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/sched/signal.h>
23 #include <linux/syscalls.h>
25 #include <linux/iomap.h>
27 #include <linux/percpu.h>
28 #include <linux/slab.h>
29 #include <linux/capability.h>
30 #include <linux/blkdev.h>
31 #include <linux/file.h>
32 #include <linux/quotaops.h>
33 #include <linux/highmem.h>
34 #include <linux/export.h>
35 #include <linux/backing-dev.h>
36 #include <linux/writeback.h>
37 #include <linux/hash.h>
38 #include <linux/suspend.h>
39 #include <linux/buffer_head.h>
40 #include <linux/task_io_accounting_ops.h>
41 #include <linux/bio.h>
42 #include <linux/notifier.h>
43 #include <linux/cpu.h>
44 #include <linux/bitops.h>
45 #include <linux/mpage.h>
46 #include <linux/bit_spinlock.h>
47 #include <linux/pagevec.h>
48 #include <trace/events/block.h>
50 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
51 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
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 mark_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 mark_buffer_write_io_error(bh);
356 clear_buffer_uptodate(bh);
360 first = page_buffers(page);
361 local_irq_save(flags);
362 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
364 clear_buffer_async_write(bh);
366 tmp = bh->b_this_page;
368 if (buffer_async_write(tmp)) {
369 BUG_ON(!buffer_locked(tmp));
372 tmp = tmp->b_this_page;
374 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
375 local_irq_restore(flags);
376 end_page_writeback(page);
380 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
381 local_irq_restore(flags);
384 EXPORT_SYMBOL(end_buffer_async_write);
387 * If a page's buffers are under async readin (end_buffer_async_read
388 * completion) then there is a possibility that another thread of
389 * control could lock one of the buffers after it has completed
390 * but while some of the other buffers have not completed. This
391 * locked buffer would confuse end_buffer_async_read() into not unlocking
392 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
393 * that this buffer is not under async I/O.
395 * The page comes unlocked when it has no locked buffer_async buffers
398 * PageLocked prevents anyone starting new async I/O reads any of
401 * PageWriteback is used to prevent simultaneous writeout of the same
404 * PageLocked prevents anyone from starting writeback of a page which is
405 * under read I/O (PageWriteback is only ever set against a locked page).
407 static void mark_buffer_async_read(struct buffer_head *bh)
409 bh->b_end_io = end_buffer_async_read;
410 set_buffer_async_read(bh);
413 static void mark_buffer_async_write_endio(struct buffer_head *bh,
414 bh_end_io_t *handler)
416 bh->b_end_io = handler;
417 set_buffer_async_write(bh);
420 void mark_buffer_async_write(struct buffer_head *bh)
422 mark_buffer_async_write_endio(bh, end_buffer_async_write);
424 EXPORT_SYMBOL(mark_buffer_async_write);
428 * fs/buffer.c contains helper functions for buffer-backed address space's
429 * fsync functions. A common requirement for buffer-based filesystems is
430 * that certain data from the backing blockdev needs to be written out for
431 * a successful fsync(). For example, ext2 indirect blocks need to be
432 * written back and waited upon before fsync() returns.
434 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
435 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
436 * management of a list of dependent buffers at ->i_mapping->private_list.
438 * Locking is a little subtle: try_to_free_buffers() will remove buffers
439 * from their controlling inode's queue when they are being freed. But
440 * try_to_free_buffers() will be operating against the *blockdev* mapping
441 * at the time, not against the S_ISREG file which depends on those buffers.
442 * So the locking for private_list is via the private_lock in the address_space
443 * which backs the buffers. Which is different from the address_space
444 * against which the buffers are listed. So for a particular address_space,
445 * mapping->private_lock does *not* protect mapping->private_list! In fact,
446 * mapping->private_list will always be protected by the backing blockdev's
449 * Which introduces a requirement: all buffers on an address_space's
450 * ->private_list must be from the same address_space: the blockdev's.
452 * address_spaces which do not place buffers at ->private_list via these
453 * utility functions are free to use private_lock and private_list for
454 * whatever they want. The only requirement is that list_empty(private_list)
455 * be true at clear_inode() time.
457 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
458 * filesystems should do that. invalidate_inode_buffers() should just go
459 * BUG_ON(!list_empty).
461 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
462 * take an address_space, not an inode. And it should be called
463 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
466 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
467 * list if it is already on a list. Because if the buffer is on a list,
468 * it *must* already be on the right one. If not, the filesystem is being
469 * silly. This will save a ton of locking. But first we have to ensure
470 * that buffers are taken *off* the old inode's list when they are freed
471 * (presumably in truncate). That requires careful auditing of all
472 * filesystems (do it inside bforget()). It could also be done by bringing
477 * The buffer's backing address_space's private_lock must be held
479 static void __remove_assoc_queue(struct buffer_head *bh)
481 list_del_init(&bh->b_assoc_buffers);
482 WARN_ON(!bh->b_assoc_map);
483 bh->b_assoc_map = NULL;
486 int inode_has_buffers(struct inode *inode)
488 return !list_empty(&inode->i_data.private_list);
492 * osync is designed to support O_SYNC io. It waits synchronously for
493 * all already-submitted IO to complete, but does not queue any new
494 * writes to the disk.
496 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
497 * you dirty the buffers, and then use osync_inode_buffers to wait for
498 * completion. Any other dirty buffers which are not yet queued for
499 * write will not be flushed to disk by the osync.
501 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
503 struct buffer_head *bh;
509 list_for_each_prev(p, list) {
511 if (buffer_locked(bh)) {
515 if (!buffer_uptodate(bh))
526 static void do_thaw_one(struct super_block *sb, void *unused)
528 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
529 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
532 static void do_thaw_all(struct work_struct *work)
534 iterate_supers(do_thaw_one, NULL);
536 printk(KERN_WARNING "Emergency Thaw complete\n");
540 * emergency_thaw_all -- forcibly thaw every frozen filesystem
542 * Used for emergency unfreeze of all filesystems via SysRq
544 void emergency_thaw_all(void)
546 struct work_struct *work;
548 work = kmalloc(sizeof(*work), GFP_ATOMIC);
550 INIT_WORK(work, do_thaw_all);
556 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
557 * @mapping: the mapping which wants those buffers written
559 * Starts I/O against the buffers at mapping->private_list, and waits upon
562 * Basically, this is a convenience function for fsync().
563 * @mapping is a file or directory which needs those buffers to be written for
564 * a successful fsync().
566 int sync_mapping_buffers(struct address_space *mapping)
568 struct address_space *buffer_mapping = mapping->private_data;
570 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
573 return fsync_buffers_list(&buffer_mapping->private_lock,
574 &mapping->private_list);
576 EXPORT_SYMBOL(sync_mapping_buffers);
579 * Called when we've recently written block `bblock', and it is known that
580 * `bblock' was for a buffer_boundary() buffer. This means that the block at
581 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
582 * dirty, schedule it for IO. So that indirects merge nicely with their data.
584 void write_boundary_block(struct block_device *bdev,
585 sector_t bblock, unsigned blocksize)
587 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
589 if (buffer_dirty(bh))
590 ll_rw_block(REQ_OP_WRITE, 0, 1, &bh);
595 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
597 struct address_space *mapping = inode->i_mapping;
598 struct address_space *buffer_mapping = bh->b_page->mapping;
600 mark_buffer_dirty(bh);
601 if (!mapping->private_data) {
602 mapping->private_data = buffer_mapping;
604 BUG_ON(mapping->private_data != buffer_mapping);
606 if (!bh->b_assoc_map) {
607 spin_lock(&buffer_mapping->private_lock);
608 list_move_tail(&bh->b_assoc_buffers,
609 &mapping->private_list);
610 bh->b_assoc_map = mapping;
611 spin_unlock(&buffer_mapping->private_lock);
614 EXPORT_SYMBOL(mark_buffer_dirty_inode);
617 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
620 * If warn is true, then emit a warning if the page is not uptodate and has
621 * not been truncated.
623 * The caller must hold lock_page_memcg().
625 static void __set_page_dirty(struct page *page, struct address_space *mapping,
630 spin_lock_irqsave(&mapping->tree_lock, flags);
631 if (page->mapping) { /* Race with truncate? */
632 WARN_ON_ONCE(warn && !PageUptodate(page));
633 account_page_dirtied(page, mapping);
634 radix_tree_tag_set(&mapping->page_tree,
635 page_index(page), PAGECACHE_TAG_DIRTY);
637 spin_unlock_irqrestore(&mapping->tree_lock, flags);
641 * Add a page to the dirty page list.
643 * It is a sad fact of life that this function is called from several places
644 * deeply under spinlocking. It may not sleep.
646 * If the page has buffers, the uptodate buffers are set dirty, to preserve
647 * dirty-state coherency between the page and the buffers. It the page does
648 * not have buffers then when they are later attached they will all be set
651 * The buffers are dirtied before the page is dirtied. There's a small race
652 * window in which a writepage caller may see the page cleanness but not the
653 * buffer dirtiness. That's fine. If this code were to set the page dirty
654 * before the buffers, a concurrent writepage caller could clear the page dirty
655 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
656 * page on the dirty page list.
658 * We use private_lock to lock against try_to_free_buffers while using the
659 * page's buffer list. Also use this to protect against clean buffers being
660 * added to the page after it was set dirty.
662 * FIXME: may need to call ->reservepage here as well. That's rather up to the
663 * address_space though.
