copied, start+len);
2212 flush_dcache_page(page);
2214 /* This could be a short (even 0-length) commit */
2215 __block_commit_write(inode, page, start, start+copied);
2219 EXPORT_SYMBOL(block_write_end);
2221 int generic_write_end(struct file *file, struct address_space *mapping,
2222 loff_t pos, unsigned len, unsigned copied,
2223 struct page *page, void *fsdata)
2225 struct inode *inode = mapping->host;
2226 loff_t old_size = inode->i_size;
2227 int i_size_changed = 0;
2229 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2232 * No need to use i_size_read() here, the i_size
2233 * cannot change under us because we hold i_mutex.
2235 * But it's important to update i_size while still holding page lock:
2236 * page writeout could otherwise come in and zero beyond i_size.
2238 if (pos+copied > inode->i_size) {
2239 i_size_write(inode, pos+copied);
2247 pagecache_isize_extended(inode, old_size, pos);
2249 * Don't mark the inode dirty under page lock. First, it unnecessarily
2250 * makes the holding time of page lock longer. Second, it forces lock
2251 * ordering of page lock and transaction start for journaling
2255 mark_inode_dirty(inode);
2259 EXPORT_SYMBOL(generic_write_end);
2262 * block_is_partially_uptodate checks whether buffers within a page are
2265 * Returns true if all buffers which correspond to a file portion
2266 * we want to read are uptodate.
2268 int block_is_partially_uptodate(struct page *page, unsigned long from,
2269 unsigned long count)
2271 unsigned block_start, block_end, blocksize;
2273 struct buffer_head *bh, *head;
2276 if (!page_has_buffers(page))
2279 head = page_buffers(page);
2280 blocksize = head->b_size;
2281 to = min_t(unsigned, PAGE_SIZE - from, count);
2283 if (from < blocksize && to > PAGE_SIZE - blocksize)
2289 block_end = block_start + blocksize;
2290 if (block_end > from && block_start < to) {
2291 if (!buffer_uptodate(bh)) {
2295 if (block_end >= to)
2298 block_start = block_end;
2299 bh = bh->b_this_page;
2300 } while (bh != head);
2304 EXPORT_SYMBOL(block_is_partially_uptodate);
2307 * Generic "read page" function for block devices that have the normal
2308 * get_block functionality. This is most of the block device filesystems.
2309 * Reads the page asynchronously --- the unlock_buffer() and
2310 * set/clear_buffer_uptodate() functions propagate buffer state into the
2311 * page struct once IO has completed.
2313 int block_read_full_page(struct page *page, get_block_t *get_block)
2315 struct inode *inode = page->mapping->host;
2316 sector_t iblock, lblock;
2317 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2318 unsigned int blocksize, bbits;
2320 int fully_mapped = 1;
2322 head = create_page_buffers(page, inode, 0);
2323 blocksize = head->b_size;
2324 bbits = block_size_bits(blocksize);
2326 iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2327 lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2333 if (buffer_uptodate(bh))
2336 if (!buffer_mapped(bh)) {
2340 if (iblock < lblock) {
2341 WARN_ON(bh->b_size != blocksize);
2342 err = get_block(inode, iblock, bh, 0);
2346 if (!buffer_mapped(bh)) {
2347 zero_user(page, i * blocksize, blocksize);
2349 set_buffer_uptodate(bh);
2353 * get_block() might have updated the buffer
2356 if (buffer_uptodate(bh))
2360 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2363 SetPageMappedToDisk(page);
2367 * All buffers are uptodate - we can set the page uptodate
2368 * as well. But not if get_block() returned an error.
2370 if (!PageError(page))
2371 SetPageUptodate(page);
2376 /* Stage two: lock the buffers */
2377 for (i = 0; i < nr; i++) {
2380 mark_buffer_async_read(bh);
2384 * Stage 3: start the IO. Check for uptodateness
2385 * inside the buffer lock in case another process reading
2386 * the underlying blockdev brought it uptodate (the sct fix).
