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/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 <linux/sched/mm.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 enum rw_hint hint, struct writeback_control *wbc);
54 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
56 inline void touch_buffer(struct buffer_head *bh)
58 trace_block_touch_buffer(bh);
59 mark_page_accessed(bh->b_page);
61 EXPORT_SYMBOL(touch_buffer);
63 void __lock_buffer(struct buffer_head *bh)
65 wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
67 EXPORT_SYMBOL(__lock_buffer);
69 void unlock_buffer(struct buffer_head *bh)
71 clear_bit_unlock(BH_Lock, &bh->b_state);
72 smp_mb__after_atomic();
73 wake_up_bit(&bh->b_state, BH_Lock);
75 EXPORT_SYMBOL(unlock_buffer);
78 * Returns if the page has dirty or writeback buffers. If all the buffers
79 * are unlocked and clean then the PageDirty information is stale. If
80 * any of the pages are locked, it is assumed they are locked for IO.
82 void buffer_check_dirty_writeback(struct page *page,
83 bool *dirty, bool *writeback)
85 struct buffer_head *head, *bh;
89 BUG_ON(!PageLocked(page));
91 if (!page_has_buffers(page))
94 if (PageWriteback(page))
97 head = page_buffers(page);
100 if (buffer_locked(bh))
103 if (buffer_dirty(bh))
106 bh = bh->b_this_page;
107 } while (bh != head);
109 EXPORT_SYMBOL(buffer_check_dirty_writeback);
112 * Block until a buffer comes unlocked. This doesn't stop it
113 * from becoming locked again - you have to lock it yourself
114 * if you want to preserve its state.
116 void __wait_on_buffer(struct buffer_head * bh)
118 wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
120 EXPORT_SYMBOL(__wait_on_buffer);
123 __clear_page_buffers(struct page *page)
125 ClearPagePrivate(page);
126 set_page_private(page, 0);
130 static void buffer_io_error(struct buffer_head *bh, char *msg)
132 if (!test_bit(BH_Quiet, &bh->b_state))
133 printk_ratelimited(KERN_ERR
134 "Buffer I/O error on dev %pg, logical block %llu%s\n",
135 bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
139 * End-of-IO handler helper function which does not touch the bh after
141 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
142 * a race there is benign: unlock_buffer() only use the bh's address for
143 * hashing after unlocking the buffer, so it doesn't actually touch the bh
146 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
149 set_buffer_uptodate(bh);
151 /* This happens, due to failed read-ahead attempts. */
152 clear_buffer_uptodate(bh);
158 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
159 * unlock the buffer. This is what ll_rw_block uses too.
161 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
163 __end_buffer_read_notouch(bh, uptodate);
166 EXPORT_SYMBOL(end_buffer_read_sync);
168 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
171 set_buffer_uptodate(bh);
173 buffer_io_error(bh, ", lost sync page write");
174 mark_buffer_write_io_error(bh);
175 clear_buffer_uptodate(bh);
180 EXPORT_SYMBOL(end_buffer_write_sync);
183 * Various filesystems appear to want __find_get_block to be non-blocking.
184 * But it's the page lock which protects the buffers. To get around this,
185 * we get exclusion from try_to_free_buffers with the blockdev mapping's
188 * Hack idea: for the blockdev mapping, private_lock contention
189 * may be quite high. This code could TryLock the page, and if that
190 * succeeds, there is no need to take private_lock.
192 static struct buffer_head *
193 __find_get_block_slow(struct block_device *bdev, sector_t block)
195 struct inode *bd_inode = bdev->bd_inode;
196 struct address_space *bd_mapping = bd_inode->i_mapping;
197 struct buffer_head *ret = NULL;
199 struct buffer_head *bh;
200 struct buffer_head *head;
203 static DEFINE_RATELIMIT_STATE(last_warned, HZ, 1);
205 index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
206 page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
210 spin_lock(&bd_mapping->private_lock);
211 if (!page_has_buffers(page))
213 head = page_buffers(page);
216 if (!buffer_mapped(bh))
218 else if (bh->b_blocknr == block) {
223 bh = bh->b_this_page;
224 } while (bh != head);
226 /* we might be here because some of the buffers on this page are
227 * not mapped. This is due to various races between
228 * file io on the block device and getblk. It gets dealt with
229 * elsewhere, don't buffer_error if we had some unmapped buffers
231 ratelimit_set_flags(&last_warned, RATELIMIT_MSG_ON_RELEASE);
232 if (all_mapped && __ratelimit(&last_warned)) {
233 printk("__find_get_block_slow() failed. block=%llu, "
234 "b_blocknr=%llu, b_state=0x%08lx, b_size=%zu, "
235 "device %pg blocksize: %d\n",
236 (unsigned long long)block,
237 (unsigned long long)bh->b_blocknr,
238 bh->b_state, bh->b_size, bdev,
239 1 << bd_inode->i_blkbits);
242 spin_unlock(&bd_mapping->private_lock);
249 * I/O completion handler for block_read_full_page() - pages
250 * which come unlocked at the end of I/O.
252 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
255 struct buffer_head *first;
256 struct buffer_head *tmp;
258 int page_uptodate = 1;
260 BUG_ON(!buffer_async_read(bh));
264 set_buffer_uptodate(bh);
266 clear_buffer_uptodate(bh);
267 buffer_io_error(bh, ", async page read");
272 * Be _very_ careful from here on. Bad things can happen if
273 * two buffer heads end IO at almost the same time and both
274 * decide that the page is now completely done.
276 first = page_buffers(page);
277 local_irq_save(flags);
278 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
279 clear_buffer_async_read(bh);
283 if (!buffer_uptodate(tmp))
285 if (buffer_async_read(tmp)) {
286 BUG_ON(!buffer_locked(tmp));
289 tmp = tmp->b_this_page;
291 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
292 local_irq_restore(flags);
295 * If none of the buffers had errors and they are all
296 * uptodate then we can set the page uptodate.
298 if (page_uptodate && !PageError(page))
299 SetPageUptodate(page);
304 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
305 local_irq_restore(flags);
310 * Completion handler for block_write_full_page() - pages which are unlocked
311 * during I/O, and which have PageWriteback cleared upon I/O completion.
313 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
316 struct buffer_head *first;
317 struct buffer_head *tmp;
320 BUG_ON(!buffer_async_write(bh));
324 set_buffer_uptodate(bh);
326 buffer_io_error(bh, ", lost async page write");
327 mark_buffer_write_io_error(bh);
328 clear_buffer_uptodate(bh);
332 first = page_buffers(page);
333 local_irq_save(flags);
334 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
336 clear_buffer_async_write(bh);
338 tmp = bh->b_this_page;
340 if (buffer_async_write(tmp)) {
341 BUG_ON(!buffer_locked(tmp));
344 tmp = tmp->b_this_page;
346 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
347 local_irq_restore(flags);
348 end_page_writeback(page);
352 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
353 local_irq_restore(flags);
356 EXPORT_SYMBOL(end_buffer_async_write);
359 * If a page's buffers are under async readin (end_buffer_async_read
360 * completion) then there is a possibility that another thread of
361 * control could lock one of the buffers after it has completed
362 * but while some of the other buffers have not completed. This
363 * locked buffer would confuse end_buffer_async_read() into not unlocking
364 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
365 * that this buffer is not under async I/O.
367 * The page comes unlocked when it has no locked buffer_async buffers
370 * PageLocked prevents anyone starting new async I/O reads any of
373 * PageWriteback is used to prevent simultaneous writeout of the same
376 * PageLocked prevents anyone from starting writeback of a page which is
377 * under read I/O (PageWriteback is only ever set against a locked page).
379 static void mark_buffer_async_read(struct buffer_head *bh)
381 bh->b_end_io = end_buffer_async_read;
382 set_buffer_async_read(bh);
385 static void mark_buffer_async_write_endio(struct buffer_head *bh,
386 bh_end_io_t *handler)
388 bh->b_end_io = handler;
389 set_buffer_async_write(bh);
392 void mark_buffer_async_write(struct buffer_head *bh)
394 mark_buffer_async_write_endio(bh, end_buffer_async_write);
396 EXPORT_SYMBOL(mark_buffer_async_write);
400 * fs/buffer.c contains helper functions for buffer-backed address space's
401 * fsync functions. A common requirement for buffer-based filesystems is
402 * that certain data from the backing blockdev needs to be written out for
403 * a successful fsync(). For example, ext2 indirect blocks need to be
404 * written back and waited upon before fsync() returns.
406 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
407 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
408 * management of a list of dependent buffers at ->i_mapping->private_list.
410 * Locking is a little subtle: try_to_free_buffers() will remove buffers
411 * from their controlling inode's queue when they are being freed. But
412 * try_to_free_buffers() will be operating against the *blockdev* mapping
413 * at the time, not against the S_ISREG file which depends on those buffers.
414 * So the locking for private_list is via the private_lock in the address_space
415 * which backs the buffers. Which is different from the address_space
416 * against which the buffers are listed. So for a particular address_space,
417 * mapping->private_lock does *not* protect mapping->private_list! In fact,
418 * mapping->private_list will always be protected by the backing blockdev's
421 * Which introduces a requirement: all buffers on an address_space's
422 * ->private_list must be from the same address_space: the blockdev's.
424 * address_spaces which do not place buffers at ->private_list via these
425 * utility functions are free to use private_lock and private_list for
426 * whatever they want. The only requirement is that list_empty(private_list)
427 * be true at clear_inode() time.
429 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
430 * filesystems should do that. invalidate_inode_buffers() should just go
431 * BUG_ON(!list_empty).
433 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
434 * take an address_space, not an inode. And it should be called
435 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
438 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
439 * list if it is already on a list. Because if the buffer is on a list,
440 * it *must* already be on the right one. If not, the filesystem is being
441 * silly. This will save a ton of locking. But first we have to ensure
442 * that buffers are taken *off* the old inode's list when they are freed
443 * (presumably in truncate). That requires careful auditing of all
444 * filesystems (do it inside bforget()). It could also be done by bringing
449 * The buffer's backing address_space's private_lock must be held
451 static void __remove_assoc_queue(struct buffer_head *bh)
453 list_del_init(&bh->b_assoc_buffers);
454 WARN_ON(!bh->b_assoc_map);
455 bh->b_assoc_map = NULL;
458 int inode_has_buffers(struct inode *inode)
460 return !list_empty(&inode->i_data.private_list);
464 * osync is designed to support O_SYNC io. It waits synchronously for
465 * all already-submitted IO to complete, but does not queue any new
466 * writes to the disk.
