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 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;
204 index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
205 page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
209 spin_lock(&bd_mapping->private_lock);
210 if (!page_has_buffers(page))
212 head = page_buffers(page);
215 if (!buffer_mapped(bh))
217 else if (bh->b_blocknr == block) {
222 bh = bh->b_this_page;
223 } while (bh != head);
225 /* we might be here because some of the buffers on this page are
226 * not mapped. This is due to various races between
227 * file io on the block device and getblk. It gets dealt with
228 * elsewhere, don't buffer_error if we had some unmapped buffers
231 printk("__find_get_block_slow() failed. "
232 "block=%llu, b_blocknr=%llu\n",
233 (unsigned long long)block,
234 (unsigned long long)bh->b_blocknr);
235 printk("b_state=0x%08lx, b_size=%zu\n",
236 bh->b_state, bh->b_size);
237 printk("device %pg blocksize: %d\n", bdev,
238 1 << bd_inode->i_blkbits);
241 spin_unlock(&bd_mapping->private_lock);
248 * I/O completion handler for block_read_full_page() - pages
249 * which come unlocked at the end of I/O.
251 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
254 struct buffer_head *first;
255 struct buffer_head *tmp;
257 int page_uptodate = 1;
259 BUG_ON(!buffer_async_read(bh));
263 set_buffer_uptodate(bh);
265 clear_buffer_uptodate(bh);
266 buffer_io_error(bh, ", async page read");
271 * Be _very_ careful from here on. Bad things can happen if
272 * two buffer heads end IO at almost the same time and both
273 * decide that the page is now completely done.
275 first = page_buffers(page);
276 local_irq_save(flags);
277 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
278 clear_buffer_async_read(bh);
282 if (!buffer_uptodate(tmp))
284 if (buffer_async_read(tmp)) {
285 BUG_ON(!buffer_locked(tmp));
288 tmp = tmp->b_this_page;
290 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
291 local_irq_restore(flags);
294 * If none of the buffers had errors and they are all
295 * uptodate then we can set the page uptodate.
297 if (page_uptodate && !PageError(page))
298 SetPageUptodate(page);
303 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
304 local_irq_restore(flags);
309 * Completion handler for block_write_full_page() - pages which are unlocked
310 * during I/O, and which have PageWriteback cleared upon I/O completion.
312 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
315 struct buffer_head *first;
316 struct buffer_head *tmp;
319 BUG_ON(!buffer_async_write(bh));
323 set_buffer_uptodate(bh);
325 buffer_io_error(bh, ", lost async page write");
326 mark_buffer_write_io_error(bh);
327 clear_buffer_uptodate(bh);
331 first = page_buffers(page);
332 local_irq_save(flags);
333 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
335 clear_buffer_async_write(bh);
337 tmp = bh->b_this_page;
339 if (buffer_async_write(tmp)) {
340 BUG_ON(!buffer_locked(tmp));
343 tmp = tmp->b_this_page;
345 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
346 local_irq_restore(flags);
347 end_page_writeback(page);
351 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
352 local_irq_restore(flags);
355 EXPORT_SYMBOL(end_buffer_async_write);
358 * If a page's buffers are under async readin (end_buffer_async_read
359 * completion) then there is a possibility that another thread of
360 * control could lock one of the buffers after it has completed
361 * but while some of the other buffers have not completed. This
362 * locked buffer would confuse end_buffer_async_read() into not unlocking
363 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
364 * that this buffer is not under async I/O.
366 * The page comes unlocked when it has no locked buffer_async buffers
369 * PageLocked prevents anyone starting new async I/O reads any of
372 * PageWriteback is used to prevent simultaneous writeout of the same
375 * PageLocked prevents anyone from starting writeback of a page which is
376 * under read I/O (PageWriteback is only ever set against a locked page).