2388 for (i = 0; i < nr; i++) {
2390 if (buffer_uptodate(bh))
2391 end_buffer_async_read(bh, 1);
2393 submit_bh(REQ_OP_READ, 0, bh);
2397 EXPORT_SYMBOL(block_read_full_page);
2399 /* utility function for filesystems that need to do work on expanding
2400 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2401 * deal with the hole.
2403 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2405 struct address_space *mapping = inode->i_mapping;
2410 err = inode_newsize_ok(inode, size);
2414 err = pagecache_write_begin(NULL, mapping, size, 0,
2415 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2420 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2426 EXPORT_SYMBOL(generic_cont_expand_simple);
2428 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2429 loff_t pos, loff_t *bytes)
2431 struct inode *inode = mapping->host;
2432 unsigned blocksize = 1 << inode->i_blkbits;
2435 pgoff_t index, curidx;
2437 unsigned zerofrom, offset, len;
2440 index = pos >> PAGE_SHIFT;
2441 offset = pos & ~PAGE_MASK;
2443 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2444 zerofrom = curpos & ~PAGE_MASK;
2445 if (zerofrom & (blocksize-1)) {
2446 *bytes |= (blocksize-1);
2449 len = PAGE_SIZE - zerofrom;
2451 err = pagecache_write_begin(file, mapping, curpos, len,
2452 AOP_FLAG_UNINTERRUPTIBLE,
2456 zero_user(page, zerofrom, len);
2457 err = pagecache_write_end(file, mapping, curpos, len, len,
2464 balance_dirty_pages_ratelimited(mapping);
2466 if (unlikely(fatal_signal_pending(current))) {
2472 /* page covers the boundary, find the boundary offset */
2473 if (index == curidx) {
2474 zerofrom = curpos & ~PAGE_MASK;
2475 /* if we will expand the thing last block will be filled */
2476 if (offset <= zerofrom) {
2479 if (zerofrom & (blocksize-1)) {
2480 *bytes |= (blocksize-1);
2483 len = offset - zerofrom;
2485 err = pagecache_write_begin(file, mapping, curpos, len,
2486 AOP_FLAG_UNINTERRUPTIBLE,
2490 zero_user(page, zerofrom, len);
2491 err = pagecache_write_end(file, mapping, curpos, len, len,
2503 * For moronic filesystems that do not allow holes in file.
2504 * We may have to extend the file.
2506 int cont_write_begin(struct file *file, struct address_space *mapping,
2507 loff_t pos, unsigned len, unsigned flags,
2508 struct page **pagep, void **fsdata,
2509 get_block_t *get_block, loff_t *bytes)
2511 struct inode *inode = mapping->host;
2512 unsigned blocksize = 1 << inode->i_blkbits;
2516 err = cont_expand_zero(file, mapping, pos, bytes);
2520 zerofrom = *bytes & ~PAGE_MASK;
2521 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2522 *bytes |= (blocksize-1);
2526 return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2528 EXPORT_SYMBOL(cont_write_begin);
2530 int block_commit_write(struct page *page, unsigned from, unsigned to)
2532 struct inode *inode = page->mapping->host;
2533 __block_commit_write(inode,page,from,to);
2536 EXPORT_SYMBOL(block_commit_write);
2539 * block_page_mkwrite() is not allowed to change the file size as it gets
2540 * called from a page fault handler when a page is first dirtied. Hence we must
2541 * be careful to check for EOF conditions here. We set the page up correctly
2542 * for a written page which means we get ENOSPC checking when writing into
2543 * holes and correct delalloc and unwritten extent mapping on filesystems that
2544 * support these features.
2546 * We are not allowed to take the i_mutex here so we have to play games to
2547 * protect against truncate races as the page could now be beyond EOF. Because
2548 * truncate writes the inode size before removing pages, once we have the
2549 * page lock we can determine safely if the page is beyond EOF. If it is not
2550 * beyond EOF, then the page is guaranteed safe against truncation until we
2553 * Direct callers of this function should protect against filesystem freezing
2554 * using sb_start_pagefault() - sb_end_pagefault() functions.