468 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
469 * you dirty the buffers, and then use osync_inode_buffers to wait for
470 * completion. Any other dirty buffers which are not yet queued for
471 * write will not be flushed to disk by the osync.
473 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
475 struct buffer_head *bh;
481 list_for_each_prev(p, list) {
483 if (buffer_locked(bh)) {
487 if (!buffer_uptodate(bh))
498 void emergency_thaw_bdev(struct super_block *sb)
500 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
501 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
505 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
506 * @mapping: the mapping which wants those buffers written
508 * Starts I/O against the buffers at mapping->private_list, and waits upon
511 * Basically, this is a convenience function for fsync().
512 * @mapping is a file or directory which needs those buffers to be written for
513 * a successful fsync().
515 int sync_mapping_buffers(struct address_space *mapping)
517 struct address_space *buffer_mapping = mapping->private_data;
519 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
522 return fsync_buffers_list(&buffer_mapping->private_lock,
523 &mapping->private_list);
525 EXPORT_SYMBOL(sync_mapping_buffers);
528 * Called when we've recently written block `bblock', and it is known that
529 * `bblock' was for a buffer_boundary() buffer. This means that the block at
530 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
531 * dirty, schedule it for IO. So that indirects merge nicely with their data.
533 void write_boundary_block(struct block_device *bdev,
534 sector_t bblock, unsigned blocksize)
536 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
538 if (buffer_dirty(bh))
539 ll_rw_block(REQ_OP_WRITE, 0, 1, &bh);
544 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
546 struct address_space *mapping = inode->i_mapping;
547 struct address_space *buffer_mapping = bh->b_page->mapping;
549 mark_buffer_dirty(bh);
550 if (!mapping->private_data) {
551 mapping->private_data = buffer_mapping;
553 BUG_ON(mapping->private_data != buffer_mapping);
555 if (!bh->b_assoc_map) {
556 spin_lock(&buffer_mapping->private_lock);
557 list_move_tail(&bh->b_assoc_buffers,
558 &mapping->private_list);
559 bh->b_assoc_map = mapping;
560 spin_unlock(&buffer_mapping->private_lock);
563 EXPORT_SYMBOL(mark_buffer_dirty_inode);
566 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
569 * If warn is true, then emit a warning if the page is not uptodate and has
570 * not been truncated.
572 * The caller must hold lock_page_memcg().
574 void __set_page_dirty(struct page *page, struct address_space *mapping,
579 xa_lock_irqsave(&mapping->i_pages, flags);
580 if (page->mapping) { /* Race with truncate? */
581 WARN_ON_ONCE(warn && !PageUptodate(page));
582 account_page_dirtied(page, mapping);
583 radix_tree_tag_set(&mapping->i_pages,
584 page_index(page), PAGECACHE_TAG_DIRTY);
586 xa_unlock_irqrestore(&mapping->i_pages, flags);
588 EXPORT_SYMBOL_GPL(__set_page_dirty);
591 * Add a page to the dirty page list.
593 * It is a sad fact of life that this function is called from several places
594 * deeply under spinlocking. It may not sleep.
596 * If the page has buffers, the uptodate buffers are set dirty, to preserve
597 * dirty-state coherency between the page and the buffers. It the page does
598 * not have buffers then when they are later attached they will all be set
601 * The buffers are dirtied before the page is dirtied. There's a small race
602 * window in which a writepage caller may see the page cleanness but not the
603 * buffer dirtiness. That's fine. If this code were to set the page dirty
604 * before the buffers, a concurrent writepage caller could clear the page dirty
605 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
606 * page on the dirty page list.
608 * We use private_lock to lock against try_to_free_buffers while using the
609 * page's buffer list. Also use this to protect against clean buffers being
610 * added to the page after it was set dirty.
612 * FIXME: may need to call ->reservepage here as well. That's rather up to the
613 * address_space though.
615 int __set_page_dirty_buffers(struct page *page)
618 struct address_space *mapping = page_mapping(page);
620 if (unlikely(!mapping))
621 return !TestSetPageDirty(page);
623 spin_lock(&mapping->private_lock);
624 if (page_has_buffers(page)) {
625 struct buffer_head *head = page_buffers(page);
626 struct buffer_head *bh = head;
629 set_buffer_dirty(bh);
630 bh = bh->b_this_page;
631 } while (bh != head);
634 * Lock out page->mem_cgroup migration to keep PageDirty
635 * synchronized with per-memcg dirty page counters.
637 lock_page_memcg(page);
638 newly_dirty = !TestSetPageDirty(page);
639 spin_unlock(&mapping->private_lock);
642 __set_page_dirty(page, mapping, 1);
644 unlock_page_memcg(page);
647 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
651 EXPORT_SYMBOL(__set_page_dirty_buffers);
654 * Write out and wait upon a list of buffers.
656 * We have conflicting pressures: we want to make sure that all
657 * initially dirty buffers get waited on, but that any subsequently
658 * dirtied buffers don't. After all, we don't want fsync to last
659 * forever if somebody is actively writing to the file.
661 * Do this in two main stages: first we copy dirty buffers to a
662 * temporary inode list, queueing the writes as we go. Then we clean
663 * up, waiting for those writes to complete.
665 * During this second stage, any subsequent updates to the file may end
666 * up refiling the buffer on the original inode's dirty list again, so
667 * there is a chance we will end up with a buffer queued for write but
668 * not yet completed on that list. So, as a final cleanup we go through
669 * the osync code to catch these locked, dirty buffers without requeuing
670 * any newly dirty buffers for write.
672 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
674 struct buffer_head *bh;
675 struct list_head tmp;
676 struct address_space *mapping;
678 struct blk_plug plug;
680 INIT_LIST_HEAD(&tmp);
681 blk_start_plug(&plug);
684 while (!list_empty(list)) {
685 bh = BH_ENTRY(list->next);
686 mapping = bh->b_assoc_map;
687 __remove_assoc_queue(bh);
688 /* Avoid race with mark_buffer_dirty_inode() which does
689 * a lockless check and we rely on seeing the dirty bit */
691 if (buffer_dirty(bh) || buffer_locked(bh)) {
692 list_add(&bh->b_assoc_buffers, &tmp);
693 bh->b_assoc_map = mapping;
694 if (buffer_dirty(bh)) {
698 * Ensure any pending I/O completes so that
699 * write_dirty_buffer() actually writes the
700 * current contents - it is a noop if I/O is
701 * still in flight on potentially older
704 write_dirty_buffer(bh, REQ_SYNC);
707 * Kick off IO for the previous mapping. Note
708 * that we will not run the very last mapping,
709 * wait_on_buffer() will do that for us
710 * through sync_buffer().
719 blk_finish_plug(&plug);
722 while (!list_empty(&tmp)) {
723 bh = BH_ENTRY(tmp.prev);
725 mapping = bh->b_assoc_map;
726 __remove_assoc_queue(bh);
727 /* Avoid race with mark_buffer_dirty_inode() which does
728 * a lockless check and we rely on seeing the dirty bit */
730 if (buffer_dirty(bh)) {
731 list_add(&bh->b_assoc_buffers,
732 &mapping->private_list);
733 bh->b_assoc_map = mapping;
737 if (!buffer_uptodate(bh))
744 err2 = osync_buffers_list(lock, list);
752 * Invalidate any and all dirty buffers on a given inode. We are
753 * probably unmounting the fs, but that doesn't mean we have already
754 * done a sync(). Just drop the buffers from the inode list.
756 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
757 * assumes that all the buffers are against the blockdev. Not true
760 void invalidate_inode_buffers(struct inode *inode)
762 if (inode_has_buffers(inode)) {
763 struct address_space *mapping = &inode->i_data;
764 struct list_head *list = &mapping->private_list;
765 struct address_space *buffer_mapping = mapping->private_data;
767 spin_lock(&buffer_mapping->private_lock);
768 while (!list_empty(list))
769 __remove_assoc_queue(BH_ENTRY(list->next));
770 spin_unlock(&buffer_mapping->private_lock);
773 EXPORT_SYMBOL(invalidate_inode_buffers);
776 * Remove any clean buffers from the inode's buffer list. This is called
777 * when we're trying to free the inode itself. Those buffers can pin it.
779 * Returns true if all buffers were removed.
781 int remove_inode_buffers(struct inode *inode)
785 if (inode_has_buffers(inode)) {
786 struct address_space *mapping = &inode->i_data;
787 struct list_head *list = &mapping->private_list;
788 struct address_space *buffer_mapping = mapping->private_data;
790 spin_lock(&buffer_mapping->private_lock);
791 while (!list_empty(list)) {
792 struct buffer_head *bh = BH_ENTRY(list->next);
793 if (buffer_dirty(bh)) {
797 __remove_assoc_queue(bh);
799 spin_unlock(&buffer_mapping->private_lock);
805 * Create the appropriate buffers when given a page for data area and
806 * the size of each buffer.. Use the bh->b_this_page linked list to
807 * follow the buffers created. Return NULL if unable to create more
810 * The retry flag is used to differentiate async IO (paging, swapping)
811 * which may not fail from ordinary buffer allocations.
813 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
816 struct buffer_head *bh, *head;
817 gfp_t gfp = GFP_NOFS | __GFP_ACCOUNT;
819 struct mem_cgroup *memcg;
824 memcg = get_mem_cgroup_from_page(page);
825 memalloc_use_memcg(memcg);
829 while ((offset -= size) >= 0) {
830 bh = alloc_buffer_head(gfp);
834 bh->b_this_page = head;
840 /* Link the buffer to its page */
841 set_bh_page(bh, page, offset);
844 memalloc_unuse_memcg();
845 mem_cgroup_put(memcg);
848 * In case anything failed, we just free everything we got.