378 static void mark_buffer_async_read(struct buffer_head *bh)
380 bh->b_end_io = end_buffer_async_read;
381 set_buffer_async_read(bh);
384 static void mark_buffer_async_write_endio(struct buffer_head *bh,
385 bh_end_io_t *handler)
387 bh->b_end_io = handler;
388 set_buffer_async_write(bh);
391 void mark_buffer_async_write(struct buffer_head *bh)
393 mark_buffer_async_write_endio(bh, end_buffer_async_write);
395 EXPORT_SYMBOL(mark_buffer_async_write);
399 * fs/buffer.c contains helper functions for buffer-backed address space's
400 * fsync functions. A common requirement for buffer-based filesystems is
401 * that certain data from the backing blockdev needs to be written out for
402 * a successful fsync(). For example, ext2 indirect blocks need to be
403 * written back and waited upon before fsync() returns.
405 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
406 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
407 * management of a list of dependent buffers at ->i_mapping->private_list.
409 * Locking is a little subtle: try_to_free_buffers() will remove buffers
410 * from their controlling inode's queue when they are being freed. But
411 * try_to_free_buffers() will be operating against the *blockdev* mapping
412 * at the time, not against the S_ISREG file which depends on those buffers.
413 * So the locking for private_list is via the private_lock in the address_space
414 * which backs the buffers. Which is different from the address_space
415 * against which the buffers are listed. So for a particular address_space,
416 * mapping->private_lock does *not* protect mapping->private_list! In fact,
417 * mapping->private_list will always be protected by the backing blockdev's
420 * Which introduces a requirement: all buffers on an address_space's
421 * ->private_list must be from the same address_space: the blockdev's.
423 * address_spaces which do not place buffers at ->private_list via these
424 * utility functions are free to use private_lock and private_list for
425 * whatever they want. The only requirement is that list_empty(private_list)
426 * be true at clear_inode() time.
428 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
429 * filesystems should do that. invalidate_inode_buffers() should just go
430 * BUG_ON(!list_empty).
432 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
433 * take an address_space, not an inode. And it should be called
434 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
437 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
438 * list if it is already on a list. Because if the buffer is on a list,
439 * it *must* already be on the right one. If not, the filesystem is being
440 * silly. This will save a ton of locking. But first we have to ensure
441 * that buffers are taken *off* the old inode's list when they are freed
442 * (presumably in truncate). That requires careful auditing of all
443 * filesystems (do it inside bforget()). It could also be done by bringing
448 * The buffer's backing address_space's private_lock must be held
450 static void __remove_assoc_queue(struct buffer_head *bh)
452 list_del_init(&bh->b_assoc_buffers);
453 WARN_ON(!bh->b_assoc_map);
454 bh->b_assoc_map = NULL;
457 int inode_has_buffers(struct inode *inode)
459 return !list_empty(&inode->i_data.private_list);
463 * osync is designed to support O_SYNC io. It waits synchronously for
464 * all already-submitted IO to complete, but does not queue any new
465 * writes to the disk.
467 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
468 * you dirty the buffers, and then use osync_inode_buffers to wait for
469 * completion. Any other dirty buffers which are not yet queued for
470 * write will not be flushed to disk by the osync.
472 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
474 struct buffer_head *bh;
480 list_for_each_prev(p, list) {
482 if (buffer_locked(bh)) {
486 if (!buffer_uptodate(bh))
497 static void do_thaw_one(struct super_block *sb, void *unused)
499 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
500 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
503 static void do_thaw_all(struct work_struct *work)
505 iterate_supers(do_thaw_one, NULL);
507 printk(KERN_WARNING "Emergency Thaw complete\n");
511 * emergency_thaw_all -- forcibly thaw every frozen filesystem
513 * Used for emergency unfreeze of all filesystems via SysRq
515 void emergency_thaw_all(void)
517 struct work_struct *work;
519 work = kmalloc(sizeof(*work), GFP_ATOMIC);
521 INIT_WORK(work, do_thaw_all);
527 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
528 * @mapping: the mapping which wants those buffers written
530 * Starts I/O against the buffers at mapping->private_list, and waits upon
533 * Basically, this is a convenience function for fsync().
534 * @mapping is a file or directory which needs those buffers to be written for
535 * a successful fsync().