665 int __set_page_dirty_buffers(struct page *page)
668 struct address_space *mapping = page_mapping(page);
670 if (unlikely(!mapping))
671 return !TestSetPageDirty(page);
673 spin_lock(&mapping->private_lock);
674 if (page_has_buffers(page)) {
675 struct buffer_head *head = page_buffers(page);
676 struct buffer_head *bh = head;
679 set_buffer_dirty(bh);
680 bh = bh->b_this_page;
681 } while (bh != head);
684 * Lock out page->mem_cgroup migration to keep PageDirty
685 * synchronized with per-memcg dirty page counters.
687 lock_page_memcg(page);
688 newly_dirty = !TestSetPageDirty(page);
689 spin_unlock(&mapping->private_lock);
692 __set_page_dirty(page, mapping, 1);
694 unlock_page_memcg(page);
697 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
701 EXPORT_SYMBOL(__set_page_dirty_buffers);
704 * Write out and wait upon a list of buffers.
706 * We have conflicting pressures: we want to make sure that all
707 * initially dirty buffers get waited on, but that any subsequently
708 * dirtied buffers don't. After all, we don't want fsync to last
709 * forever if somebody is actively writing to the file.
711 * Do this in two main stages: first we copy dirty buffers to a
712 * temporary inode list, queueing the writes as we go. Then we clean
713 * up, waiting for those writes to complete.
715 * During this second stage, any subsequent updates to the file may end
716 * up refiling the buffer on the original inode's dirty list again, so
717 * there is a chance we will end up with a buffer queued for write but
718 * not yet completed on that list. So, as a final cleanup we go through
719 * the osync code to catch these locked, dirty buffers without requeuing
720 * any newly dirty buffers for write.
722 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
724 struct buffer_head *bh;
725 struct list_head tmp;
726 struct address_space *mapping;
728 struct blk_plug plug;
730 INIT_LIST_HEAD(&tmp);
731 blk_start_plug(&plug);
734 while (!list_empty(list)) {
735 bh = BH_ENTRY(list->next);
736 mapping = bh->b_assoc_map;
737 __remove_assoc_queue(bh);
738 /* Avoid race with mark_buffer_dirty_inode() which does
739 * a lockless check and we rely on seeing the dirty bit */
741 if (buffer_dirty(bh) || buffer_locked(bh)) {
742 list_add(&bh->b_assoc_buffers, &tmp);
743 bh->b_assoc_map = mapping;
744 if (buffer_dirty(bh)) {
748 * Ensure any pending I/O completes so that
749 * write_dirty_buffer() actually writes the
750 * current contents - it is a noop if I/O is
751 * still in flight on potentially older
754 write_dirty_buffer(bh, REQ_SYNC);
757 * Kick off IO for the previous mapping. Note
758 * that we will not run the very last mapping,
759 * wait_on_buffer() will do that for us
760 * through sync_buffer().
769 blk_finish_plug(&plug);
772 while (!list_empty(&tmp)) {
773 bh = BH_ENTRY(tmp.prev);
775 mapping = bh->b_assoc_map;
776 __remove_assoc_queue(bh);
777 /* Avoid race with mark_buffer_dirty_inode() which does
778 * a lockless check and we rely on seeing the dirty bit */
780 if (buffer_dirty(bh)) {
781 list_add(&bh->b_assoc_buffers,
782 &mapping->private_list);
783 bh->b_assoc_map = mapping;
787 if (!buffer_uptodate(bh))
794 err2 = osync_buffers_list(lock, list);
802 * Invalidate any and all dirty buffers on a given inode. We are
803 * probably unmounting the fs, but that doesn't mean we have already
804 * done a sync(). Just drop the buffers from the inode list.
806 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
807 * assumes that all the buffers are against the blockdev. Not true
810 void invalidate_inode_buffers(struct inode *inode)
812 if (inode_has_buffers(inode)) {
813 struct address_space *mapping = &inode->i_data;
814 struct list_head *list = &mapping->private_list;
815 struct address_space *buffer_mapping = mapping->private_data;
817 spin_lock(&buffer_mapping->private_lock);
818 while (!list_empty(list))
819 __remove_assoc_queue(BH_ENTRY(list->next));
820 spin_unlock(&buffer_mapping->private_lock);
823 EXPORT_SYMBOL(invalidate_inode_buffers);
826 * Remove any clean buffers from the inode's buffer list. This is called
827 * when we're trying to free the inode itself. Those buffers can pin it.
829 * Returns true if all buffers were removed.
831 int remove_inode_buffers(struct inode *inode)
835 if (inode_has_buffers(inode)) {
836 struct address_space *mapping = &inode->i_data;
837 struct list_head *list = &mapping->private_list;
838 struct address_space *buffer_mapping = mapping->private_data;
840 spin_lock(&buffer_mapping->private_lock);
841 while (!list_empty(list)) {
842 struct buffer_head *bh = BH_ENTRY(list->next);
843 if (buffer_dirty(bh)) {
847 __remove_assoc_queue(bh);
849 spin_unlock(&buffer_mapping->private_lock);
855 * Create the appropriate buffers when given a page for data area and
856 * the size of each buffer.. Use the bh->b_this_page linked list to
857 * follow the buffers created. Return NULL if unable to create more
860 * The retry flag is used to differentiate async IO (paging, swapping)
861 * which may not fail from ordinary buffer allocations.
863 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
866 struct buffer_head *bh, *head;
872 while ((offset -= size) >= 0) {
873 bh = alloc_buffer_head(GFP_NOFS);
877 bh->b_this_page = head;
883 /* Link the buffer to its page */
884 set_bh_page(bh, page, offset);
888 * In case anything failed, we just free everything we got.
894 head = head->b_this_page;
895 free_buffer_head(bh);
900 * Return failure for non-async IO requests. Async IO requests
901 * are not allowed to fail, so we have to wait until buffer heads
902 * become available. But we don't want tasks sleeping with
903 * partially complete buffers, so all were released above.
908 /* We're _really_ low on memory. Now we just
909 * wait for old buffer heads to become free due to
910 * finishing IO. Since this is an async request and
911 * the reserve list is empty, we're sure there are
912 * async buffer heads in use.
917 EXPORT_SYMBOL_GPL(alloc_page_buffers);
920 link_dev_buffers(struct page *page, struct buffer_head *head)
922 struct buffer_head *bh, *tail;
927 bh = bh->b_this_page;
929 tail->b_this_page = head;
930 attach_page_buffers(page, head);
933 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
935 sector_t retval = ~((sector_t)0);
936 loff_t sz = i_size_read(bdev->bd_inode);
939 unsigned int sizebits = blksize_bits(size);
940 retval = (sz >> sizebits);
946 * Initialise the state of a blockdev page's buffers.
949 init_page_buffers(struct page *page, struct block_device *bdev,
950 sector_t block, int size)
952 struct buffer_head *head = page_buffers(page);
953 struct buffer_head *bh = head;
954 int uptodate = PageUptodate(page);
955 sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
958 if (!buffer_mapped(bh)) {
959 init_buffer(bh, NULL, NULL);
961 bh->b_blocknr = block;
963 set_buffer_uptodate(bh);
964 if (block < end_block)
965 set_buffer_mapped(bh);
968 bh = bh->b_this_page;
969 } while (bh != head);
972 * Caller needs to validate requested block against end of device.
978 * Create the page-cache page that contains the requested block.
980 * This is used purely for blockdev mappings.
983 grow_dev_page(struct block_device *bdev, sector_t block,
984 pgoff_t index, int size, int sizebits, gfp_t gfp)
986 struct inode *inode = bdev->bd_inode;
988 struct buffer_head *bh;
990 int ret = 0; /* Will call free_more_memory() */
993 gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
996 * XXX: __getblk_slow() can not really deal with failure and
997 * will endlessly loop on improvised global reclaim. Prefer
998 * looping in the allocator rather than here, at least that
999 * code knows what it's doing.
1001 gfp_mask |= __GFP_NOFAIL;
1003 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
1007 BUG_ON(!PageLocked(page));
1009 if (page_has_buffers(page)) {
1010 bh = page_buffers(page);
1011 if (bh->b_size == size) {
1012 end_block = init_page_buffers(page, bdev,
1013 (sector_t)index << sizebits,
1017 if (!try_to_free_buffers(page))
1022 * Allocate some buffers for this page
1024 bh = alloc_page_buffers(page, size, 0);
1029 * Link the page to the buffers and initialise them. Take the
1030 * lock to be atomic wrt __find_get_block(), which does not
1031 * run under the page lock.
1033 spin_lock(&inode->i_mapping->private_lock);
1034 link_dev_buffers(page, bh);
1035 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1037 spin_unlock(&inode->i_mapping->private_lock);
1039 ret = (block < end_block) ? 1 : -ENXIO;
1047 * Create buffers for the specified block device block's page. If
1048 * that page was dirty, the buffers are set dirty also.
1051 grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
1059 } while ((size << sizebits) < PAGE_SIZE);
1061 index = block >> sizebits;
1064 * Check for a block which wants to lie outside our maximum possible
1065 * pagecache index. (this comparison is done using sector_t types).
1067 if (unlikely(index != block >> sizebits)) {
1068 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1070 __func__, (unsigned long long)block,
1075 /* Create a page with the proper size buffers.. */
1076 return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1079 static struct buffer_head *
1080 __getblk_slow(struct block_device *bdev, sector_t block,
1081 unsigned size, gfp_t gfp)
1083 /* Size must be multiple of hard sectorsize */
1084 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1085 (size < 512 || size > PAGE_SIZE))) {
1086 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1088 printk(KERN_ERR "logical block size: %d\n",
1089 bdev_logical_block_size(bdev));
1096 struct buffer_head *bh;
1099 bh = __find_get_block(bdev, block, size);
1103 ret = grow_buffers(bdev, block, size, gfp);
1112 * The relationship between dirty buffers and dirty pages:
1114 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1115 * the page is tagged dirty in its radix tree.