854 head = head->b_this_page;
855 free_buffer_head(bh);
861 EXPORT_SYMBOL_GPL(alloc_page_buffers);
864 link_dev_buffers(struct page *page, struct buffer_head *head)
866 struct buffer_head *bh, *tail;
871 bh = bh->b_this_page;
873 tail->b_this_page = head;
874 attach_page_buffers(page, head);
877 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
879 sector_t retval = ~((sector_t)0);
880 loff_t sz = i_size_read(bdev->bd_inode);
883 unsigned int sizebits = blksize_bits(size);
884 retval = (sz >> sizebits);
890 * Initialise the state of a blockdev page's buffers.
893 init_page_buffers(struct page *page, struct block_device *bdev,
894 sector_t block, int size)
896 struct buffer_head *head = page_buffers(page);
897 struct buffer_head *bh = head;
898 int uptodate = PageUptodate(page);
899 sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
902 if (!buffer_mapped(bh)) {
904 bh->b_private = NULL;
906 bh->b_blocknr = block;
908 set_buffer_uptodate(bh);
909 if (block < end_block)
910 set_buffer_mapped(bh);
913 bh = bh->b_this_page;
914 } while (bh != head);
917 * Caller needs to validate requested block against end of device.
923 * Create the page-cache page that contains the requested block.
925 * This is used purely for blockdev mappings.
928 grow_dev_page(struct block_device *bdev, sector_t block,
929 pgoff_t index, int size, int sizebits, gfp_t gfp)
931 struct inode *inode = bdev->bd_inode;
933 struct buffer_head *bh;
935 int ret = 0; /* Will call free_more_memory() */
938 gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
941 * XXX: __getblk_slow() can not really deal with failure and
942 * will endlessly loop on improvised global reclaim. Prefer
943 * looping in the allocator rather than here, at least that
944 * code knows what it's doing.
946 gfp_mask |= __GFP_NOFAIL;
948 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
950 BUG_ON(!PageLocked(page));
952 if (page_has_buffers(page)) {
953 bh = page_buffers(page);
954 if (bh->b_size == size) {
955 end_block = init_page_buffers(page, bdev,
956 (sector_t)index << sizebits,
960 if (!try_to_free_buffers(page))
965 * Allocate some buffers for this page
967 bh = alloc_page_buffers(page, size, true);
970 * Link the page to the buffers and initialise them. Take the
971 * lock to be atomic wrt __find_get_block(), which does not
972 * run under the page lock.
974 spin_lock(&inode->i_mapping->private_lock);
975 link_dev_buffers(page, bh);
976 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
978 spin_unlock(&inode->i_mapping->private_lock);
980 ret = (block < end_block) ? 1 : -ENXIO;
988 * Create buffers for the specified block device block's page. If
989 * that page was dirty, the buffers are set dirty also.
992 grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
1000 } while ((size << sizebits) < PAGE_SIZE);
1002 index = block >> sizebits;
1005 * Check for a block which wants to lie outside our maximum possible
1006 * pagecache index. (this comparison is done using sector_t types).
1008 if (unlikely(index != block >> sizebits)) {
1009 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1011 __func__, (unsigned long long)block,
1016 /* Create a page with the proper size buffers.. */
1017 return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1020 static struct buffer_head *
1021 __getblk_slow(struct block_device *bdev, sector_t block,
1022 unsigned size, gfp_t gfp)
1024 /* Size must be multiple of hard sectorsize */
1025 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1026 (size < 512 || size > PAGE_SIZE))) {
1027 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1029 printk(KERN_ERR "logical block size: %d\n",
1030 bdev_logical_block_size(bdev));
1037 struct buffer_head *bh;
1040 bh = __find_get_block(bdev, block, size);
1044 ret = grow_buffers(bdev, block, size, gfp);
1051 * The relationship between dirty buffers and dirty pages:
1053 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1054 * the page is tagged dirty in its radix tree.
1056 * At all times, the dirtiness of the buffers represents the dirtiness of
1057 * subsections of the page. If the page has buffers, the page dirty bit is
1058 * merely a hint about the true dirty state.
1060 * When a page is set dirty in its entirety, all its buffers are marked dirty
1061 * (if the page has buffers).
1063 * When a buffer is marked dirty, its page is dirtied, but the page's other
1066 * Also. When blockdev buffers are explicitly read with bread(), they
1067 * individually become uptodate. But their backing page remains not
1068 * uptodate - even if all of its buffers are uptodate. A subsequent
1069 * block_read_full_page() against that page will discover all the uptodate
1070 * buffers, will set the page uptodate and will perform no I/O.
1074 * mark_buffer_dirty - mark a buffer_head as needing writeout
1075 * @bh: the buffer_head to mark dirty
1077 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1078 * backing page dirty, then tag the page as dirty in its address_space's radix
1079 * tree and then attach the address_space's inode to its superblock's dirty
1082 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1083 * i_pages lock and mapping->host->i_lock.
1085 void mark_buffer_dirty(struct buffer_head *bh)
1087 WARN_ON_ONCE(!buffer_uptodate(bh));
1089 trace_block_dirty_buffer(bh);
1092 * Very *carefully* optimize the it-is-already-dirty case.
1094 * Don't let the final "is it dirty" escape to before we
1095 * perhaps modified the buffer.
1097 if (buffer_dirty(bh)) {
1099 if (buffer_dirty(bh))
1103 if (!test_set_buffer_dirty(bh)) {
1104 struct page *page = bh->b_page;
1105 struct address_space *mapping = NULL;
1107 lock_page_memcg(page);
1108 if (!TestSetPageDirty(page)) {
1109 mapping = page_mapping(page);
1111 __set_page_dirty(page, mapping, 0);
1113 unlock_page_memcg(page);
1115 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1118 EXPORT_SYMBOL(mark_buffer_dirty);
1120 void mark_buffer_write_io_error(struct buffer_head *bh)
1122 set_buffer_write_io_error(bh);
1123 /* FIXME: do we need to set this in both places? */
1124 if (bh->b_page && bh->b_page->mapping)
1125 mapping_set_error(bh->b_page->mapping, -EIO);
1126 if (bh->b_assoc_map)
1127 mapping_set_error(bh->b_assoc_map, -EIO);
1129 EXPORT_SYMBOL(mark_buffer_write_io_error);
1132 * Decrement a buffer_head's reference count. If all buffers against a page
1133 * have zero reference count, are clean and unlocked, and if the page is clean
1134 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1135 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1136 * a page but it ends up not being freed, and buffers may later be reattached).
1138 void __brelse(struct buffer_head * buf)
1140 if (atomic_read(&buf->b_count)) {
1144 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1146 EXPORT_SYMBOL(__brelse);
1149 * bforget() is like brelse(), except it discards any
1150 * potentially dirty data.
1152 void __bforget(struct buffer_head *bh)
1154 clear_buffer_dirty(bh);
1155 if (bh->b_assoc_map) {
1156 struct address_space *buffer_mapping = bh->b_page->mapping;
1158 spin_lock(&buffer_mapping->private_lock);
1159 list_del_init(&bh->b_assoc_buffers);
1160 bh->b_assoc_map = NULL;
1161 spin_unlock(&buffer_mapping->private_lock);
1165 EXPORT_SYMBOL(__bforget);
1167 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1170 if (buffer_uptodate(bh)) {
1175 bh->b_end_io = end_buffer_read_sync;
1176 submit_bh(REQ_OP_READ, 0, bh);
1178 if (buffer_uptodate(bh))
1186 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1187 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1188 * refcount elevated by one when they're in an LRU. A buffer can only appear
1189 * once in a particular CPU's LRU. A single buffer can be present in multiple
1190 * CPU's LRUs at the same time.
1192 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1193 * sb_find_get_block().
1195 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1196 * a local interrupt disable for that.
1199 #define BH_LRU_SIZE 16
1202 struct buffer_head *bhs[BH_LRU_SIZE];
1205 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1208 #define bh_lru_lock() local_irq_disable()
1209 #define bh_lru_unlock() local_irq_enable()
1211 #define bh_lru_lock() preempt_disable()
1212 #define bh_lru_unlock() preempt_enable()
1215 static inline void check_irqs_on(void)
1217 #ifdef irqs_disabled
1218 BUG_ON(irqs_disabled());
1223 * Install a buffer_head into this cpu's LRU. If not already in the LRU, it is
1224 * inserted at the front, and the buffer_head at the back if any is evicted.
1225 * Or, if already in the LRU it is moved to the front.
1227 static void bh_lru_install(struct buffer_head *bh)
1229 struct buffer_head *evictee = bh;
1236 b = this_cpu_ptr(&bh_lrus);
1237 for (i = 0; i < BH_LRU_SIZE; i++) {
1238 swap(evictee, b->bhs[i]);
1239 if (evictee == bh) {
1251 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1253 static struct buffer_head *
1254 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1256 struct buffer_head *ret = NULL;
1261 for (i = 0; i < BH_LRU_SIZE; i++) {
1262 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1264 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1265 bh->b_size == size) {
1268 __this_cpu_write(bh_lrus.bhs[i],
1269 __this_cpu_read(bh_lrus.bhs[i - 1]));
1272 __this_cpu_write(bh_lrus.bhs[0], bh);
1284 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1285 * it in the LRU and mark it as accessed. If it is not present then return
1288 struct buffer_head *
1289 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1291 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1294 /* __find_get_block_slow will mark the page accessed */
1295 bh = __find_get_block_slow(bdev, block);
1303 EXPORT_SYMBOL(__find_get_block);
1306 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1307 * which corresponds to the passed block_device, block and size. The
1308 * returned buffer has its reference count incremented.
1310 * __getblk_gfp() will lock up the machine if grow_dev_page's
1311 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1313 struct buffer_head *
1314 __getblk_gfp(struct block_device *bdev, sector_t block,
1315 unsigned size, gfp_t gfp)
1317 struct buffer_head *bh = __find_get_block(bdev, block, size);
1321 bh = __getblk_slow(bdev, block, size, gfp);
1324 EXPORT_SYMBOL(__getblk_gfp);
1327 * Do async read-ahead on a buffer..
1329 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1331 struct buffer_head *bh = __getblk(bdev, block, size);
1333 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1337 EXPORT_SYMBOL(__breadahead);
1339 void __breadahead_gfp(struct block_device *bdev, sector_t block, unsigned size,
1342 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1344 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1348 EXPORT_SYMBOL(__breadahead_gfp);
1351 * __bread_gfp() - reads a specified block and returns the bh
1352 * @bdev: the block_device to read from
1353 * @block: number of block
1354 * @size: size (in bytes) to read
1355 * @gfp: page allocation flag
1357 * Reads a specified block, and returns buffer head that contains it.