2556 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2557 get_block_t get_block)
2559 struct page *page = vmf->page;
2560 struct inode *inode = file_inode(vma->vm_file);
2566 size = i_size_read(inode);
2567 if ((page->mapping != inode->i_mapping) ||
2568 (page_offset(page) > size)) {
2569 /* We overload EFAULT to mean page got truncated */
2574 /* page is wholly or partially inside EOF */
2575 if (((page->index + 1) << PAGE_SHIFT) > size)
2576 end = size & ~PAGE_MASK;
2580 ret = __block_write_begin(page, 0, end, get_block);
2582 ret = block_commit_write(page, 0, end);
2584 if (unlikely(ret < 0))
2586 set_page_dirty(page);
2587 wait_for_stable_page(page);
2593 EXPORT_SYMBOL(block_page_mkwrite);
2596 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2597 * immediately, while under the page lock. So it needs a special end_io
2598 * handler which does not touch the bh after unlocking it.
2600 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2602 __end_buffer_read_notouch(bh, uptodate);
2606 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2607 * the page (converting it to circular linked list and taking care of page
2610 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2612 struct buffer_head *bh;
2614 BUG_ON(!PageLocked(page));
2616 spin_lock(&page->mapping->private_lock);
2619 if (PageDirty(page))
2620 set_buffer_dirty(bh);
2621 if (!bh->b_this_page)
2622 bh->b_this_page = head;
2623 bh = bh->b_this_page;
2624 } while (bh != head);
2625 attach_page_buffers(page, head);
2626 spin_unlock(&page->mapping->private_lock);
2630 * On entry, the page is fully not uptodate.
2631 * On exit the page is fully uptodate in the areas outside (from,to)
2632 * The filesystem needs to handle block truncation upon failure.
2634 int nobh_write_begin(struct address_space *mapping,
2635 loff_t pos, unsigned len, unsigned flags,
2636 struct page **pagep, void **fsdata,
2637 get_block_t *get_block)
2639 struct inode *inode = mapping->host;
2640 const unsigned blkbits = inode->i_blkbits;
2641 const unsigned blocksize = 1 << blkbits;
2642 struct buffer_head *head, *bh;
2646 unsigned block_in_page;
2647 unsigned block_start, block_end;
2648 sector_t block_in_file;
2651 int is_mapped_to_disk = 1;
2653 index = pos >> PAGE_SHIFT;
2654 from = pos & (PAGE_SIZE - 1);
2657 page = grab_cache_page_write_begin(mapping, index, flags);
2663 if (page_has_buffers(page)) {
2664 ret = __block_write_begin(page, pos, len, get_block);
2670 if (PageMappedToDisk(page))
2674 * Allocate buffers so that we can keep track of state, and potentially
2675 * attach them to the page if an error occurs. In the common case of
2676 * no error, they will just be freed again without ever being attached
2677 * to the page (which is all OK, because we're under the page lock).
2679 * Be careful: the buffer linked list is a NULL terminated one, rather
2680 * than the circular one we're used to.
2682 head = alloc_page_buffers(page, blocksize, 0);
2688 block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2691 * We loop across all blocks in the page, whether or not they are
2692 * part of the affected region. This is so we can discover if the
2693 * page is fully mapped-to-disk.
2695 for (block_start = 0, block_in_page = 0, bh = head;
2696 block_start < PAGE_SIZE;
2697 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2700 block_end = block_start + blocksize;
2703 if (block_start >= to)
2705 ret = get_block(inode, block_in_file + block_in_page,
2709 if (!buffer_mapped(bh))
2710 is_mapped_to_disk = 0;
2712 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2713 if (PageUptodate(page)) {
2714 set_buffer_uptodate(bh);
2717 if (buffer_new(bh) || !buffer_mapped(bh)) {
2718 zero_user_segments(page, block_start, from,
2722 if (buffer_uptodate(bh))
2723 continue; /* reiserfs does this */
2724 if (block_start < from || block_end > to) {
2726 bh->b_end_io = end_buffer_read_nobh;
2727 submit_bh(REQ_OP_READ, 0, bh);
2734 * The page is locked, so these buffers are protected from
2735 * any VM or truncate activity. Hence we don't need to care
2736 * for the buffer_head refcounts.