1117 * At all times, the dirtiness of the buffers represents the dirtiness of
1118 * subsections of the page. If the page has buffers, the page dirty bit is
1119 * merely a hint about the true dirty state.
1121 * When a page is set dirty in its entirety, all its buffers are marked dirty
1122 * (if the page has buffers).
1124 * When a buffer is marked dirty, its page is dirtied, but the page's other
1127 * Also. When blockdev buffers are explicitly read with bread(), they
1128 * individually become uptodate. But their backing page remains not
1129 * uptodate - even if all of its buffers are uptodate. A subsequent
1130 * block_read_full_page() against that page will discover all the uptodate
1131 * buffers, will set the page uptodate and will perform no I/O.
1135 * mark_buffer_dirty - mark a buffer_head as needing writeout
1136 * @bh: the buffer_head to mark dirty
1138 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1139 * backing page dirty, then tag the page as dirty in its address_space's radix
1140 * tree and then attach the address_space's inode to its superblock's dirty
1143 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1144 * mapping->tree_lock and mapping->host->i_lock.
1146 void mark_buffer_dirty(struct buffer_head *bh)
1148 WARN_ON_ONCE(!buffer_uptodate(bh));
1150 trace_block_dirty_buffer(bh);
1153 * Very *carefully* optimize the it-is-already-dirty case.
1155 * Don't let the final "is it dirty" escape to before we
1156 * perhaps modified the buffer.
1158 if (buffer_dirty(bh)) {
1160 if (buffer_dirty(bh))
1164 if (!test_set_buffer_dirty(bh)) {
1165 struct page *page = bh->b_page;
1166 struct address_space *mapping = NULL;
1168 lock_page_memcg(page);
1169 if (!TestSetPageDirty(page)) {
1170 mapping = page_mapping(page);
1172 __set_page_dirty(page, mapping, 0);
1174 unlock_page_memcg(page);
1176 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1179 EXPORT_SYMBOL(mark_buffer_dirty);
1181 void mark_buffer_write_io_error(struct buffer_head *bh)
1183 set_buffer_write_io_error(bh);
1184 /* FIXME: do we need to set this in both places? */
1185 if (bh->b_page && bh->b_page->mapping)
1186 mapping_set_error(bh->b_page->mapping, -EIO);
1187 if (bh->b_assoc_map)
1188 mapping_set_error(bh->b_assoc_map, -EIO);
1190 EXPORT_SYMBOL(mark_buffer_write_io_error);
1193 * Decrement a buffer_head's reference count. If all buffers against a page
1194 * have zero reference count, are clean and unlocked, and if the page is clean
1195 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1196 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1197 * a page but it ends up not being freed, and buffers may later be reattached).
1199 void __brelse(struct buffer_head * buf)
1201 if (atomic_read(&buf->b_count)) {
1205 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1207 EXPORT_SYMBOL(__brelse);
1210 * bforget() is like brelse(), except it discards any
1211 * potentially dirty data.
1213 void __bforget(struct buffer_head *bh)
1215 clear_buffer_dirty(bh);
1216 if (bh->b_assoc_map) {
1217 struct address_space *buffer_mapping = bh->b_page->mapping;
1219 spin_lock(&buffer_mapping->private_lock);
1220 list_del_init(&bh->b_assoc_buffers);
1221 bh->b_assoc_map = NULL;
1222 spin_unlock(&buffer_mapping->private_lock);
1226 EXPORT_SYMBOL(__bforget);
1228 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1231 if (buffer_uptodate(bh)) {
1236 bh->b_end_io = end_buffer_read_sync;
1237 submit_bh(REQ_OP_READ, 0, bh);
1239 if (buffer_uptodate(bh))
1247 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1248 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1249 * refcount elevated by one when they're in an LRU. A buffer can only appear
1250 * once in a particular CPU's LRU. A single buffer can be present in multiple
1251 * CPU's LRUs at the same time.
1253 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1254 * sb_find_get_block().
1256 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1257 * a local interrupt disable for that.
1260 #define BH_LRU_SIZE 16
1263 struct buffer_head *bhs[BH_LRU_SIZE];
1266 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1269 #define bh_lru_lock() local_irq_disable()
1270 #define bh_lru_unlock() local_irq_enable()
1272 #define bh_lru_lock() preempt_disable()
1273 #define bh_lru_unlock() preempt_enable()
1276 static inline void check_irqs_on(void)
1278 #ifdef irqs_disabled
1279 BUG_ON(irqs_disabled());
1284 * The LRU management algorithm is dopey-but-simple. Sorry.
1286 static void bh_lru_install(struct buffer_head *bh)
1288 struct buffer_head *evictee = NULL;
1292 if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1293 struct buffer_head *bhs[BH_LRU_SIZE];
1299 for (in = 0; in < BH_LRU_SIZE; in++) {
1300 struct buffer_head *bh2 =
1301 __this_cpu_read(bh_lrus.bhs[in]);
1306 if (out >= BH_LRU_SIZE) {
1307 BUG_ON(evictee != NULL);
1314 while (out < BH_LRU_SIZE)
1316 memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1325 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1327 static struct buffer_head *
1328 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1330 struct buffer_head *ret = NULL;
1335 for (i = 0; i < BH_LRU_SIZE; i++) {
1336 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1338 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1339 bh->b_size == size) {
1342 __this_cpu_write(bh_lrus.bhs[i],
1343 __this_cpu_read(bh_lrus.bhs[i - 1]));
1346 __this_cpu_write(bh_lrus.bhs[0], bh);
1358 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1359 * it in the LRU and mark it as accessed. If it is not present then return
1362 struct buffer_head *
1363 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1365 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1368 /* __find_get_block_slow will mark the page accessed */
1369 bh = __find_get_block_slow(bdev, block);
1377 EXPORT_SYMBOL(__find_get_block);
1380 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1381 * which corresponds to the passed block_device, block and size. The
1382 * returned buffer has its reference count incremented.
1384 * __getblk_gfp() will lock up the machine if grow_dev_page's
1385 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1387 struct buffer_head *
1388 __getblk_gfp(struct block_device *bdev, sector_t block,
1389 unsigned size, gfp_t gfp)
1391 struct buffer_head *bh = __find_get_block(bdev, block, size);
1395 bh = __getblk_slow(bdev, block, size, gfp);
1398 EXPORT_SYMBOL(__getblk_gfp);
1401 * Do async read-ahead on a buffer..
1403 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1405 struct buffer_head *bh = __getblk(bdev, block, size);
1407 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1411 EXPORT_SYMBOL(__breadahead);
1414 * __bread_gfp() - reads a specified block and returns the bh
1415 * @bdev: the block_device to read from
1416 * @block: number of block
1417 * @size: size (in bytes) to read
1418 * @gfp: page allocation flag
1420 * Reads a specified block, and returns buffer head that contains it.
1421 * The page cache can be allocated from non-movable area
1422 * not to prevent page migration if you set gfp to zero.
1423 * It returns NULL if the block was unreadable.
1425 struct buffer_head *
1426 __bread_gfp(struct block_device *bdev, sector_t block,
1427 unsigned size, gfp_t gfp)
1429 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1431 if (likely(bh) && !buffer_uptodate(bh))
1432 bh = __bread_slow(bh);
1435 EXPORT_SYMBOL(__bread_gfp);
1438 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1439 * This doesn't race because it runs in each cpu either in irq
1440 * or with preempt disabled.
1442 static void invalidate_bh_lru(void *arg)
1444 struct bh_lru *b = &get_cpu_var(bh_lrus);
1447 for (i = 0; i < BH_LRU_SIZE; i++) {
1451 put_cpu_var(bh_lrus);
1454 static bool has_bh_in_lru(int cpu, void *dummy)
1456 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1459 for (i = 0; i < BH_LRU_SIZE; i++) {
1467 void invalidate_bh_lrus(void)
1469 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1471 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1473 void set_bh_page(struct buffer_head *bh,
1474 struct page *page, unsigned long offset)
1477 BUG_ON(offset >= PAGE_SIZE);
1478 if (PageHighMem(page))
1480 * This catches illegal uses and preserves the offset:
1482 bh->b_data = (char *)(0 + offset);
1484 bh->b_data = page_address(page) + offset;
1486 EXPORT_SYMBOL(set_bh_page);
1489 * Called when truncating a buffer on a page completely.
1492 /* Bits that are cleared during an invalidate */
1493 #define BUFFER_FLAGS_DISCARD \
1494 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1495 1 << BH_Delay | 1 << BH_Unwritten)
1497 static void discard_buffer(struct buffer_head * bh)
1499 unsigned long b_state, b_state_old;
1502 clear_buffer_dirty(bh);
1504 b_state = bh->b_state;
1506 b_state_old = cmpxchg(&bh->b_state, b_state,
1507 (b_state & ~BUFFER_FLAGS_DISCARD));
1508 if (b_state_old == b_state)
1510 b_state = b_state_old;
1516 * block_invalidatepage - invalidate part or all of a buffer-backed page
1518 * @page: the page which is affected
1519 * @offset: start of the range to invalidate
1520 * @length: length of the range to invalidate
1522 * block_invalidatepage() is called when all or part of the page has become
1523 * invalidated by a truncate operation.