1358 * The page cache can be allocated from non-movable area
1359 * not to prevent page migration if you set gfp to zero.
1360 * It returns NULL if the block was unreadable.
1362 struct buffer_head *
1363 __bread_gfp(struct block_device *bdev, sector_t block,
1364 unsigned size, gfp_t gfp)
1366 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1368 if (likely(bh) && !buffer_uptodate(bh))
1369 bh = __bread_slow(bh);
1372 EXPORT_SYMBOL(__bread_gfp);
1375 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1376 * This doesn't race because it runs in each cpu either in irq
1377 * or with preempt disabled.
1379 static void invalidate_bh_lru(void *arg)
1381 struct bh_lru *b = &get_cpu_var(bh_lrus);
1384 for (i = 0; i < BH_LRU_SIZE; i++) {
1388 put_cpu_var(bh_lrus);
1391 static bool has_bh_in_lru(int cpu, void *dummy)
1393 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1396 for (i = 0; i < BH_LRU_SIZE; i++) {
1404 void invalidate_bh_lrus(void)
1406 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1408 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1410 void set_bh_page(struct buffer_head *bh,
1411 struct page *page, unsigned long offset)
1414 BUG_ON(offset >= PAGE_SIZE);
1415 if (PageHighMem(page))
1417 * This catches illegal uses and preserves the offset:
1419 bh->b_data = (char *)(0 + offset);
1421 bh->b_data = page_address(page) + offset;
1423 EXPORT_SYMBOL(set_bh_page);
1426 * Called when truncating a buffer on a page completely.
1429 /* Bits that are cleared during an invalidate */
1430 #define BUFFER_FLAGS_DISCARD \
1431 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1432 1 << BH_Delay | 1 << BH_Unwritten)
1434 static void discard_buffer(struct buffer_head * bh)
1436 unsigned long b_state, b_state_old;
1439 clear_buffer_dirty(bh);
1441 b_state = bh->b_state;
1443 b_state_old = cmpxchg(&bh->b_state, b_state,
1444 (b_state & ~BUFFER_FLAGS_DISCARD));
1445 if (b_state_old == b_state)
1447 b_state = b_state_old;
1453 * block_invalidatepage - invalidate part or all of a buffer-backed page
1455 * @page: the page which is affected
1456 * @offset: start of the range to invalidate
1457 * @length: length of the range to invalidate
1459 * block_invalidatepage() is called when all or part of the page has become
1460 * invalidated by a truncate operation.
1462 * block_invalidatepage() does not have to release all buffers, but it must
1463 * ensure that no dirty buffer is left outside @offset and that no I/O
1464 * is underway against any of the blocks which are outside the truncation
1465 * point. Because the caller is about to free (and possibly reuse) those
1468 void block_invalidatepage(struct page *page, unsigned int offset,
1469 unsigned int length)
1471 struct buffer_head *head, *bh, *next;
1472 unsigned int curr_off = 0;
1473 unsigned int stop = length + offset;
1475 BUG_ON(!PageLocked(page));
1476 if (!page_has_buffers(page))
1480 * Check for overflow
1482 BUG_ON(stop > PAGE_SIZE || stop < length);
1484 head = page_buffers(page);
1487 unsigned int next_off = curr_off + bh->b_size;
1488 next = bh->b_this_page;
1491 * Are we still fully in range ?
1493 if (next_off > stop)
1497 * is this block fully invalidated?
1499 if (offset <= curr_off)
1501 curr_off = next_off;
1503 } while (bh != head);
1506 * We release buffers only if the entire page is being invalidated.
1507 * The get_block cached value has been unconditionally invalidated,
1508 * so real IO is not possible anymore.
1510 if (length == PAGE_SIZE)
1511 try_to_release_page(page, 0);
1515 EXPORT_SYMBOL(block_invalidatepage);
1519 * We attach and possibly dirty the buffers atomically wrt
1520 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1521 * is already excluded via the page lock.
1523 void create_empty_buffers(struct page *page,
1524 unsigned long blocksize, unsigned long b_state)
1526 struct buffer_head *bh, *head, *tail;
1528 head = alloc_page_buffers(page, blocksize, true);
1531 bh->b_state |= b_state;
1533 bh = bh->b_this_page;
1535 tail->b_this_page = head;
1537 spin_lock(&page->mapping->private_lock);
1538 if (PageUptodate(page) || PageDirty(page)) {
1541 if (PageDirty(page))
1542 set_buffer_dirty(bh);
1543 if (PageUptodate(page))
1544 set_buffer_uptodate(bh);
1545 bh = bh->b_this_page;
1546 } while (bh != head);
1548 attach_page_buffers(page, head);
1549 spin_unlock(&page->mapping->private_lock);
1551 EXPORT_SYMBOL(create_empty_buffers);
1554 * clean_bdev_aliases: clean a range of buffers in block device
1555 * @bdev: Block device to clean buffers in
1556 * @block: Start of a range of blocks to clean
1557 * @len: Number of blocks to clean
1559 * We are taking a range of blocks for data and we don't want writeback of any
1560 * buffer-cache aliases starting from return from this function and until the
1561 * moment when something will explicitly mark the buffer dirty (hopefully that
1562 * will not happen until we will free that block ;-) We don't even need to mark
1563 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1564 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1565 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1566 * would confuse anyone who might pick it with bread() afterwards...
1568 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1569 * writeout I/O going on against recently-freed buffers. We don't wait on that
1570 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1571 * need to. That happens here.
1573 void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1575 struct inode *bd_inode = bdev->bd_inode;
1576 struct address_space *bd_mapping = bd_inode->i_mapping;
1577 struct pagevec pvec;
1578 pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1581 struct buffer_head *bh;
1582 struct buffer_head *head;
1584 end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1585 pagevec_init(&pvec);
1586 while (pagevec_lookup_range(&pvec, bd_mapping, &index, end)) {
1587 count = pagevec_count(&pvec);
1588 for (i = 0; i < count; i++) {
1589 struct page *page = pvec.pages[i];
1591 if (!page_has_buffers(page))
1594 * We use page lock instead of bd_mapping->private_lock
1595 * to pin buffers here since we can afford to sleep and
1596 * it scales better than a global spinlock lock.
1599 /* Recheck when the page is locked which pins bhs */
1600 if (!page_has_buffers(page))
1602 head = page_buffers(page);
1605 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1607 if (bh->b_blocknr >= block + len)
1609 clear_buffer_dirty(bh);
1611 clear_buffer_req(bh);
1613 bh = bh->b_this_page;
1614 } while (bh != head);
1618 pagevec_release(&pvec);
1620 /* End of range already reached? */
1621 if (index > end || !index)
1625 EXPORT_SYMBOL(clean_bdev_aliases);
1628 * Size is a power-of-two in the range 512..PAGE_SIZE,
1629 * and the case we care about most is PAGE_SIZE.
1631 * So this *could* possibly be written with those
1632 * constraints in mind (relevant mostly if some
1633 * architecture has a slow bit-scan instruction)
1635 static inline int block_size_bits(unsigned int blocksize)
1637 return ilog2(blocksize);
1640 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1642 BUG_ON(!PageLocked(page));
1644 if (!page_has_buffers(page))
1645 create_empty_buffers(page, 1 << READ_ONCE(inode->i_blkbits),
1647 return page_buffers(page);
1651 * NOTE! All mapped/uptodate combinations are valid:
1653 * Mapped Uptodate Meaning
1655 * No No "unknown" - must do get_block()
1656 * No Yes "hole" - zero-filled
1657 * Yes No "allocated" - allocated on disk, not read in
1658 * Yes Yes "valid" - allocated and up-to-date in memory.
1660 * "Dirty" is valid only with the last case (mapped+uptodate).
1664 * While block_write_full_page is writing back the dirty buffers under
1665 * the page lock, whoever dirtied the buffers may decide to clean them
1666 * again at any time. We handle that by only looking at the buffer
1667 * state inside lock_buffer().
1669 * If block_write_full_page() is called for regular writeback
1670 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1671 * locked buffer. This only can happen if someone has written the buffer
1672 * directly, with submit_bh(). At the address_space level PageWriteback
1673 * prevents this contention from occurring.
1675 * If block_write_full_page() is called with wbc->sync_mode ==
1676 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1677 * causes the writes to be flagged as synchronous writes.
1679 int __block_write_full_page(struct inode *inode, struct page *page,
1680 get_block_t *get_block, struct writeback_control *wbc,
1681 bh_end_io_t *handler)
1685 sector_t last_block;
1686 struct buffer_head *bh, *head;
1687 unsigned int blocksize, bbits;
1688 int nr_underway = 0;
1689 int write_flags = wbc_to_write_flags(wbc);
1691 head = create_page_buffers(page, inode,
1692 (1 << BH_Dirty)|(1 << BH_Uptodate));
1695 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1696 * here, and the (potentially unmapped) buffers may become dirty at
1697 * any time. If a buffer becomes dirty here after we've inspected it
1698 * then we just miss that fact, and the page stays dirty.
1700 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1701 * handle that here by just cleaning them.
1705 blocksize = bh->b_size;
1706 bbits = block_size_bits(blocksize);
1708 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1709 last_block = (i_size_read(inode) - 1) >> bbits;
1712 * Get all the dirty buffers mapped to disk addresses and
1713 * handle any aliases from the underlying blockdev's mapping.
1716 if (block > last_block) {
1718 * mapped buffers outside i_size will occur, because
1719 * this page can be outside i_size when there is a
1720 * truncate in progress.
1723 * The buffer was zeroed by block_write_full_page()
1725 clear_buffer_dirty(bh);
1726 set_buffer_uptodate(bh);
1727 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1729 WARN_ON(bh->b_size != blocksize);
1730 err = get_block(inode, block, bh, 1);
1733 clear_buffer_delay(bh);
1734 if (buffer_new(bh)) {
1735 /* blockdev mappings never come here */
1736 clear_buffer_new(bh);
1737 clean_bdev_bh_alias(bh);
1740 bh = bh->b_this_page;
1742 } while (bh != head);
1745 if (!buffer_mapped(bh))
1748 * If it's a fully non-blocking write attempt and we cannot
1749 * lock the buffer then redirty the page. Note that this can
1750 * potentially cause a busy-wait loop from writeback threads
1751 * and kswapd activity, but those code paths have their own
1752 * higher-level throttling.