2738 for (bh = head; bh; bh = bh->b_this_page) {
2740 if (!buffer_uptodate(bh))
2747 if (is_mapped_to_disk)
2748 SetPageMappedToDisk(page);
2750 *fsdata = head; /* to be released by nobh_write_end */
2757 * Error recovery is a bit difficult. We need to zero out blocks that
2758 * were newly allocated, and dirty them to ensure they get written out.
2759 * Buffers need to be attached to the page at this point, otherwise
2760 * the handling of potential IO errors during writeout would be hard
2761 * (could try doing synchronous writeout, but what if that fails too?)
2763 attach_nobh_buffers(page, head);
2764 page_zero_new_buffers(page, from, to);
2773 EXPORT_SYMBOL(nobh_write_begin);
2775 int nobh_write_end(struct file *file, struct address_space *mapping,
2776 loff_t pos, unsigned len, unsigned copied,
2777 struct page *page, void *fsdata)
2779 struct inode *inode = page->mapping->host;
2780 struct buffer_head *head = fsdata;
2781 struct buffer_head *bh;
2782 BUG_ON(fsdata != NULL && page_has_buffers(page));
2784 if (unlikely(copied < len) && head)
2785 attach_nobh_buffers(page, head);
2786 if (page_has_buffers(page))
2787 return generic_write_end(file, mapping, pos, len,
2788 copied, page, fsdata);
2790 SetPageUptodate(page);
2791 set_page_dirty(page);
2792 if (pos+copied > inode->i_size) {
2793 i_size_write(inode, pos+copied);
2794 mark_inode_dirty(inode);
2802 head = head->b_this_page;
2803 free_buffer_head(bh);
2808 EXPORT_SYMBOL(nobh_write_end);
2811 * nobh_writepage() - based on block_full_write_page() except
2812 * that it tries to operate without attaching bufferheads to
2815 int nobh_writepage(struct page *page, get_block_t *get_block,
2816 struct writeback_control *wbc)
2818 struct inode * const inode = page->mapping->host;
2819 loff_t i_size = i_size_read(inode);
2820 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2824 /* Is the page fully inside i_size? */
2825 if (page->index < end_index)
2828 /* Is the page fully outside i_size? (truncate in progress) */
2829 offset = i_size & (PAGE_SIZE-1);
2830 if (page->index >= end_index+1 || !offset) {
2832 * The page may have dirty, unmapped buffers. For example,
2833 * they may have been added in ext3_writepage(). Make them
2834 * freeable here, so the page does not leak.
2837 /* Not really sure about this - do we need this ? */
2838 if (page->mapping->a_ops->invalidatepage)
2839 page->mapping->a_ops->invalidatepage(page, offset);
2842 return 0; /* don't care */
2846 * The page straddles i_size. It must be zeroed out on each and every
2847 * writepage invocation because it may be mmapped. "A file is mapped
2848 * in multiples of the page size. For a file that is not a multiple of
2849 * the page size, the remaining memory is zeroed when mapped, and
2850 * writes to that region are not written out to the file."