1525 * block_invalidatepage() does not have to release all buffers, but it must
1526 * ensure that no dirty buffer is left outside @offset and that no I/O
1527 * is underway against any of the blocks which are outside the truncation
1528 * point. Because the caller is about to free (and possibly reuse) those
1531 void block_invalidatepage(struct page *page, unsigned int offset,
1532 unsigned int length)
1534 struct buffer_head *head, *bh, *next;
1535 unsigned int curr_off = 0;
1536 unsigned int stop = length + offset;
1538 BUG_ON(!PageLocked(page));
1539 if (!page_has_buffers(page))
1543 * Check for overflow
1545 BUG_ON(stop > PAGE_SIZE || stop < length);
1547 head = page_buffers(page);
1550 unsigned int next_off = curr_off + bh->b_size;
1551 next = bh->b_this_page;
1554 * Are we still fully in range ?
1556 if (next_off > stop)
1560 * is this block fully invalidated?
1562 if (offset <= curr_off)
1564 curr_off = next_off;
1566 } while (bh != head);
1569 * We release buffers only if the entire page is being invalidated.
1570 * The get_block cached value has been unconditionally invalidated,
1571 * so real IO is not possible anymore.
1574 try_to_release_page(page, 0);
1578 EXPORT_SYMBOL(block_invalidatepage);
1582 * We attach and possibly dirty the buffers atomically wrt
1583 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1584 * is already excluded via the page lock.
1586 void create_empty_buffers(struct page *page,
1587 unsigned long blocksize, unsigned long b_state)
1589 struct buffer_head *bh, *head, *tail;
1591 head = alloc_page_buffers(page, blocksize, 1);
1594 bh->b_state |= b_state;
1596 bh = bh->b_this_page;
1598 tail->b_this_page = head;
1600 spin_lock(&page->mapping->private_lock);
1601 if (PageUptodate(page) || PageDirty(page)) {
1604 if (PageDirty(page))
1605 set_buffer_dirty(bh);
1606 if (PageUptodate(page))
1607 set_buffer_uptodate(bh);
1608 bh = bh->b_this_page;
1609 } while (bh != head);
1611 attach_page_buffers(page, head);
1612 spin_unlock(&page->mapping->private_lock);
1614 EXPORT_SYMBOL(create_empty_buffers);
1617 * clean_bdev_aliases: clean a range of buffers in block device
1618 * @bdev: Block device to clean buffers in
1619 * @block: Start of a range of blocks to clean
1620 * @len: Number of blocks to clean
1622 * We are taking a range of blocks for data and we don't want writeback of any
1623 * buffer-cache aliases starting from return from this function and until the
1624 * moment when something will explicitly mark the buffer dirty (hopefully that
1625 * will not happen until we will free that block ;-) We don't even need to mark
1626 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1627 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1628 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1629 * would confuse anyone who might pick it with bread() afterwards...
1631 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1632 * writeout I/O going on against recently-freed buffers. We don't wait on that
1633 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1634 * need to. That happens here.
1636 void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1638 struct inode *bd_inode = bdev->bd_inode;
1639 struct address_space *bd_mapping = bd_inode->i_mapping;
1640 struct pagevec pvec;
1641 pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1644 struct buffer_head *bh;
1645 struct buffer_head *head;
1647 end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1648 pagevec_init(&pvec, 0);
1649 while (index <= end && pagevec_lookup(&pvec, bd_mapping, index,
1650 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
1651 for (i = 0; i < pagevec_count(&pvec); i++) {
1652 struct page *page = pvec.pages[i];
1654 index = page->index;
1657 if (!page_has_buffers(page))
1660 * We use page lock instead of bd_mapping->private_lock
1661 * to pin buffers here since we can afford to sleep and
1662 * it scales better than a global spinlock lock.
1665 /* Recheck when the page is locked which pins bhs */
1666 if (!page_has_buffers(page))
1668 head = page_buffers(page);
1671 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1673 if (bh->b_blocknr >= block + len)
1675 clear_buffer_dirty(bh);
1677 clear_buffer_req(bh);
1679 bh = bh->b_this_page;
1680 } while (bh != head);
1684 pagevec_release(&pvec);
1689 EXPORT_SYMBOL(clean_bdev_aliases);
1692 * Size is a power-of-two in the range 512..PAGE_SIZE,
1693 * and the case we care about most is PAGE_SIZE.
1695 * So this *could* possibly be written with those
1696 * constraints in mind (relevant mostly if some
1697 * architecture has a slow bit-scan instruction)
1699 static inline int block_size_bits(unsigned int blocksize)
1701 return ilog2(blocksize);
1704 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1706 BUG_ON(!PageLocked(page));
1708 if (!page_has_buffers(page))
1709 create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1710 return page_buffers(page);
1714 * NOTE! All mapped/uptodate combinations are valid:
1716 * Mapped Uptodate Meaning
1718 * No No "unknown" - must do get_block()
1719 * No Yes "hole" - zero-filled
1720 * Yes No "allocated" - allocated on disk, not read in
1721 * Yes Yes "valid" - allocated and up-to-date in memory.
1723 * "Dirty" is valid only with the last case (mapped+uptodate).
1727 * While block_write_full_page is writing back the dirty buffers under
1728 * the page lock, whoever dirtied the buffers may decide to clean them
1729 * again at any time. We handle that by only looking at the buffer
1730 * state inside lock_buffer().
1732 * If block_write_full_page() is called for regular writeback
1733 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1734 * locked buffer. This only can happen if someone has written the buffer
1735 * directly, with submit_bh(). At the address_space level PageWriteback
1736 * prevents this contention from occurring.
1738 * If block_write_full_page() is called with wbc->sync_mode ==
1739 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1740 * causes the writes to be flagged as synchronous writes.
1742 int __block_write_full_page(struct inode *inode, struct page *page,
1743 get_block_t *get_block, struct writeback_control *wbc,
1744 bh_end_io_t *handler)
1748 sector_t last_block;
1749 struct buffer_head *bh, *head;
1750 unsigned int blocksize, bbits;
1751 int nr_underway = 0;
1752 int write_flags = wbc_to_write_flags(wbc);
1754 head = create_page_buffers(page, inode,
1755 (1 << BH_Dirty)|(1 << BH_Uptodate));
1758 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1759 * here, and the (potentially unmapped) buffers may become dirty at
1760 * any time. If a buffer becomes dirty here after we've inspected it
1761 * then we just miss that fact, and the page stays dirty.
1763 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1764 * handle that here by just cleaning them.
1768 blocksize = bh->b_size;
1769 bbits = block_size_bits(blocksize);
1771 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1772 last_block = (i_size_read(inode) - 1) >> bbits;
1775 * Get all the dirty buffers mapped to disk addresses and
1776 * handle any aliases from the underlying blockdev's mapping.
1779 if (block > last_block) {
1781 * mapped buffers outside i_size will occur, because
1782 * this page can be outside i_size when there is a
1783 * truncate in progress.
1786 * The buffer was zeroed by block_write_full_page()
1788 clear_buffer_dirty(bh);
1789 set_buffer_uptodate(bh);
1790 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1792 WARN_ON(bh->b_size != blocksize);
1793 err = get_block(inode, block, bh, 1);
1796 clear_buffer_delay(bh);
1797 if (buffer_new(bh)) {
1798 /* blockdev mappings never come here */
1799 clear_buffer_new(bh);
1800 clean_bdev_bh_alias(bh);
1803 bh = bh->b_this_page;
1805 } while (bh != head);
1808 if (!buffer_mapped(bh))
1811 * If it's a fully non-blocking write attempt and we cannot
1812 * lock the buffer then redirty the page. Note that this can
1813 * potentially cause a busy-wait loop from writeback threads
1814 * and kswapd activity, but those code paths have their own
1815 * higher-level throttling.
1817 if (wbc->sync_mode != WB_SYNC_NONE) {
1819 } else if (!trylock_buffer(bh)) {
1820 redirty_page_for_writepage(wbc, page);
1823 if (test_clear_buffer_dirty(bh)) {
1824 mark_buffer_async_write_endio(bh, handler);
1828 } while ((bh = bh->b_this_page) != head);
1831 * The page and its buffers are protected by PageWriteback(), so we can
1832 * drop the bh refcounts early.
1834 BUG_ON(PageWriteback(page));
1835 set_page_writeback(page);
1838 struct buffer_head *next = bh->b_this_page;
1839 if (buffer_async_write(bh)) {
1840 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh, wbc);
1844 } while (bh != head);
1849 if (nr_underway == 0) {
1851 * The page was marked dirty, but the buffers were
1852 * clean. Someone wrote them back by hand with
1853 * ll_rw_block/submit_bh. A rare case.
1855 end_page_writeback(page);
1858 * The page and buffer_heads can be released at any time from
1866 * ENOSPC, or some other error. We may already have added some
1867 * blocks to the file, so we need to write these out to avoid
1868 * exposing stale data.
1869 * The page is currently locked and not marked for writeback
1872 /* Recovery: lock and submit the mapped buffers */
1874 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1875 !buffer_delay(bh)) {
1877 mark_buffer_async_write_endio(bh, handler);
1880 * The buffer may have been set dirty during
1881 * attachment to a dirty page.
1883 clear_buffer_dirty(bh);
1885 } while ((bh = bh->b_this_page) != head);
1887 BUG_ON(PageWriteback(page));
1888 mapping_set_error(page->mapping, err);
1889 set_page_writeback(page);
1891 struct buffer_head *next = bh->b_this_page;
1892 if (buffer_async_write(bh)) {
1893 clear_buffer_dirty(bh);
1894 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh, wbc);
1898 } while (bh != head);
1902 EXPORT_SYMBOL(__block_write_full_page);
1905 * If a page has any new buffers, zero them out here, and mark them uptodate
1906 * and dirty so they'll be written out (in order to prevent uninitialised
1907 * block data from leaking). And clear the new bit.