1754 if (wbc->sync_mode != WB_SYNC_NONE) {
1756 } else if (!trylock_buffer(bh)) {
1757 redirty_page_for_writepage(wbc, page);
1760 if (test_clear_buffer_dirty(bh)) {
1761 mark_buffer_async_write_endio(bh, handler);
1765 } while ((bh = bh->b_this_page) != head);
1768 * The page and its buffers are protected by PageWriteback(), so we can
1769 * drop the bh refcounts early.
1771 BUG_ON(PageWriteback(page));
1772 set_page_writeback(page);
1775 struct buffer_head *next = bh->b_this_page;
1776 if (buffer_async_write(bh)) {
1777 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1778 inode->i_write_hint, wbc);
1782 } while (bh != head);
1787 if (nr_underway == 0) {
1789 * The page was marked dirty, but the buffers were
1790 * clean. Someone wrote them back by hand with
1791 * ll_rw_block/submit_bh. A rare case.
1793 end_page_writeback(page);
1796 * The page and buffer_heads can be released at any time from
1804 * ENOSPC, or some other error. We may already have added some
1805 * blocks to the file, so we need to write these out to avoid
1806 * exposing stale data.
1807 * The page is currently locked and not marked for writeback
1810 /* Recovery: lock and submit the mapped buffers */
1812 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1813 !buffer_delay(bh)) {
1815 mark_buffer_async_write_endio(bh, handler);
1818 * The buffer may have been set dirty during
1819 * attachment to a dirty page.
1821 clear_buffer_dirty(bh);
1823 } while ((bh = bh->b_this_page) != head);
1825 BUG_ON(PageWriteback(page));
1826 mapping_set_error(page->mapping, err);
1827 set_page_writeback(page);
1829 struct buffer_head *next = bh->b_this_page;
1830 if (buffer_async_write(bh)) {
1831 clear_buffer_dirty(bh);
1832 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1833 inode->i_write_hint, wbc);
1837 } while (bh != head);
1841 EXPORT_SYMBOL(__block_write_full_page);
1844 * If a page has any new buffers, zero them out here, and mark them uptodate
1845 * and dirty so they'll be written out (in order to prevent uninitialised
1846 * block data from leaking). And clear the new bit.
1848 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1850 unsigned int block_start, block_end;
1851 struct buffer_head *head, *bh;
1853 BUG_ON(!PageLocked(page));
1854 if (!page_has_buffers(page))
1857 bh = head = page_buffers(page);
1860 block_end = block_start + bh->b_size;
1862 if (buffer_new(bh)) {
1863 if (block_end > from && block_start < to) {
1864 if (!PageUptodate(page)) {
1865 unsigned start, size;
1867 start = max(from, block_start);
1868 size = min(to, block_end) - start;
1870 zero_user(page, start, size);
1871 set_buffer_uptodate(bh);
1874 clear_buffer_new(bh);
1875 mark_buffer_dirty(bh);
1879 block_start = block_end;
1880 bh = bh->b_this_page;
1881 } while (bh != head);
1883 EXPORT_SYMBOL(page_zero_new_buffers);
1886 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1887 struct iomap *iomap)
1889 loff_t offset = block << inode->i_blkbits;
1891 bh->b_bdev = iomap->bdev;
1894 * Block points to offset in file we need to map, iomap contains
1895 * the offset at which the map starts. If the map ends before the
1896 * current block, then do not map the buffer and let the caller
1899 BUG_ON(offset >= iomap->offset + iomap->length);
1901 switch (iomap->type) {
1904 * If the buffer is not up to date or beyond the current EOF,
1905 * we need to mark it as new to ensure sub-block zeroing is
1906 * executed if necessary.
1908 if (!buffer_uptodate(bh) ||
1909 (offset >= i_size_read(inode)))
1912 case IOMAP_DELALLOC:
1913 if (!buffer_uptodate(bh) ||
1914 (offset >= i_size_read(inode)))
1916 set_buffer_uptodate(bh);
1917 set_buffer_mapped(bh);
1918 set_buffer_delay(bh);
1920 case IOMAP_UNWRITTEN:
1922 * For unwritten regions, we always need to ensure that regions
1923 * in the block we are not writing to are zeroed. Mark the
1924 * buffer as new to ensure this.
1927 set_buffer_unwritten(bh);
1930 if ((iomap->flags & IOMAP_F_NEW) ||
1931 offset >= i_size_read(inode))
1933 bh->b_blocknr = (iomap->addr + offset - iomap->offset) >>
1935 set_buffer_mapped(bh);
1940 int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
1941 get_block_t *get_block, struct iomap *iomap)
1943 unsigned from = pos & (PAGE_SIZE - 1);
1944 unsigned to = from + len;
1945 struct inode *inode = page->mapping->host;
1946 unsigned block_start, block_end;
1949 unsigned blocksize, bbits;
1950 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1952 BUG_ON(!PageLocked(page));
1953 BUG_ON(from > PAGE_SIZE);
1954 BUG_ON(to > PAGE_SIZE);
1957 head = create_page_buffers(page, inode, 0);
1958 blocksize = head->b_size;
1959 bbits = block_size_bits(blocksize);
1961 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1963 for(bh = head, block_start = 0; bh != head || !block_start;
1964 block++, block_start=block_end, bh = bh->b_this_page) {
1965 block_end = block_start + blocksize;
1966 if (block_end <= from || block_start >= to) {
1967 if (PageUptodate(page)) {
1968 if (!buffer_uptodate(bh))
1969 set_buffer_uptodate(bh);
1974 clear_buffer_new(bh);
1975 if (!buffer_mapped(bh)) {
1976 WARN_ON(bh->b_size != blocksize);
1978 err = get_block(inode, block, bh, 1);
1982 iomap_to_bh(inode, block, bh, iomap);
1985 if (buffer_new(bh)) {
1986 clean_bdev_bh_alias(bh);
1987 if (PageUptodate(page)) {
1988 clear_buffer_new(bh);
1989 set_buffer_uptodate(bh);
1990 mark_buffer_dirty(bh);
1993 if (block_end > to || block_start < from)
1994 zero_user_segments(page,
2000 if (PageUptodate(page)) {
2001 if (!buffer_uptodate(bh))
2002 set_buffer_uptodate(bh);
2005 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2006 !buffer_unwritten(bh) &&
2007 (block_start < from || block_end > to)) {
2008 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2013 * If we issued read requests - let them complete.
2015 while(wait_bh > wait) {
2016 wait_on_buffer(*--wait_bh);
2017 if (!buffer_uptodate(*wait_bh))
2021 page_zero_new_buffers(page, from, to);
2025 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2026 get_block_t *get_block)
2028 return __block_write_begin_int(page, pos, len, get_block, NULL);
2030 EXPORT_SYMBOL(__block_write_begin);
2032 static int __block_commit_write(struct inode *inode, struct page *page,
2033 unsigned from, unsigned to)
2035 unsigned block_start, block_end;
2038 struct buffer_head *bh, *head;
2040 bh = head = page_buffers(page);
2041 blocksize = bh->b_size;
2045 block_end = block_start + blocksize;
2046 if (block_end <= from || block_start >= to) {
2047 if (!buffer_uptodate(bh))
2050 set_buffer_uptodate(bh);
2051 mark_buffer_dirty(bh);
2053 clear_buffer_new(bh);
2055 block_start = block_end;
2056 bh = bh->b_this_page;
2057 } while (bh != head);
2060 * If this is a partial write which happened to make all buffers
2061 * uptodate then we can optimize away a bogus readpage() for
2062 * the next read(). Here we 'discover' whether the page went
2063 * uptodate as a result of this (potentially partial) write.
2066 SetPageUptodate(page);
2071 * block_write_begin takes care of the basic task of block allocation and
2072 * bringing partial write blocks uptodate first.
2074 * The filesystem needs to handle block truncation upon failure.
2076 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2077 unsigned flags, struct page **pagep, get_block_t *get_block)
2079 pgoff_t index = pos >> PAGE_SHIFT;
2083 page = grab_cache_page_write_begin(mapping, index, flags);
2087 status = __block_write_begin(page, pos, len, get_block);
2088 if (unlikely(status)) {
2097 EXPORT_SYMBOL(block_write_begin);
2099 int __generic_write_end(struct inode *inode, loff_t pos, unsigned copied,
2102 loff_t old_size = inode->i_size;
2103 bool i_size_changed = false;
2106 * No need to use i_size_read() here, the i_size cannot change under us
2107 * because we hold i_rwsem.
2109 * But it's important to update i_size while still holding page lock:
2110 * page writeout could otherwise come in and zero beyond i_size.
2112 if (pos + copied > inode->i_size) {
2113 i_size_write(inode, pos + copied);
2114 i_size_changed = true;
2121 pagecache_isize_extended(inode, old_size, pos);
2123 * Don't mark the inode dirty under page lock. First, it unnecessarily
2124 * makes the holding time of page lock longer. Second, it forces lock
2125 * ordering of page lock and transaction start for journaling
2129 mark_inode_dirty(inode);
2133 int block_write_end(struct file *file, struct address_space *mapping,
2134 loff_t pos, unsigned len, unsigned copied,
2135 struct page *page, void *fsdata)
2137 struct inode *inode = mapping->host;
2140 start = pos & (PAGE_SIZE - 1);
2142 if (unlikely(copied < len)) {
2144 * The buffers that were written will now be uptodate, so we
2145 * don't have to worry about a readpage reading them and
2146 * overwriting a partial write. However if we have encountered
2147 * a short write and only partially written into a buffer, it
2148 * will not be marked uptodate, so a readpage might come in and
2149 * destroy our partial write.
2151 * Do the simplest thing, and just treat any short write to a
2152 * non uptodate page as a zero-length write, and force the
2153 * caller to redo the whole thing.