2852 zero_user_segment(page, offset, PAGE_SIZE);
2854 ret = mpage_writepage(page, get_block, wbc);
2856 ret = __block_write_full_page(inode, page, get_block, wbc,
2857 end_buffer_async_write);
2860 EXPORT_SYMBOL(nobh_writepage);
2862 int nobh_truncate_page(struct address_space *mapping,
2863 loff_t from, get_block_t *get_block)
2865 pgoff_t index = from >> PAGE_SHIFT;
2866 unsigned offset = from & (PAGE_SIZE-1);
2869 unsigned length, pos;
2870 struct inode *inode = mapping->host;
2872 struct buffer_head map_bh;
2875 blocksize = 1 << inode->i_blkbits;
2876 length = offset & (blocksize - 1);
2878 /* Block boundary? Nothing to do */
2882 length = blocksize - length;
2883 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2885 page = grab_cache_page(mapping, index);
2890 if (page_has_buffers(page)) {
2894 return block_truncate_page(mapping, from, get_block);
2897 /* Find the buffer that contains "offset" */
2899 while (offset >= pos) {
2904 map_bh.b_size = blocksize;
2906 err = get_block(inode, iblock, &map_bh, 0);
2909 /* unmapped? It's a hole - nothing to do */
2910 if (!buffer_mapped(&map_bh))
2913 /* Ok, it's mapped. Make sure it's up-to-date */
2914 if (!PageUptodate(page)) {
2915 err = mapping->a_ops->readpage(NULL, page);
2921 if (!PageUptodate(page)) {
2925 if (page_has_buffers(page))
2928 zero_user(page, offset, length);
2929 set_page_dirty(page);
2938 EXPORT_SYMBOL(nobh_truncate_page);
2940 int block_truncate_page(struct address_space *mapping,
2941 loff_t from, get_block_t *get_block)
2943 pgoff_t index = from >> PAGE_SHIFT;
2944 unsigned offset = from & (PAGE_SIZE-1);
2947 unsigned length, pos;
2948 struct inode *inode = mapping->host;
2950 struct buffer_head *bh;
2953 blocksize = 1 << inode->i_blkbits;
2954 length = offset & (blocksize - 1);
2956 /* Block boundary? Nothing to do */
2960 length = blocksize - length;
2961 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2963 page = grab_cache_page(mapping, index);
2968 if (!page_has_buffers(page))
2969 create_empty_buffers(page, blocksize, 0);
2971 /* Find the buffer that contains "offset" */
2972 bh = page_buffers(page);
2974 while (offset >= pos) {
2975 bh = bh->b_this_page;
2981 if (!buffer_mapped(bh)) {
2982 WARN_ON(bh->b_size != blocksize);
2983 err = get_block(inode, iblock, bh, 0);
2986 /* unmapped? It's a hole - nothing to do */
2987 if (!buffer_mapped(bh))
2991 /* Ok, it's mapped. Make sure it's up-to-date */
2992 if (PageUptodate(page))
2993 set_buffer_uptodate(bh);
2995 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2997 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2999 /* Uhhuh. Read error. Complain and punt. */
3000 if (!buffer_uptodate(bh))
3004 zero_user(page, offset, length);
3005 mark_buffer_dirty(bh);
3014 EXPORT_SYMBOL(block_truncate_page);
3017 * The generic ->writepage function for buffer-backed address_spaces
3019 int block_write_full_page(struct page *page, get_block_t *get_block,
3020 struct writeback_control *wbc)
3022 struct inode * const inode = page->mapping->host;
3023 loff_t i_size = i_size_read(inode);
3024 const pgoff_t end_index = i_size >> PAGE_SHIFT;
3027 /* Is the page fully inside i_size? */
3028 if (page->index < end_index)
3029 return __block_write_full_page(inode, page, get_block, wbc,
3030 end_buffer_async_write);
3032 /* Is the page fully outside i_size? (truncate in progress) */
3033 offset = i_size & (PAGE_SIZE-1);
3034 if (page->index >= end_index+1 || !offset) {
3036 * The page may have dirty, unmapped buffers. For example,
3037 * they may have been added in ext3_writepage(). Make them
3038 * freeable here, so the page does not leak.
3040 do_invalidatepage(page, 0, PAGE_SIZE);
3042 return 0; /* don't care */
3046 * The page straddles i_size. It must be zeroed out on each and every
3047 * writepage invocation because it may be mmapped. "A file is mapped
3048 * in multiples of the page size. For a file that is not a multiple of
3049 * the page size, the remaining memory is zeroed when mapped, and
3050 * writes to that region are not written out to the file."
3052 zero_user_segment(page, offset, PAGE_SIZE);
3053 return __block_write_full_page(inode, page, get_block, wbc,
3054 end_buffer_async_write);
3056 EXPORT_SYMBOL(block_write_full_page);
3058 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
3059 get_block_t *get_block)
3061 struct buffer_head tmp;
3062 struct inode *inode = mapping->host;
3065 tmp.b_size = 1 << inode->i_blkbits;
3066 get_block(inode, block, &tmp, 0);
3067 return tmp.b_blocknr;
3069 EXPORT_SYMBOL(generic_block_bmap);
3071 static void end_bio_bh_io_sync(struct bio *bio)
3073 struct buffer_head *bh = bio->bi_private;
3075 if (unlikely(bio_flagged(bio, BIO_QUIET)))
3076 set_bit(BH_Quiet, &bh->b_state);
3078 bh->b_end_io(bh, !bio->bi_error);
3083 * This allows us to do IO even on the odd last sectors
3084 * of a device, even if the block size is some multiple
3085 * of the physical sector size.