1909 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1911 unsigned int block_start, block_end;
1912 struct buffer_head *head, *bh;
1914 BUG_ON(!PageLocked(page));
1915 if (!page_has_buffers(page))
1918 bh = head = page_buffers(page);
1921 block_end = block_start + bh->b_size;
1923 if (buffer_new(bh)) {
1924 if (block_end > from && block_start < to) {
1925 if (!PageUptodate(page)) {
1926 unsigned start, size;
1928 start = max(from, block_start);
1929 size = min(to, block_end) - start;
1931 zero_user(page, start, size);
1932 set_buffer_uptodate(bh);
1935 clear_buffer_new(bh);
1936 mark_buffer_dirty(bh);
1940 block_start = block_end;
1941 bh = bh->b_this_page;
1942 } while (bh != head);
1944 EXPORT_SYMBOL(page_zero_new_buffers);
1947 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1948 struct iomap *iomap)
1950 loff_t offset = block << inode->i_blkbits;
1952 bh->b_bdev = iomap->bdev;
1955 * Block points to offset in file we need to map, iomap contains
1956 * the offset at which the map starts. If the map ends before the
1957 * current block, then do not map the buffer and let the caller
1960 BUG_ON(offset >= iomap->offset + iomap->length);
1962 switch (iomap->type) {
1965 * If the buffer is not up to date or beyond the current EOF,
1966 * we need to mark it as new to ensure sub-block zeroing is
1967 * executed if necessary.
1969 if (!buffer_uptodate(bh) ||
1970 (offset >= i_size_read(inode)))
1973 case IOMAP_DELALLOC:
1974 if (!buffer_uptodate(bh) ||
1975 (offset >= i_size_read(inode)))
1977 set_buffer_uptodate(bh);
1978 set_buffer_mapped(bh);
1979 set_buffer_delay(bh);
1981 case IOMAP_UNWRITTEN:
1983 * For unwritten regions, we always need to ensure that
1984 * sub-block writes cause the regions in the block we are not
1985 * writing to are zeroed. Set the buffer as new to ensure this.
1988 set_buffer_unwritten(bh);
1991 if (offset >= i_size_read(inode))
1993 bh->b_blocknr = (iomap->blkno >> (inode->i_blkbits - 9)) +
1994 ((offset - iomap->offset) >> inode->i_blkbits);
1995 set_buffer_mapped(bh);
2000 int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
2001 get_block_t *get_block, struct iomap *iomap)
2003 unsigned from = pos & (PAGE_SIZE - 1);
2004 unsigned to = from + len;
2005 struct inode *inode = page->mapping->host;
2006 unsigned block_start, block_end;
2009 unsigned blocksize, bbits;
2010 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
2012 BUG_ON(!PageLocked(page));
2013 BUG_ON(from > PAGE_SIZE);
2014 BUG_ON(to > PAGE_SIZE);
2017 head = create_page_buffers(page, inode, 0);
2018 blocksize = head->b_size;
2019 bbits = block_size_bits(blocksize);
2021 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
2023 for(bh = head, block_start = 0; bh != head || !block_start;
2024 block++, block_start=block_end, bh = bh->b_this_page) {
2025 block_end = block_start + blocksize;
2026 if (block_end <= from || block_start >= to) {
2027 if (PageUptodate(page)) {
2028 if (!buffer_uptodate(bh))
2029 set_buffer_uptodate(bh);
2034 clear_buffer_new(bh);
2035 if (!buffer_mapped(bh)) {
2036 WARN_ON(bh->b_size != blocksize);
2038 err = get_block(inode, block, bh, 1);
2042 iomap_to_bh(inode, block, bh, iomap);
2045 if (buffer_new(bh)) {
2046 clean_bdev_bh_alias(bh);
2047 if (PageUptodate(page)) {
2048 clear_buffer_new(bh);
2049 set_buffer_uptodate(bh);
2050 mark_buffer_dirty(bh);
2053 if (block_end > to || block_start < from)
2054 zero_user_segments(page,
2060 if (PageUptodate(page)) {
2061 if (!buffer_uptodate(bh))
2062 set_buffer_uptodate(bh);
2065 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2066 !buffer_unwritten(bh) &&
2067 (block_start < from || block_end > to)) {
2068 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2073 * If we issued read requests - let them complete.
2075 while(wait_bh > wait) {
2076 wait_on_buffer(*--wait_bh);
2077 if (!buffer_uptodate(*wait_bh))
2081 page_zero_new_buffers(page, from, to);
2085 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2086 get_block_t *get_block)
2088 return __block_write_begin_int(page, pos, len, get_block, NULL);
2090 EXPORT_SYMBOL(__block_write_begin);
2092 static int __block_commit_write(struct inode *inode, struct page *page,
2093 unsigned from, unsigned to)
2095 unsigned block_start, block_end;
2098 struct buffer_head *bh, *head;
2100 bh = head = page_buffers(page);
2101 blocksize = bh->b_size;
2105 block_end = block_start + blocksize;
2106 if (block_end <= from || block_start >= to) {
2107 if (!buffer_uptodate(bh))
2110 set_buffer_uptodate(bh);
2111 mark_buffer_dirty(bh);
2113 clear_buffer_new(bh);
2115 block_start = block_end;
2116 bh = bh->b_this_page;
2117 } while (bh != head);
2120 * If this is a partial write which happened to make all buffers
2121 * uptodate then we can optimize away a bogus readpage() for
2122 * the next read(). Here we 'discover' whether the page went
2123 * uptodate as a result of this (potentially partial) write.
2126 SetPageUptodate(page);
2131 * block_write_begin takes care of the basic task of block allocation and
2132 * bringing partial write blocks uptodate first.
2134 * The filesystem needs to handle block truncation upon failure.
2136 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2137 unsigned flags, struct page **pagep, get_block_t *get_block)
2139 pgoff_t index = pos >> PAGE_SHIFT;
2143 page = grab_cache_page_write_begin(mapping, index, flags);
2147 status = __block_write_begin(page, pos, len, get_block);
2148 if (unlikely(status)) {
2157 EXPORT_SYMBOL(block_write_begin);
2159 int block_write_end(struct file *file, struct address_space *mapping,
2160 loff_t pos, unsigned len, unsigned copied,
2161 struct page *page, void *fsdata)
2163 struct inode *inode = mapping->host;
2166 start = pos & (PAGE_SIZE - 1);
2168 if (unlikely(copied < len)) {
2170 * The buffers that were written will now be uptodate, so we
2171 * don't have to worry about a readpage reading them and
2172 * overwriting a partial write. However if we have encountered
2173 * a short write and only partially written into a buffer, it
2174 * will not be marked uptodate, so a readpage might come in and
2175 * destroy our partial write.
2177 * Do the simplest thing, and just treat any short write to a
2178 * non uptodate page as a zero-length write, and force the
2179 * caller to redo the whole thing.
2181 if (!PageUptodate(page))
2184 page_zero_new_buffers(page, start+copied, start+len);
2186 flush_dcache_page(page);
2188 /* This could be a short (even 0-length) commit */
2189 __block_commit_write(inode, page, start, start+copied);
2193 EXPORT_SYMBOL(block_write_end);
2195 int generic_write_end(struct file *file, struct address_space *mapping,
2196 loff_t pos, unsigned len, unsigned copied,
2197 struct page *page, void *fsdata)
2199 struct inode *inode = mapping->host;
2200 loff_t old_size = inode->i_size;
2201 int i_size_changed = 0;
2203 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2206 * No need to use i_size_read() here, the i_size
2207 * cannot change under us because we hold i_mutex.
2209 * But it's important to update i_size while still holding page lock:
2210 * page writeout could otherwise come in and zero beyond i_size.
2212 if (pos+copied > inode->i_size) {
2213 i_size_write(inode, pos+copied);
2221 pagecache_isize_extended(inode, old_size, pos);
2223 * Don't mark the inode dirty under page lock. First, it unnecessarily
2224 * makes the holding time of page lock longer. Second, it forces lock
2225 * ordering of page lock and transaction start for journaling
2229 mark_inode_dirty(inode);
2233 EXPORT_SYMBOL(generic_write_end);
2236 * block_is_partially_uptodate checks whether buffers within a page are
2239 * Returns true if all buffers which correspond to a file portion
2240 * we want to read are uptodate.
2242 int block_is_partially_uptodate(struct page *page, unsigned long from,
2243 unsigned long count)
2245 unsigned block_start, block_end, blocksize;
2247 struct buffer_head *bh, *head;
2250 if (!page_has_buffers(page))
2253 head = page_buffers(page);
2254 blocksize = head->b_size;
2255 to = min_t(unsigned, PAGE_SIZE - from, count);
2257 if (from < blocksize && to > PAGE_SIZE - blocksize)
2263 block_end = block_start + blocksize;
2264 if (block_end > from && block_start < to) {
2265 if (!buffer_uptodate(bh)) {
2269 if (block_end >= to)
2272 block_start = block_end;
2273 bh = bh->b_this_page;
2274 } while (bh != head);
2278 EXPORT_SYMBOL(block_is_partially_uptodate);
2281 * Generic "read page" function for block devices that have the normal
2282 * get_block functionality. This is most of the block device filesystems.
2283 * Reads the page asynchronously --- the unlock_buffer() and
2284 * set/clear_buffer_uptodate() functions propagate buffer state into the
2285 * page struct once IO has completed.