2155 if (!PageUptodate(page))
2158 page_zero_new_buffers(page, start+copied, start+len);
2160 flush_dcache_page(page);
2162 /* This could be a short (even 0-length) commit */
2163 __block_commit_write(inode, page, start, start+copied);
2167 EXPORT_SYMBOL(block_write_end);
2169 int generic_write_end(struct file *file, struct address_space *mapping,
2170 loff_t pos, unsigned len, unsigned copied,
2171 struct page *page, void *fsdata)
2173 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2174 return __generic_write_end(mapping->host, pos, copied, page);
2176 EXPORT_SYMBOL(generic_write_end);
2179 * block_is_partially_uptodate checks whether buffers within a page are
2182 * Returns true if all buffers which correspond to a file portion
2183 * we want to read are uptodate.
2185 int block_is_partially_uptodate(struct page *page, unsigned long from,
2186 unsigned long count)
2188 unsigned block_start, block_end, blocksize;
2190 struct buffer_head *bh, *head;
2193 if (!page_has_buffers(page))
2196 head = page_buffers(page);
2197 blocksize = head->b_size;
2198 to = min_t(unsigned, PAGE_SIZE - from, count);
2200 if (from < blocksize && to > PAGE_SIZE - blocksize)
2206 block_end = block_start + blocksize;
2207 if (block_end > from && block_start < to) {
2208 if (!buffer_uptodate(bh)) {
2212 if (block_end >= to)
2215 block_start = block_end;
2216 bh = bh->b_this_page;
2217 } while (bh != head);
2221 EXPORT_SYMBOL(block_is_partially_uptodate);
2224 * Generic "read page" function for block devices that have the normal
2225 * get_block functionality. This is most of the block device filesystems.
2226 * Reads the page asynchronously --- the unlock_buffer() and
2227 * set/clear_buffer_uptodate() functions propagate buffer state into the
2228 * page struct once IO has completed.
2230 int block_read_full_page(struct page *page, get_block_t *get_block)
2232 struct inode *inode = page->mapping->host;
2233 sector_t iblock, lblock;
2234 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2235 unsigned int blocksize, bbits;
2237 int fully_mapped = 1;
2239 head = create_page_buffers(page, inode, 0);
2240 blocksize = head->b_size;
2241 bbits = block_size_bits(blocksize);
2243 iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2244 lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2250 if (buffer_uptodate(bh))
2253 if (!buffer_mapped(bh)) {
2257 if (iblock < lblock) {
2258 WARN_ON(bh->b_size != blocksize);
2259 err = get_block(inode, iblock, bh, 0);
2263 if (!buffer_mapped(bh)) {
2264 zero_user(page, i * blocksize, blocksize);
2266 set_buffer_uptodate(bh);
2270 * get_block() might have updated the buffer
2273 if (buffer_uptodate(bh))
2277 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2280 SetPageMappedToDisk(page);
2284 * All buffers are uptodate - we can set the page uptodate
2285 * as well. But not if get_block() returned an error.
2287 if (!PageError(page))
2288 SetPageUptodate(page);
2293 /* Stage two: lock the buffers */
2294 for (i = 0; i < nr; i++) {
2297 mark_buffer_async_read(bh);
2301 * Stage 3: start the IO. Check for uptodateness
2302 * inside the buffer lock in case another process reading
2303 * the underlying blockdev brought it uptodate (the sct fix).
2305 for (i = 0; i < nr; i++) {
2307 if (buffer_uptodate(bh))
2308 end_buffer_async_read(bh, 1);
2310 submit_bh(REQ_OP_READ, 0, bh);
2314 EXPORT_SYMBOL(block_read_full_page);
2316 /* utility function for filesystems that need to do work on expanding
2317 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2318 * deal with the hole.
2320 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2322 struct address_space *mapping = inode->i_mapping;
2327 err = inode_newsize_ok(inode, size);
2331 err = pagecache_write_begin(NULL, mapping, size, 0,
2332 AOP_FLAG_CONT_EXPAND, &page, &fsdata);
2336 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2342 EXPORT_SYMBOL(generic_cont_expand_simple);
2344 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2345 loff_t pos, loff_t *bytes)
2347 struct inode *inode = mapping->host;
2348 unsigned int blocksize = i_blocksize(inode);
2351 pgoff_t index, curidx;
2353 unsigned zerofrom, offset, len;
2356 index = pos >> PAGE_SHIFT;
2357 offset = pos & ~PAGE_MASK;
2359 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2360 zerofrom = curpos & ~PAGE_MASK;
2361 if (zerofrom & (blocksize-1)) {
2362 *bytes |= (blocksize-1);
2365 len = PAGE_SIZE - zerofrom;
2367 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2371 zero_user(page, zerofrom, len);
2372 err = pagecache_write_end(file, mapping, curpos, len, len,
2379 balance_dirty_pages_ratelimited(mapping);
2381 if (unlikely(fatal_signal_pending(current))) {
2387 /* page covers the boundary, find the boundary offset */
2388 if (index == curidx) {
2389 zerofrom = curpos & ~PAGE_MASK;
2390 /* if we will expand the thing last block will be filled */
2391 if (offset <= zerofrom) {
2394 if (zerofrom & (blocksize-1)) {
2395 *bytes |= (blocksize-1);
2398 len = offset - zerofrom;
2400 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2404 zero_user(page, zerofrom, len);
2405 err = pagecache_write_end(file, mapping, curpos, len, len,
2417 * For moronic filesystems that do not allow holes in file.
2418 * We may have to extend the file.
2420 int cont_write_begin(struct file *file, struct address_space *mapping,
2421 loff_t pos, unsigned len, unsigned flags,
2422 struct page **pagep, void **fsdata,
2423 get_block_t *get_block, loff_t *bytes)
2425 struct inode *inode = mapping->host;
2426 unsigned int blocksize = i_blocksize(inode);
2427 unsigned int zerofrom;
2430 err = cont_expand_zero(file, mapping, pos, bytes);
2434 zerofrom = *bytes & ~PAGE_MASK;
2435 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2436 *bytes |= (blocksize-1);
2440 return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2442 EXPORT_SYMBOL(cont_write_begin);
2444 int block_commit_write(struct page *page, unsigned from, unsigned to)
2446 struct inode *inode = page->mapping->host;
2447 __block_commit_write(inode,page,from,to);
2450 EXPORT_SYMBOL(block_commit_write);
2453 * block_page_mkwrite() is not allowed to change the file size as it gets
2454 * called from a page fault handler when a page is first dirtied. Hence we must
2455 * be careful to check for EOF conditions here. We set the page up correctly
2456 * for a written page which means we get ENOSPC checking when writing into
2457 * holes and correct delalloc and unwritten extent mapping on filesystems that
2458 * support these features.
2460 * We are not allowed to take the i_mutex here so we have to play games to
2461 * protect against truncate races as the page could now be beyond EOF. Because
2462 * truncate writes the inode size before removing pages, once we have the
2463 * page lock we can determine safely if the page is beyond EOF. If it is not
2464 * beyond EOF, then the page is guaranteed safe against truncation until we
2467 * Direct callers of this function should protect against filesystem freezing
2468 * using sb_start_pagefault() - sb_end_pagefault() functions.
2470 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2471 get_block_t get_block)
2473 struct page *page = vmf->page;
2474 struct inode *inode = file_inode(vma->vm_file);
2480 size = i_size_read(inode);
2481 if ((page->mapping != inode->i_mapping) ||
2482 (page_offset(page) > size)) {
2483 /* We overload EFAULT to mean page got truncated */
2488 /* page is wholly or partially inside EOF */
2489 if (((page->index + 1) << PAGE_SHIFT) > size)
2490 end = size & ~PAGE_MASK;
2494 ret = __block_write_begin(page, 0, end, get_block);
2496 ret = block_commit_write(page, 0, end);
2498 if (unlikely(ret < 0))
2500 set_page_dirty(page);
2501 wait_for_stable_page(page);
2507 EXPORT_SYMBOL(block_page_mkwrite);
2510 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2511 * immediately, while under the page lock. So it needs a special end_io
2512 * handler which does not touch the bh after unlocking it.
2514 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2516 __end_buffer_read_notouch(bh, uptodate);
2520 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2521 * the page (converting it to circular linked list and taking care of page
2524 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2526 struct buffer_head *bh;
2528 BUG_ON(!PageLocked(page));
2530 spin_lock(&page->mapping->private_lock);
2533 if (PageDirty(page))
2534 set_buffer_dirty(bh);
2535 if (!bh->b_this_page)
2536 bh->b_this_page = head;
2537 bh = bh->b_this_page;
2538 } while (bh != head);
2539 attach_page_buffers(page, head);
2540 spin_unlock(&page->mapping->private_lock);
2544 * On entry, the page is fully not uptodate.
2545 * On exit the page is fully uptodate in the areas outside (from,to)
2546 * The filesystem needs to handle block truncation upon failure.
2548 int nobh_write_begin(struct address_space *mapping,
2549 loff_t pos, unsigned len, unsigned flags,
2550 struct page **pagep, void **fsdata,
2551 get_block_t *get_block)
2553 struct inode *inode = mapping->host;
2554 const unsigned blkbits = inode->i_blkbits;
2555 const unsigned blocksize = 1 << blkbits;
2556 struct buffer_head *head, *bh;
2560 unsigned block_in_page;
2561 unsigned block_start, block_end;
2562 sector_t block_in_file;
2565 int is_mapped_to_disk = 1;
2567 index = pos >> PAGE_SHIFT;
2568 from = pos & (PAGE_SIZE - 1);
2571 page = grab_cache_page_write_begin(mapping, index, flags);
2577 if (page_has_buffers(page)) {
2578 ret = __block_write_begin(page, pos, len, get_block);
2584 if (PageMappedToDisk(page))
2588 * Allocate buffers so that we can keep track of state, and potentially
2589 * attach them to the page if an error occurs. In the common case of
2590 * no error, they will just be freed again without ever being attached
2591 * to the page (which is all OK, because we're under the page lock).
2593 * Be careful: the buffer linked list is a NULL terminated one, rather
2594 * than the circular one we're used to.
2596 head = alloc_page_buffers(page, blocksize, false);
2602 block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2605 * We loop across all blocks in the page, whether or not they are
2606 * part of the affected region. This is so we can discover if the
2607 * page is fully mapped-to-disk.