3087 * We'll just truncate the bio to the size of the device,
3088 * and clear the end of the buffer head manually.
3090 * Truly out-of-range accesses will turn into actual IO
3091 * errors, this only handles the "we need to be able to
3092 * do IO at the final sector" case.
3094 void guard_bio_eod(int op, struct bio *bio)
3097 struct bio_vec *bvec = &bio->bi_io_vec[bio->bi_vcnt - 1];
3098 unsigned truncated_bytes;
3100 maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
3105 * If the *whole* IO is past the end of the device,
3106 * let it through, and the IO layer will turn it into
3109 if (unlikely(bio->bi_iter.bi_sector >= maxsector))
3112 maxsector -= bio->bi_iter.bi_sector;
3113 if (likely((bio->bi_iter.bi_size >> 9) <= maxsector))
3116 /* Uhhuh. We've got a bio that straddles the device size! */
3117 truncated_bytes = bio->bi_iter.bi_size - (maxsector << 9);
3119 /* Truncate the bio.. */
3120 bio->bi_iter.bi_size -= truncated_bytes;
3121 bvec->bv_len -= truncated_bytes;
3123 /* ..and clear the end of the buffer for reads */
3124 if (op == REQ_OP_READ) {
3125 zero_user(bvec->bv_page, bvec->bv_offset + bvec->bv_len,
3130 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3131 unsigned long bio_flags, struct writeback_control *wbc)
3135 BUG_ON(!buffer_locked(bh));
3136 BUG_ON(!buffer_mapped(bh));
3137 BUG_ON(!bh->b_end_io);
3138 BUG_ON(buffer_delay(bh));
3139 BUG_ON(buffer_unwritten(bh));
3142 * Only clear out a write error when rewriting
3144 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3145 clear_buffer_write_io_error(bh);
3148 * from here on down, it's all bio -- do the initial mapping,
3149 * submit_bio -> generic_make_request may further map this bio around
3151 bio = bio_alloc(GFP_NOIO, 1);
3154 wbc_init_bio(wbc, bio);
3155 wbc_account_io(wbc, bh->b_page, bh->b_size);
3158 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3159 bio->bi_bdev = bh->b_bdev;
3161 bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3162 BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3164 bio->bi_end_io = end_bio_bh_io_sync;
3165 bio->bi_private = bh;
3166 bio->bi_flags |= bio_flags;
3168 /* Take care of bh's that straddle the end of the device */
3169 guard_bio_eod(op, bio);
3171 if (buffer_meta(bh))
3172 op_flags |= REQ_META;
3173 if (buffer_prio(bh))
3174 op_flags |= REQ_PRIO;
3175 bio_set_op_attrs(bio, op, op_flags);
3181 int _submit_bh(int op, int op_flags, struct buffer_head *bh,
3182 unsigned long bio_flags)
3184 return submit_bh_wbc(op, op_flags, bh, bio_flags, NULL);
3186 EXPORT_SYMBOL_GPL(_submit_bh);
3188 int submit_bh(int op, int op_flags, struct buffer_head *bh)
3190 return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3192 EXPORT_SYMBOL(submit_bh);
3195 * ll_rw_block: low-level access to block devices (DEPRECATED)
3196 * @op: whether to %READ or %WRITE
3197 * @op_flags: req_flag_bits
3198 * @nr: number of &struct buffer_heads in the array
3199 * @bhs: array of pointers to &struct buffer_head
3201 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3202 * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3203 * @op_flags contains flags modifying the detailed I/O behavior, most notably
3206 * This function drops any buffer that it cannot get a lock on (with the
3207 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3208 * request, and any buffer that appears to be up-to-date when doing read
3209 * request. Further it marks as clean buffers that are processed for
3210 * writing (the buffer cache won't assume that they are actually clean
3211 * until the buffer gets unlocked).