2287 int block_read_full_page(struct page *page, get_block_t *get_block)
2289 struct inode *inode = page->mapping->host;
2290 sector_t iblock, lblock;
2291 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2292 unsigned int blocksize, bbits;
2294 int fully_mapped = 1;
2296 head = create_page_buffers(page, inode, 0);
2297 blocksize = head->b_size;
2298 bbits = block_size_bits(blocksize);
2300 iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2301 lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2307 if (buffer_uptodate(bh))
2310 if (!buffer_mapped(bh)) {
2314 if (iblock < lblock) {
2315 WARN_ON(bh->b_size != blocksize);
2316 err = get_block(inode, iblock, bh, 0);
2320 if (!buffer_mapped(bh)) {
2321 zero_user(page, i * blocksize, blocksize);
2323 set_buffer_uptodate(bh);
2327 * get_block() might have updated the buffer
2330 if (buffer_uptodate(bh))
2334 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2337 SetPageMappedToDisk(page);
2341 * All buffers are uptodate - we can set the page uptodate
2342 * as well. But not if get_block() returned an error.
2344 if (!PageError(page))
2345 SetPageUptodate(page);
2350 /* Stage two: lock the buffers */
2351 for (i = 0; i < nr; i++) {
2354 mark_buffer_async_read(bh);
2358 * Stage 3: start the IO. Check for uptodateness
2359 * inside the buffer lock in case another process reading
2360 * the underlying blockdev brought it uptodate (the sct fix).
2362 for (i = 0; i < nr; i++) {
2364 if (buffer_uptodate(bh))
2365 end_buffer_async_read(bh, 1);
2367 submit_bh(REQ_OP_READ, 0, bh);
2371 EXPORT_SYMBOL(block_read_full_page);
2373 /* utility function for filesystems that need to do work on expanding
2374 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2375 * deal with the hole.
2377 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2379 struct address_space *mapping = inode->i_mapping;
2384 err = inode_newsize_ok(inode, size);
2388 err = pagecache_write_begin(NULL, mapping, size, 0,
2389 AOP_FLAG_CONT_EXPAND, &page, &fsdata);
2393 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2399 EXPORT_SYMBOL(generic_cont_expand_simple);
2401 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2402 loff_t pos, loff_t *bytes)
2404 struct inode *inode = mapping->host;
2405 unsigned int blocksize = i_blocksize(inode);
2408 pgoff_t index, curidx;
2410 unsigned zerofrom, offset, len;
2413 index = pos >> PAGE_SHIFT;
2414 offset = pos & ~PAGE_MASK;
2416 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2417 zerofrom = curpos & ~PAGE_MASK;
2418 if (zerofrom & (blocksize-1)) {
2419 *bytes |= (blocksize-1);
2422 len = PAGE_SIZE - zerofrom;
2424 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2428 zero_user(page, zerofrom, len);
2429 err = pagecache_write_end(file, mapping, curpos, len, len,
2436 balance_dirty_pages_ratelimited(mapping);
2438 if (unlikely(fatal_signal_pending(current))) {
2444 /* page covers the boundary, find the boundary offset */
2445 if (index == curidx) {
2446 zerofrom = curpos & ~PAGE_MASK;
2447 /* if we will expand the thing last block will be filled */
2448 if (offset <= zerofrom) {
2451 if (zerofrom & (blocksize-1)) {
2452 *bytes |= (blocksize-1);
2455 len = offset - zerofrom;
2457 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2461 zero_user(page, zerofrom, len);
2462 err = pagecache_write_end(file, mapping, curpos, len, len,
2474 * For moronic filesystems that do not allow holes in file.
2475 * We may have to extend the file.
2477 int cont_write_begin(struct file *file, struct address_space *mapping,
2478 loff_t pos, unsigned len, unsigned flags,
2479 struct page **pagep, void **fsdata,
2480 get_block_t *get_block, loff_t *bytes)
2482 struct inode *inode = mapping->host;
2483 unsigned int blocksize = i_blocksize(inode);
2484 unsigned int zerofrom;
2487 err = cont_expand_zero(file, mapping, pos, bytes);
2491 zerofrom = *bytes & ~PAGE_MASK;
2492 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2493 *bytes |= (blocksize-1);
2497 return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2499 EXPORT_SYMBOL(cont_write_begin);
2501 int block_commit_write(struct page *page, unsigned from, unsigned to)
2503 struct inode *inode = page->mapping->host;
2504 __block_commit_write(inode,page,from,to);
2507 EXPORT_SYMBOL(block_commit_write);
2510 * block_page_mkwrite() is not allowed to change the file size as it gets
2511 * called from a page fault handler when a page is first dirtied. Hence we must
2512 * be careful to check for EOF conditions here. We set the page up correctly
2513 * for a written page which means we get ENOSPC checking when writing into
2514 * holes and correct delalloc and unwritten extent mapping on filesystems that
2515 * support these features.
2517 * We are not allowed to take the i_mutex here so we have to play games to
2518 * protect against truncate races as the page could now be beyond EOF. Because
2519 * truncate writes the inode size before removing pages, once we have the
2520 * page lock we can determine safely if the page is beyond EOF. If it is not
2521 * beyond EOF, then the page is guaranteed safe against truncation until we
2524 * Direct callers of this function should protect against filesystem freezing
2525 * using sb_start_pagefault() - sb_end_pagefault() functions.
2527 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2528 get_block_t get_block)
2530 struct page *page = vmf->page;
2531 struct inode *inode = file_inode(vma->vm_file);
2537 size = i_size_read(inode);
2538 if ((page->mapping != inode->i_mapping) ||
2539 (page_offset(page) > size)) {
2540 /* We overload EFAULT to mean page got truncated */
2545 /* page is wholly or partially inside EOF */
2546 if (((page->index + 1) << PAGE_SHIFT) > size)
2547 end = size & ~PAGE_MASK;
2551 ret = __block_write_begin(page, 0, end, get_block);
2553 ret = block_commit_write(page, 0, end);
2555 if (unlikely(ret < 0))
2557 set_page_dirty(page);
2558 wait_for_stable_page(page);
2564 EXPORT_SYMBOL(block_page_mkwrite);
2567 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2568 * immediately, while under the page lock. So it needs a special end_io
2569 * handler which does not touch the bh after unlocking it.
2571 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2573 __end_buffer_read_notouch(bh, uptodate);
2577 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2578 * the page (converting it to circular linked list and taking care of page
2581 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2583 struct buffer_head *bh;
2585 BUG_ON(!PageLocked(page));
2587 spin_lock(&page->mapping->private_lock);
2590 if (PageDirty(page))
2591 set_buffer_dirty(bh);
2592 if (!bh->b_this_page)
2593 bh->b_this_page = head;
2594 bh = bh->b_this_page;
2595 } while (bh != head);
2596 attach_page_buffers(page, head);
2597 spin_unlock(&page->mapping->private_lock);
2601 * On entry, the page is fully not uptodate.
2602 * On exit the page is fully uptodate in the areas outside (from,to)
2603 * The filesystem needs to handle block truncation upon failure.
2605 int nobh_write_begin(struct address_space *mapping,
2606 loff_t pos, unsigned len, unsigned flags,
2607 struct page **pagep, void **fsdata,
2608 get_block_t *get_block)
2610 struct inode *inode = mapping->host;
2611 const unsigned blkbits = inode->i_blkbits;
2612 const unsigned blocksize = 1 << blkbits;
2613 struct buffer_head *head, *bh;
2617 unsigned block_in_page;
2618 unsigned block_start, block_end;
2619 sector_t block_in_file;
2622 int is_mapped_to_disk = 1;
2624 index = pos >> PAGE_SHIFT;
2625 from = pos & (PAGE_SIZE - 1);
2628 page = grab_cache_page_write_begin(mapping, index, flags);
2634 if (page_has_buffers(page)) {
2635 ret = __block_write_begin(page, pos, len, get_block);
2641 if (PageMappedToDisk(page))
2645 * Allocate buffers so that we can keep track of state, and potentially
2646 * attach them to the page if an error occurs. In the common case of
2647 * no error, they will just be freed again without ever being attached
2648 * to the page (which is all OK, because we're under the page lock).
2650 * Be careful: the buffer linked list is a NULL terminated one, rather
2651 * than the circular one we're used to.
2653 head = alloc_page_buffers(page, blocksize, 0);
2659 block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2662 * We loop across all blocks in the page, whether or not they are
2663 * part of the affected region. This is so we can discover if the
2664 * page is fully mapped-to-disk.
2666 for (block_start = 0, block_in_page = 0, bh = head;
2667 block_start < PAGE_SIZE;
2668 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2671 block_end = block_start + blocksize;
2674 if (block_start >= to)
2676 ret = get_block(inode, block_in_file + block_in_page,
2680 if (!buffer_mapped(bh))
2681 is_mapped_to_disk = 0;
2683 clean_bdev_bh_alias(bh);
2684 if (PageUptodate(page)) {
2685 set_buffer_uptodate(bh);
2688 if (buffer_new(bh) || !buffer_mapped(bh)) {
2689 zero_user_segments(page, block_start, from,
2693 if (buffer_uptodate(bh))
2694 continue; /* reiserfs does this */
2695 if (block_start < from || block_end > to) {
2697 bh->b_end_io = end_buffer_read_nobh;
2698 submit_bh(REQ_OP_READ, 0, bh);
2705 * The page is locked, so these buffers are protected from
2706 * any VM or truncate activity. Hence we don't need to care
2707 * for the buffer_head refcounts.
2709 for (bh = head; bh; bh = bh->b_this_page) {
2711 if (!buffer_uptodate(bh))
2718 if (is_mapped_to_disk)
2719 SetPageMappedToDisk(page);
2721 *fsdata = head; /* to be released by nobh_write_end */
2728 * Error recovery is a bit difficult. We need to zero out blocks that
2729 * were newly allocated, and dirty them to ensure they get written out.