2609 for (block_start = 0, block_in_page = 0, bh = head;
2610 block_start < PAGE_SIZE;
2611 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2614 block_end = block_start + blocksize;
2617 if (block_start >= to)
2619 ret = get_block(inode, block_in_file + block_in_page,
2623 if (!buffer_mapped(bh))
2624 is_mapped_to_disk = 0;
2626 clean_bdev_bh_alias(bh);
2627 if (PageUptodate(page)) {
2628 set_buffer_uptodate(bh);
2631 if (buffer_new(bh) || !buffer_mapped(bh)) {
2632 zero_user_segments(page, block_start, from,
2636 if (buffer_uptodate(bh))
2637 continue; /* reiserfs does this */
2638 if (block_start < from || block_end > to) {
2640 bh->b_end_io = end_buffer_read_nobh;
2641 submit_bh(REQ_OP_READ, 0, bh);
2648 * The page is locked, so these buffers are protected from
2649 * any VM or truncate activity. Hence we don't need to care
2650 * for the buffer_head refcounts.
2652 for (bh = head; bh; bh = bh->b_this_page) {
2654 if (!buffer_uptodate(bh))
2661 if (is_mapped_to_disk)
2662 SetPageMappedToDisk(page);
2664 *fsdata = head; /* to be released by nobh_write_end */
2671 * Error recovery is a bit difficult. We need to zero out blocks that
2672 * were newly allocated, and dirty them to ensure they get written out.
2673 * Buffers need to be attached to the page at this point, otherwise
2674 * the handling of potential IO errors during writeout would be hard
2675 * (could try doing synchronous writeout, but what if that fails too?)
2677 attach_nobh_buffers(page, head);
2678 page_zero_new_buffers(page, from, to);
2687 EXPORT_SYMBOL(nobh_write_begin);
2689 int nobh_write_end(struct file *file, struct address_space *mapping,
2690 loff_t pos, unsigned len, unsigned copied,
2691 struct page *page, void *fsdata)
2693 struct inode *inode = page->mapping->host;
2694 struct buffer_head *head = fsdata;
2695 struct buffer_head *bh;
2696 BUG_ON(fsdata != NULL && page_has_buffers(page));
2698 if (unlikely(copied < len) && head)
2699 attach_nobh_buffers(page, head);
2700 if (page_has_buffers(page))
2701 return generic_write_end(file, mapping, pos, len,
2702 copied, page, fsdata);
2704 SetPageUptodate(page);
2705 set_page_dirty(page);
2706 if (pos+copied > inode->i_size) {
2707 i_size_write(inode, pos+copied);
2708 mark_inode_dirty(inode);
2716 head = head->b_this_page;
2717 free_buffer_head(bh);
2722 EXPORT_SYMBOL(nobh_write_end);
2725 * nobh_writepage() - based on block_full_write_page() except
2726 * that it tries to operate without attaching bufferheads to
2729 int nobh_writepage(struct page *page, get_block_t *get_block,
2730 struct writeback_control *wbc)
2732 struct inode * const inode = page->mapping->host;
2733 loff_t i_size = i_size_read(inode);
2734 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2738 /* Is the page fully inside i_size? */
2739 if (page->index < end_index)
2742 /* Is the page fully outside i_size? (truncate in progress) */
2743 offset = i_size & (PAGE_SIZE-1);
2744 if (page->index >= end_index+1 || !offset) {
2746 return 0; /* don't care */
2750 * The page straddles i_size. It must be zeroed out on each and every
2751 * writepage invocation because it may be mmapped. "A file is mapped
2752 * in multiples of the page size. For a file that is not a multiple of
2753 * the page size, the remaining memory is zeroed when mapped, and
2754 * writes to that region are not written out to the file."
2756 zero_user_segment(page, offset, PAGE_SIZE);
2758 ret = mpage_writepage(page, get_block, wbc);
2760 ret = __block_write_full_page(inode, page, get_block, wbc,
2761 end_buffer_async_write);
2764 EXPORT_SYMBOL(nobh_writepage);
2766 int nobh_truncate_page(struct address_space *mapping,
2767 loff_t from, get_block_t *get_block)
2769 pgoff_t index = from >> PAGE_SHIFT;
2770 unsigned offset = from & (PAGE_SIZE-1);
2773 unsigned length, pos;
2774 struct inode *inode = mapping->host;
2776 struct buffer_head map_bh;
2779 blocksize = i_blocksize(inode);
2780 length = offset & (blocksize - 1);
2782 /* Block boundary? Nothing to do */
2786 length = blocksize - length;
2787 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2789 page = grab_cache_page(mapping, index);
2794 if (page_has_buffers(page)) {
2798 return block_truncate_page(mapping, from, get_block);
2801 /* Find the buffer that contains "offset" */
2803 while (offset >= pos) {
2808 map_bh.b_size = blocksize;
2810 err = get_block(inode, iblock, &map_bh, 0);
2813 /* unmapped? It's a hole - nothing to do */
2814 if (!buffer_mapped(&map_bh))
2817 /* Ok, it's mapped. Make sure it's up-to-date */
2818 if (!PageUptodate(page)) {
2819 err = mapping->a_ops->readpage(NULL, page);
2825 if (!PageUptodate(page)) {
2829 if (page_has_buffers(page))
2832 zero_user(page, offset, length);
2833 set_page_dirty(page);
2842 EXPORT_SYMBOL(nobh_truncate_page);
2844 int block_truncate_page(struct address_space *mapping,
2845 loff_t from, get_block_t *get_block)
2847 pgoff_t index = from >> PAGE_SHIFT;
2848 unsigned offset = from & (PAGE_SIZE-1);
2851 unsigned length, pos;
2852 struct inode *inode = mapping->host;
2854 struct buffer_head *bh;
2857 blocksize = i_blocksize(inode);
2858 length = offset & (blocksize - 1);
2860 /* Block boundary? Nothing to do */
2864 length = blocksize - length;
2865 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2867 page = grab_cache_page(mapping, index);
2872 if (!page_has_buffers(page))
2873 create_empty_buffers(page, blocksize, 0);
2875 /* Find the buffer that contains "offset" */
2876 bh = page_buffers(page);
2878 while (offset >= pos) {
2879 bh = bh->b_this_page;
2885 if (!buffer_mapped(bh)) {
2886 WARN_ON(bh->b_size != blocksize);
2887 err = get_block(inode, iblock, bh, 0);
2890 /* unmapped? It's a hole - nothing to do */
2891 if (!buffer_mapped(bh))
2895 /* Ok, it's mapped. Make sure it's up-to-date */
2896 if (PageUptodate(page))
2897 set_buffer_uptodate(bh);
2899 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2901 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2903 /* Uhhuh. Read error. Complain and punt. */
2904 if (!buffer_uptodate(bh))
2908 zero_user(page, offset, length);
2909 mark_buffer_dirty(bh);
2918 EXPORT_SYMBOL(block_truncate_page);
2921 * The generic ->writepage function for buffer-backed address_spaces
2923 int block_write_full_page(struct page *page, get_block_t *get_block,
2924 struct writeback_control *wbc)
2926 struct inode * const inode = page->mapping->host;
2927 loff_t i_size = i_size_read(inode);
2928 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2931 /* Is the page fully inside i_size? */
2932 if (page->index < end_index)
2933 return __block_write_full_page(inode, page, get_block, wbc,
2934 end_buffer_async_write);
2936 /* Is the page fully outside i_size? (truncate in progress) */
2937 offset = i_size & (PAGE_SIZE-1);
2938 if (page->index >= end_index+1 || !offset) {
2940 return 0; /* don't care */
2944 * The page straddles i_size. It must be zeroed out on each and every
2945 * writepage invocation because it may be mmapped. "A file is mapped
2946 * in multiples of the page size. For a file that is not a multiple of
2947 * the page size, the remaining memory is zeroed when mapped, and
2948 * writes to that region are not written out to the file."
2950 zero_user_segment(page, offset, PAGE_SIZE);
2951 return __block_write_full_page(inode, page, get_block, wbc,
2952 end_buffer_async_write);
2954 EXPORT_SYMBOL(block_write_full_page);
2956 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2957 get_block_t *get_block)
2959 struct inode *inode = mapping->host;
2960 struct buffer_head tmp = {
2961 .b_size = i_blocksize(inode),
2964 get_block(inode, block, &tmp, 0);
2965 return tmp.b_blocknr;
2967 EXPORT_SYMBOL(generic_block_bmap);
2969 static void end_bio_bh_io_sync(struct bio *bio)
2971 struct buffer_head *bh = bio->bi_private;
2973 if (unlikely(bio_flagged(bio, BIO_QUIET)))
2974 set_bit(BH_Quiet, &bh->b_state);
2976 bh->b_end_io(bh, !bio->bi_status);
2981 * This allows us to do IO even on the odd last sectors
2982 * of a device, even if the block size is some multiple
2983 * of the physical sector size.
2985 * We'll just truncate the bio to the size of the device,
2986 * and clear the end of the buffer head manually.
2988 * Truly out-of-range accesses will turn into actual IO
2989 * errors, this only handles the "we need to be able to
2990 * do IO at the final sector" case.