3213 * ll_rw_block sets b_end_io to simple completion handler that marks
3214 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3217 * All of the buffers must be for the same device, and must also be a
3218 * multiple of the current approved size for the device.
3220 void ll_rw_block(int op, int op_flags, int nr, struct buffer_head *bhs[])
3224 for (i = 0; i < nr; i++) {
3225 struct buffer_head *bh = bhs[i];
3227 if (!trylock_buffer(bh))
3230 if (test_clear_buffer_dirty(bh)) {
3231 bh->b_end_io = end_buffer_write_sync;
3233 submit_bh(op, op_flags, bh);
3237 if (!buffer_uptodate(bh)) {
3238 bh->b_end_io = end_buffer_read_sync;
3240 submit_bh(op, op_flags, bh);
3247 EXPORT_SYMBOL(ll_rw_block);
3249 void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3252 if (!test_clear_buffer_dirty(bh)) {
3256 bh->b_end_io = end_buffer_write_sync;
3258 submit_bh(REQ_OP_WRITE, op_flags, bh);
3260 EXPORT_SYMBOL(write_dirty_buffer);
3263 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3264 * and then start new I/O and then wait upon it. The caller must have a ref on
3267 int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3271 WARN_ON(atomic_read(&bh->b_count) < 1);
3273 if (test_clear_buffer_dirty(bh)) {
3275 bh->b_end_io = end_buffer_write_sync;
3276 ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3278 if (!ret && !buffer_uptodate(bh))
3285 EXPORT_SYMBOL(__sync_dirty_buffer);
3287 int sync_dirty_buffer(struct buffer_head *bh)
3289 return __sync_dirty_buffer(bh, REQ_SYNC);
3291 EXPORT_SYMBOL(sync_dirty_buffer);
3294 * try_to_free_buffers() checks if all the buffers on this particular page
3295 * are unused, and releases them if so.
3297 * Exclusion against try_to_free_buffers may be obtained by either
3298 * locking the page or by holding its mapping's private_lock.
3300 * If the page is dirty but all the buffers are clean then we need to
3301 * be sure to mark the page clean as well. This is because the page
3302 * may be against a block device, and a later reattachment of buffers
3303 * to a dirty page will set *all* buffers dirty. Which would corrupt
3304 * filesystem data on the same device.
3306 * The same applies to regular filesystem pages: if all the buffers are
3307 * clean then we set the page clean and proceed. To do that, we require
3308 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3311 * try_to_free_buffers() is non-blocking.
3313 static inline int buffer_busy(struct buffer_head *bh)
3315 return atomic_read(&bh->b_count) |
3316 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3320 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3322 struct buffer_head *head = page_buffers(page);
3323 struct buffer_head *bh;
3327 if (buffer_write_io_error(bh) && page->mapping)
3328 mapping_set_error(page->mapping, -EIO);
3329 if (buffer_busy(bh))
3331 bh = bh->b_this_page;
3332 } while (bh != head);
3335 struct buffer_head *next = bh->b_this_page;
3337 if (bh->b_assoc_map)
3338 __remove_assoc_queue(bh);
3340 } while (bh != head);
3341 *buffers_to_free = head;
3342 __clear_page_buffers(page);
3348 int try_to_free_buffers(struct page *page)
3350 struct address_space * const mapping = page->mapping;
3351 struct buffer_head *buffers_to_free = NULL;
3354 BUG_ON(!PageLocked(page));
3355 if (PageWriteback(page))
3358 if (mapping == NULL) { /* can this still happen? */
3359 ret = drop_buffers(page, &buffers_to_free);
3363 spin_lock(&mapping->private_lock);
3364 ret = drop_buffers(page, &buffers_to_free);
3367 * If the filesystem writes its buffers by hand (eg ext3)
3368 * then we can have clean buffers against a dirty page. We
3369 * clean the page here; otherwise the VM will never notice
3370 * that the filesystem did any IO at all.