2730 * Buffers need to be attached to the page at this point, otherwise
2731 * the handling of potential IO errors during writeout would be hard
2732 * (could try doing synchronous writeout, but what if that fails too?)
2734 attach_nobh_buffers(page, head);
2735 page_zero_new_buffers(page, from, to);
2744 EXPORT_SYMBOL(nobh_write_begin);
2746 int nobh_write_end(struct file *file, struct address_space *mapping,
2747 loff_t pos, unsigned len, unsigned copied,
2748 struct page *page, void *fsdata)
2750 struct inode *inode = page->mapping->host;
2751 struct buffer_head *head = fsdata;
2752 struct buffer_head *bh;
2753 BUG_ON(fsdata != NULL && page_has_buffers(page));
2755 if (unlikely(copied < len) && head)
2756 attach_nobh_buffers(page, head);
2757 if (page_has_buffers(page))
2758 return generic_write_end(file, mapping, pos, len,
2759 copied, page, fsdata);
2761 SetPageUptodate(page);
2762 set_page_dirty(page);
2763 if (pos+copied > inode->i_size) {
2764 i_size_write(inode, pos+copied);
2765 mark_inode_dirty(inode);
2773 head = head->b_this_page;
2774 free_buffer_head(bh);
2779 EXPORT_SYMBOL(nobh_write_end);
2782 * nobh_writepage() - based on block_full_write_page() except
2783 * that it tries to operate without attaching bufferheads to
2786 int nobh_writepage(struct page *page, get_block_t *get_block,
2787 struct writeback_control *wbc)
2789 struct inode * const inode = page->mapping->host;
2790 loff_t i_size = i_size_read(inode);
2791 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2795 /* Is the page fully inside i_size? */
2796 if (page->index < end_index)
2799 /* Is the page fully outside i_size? (truncate in progress) */
2800 offset = i_size & (PAGE_SIZE-1);
2801 if (page->index >= end_index+1 || !offset) {
2803 * The page may have dirty, unmapped buffers. For example,
2804 * they may have been added in ext3_writepage(). Make them
2805 * freeable here, so the page does not leak.
2808 /* Not really sure about this - do we need this ? */
2809 if (page->mapping->a_ops->invalidatepage)
2810 page->mapping->a_ops->invalidatepage(page, offset);
2813 return 0; /* don't care */
2817 * The page straddles i_size. It must be zeroed out on each and every
2818 * writepage invocation because it may be mmapped. "A file is mapped
2819 * in multiples of the page size. For a file that is not a multiple of
2820 * the page size, the remaining memory is zeroed when mapped, and
2821 * writes to that region are not written out to the file."
2823 zero_user_segment(page, offset, PAGE_SIZE);
2825 ret = mpage_writepage(page, get_block, wbc);
2827 ret = __block_write_full_page(inode, page, get_block, wbc,
2828 end_buffer_async_write);
2831 EXPORT_SYMBOL(nobh_writepage);
2833 int nobh_truncate_page(struct address_space *mapping,
2834 loff_t from, get_block_t *get_block)
2836 pgoff_t index = from >> PAGE_SHIFT;
2837 unsigned offset = from & (PAGE_SIZE-1);
2840 unsigned length, pos;
2841 struct inode *inode = mapping->host;
2843 struct buffer_head map_bh;
2846 blocksize = i_blocksize(inode);
2847 length = offset & (blocksize - 1);
2849 /* Block boundary? Nothing to do */
2853 length = blocksize - length;
2854 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2856 page = grab_cache_page(mapping, index);
2861 if (page_has_buffers(page)) {
2865 return block_truncate_page(mapping, from, get_block);
2868 /* Find the buffer that contains "offset" */
2870 while (offset >= pos) {
2875 map_bh.b_size = blocksize;
2877 err = get_block(inode, iblock, &map_bh, 0);
2880 /* unmapped? It's a hole - nothing to do */
2881 if (!buffer_mapped(&map_bh))
2884 /* Ok, it's mapped. Make sure it's up-to-date */
2885 if (!PageUptodate(page)) {
2886 err = mapping->a_ops->readpage(NULL, page);
2892 if (!PageUptodate(page)) {
2896 if (page_has_buffers(page))
2899 zero_user(page, offset, length);
2900 set_page_dirty(page);
2909 EXPORT_SYMBOL(nobh_truncate_page);
2911 int block_truncate_page(struct address_space *mapping,
2912 loff_t from, get_block_t *get_block)
2914 pgoff_t index = from >> PAGE_SHIFT;
2915 unsigned offset = from & (PAGE_SIZE-1);
2918 unsigned length, pos;
2919 struct inode *inode = mapping->host;
2921 struct buffer_head *bh;
2924 blocksize = i_blocksize(inode);
2925 length = offset & (blocksize - 1);
2927 /* Block boundary? Nothing to do */
2931 length = blocksize - length;
2932 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2934 page = grab_cache_page(mapping, index);
2939 if (!page_has_buffers(page))
2940 create_empty_buffers(page, blocksize, 0);
2942 /* Find the buffer that contains "offset" */
2943 bh = page_buffers(page);
2945 while (offset >= pos) {
2946 bh = bh->b_this_page;
2952 if (!buffer_mapped(bh)) {
2953 WARN_ON(bh->b_size != blocksize);
2954 err = get_block(inode, iblock, bh, 0);
2957 /* unmapped? It's a hole - nothing to do */
2958 if (!buffer_mapped(bh))
2962 /* Ok, it's mapped. Make sure it's up-to-date */
2963 if (PageUptodate(page))
2964 set_buffer_uptodate(bh);
2966 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2968 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2970 /* Uhhuh. Read error. Complain and punt. */
2971 if (!buffer_uptodate(bh))
2975 zero_user(page, offset, length);
2976 mark_buffer_dirty(bh);
2985 EXPORT_SYMBOL(block_truncate_page);
2988 * The generic ->writepage function for buffer-backed address_spaces
2990 int block_write_full_page(struct page *page, get_block_t *get_block,
2991 struct writeback_control *wbc)
2993 struct inode * const inode = page->mapping->host;
2994 loff_t i_size = i_size_read(inode);
2995 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2998 /* Is the page fully inside i_size? */
2999 if (page->index < end_index)
3000 return __block_write_full_page(inode, page, get_block, wbc,
3001 end_buffer_async_write);
3003 /* Is the page fully outside i_size? (truncate in progress) */
3004 offset = i_size & (PAGE_SIZE-1);
3005 if (page->index >= end_index+1 || !offset) {
3007 * The page may have dirty, unmapped buffers. For example,
3008 * they may have been added in ext3_writepage(). Make them
3009 * freeable here, so the page does not leak.
3011 do_invalidatepage(page, 0, PAGE_SIZE);
3013 return 0; /* don't care */
3017 * The page straddles i_size. It must be zeroed out on each and every
3018 * writepage invocation because it may be mmapped. "A file is mapped
3019 * in multiples of the page size. For a file that is not a multiple of
3020 * the page size, the remaining memory is zeroed when mapped, and
3021 * writes to that region are not written out to the file."
3023 zero_user_segment(page, offset, PAGE_SIZE);
3024 return __block_write_full_page(inode, page, get_block, wbc,
3025 end_buffer_async_write);
3027 EXPORT_SYMBOL(block_write_full_page);
3029 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
3030 get_block_t *get_block)
3032 struct buffer_head tmp;
3033 struct inode *inode = mapping->host;
3036 tmp.b_size = i_blocksize(inode);
3037 get_block(inode, block, &tmp, 0);
3038 return tmp.b_blocknr;
3040 EXPORT_SYMBOL(generic_block_bmap);
3042 static void end_bio_bh_io_sync(struct bio *bio)
3044 struct buffer_head *bh = bio->bi_private;
3046 if (unlikely(bio_flagged(bio, BIO_QUIET)))
3047 set_bit(BH_Quiet, &bh->b_state);
3049 bh->b_end_io(bh, !bio->bi_error);
3054 * This allows us to do IO even on the odd last sectors
3055 * of a device, even if the block size is some multiple
3056 * of the physical sector size.
3058 * We'll just truncate the bio to the size of the device,
3059 * and clear the end of the buffer head manually.
3061 * Truly out-of-range accesses will turn into actual IO
3062 * errors, this only handles the "we need to be able to
3063 * do IO at the final sector" case.