2992 void guard_bio_eod(int op, struct bio *bio)
2995 struct bio_vec *bvec = bio_last_bvec_all(bio);
2996 unsigned truncated_bytes;
2997 struct hd_struct *part;
3000 part = __disk_get_part(bio->bi_disk, bio->bi_partno);
3002 maxsector = part_nr_sects_read(part);
3004 maxsector = get_capacity(bio->bi_disk);
3011 * If the *whole* IO is past the end of the device,
3012 * let it through, and the IO layer will turn it into
3015 if (unlikely(bio->bi_iter.bi_sector >= maxsector))
3018 maxsector -= bio->bi_iter.bi_sector;
3019 if (likely((bio->bi_iter.bi_size >> 9) <= maxsector))
3022 /* Uhhuh. We've got a bio that straddles the device size! */
3023 truncated_bytes = bio->bi_iter.bi_size - (maxsector << 9);
3026 * The bio contains more than one segment which spans EOD, just return
3027 * and let IO layer turn it into an EIO
3029 if (truncated_bytes > bvec->bv_len)
3032 /* Truncate the bio.. */
3033 bio->bi_iter.bi_size -= truncated_bytes;
3034 bvec->bv_len -= truncated_bytes;
3036 /* ..and clear the end of the buffer for reads */
3037 if (op == REQ_OP_READ) {
3038 zero_user(bvec->bv_page, bvec->bv_offset + bvec->bv_len,
3043 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3044 enum rw_hint write_hint, struct writeback_control *wbc)
3048 BUG_ON(!buffer_locked(bh));
3049 BUG_ON(!buffer_mapped(bh));
3050 BUG_ON(!bh->b_end_io);
3051 BUG_ON(buffer_delay(bh));
3052 BUG_ON(buffer_unwritten(bh));
3055 * Only clear out a write error when rewriting
3057 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3058 clear_buffer_write_io_error(bh);
3061 * from here on down, it's all bio -- do the initial mapping,
3062 * submit_bio -> generic_make_request may further map this bio around
3064 bio = bio_alloc(GFP_NOIO, 1);
3067 wbc_init_bio(wbc, bio);
3068 wbc_account_io(wbc, bh->b_page, bh->b_size);
3071 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3072 bio_set_dev(bio, bh->b_bdev);
3073 bio->bi_write_hint = write_hint;
3075 bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3076 BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3078 bio->bi_end_io = end_bio_bh_io_sync;
3079 bio->bi_private = bh;
3081 /* Take care of bh's that straddle the end of the device */
3082 guard_bio_eod(op, bio);
3084 if (buffer_meta(bh))
3085 op_flags |= REQ_META;
3086 if (buffer_prio(bh))
3087 op_flags |= REQ_PRIO;
3088 bio_set_op_attrs(bio, op, op_flags);
3094 int submit_bh(int op, int op_flags, struct buffer_head *bh)
3096 return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3098 EXPORT_SYMBOL(submit_bh);
3101 * ll_rw_block: low-level access to block devices (DEPRECATED)
3102 * @op: whether to %READ or %WRITE
3103 * @op_flags: req_flag_bits
3104 * @nr: number of &struct buffer_heads in the array
3105 * @bhs: array of pointers to &struct buffer_head
3107 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3108 * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3109 * @op_flags contains flags modifying the detailed I/O behavior, most notably
3112 * This function drops any buffer that it cannot get a lock on (with the
3113 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3114 * request, and any buffer that appears to be up-to-date when doing read
3115 * request. Further it marks as clean buffers that are processed for
3116 * writing (the buffer cache won't assume that they are actually clean
3117 * until the buffer gets unlocked).
3119 * ll_rw_block sets b_end_io to simple completion handler that marks
3120 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3123 * All of the buffers must be for the same device, and must also be a
3124 * multiple of the current approved size for the device.
3126 void ll_rw_block(int op, int op_flags, int nr, struct buffer_head *bhs[])
3130 for (i = 0; i < nr; i++) {
3131 struct buffer_head *bh = bhs[i];
3133 if (!trylock_buffer(bh))
3136 if (test_clear_buffer_dirty(bh)) {
3137 bh->b_end_io = end_buffer_write_sync;
3139 submit_bh(op, op_flags, bh);
3143 if (!buffer_uptodate(bh)) {
3144 bh->b_end_io = end_buffer_read_sync;
3146 submit_bh(op, op_flags, bh);
3153 EXPORT_SYMBOL(ll_rw_block);
3155 void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3158 if (!test_clear_buffer_dirty(bh)) {
3162 bh->b_end_io = end_buffer_write_sync;
3164 submit_bh(REQ_OP_WRITE, op_flags, bh);
3166 EXPORT_SYMBOL(write_dirty_buffer);
3169 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3170 * and then start new I/O and then wait upon it. The caller must have a ref on
3173 int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3177 WARN_ON(atomic_read(&bh->b_count) < 1);
3179 if (test_clear_buffer_dirty(bh)) {
3181 * The bh should be mapped, but it might not be if the
3182 * device was hot-removed. Not much we can do but fail the I/O.
3184 if (!buffer_mapped(bh)) {
3190 bh->b_end_io = end_buffer_write_sync;
3191 ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3193 if (!ret && !buffer_uptodate(bh))
3200 EXPORT_SYMBOL(__sync_dirty_buffer);
3202 int sync_dirty_buffer(struct buffer_head *bh)
3204 return __sync_dirty_buffer(bh, REQ_SYNC);
3206 EXPORT_SYMBOL(sync_dirty_buffer);
3209 * try_to_free_buffers() checks if all the buffers on this particular page
3210 * are unused, and releases them if so.
3212 * Exclusion against try_to_free_buffers may be obtained by either
3213 * locking the page or by holding its mapping's private_lock.
3215 * If the page is dirty but all the buffers are clean then we need to
3216 * be sure to mark the page clean as well. This is because the page
3217 * may be against a block device, and a later reattachment of buffers
3218 * to a dirty page will set *all* buffers dirty. Which would corrupt
3219 * filesystem data on the same device.
3221 * The same applies to regular filesystem pages: if all the buffers are
3222 * clean then we set the page clean and proceed. To do that, we require
3223 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3226 * try_to_free_buffers() is non-blocking.
3228 static inline int buffer_busy(struct buffer_head *bh)
3230 return atomic_read(&bh->b_count) |
3231 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3235 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3237 struct buffer_head *head = page_buffers(page);
3238 struct buffer_head *bh;
3242 if (buffer_busy(bh))
3244 bh = bh->b_this_page;
3245 } while (bh != head);
3248 struct buffer_head *next = bh->b_this_page;
3250 if (bh->b_assoc_map)
3251 __remove_assoc_queue(bh);
3253 } while (bh != head);
3254 *buffers_to_free = head;
3255 __clear_page_buffers(page);
3261 int try_to_free_buffers(struct page *page)
3263 struct address_space * const mapping = page->mapping;
3264 struct buffer_head *buffers_to_free = NULL;
3267 BUG_ON(!PageLocked(page));
3268 if (PageWriteback(page))
3271 if (mapping == NULL) { /* can this still happen? */
3272 ret = drop_buffers(page, &buffers_to_free);
3276 spin_lock(&mapping->private_lock);
3277 ret = drop_buffers(page, &buffers_to_free);
3280 * If the filesystem writes its buffers by hand (eg ext3)
3281 * then we can have clean buffers against a dirty page. We
3282 * clean the page here; otherwise the VM will never notice
3283 * that the filesystem did any IO at all.
3285 * Also, during truncate, discard_buffer will have marked all
3286 * the page's buffers clean. We discover that here and clean
3289 * private_lock must be held over this entire operation in order
3290 * to synchronise against __set_page_dirty_buffers and prevent the
3291 * dirty bit from being lost.
3294 cancel_dirty_page(page);
3295 spin_unlock(&mapping->private_lock);
3297 if (buffers_to_free) {
3298 struct buffer_head *bh = buffers_to_free;
3301 struct buffer_head *next = bh->b_this_page;
3302 free_buffer_head(bh);
3304 } while (bh != buffers_to_free);
3308 EXPORT_SYMBOL(try_to_free_buffers);
3311 * There are no bdflush tunables left. But distributions are
3312 * still running obsolete flush daemons, so we terminate them here.
3314 * Use of bdflush() is deprecated and will be removed in a future kernel.
3315 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3317 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3319 static int msg_count;
3321 if (!capable(CAP_SYS_ADMIN))
3324 if (msg_count < 5) {
3327 "warning: process `%s' used the obsolete bdflush"
3328 " system call\n", current->comm);
3329 printk(KERN_INFO "Fix your initscripts?\n");
3338 * Buffer-head allocation
3340 static struct kmem_cache *bh_cachep __read_mostly;
3343 * Once the number of bh's in the machine exceeds this level, we start
3344 * stripping them in writeback.
3346 static unsigned long max_buffer_heads;
3348 int buffer_heads_over_limit;
3350 struct bh_accounting {
3351 int nr; /* Number of live bh's */
3352 int ratelimit; /* Limit cacheline bouncing */
3355 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3357 static void recalc_bh_state(void)
3362 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3364 __this_cpu_write(bh_accounting.ratelimit, 0);
3365 for_each_online_cpu(i)
3366 tot += per_cpu(bh_accounting, i).nr;
3367 buffer_heads_over_limit = (tot > max_buffer_heads);
3370 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3372 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3374 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3376 __this_cpu_inc(bh_accounting.nr);
3382 EXPORT_SYMBOL(alloc_buffer_head);
3384 void free_buffer_head(struct buffer_head *bh)
3386 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3387 kmem_cache_free(bh_cachep, bh);
3389 __this_cpu_dec(bh_accounting.nr);
3393 EXPORT_SYMBOL(free_buffer_head);
3395 static int buffer_exit_cpu_dead(unsigned int cpu)
3398 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3400 for (i = 0; i < BH_LRU_SIZE; i++) {
3404 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3405 per_cpu(bh_accounting, cpu).nr = 0;
3410 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3411 * @bh: struct buffer_head
3413 * Return true if the buffer is up-to-date and false,
3414 * with the buffer locked, if not.
3416 int bh_uptodate_or_lock(struct buffer_head *bh)
3418 if (!buffer_uptodate(bh)) {
3420 if (!buffer_uptodate(bh))
3426 EXPORT_SYMBOL(bh_uptodate_or_lock);
3429 * bh_submit_read - Submit a locked buffer for reading
3430 * @bh: struct buffer_head
3432 * Returns zero on success and -EIO on error.
3434 int bh_submit_read(struct buffer_head *bh)
3436 BUG_ON(!buffer_locked(bh));
3438 if (buffer_uptodate(bh)) {
3444 bh->b_end_io = end_buffer_read_sync;
3445 submit_bh(REQ_OP_READ, 0, bh);
3447 if (buffer_uptodate(bh))
3451 EXPORT_SYMBOL(bh_submit_read);
3453 void __init buffer_init(void)
3455 unsigned long nrpages;
3458 bh_cachep = kmem_cache_create("buffer_head",
3459 sizeof(struct buffer_head), 0,
3460 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3465 * Limit the bh occupancy to 10% of ZONE_NORMAL
3467 nrpages = (nr_free_buffer_pages() * 10) / 100;
3468 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3469 ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3470 NULL, buffer_exit_cpu_dead);