3372 * Also, during truncate, discard_buffer will have marked all
3373 * the page's buffers clean. We discover that here and clean
3376 * private_lock must be held over this entire operation in order
3377 * to synchronise against __set_page_dirty_buffers and prevent the
3378 * dirty bit from being lost.
3381 cancel_dirty_page(page);
3382 spin_unlock(&mapping->private_lock);
3384 if (buffers_to_free) {
3385 struct buffer_head *bh = buffers_to_free;
3388 struct buffer_head *next = bh->b_this_page;
3389 free_buffer_head(bh);
3391 } while (bh != buffers_to_free);
3395 EXPORT_SYMBOL(try_to_free_buffers);
3398 * There are no bdflush tunables left. But distributions are
3399 * still running obsolete flush daemons, so we terminate them here.
3401 * Use of bdflush() is deprecated and will be removed in a future kernel.
3402 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3404 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3406 static int msg_count;
3408 if (!capable(CAP_SYS_ADMIN))
3411 if (msg_count < 5) {
3414 "warning: process `%s' used the obsolete bdflush"
3415 " system call\n", current->comm);
3416 printk(KERN_INFO "Fix your initscripts?\n");
3425 * Buffer-head allocation
3427 static struct kmem_cache *bh_cachep __read_mostly;
3430 * Once the number of bh's in the machine exceeds this level, we start
3431 * stripping them in writeback.
3433 static unsigned long max_buffer_heads;
3435 int buffer_heads_over_limit;
3437 struct bh_accounting {
3438 int nr; /* Number of live bh's */
3439 int ratelimit; /* Limit cacheline bouncing */
3442 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3444 static void recalc_bh_state(void)
3449 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3451 __this_cpu_write(bh_accounting.ratelimit, 0);
3452 for_each_online_cpu(i)
3453 tot += per_cpu(bh_accounting, i).nr;
3454 buffer_heads_over_limit = (tot > max_buffer_heads);
3457 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3459 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3461 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3463 __this_cpu_inc(bh_accounting.nr);
3469 EXPORT_SYMBOL(alloc_buffer_head);
3471 void free_buffer_head(struct buffer_head *bh)
3473 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3474 kmem_cache_free(bh_cachep, bh);
3476 __this_cpu_dec(bh_accounting.nr);
3480 EXPORT_SYMBOL(free_buffer_head);
3482 static void buffer_exit_cpu(int cpu)
3485 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3487 for (i = 0; i < BH_LRU_SIZE; i++) {
3491 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3492 per_cpu(bh_accounting, cpu).nr = 0;
3495 static int buffer_cpu_notify(struct notifier_block *self,
3496 unsigned long action, void *hcpu)
3498 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3499 buffer_exit_cpu((unsigned long)hcpu);
3504 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3505 * @bh: struct buffer_head
3507 * Return true if the buffer is up-to-date and false,
3508 * with the buffer locked, if not.
3510 int bh_uptodate_or_lock(struct buffer_head *bh)
3512 if (!buffer_uptodate(bh)) {
3514 if (!buffer_uptodate(bh))
3520 EXPORT_SYMBOL(bh_uptodate_or_lock);
3523 * bh_submit_read - Submit a locked buffer for reading
3524 * @bh: struct buffer_head
3526 * Returns zero on success and -EIO on error.
3528 int bh_submit_read(struct buffer_head *bh)
3530 BUG_ON(!buffer_locked(bh));
3532 if (buffer_uptodate(bh)) {
3538 bh->b_end_io = end_buffer_read_sync;
3539 submit_bh(REQ_OP_READ, 0, bh);
3541 if (buffer_uptodate(bh))
3545 EXPORT_SYMBOL(bh_submit_read);
3547 void __init buffer_init(void)
3549 unsigned long nrpages;
3551 bh_cachep = kmem_cache_create("buffer_head",
3552 sizeof(struct buffer_head), 0,
3553 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3558 * Limit the bh occupancy to 10% of ZONE_NORMAL
3560 nrpages = (nr_free_buffer_pages() * 10) / 100;
3561 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3562 hotcpu_notifier(buffer_cpu_notify, 0);