3065 void guard_bio_eod(int op, struct bio *bio)
3068 struct bio_vec *bvec = &bio->bi_io_vec[bio->bi_vcnt - 1];
3069 unsigned truncated_bytes;
3071 maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
3076 * If the *whole* IO is past the end of the device,
3077 * let it through, and the IO layer will turn it into
3080 if (unlikely(bio->bi_iter.bi_sector >= maxsector))
3083 maxsector -= bio->bi_iter.bi_sector;
3084 if (likely((bio->bi_iter.bi_size >> 9) <= maxsector))
3087 /* Uhhuh. We've got a bio that straddles the device size! */
3088 truncated_bytes = bio->bi_iter.bi_size - (maxsector << 9);
3090 /* Truncate the bio.. */
3091 bio->bi_iter.bi_size -= truncated_bytes;
3092 bvec->bv_len -= truncated_bytes;
3094 /* ..and clear the end of the buffer for reads */
3095 if (op == REQ_OP_READ) {
3096 zero_user(bvec->bv_page, bvec->bv_offset + bvec->bv_len,
3101 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3102 struct writeback_control *wbc)
3106 BUG_ON(!buffer_locked(bh));
3107 BUG_ON(!buffer_mapped(bh));
3108 BUG_ON(!bh->b_end_io);
3109 BUG_ON(buffer_delay(bh));
3110 BUG_ON(buffer_unwritten(bh));
3113 * Only clear out a write error when rewriting
3115 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3116 clear_buffer_write_io_error(bh);
3119 * from here on down, it's all bio -- do the initial mapping,
3120 * submit_bio -> generic_make_request may further map this bio around
3122 bio = bio_alloc(GFP_NOIO, 1);
3125 wbc_init_bio(wbc, bio);
3126 wbc_account_io(wbc, bh->b_page, bh->b_size);
3129 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3130 bio->bi_bdev = bh->b_bdev;
3132 bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3133 BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3135 bio->bi_end_io = end_bio_bh_io_sync;
3136 bio->bi_private = bh;
3138 /* Take care of bh's that straddle the end of the device */
3139 guard_bio_eod(op, bio);
3141 if (buffer_meta(bh))
3142 op_flags |= REQ_META;
3143 if (buffer_prio(bh))
3144 op_flags |= REQ_PRIO;
3145 bio_set_op_attrs(bio, op, op_flags);
3151 int submit_bh(int op, int op_flags, struct buffer_head *bh)
3153 return submit_bh_wbc(op, op_flags, bh, NULL);
3155 EXPORT_SYMBOL(submit_bh);
3158 * ll_rw_block: low-level access to block devices (DEPRECATED)
3159 * @op: whether to %READ or %WRITE
3160 * @op_flags: req_flag_bits
3161 * @nr: number of &struct buffer_heads in the array
3162 * @bhs: array of pointers to &struct buffer_head
3164 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3165 * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3166 * @op_flags contains flags modifying the detailed I/O behavior, most notably
3169 * This function drops any buffer that it cannot get a lock on (with the
3170 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3171 * request, and any buffer that appears to be up-to-date when doing read
3172 * request. Further it marks as clean buffers that are processed for
3173 * writing (the buffer cache won't assume that they are actually clean
3174 * until the buffer gets unlocked).
3176 * ll_rw_block sets b_end_io to simple completion handler that marks
3177 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3180 * All of the buffers must be for the same device, and must also be a
3181 * multiple of the current approved size for the device.
3183 void ll_rw_block(int op, int op_flags, int nr, struct buffer_head *bhs[])
3187 for (i = 0; i < nr; i++) {
3188 struct buffer_head *bh = bhs[i];
3190 if (!trylock_buffer(bh))
3193 if (test_clear_buffer_dirty(bh)) {
3194 bh->b_end_io = end_buffer_write_sync;
3196 submit_bh(op, op_flags, bh);
3200 if (!buffer_uptodate(bh)) {
3201 bh->b_end_io = end_buffer_read_sync;
3203 submit_bh(op, op_flags, bh);
3210 EXPORT_SYMBOL(ll_rw_block);
3212 void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3215 if (!test_clear_buffer_dirty(bh)) {
3219 bh->b_end_io = end_buffer_write_sync;
3221 submit_bh(REQ_OP_WRITE, op_flags, bh);
3223 EXPORT_SYMBOL(write_dirty_buffer);
3226 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3227 * and then start new I/O and then wait upon it. The caller must have a ref on
3230 int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3234 WARN_ON(atomic_read(&bh->b_count) < 1);
3236 if (test_clear_buffer_dirty(bh)) {
3238 bh->b_end_io = end_buffer_write_sync;
3239 ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3241 if (!ret && !buffer_uptodate(bh))
3248 EXPORT_SYMBOL(__sync_dirty_buffer);
3250 int sync_dirty_buffer(struct buffer_head *bh)
3252 return __sync_dirty_buffer(bh, REQ_SYNC);
3254 EXPORT_SYMBOL(sync_dirty_buffer);
3257 * try_to_free_buffers() checks if all the buffers on this particular page
3258 * are unused, and releases them if so.
3260 * Exclusion against try_to_free_buffers may be obtained by either
3261 * locking the page or by holding its mapping's private_lock.
3263 * If the page is dirty but all the buffers are clean then we need to
3264 * be sure to mark the page clean as well. This is because the page
3265 * may be against a block device, and a later reattachment of buffers
3266 * to a dirty page will set *all* buffers dirty. Which would corrupt
3267 * filesystem data on the same device.
3269 * The same applies to regular filesystem pages: if all the buffers are
3270 * clean then we set the page clean and proceed. To do that, we require
3271 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3274 * try_to_free_buffers() is non-blocking.
3276 static inline int buffer_busy(struct buffer_head *bh)
3278 return atomic_read(&bh->b_count) |
3279 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3283 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3285 struct buffer_head *head = page_buffers(page);
3286 struct buffer_head *bh;
3290 if (buffer_busy(bh))
3292 bh = bh->b_this_page;
3293 } while (bh != head);
3296 struct buffer_head *next = bh->b_this_page;
3298 if (bh->b_assoc_map)
3299 __remove_assoc_queue(bh);
3301 } while (bh != head);
3302 *buffers_to_free = head;
3303 __clear_page_buffers(page);
3309 int try_to_free_buffers(struct page *page)
3311 struct address_space * const mapping = page->mapping;
3312 struct buffer_head *buffers_to_free = NULL;
3315 BUG_ON(!PageLocked(page));
3316 if (PageWriteback(page))
3319 if (mapping == NULL) { /* can this still happen? */
3320 ret = drop_buffers(page, &buffers_to_free);
3324 spin_lock(&mapping->private_lock);
3325 ret = drop_buffers(page, &buffers_to_free);
3328 * If the filesystem writes its buffers by hand (eg ext3)
3329 * then we can have clean buffers against a dirty page. We
3330 * clean the page here; otherwise the VM will never notice
3331 * that the filesystem did any IO at all.
3333 * Also, during truncate, discard_buffer will have marked all
3334 * the page's buffers clean. We discover that here and clean
3337 * private_lock must be held over this entire operation in order
3338 * to synchronise against __set_page_dirty_buffers and prevent the
3339 * dirty bit from being lost.
3342 cancel_dirty_page(page);
3343 spin_unlock(&mapping->private_lock);
3345 if (buffers_to_free) {
3346 struct buffer_head *bh = buffers_to_free;
3349 struct buffer_head *next = bh->b_this_page;
3350 free_buffer_head(bh);
3352 } while (bh != buffers_to_free);
3356 EXPORT_SYMBOL(try_to_free_buffers);
3359 * There are no bdflush tunables left. But distributions are
3360 * still running obsolete flush daemons, so we terminate them here.
3362 * Use of bdflush() is deprecated and will be removed in a future kernel.
3363 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3365 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3367 static int msg_count;
3369 if (!capable(CAP_SYS_ADMIN))
3372 if (msg_count < 5) {
3375 "warning: process `%s' used the obsolete bdflush"
3376 " system call\n", current->comm);
3377 printk(KERN_INFO "Fix your initscripts?\n");
3386 * Buffer-head allocation
3388 static struct kmem_cache *bh_cachep __read_mostly;
3391 * Once the number of bh's in the machine exceeds this level, we start
3392 * stripping them in writeback.
3394 static unsigned long max_buffer_heads;
3396 int buffer_heads_over_limit;
3398 struct bh_accounting {
3399 int nr; /* Number of live bh's */
3400 int ratelimit; /* Limit cacheline bouncing */
3403 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3405 static void recalc_bh_state(void)
3410 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3412 __this_cpu_write(bh_accounting.ratelimit, 0);
3413 for_each_online_cpu(i)
3414 tot += per_cpu(bh_accounting, i).nr;
3415 buffer_heads_over_limit = (tot > max_buffer_heads);
3418 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3420 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3422 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3424 __this_cpu_inc(bh_accounting.nr);
3430 EXPORT_SYMBOL(alloc_buffer_head);
3432 void free_buffer_head(struct buffer_head *bh)
3434 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3435 kmem_cache_free(bh_cachep, bh);
3437 __this_cpu_dec(bh_accounting.nr);
3441 EXPORT_SYMBOL(free_buffer_head);
3443 static int buffer_exit_cpu_dead(unsigned int cpu)
3446 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3448 for (i = 0; i < BH_LRU_SIZE; i++) {
3452 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3453 per_cpu(bh_accounting, cpu).nr = 0;
3458 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3459 * @bh: struct buffer_head
3461 * Return true if the buffer is up-to-date and false,
3462 * with the buffer locked, if not.
3464 int bh_uptodate_or_lock(struct buffer_head *bh)
3466 if (!buffer_uptodate(bh)) {
3468 if (!buffer_uptodate(bh))
3474 EXPORT_SYMBOL(bh_uptodate_or_lock);
3477 * bh_submit_read - Submit a locked buffer for reading
3478 * @bh: struct buffer_head
3480 * Returns zero on success and -EIO on error.
3482 int bh_submit_read(struct buffer_head *bh)
3484 BUG_ON(!buffer_locked(bh));
3486 if (buffer_uptodate(bh)) {
3492 bh->b_end_io = end_buffer_read_sync;
3493 submit_bh(REQ_OP_READ, 0, bh);
3495 if (buffer_uptodate(bh))
3499 EXPORT_SYMBOL(bh_submit_read);
3501 void __init buffer_init(void)
3503 unsigned long nrpages;
3506 bh_cachep = kmem_cache_create("buffer_head",
3507 sizeof(struct buffer_head), 0,
3508 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3513 * Limit the bh occupancy to 10% of ZONE_NORMAL
3515 nrpages = (nr_free_buffer_pages() * 10) / 100;
3516 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3517 ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3518 NULL, buffer_exit_cpu_dead);