537 int sync_mapping_buffers(struct address_space *mapping)
539 struct address_space *buffer_mapping = mapping->private_data;
541 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
544 return fsync_buffers_list(&buffer_mapping->private_lock,
545 &mapping->private_list);
547 EXPORT_SYMBOL(sync_mapping_buffers);
550 * Called when we've recently written block `bblock', and it is known that
551 * `bblock' was for a buffer_boundary() buffer. This means that the block at
552 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
553 * dirty, schedule it for IO. So that indirects merge nicely with their data.
555 void write_boundary_block(struct block_device *bdev,
556 sector_t bblock, unsigned blocksize)
558 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
560 if (buffer_dirty(bh))
561 ll_rw_block(REQ_OP_WRITE, 0, 1, &bh);
566 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
568 struct address_space *mapping = inode->i_mapping;
569 struct address_space *buffer_mapping = bh->b_page->mapping;
571 mark_buffer_dirty(bh);
572 if (!mapping->private_data) {
573 mapping->private_data = buffer_mapping;
575 BUG_ON(mapping->private_data != buffer_mapping);
577 if (!bh->b_assoc_map) {
578 spin_lock(&buffer_mapping->private_lock);
579 list_move_tail(&bh->b_assoc_buffers,
580 &mapping->private_list);
581 bh->b_assoc_map = mapping;
582 spin_unlock(&buffer_mapping->private_lock);
585 EXPORT_SYMBOL(mark_buffer_dirty_inode);
588 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
591 * If warn is true, then emit a warning if the page is not uptodate and has
592 * not been truncated.
594 * The caller must hold lock_page_memcg().
596 void __set_page_dirty(struct page *page, struct address_space *mapping,
601 xa_lock_irqsave(&mapping->i_pages, flags);
602 if (page->mapping) { /* Race with truncate? */
603 WARN_ON_ONCE(warn && !PageUptodate(page));
604 account_page_dirtied(page, mapping);
605 radix_tree_tag_set(&mapping->i_pages,
606 page_index(page), PAGECACHE_TAG_DIRTY);
608 xa_unlock_irqrestore(&mapping->i_pages, flags);
610 EXPORT_SYMBOL_GPL(__set_page_dirty);
613 * Add a page to the dirty page list.
615 * It is a sad fact of life that this function is called from several places
616 * deeply under spinlocking. It may not sleep.
618 * If the page has buffers, the uptodate buffers are set dirty, to preserve
619 * dirty-state coherency between the page and the buffers. It the page does
620 * not have buffers then when they are later attached they will all be set
623 * The buffers are dirtied before the page is dirtied. There's a small race
624 * window in which a writepage caller may see the page cleanness but not the
625 * buffer dirtiness. That's fine. If this code were to set the page dirty
626 * before the buffers, a concurrent writepage caller could clear the page dirty
627 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
628 * page on the dirty page list.
630 * We use private_lock to lock against try_to_free_buffers while using the
631 * page's buffer list. Also use this to protect against clean buffers being
632 * added to the page after it was set dirty.
634 * FIXME: may need to call ->reservepage here as well. That's rather up to the
635 * address_space though.
637 int __set_page_dirty_buffers(struct page *page)
640 struct address_space *mapping = page_mapping(page);
642 if (unlikely(!mapping))
643 return !TestSetPageDirty(page);
645 spin_lock(&mapping->private_lock);
646 if (page_has_buffers(page)) {
647 struct buffer_head *head = page_buffers(page);
648 struct buffer_head *bh = head;
651 set_buffer_dirty(bh);
652 bh = bh->b_this_page;
653 } while (bh != head);
656 * Lock out page->mem_cgroup migration to keep PageDirty
657 * synchronized with per-memcg dirty page counters.
659 lock_page_memcg(page);
660 newly_dirty = !TestSetPageDirty(page);
661 spin_unlock(&mapping->private_lock);
664 __set_page_dirty(page, mapping, 1);
666 unlock_page_memcg(page);
669 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
673 EXPORT_SYMBOL(__set_page_dirty_buffers);
676 * Write out and wait upon a list of buffers.
678 * We have conflicting pressures: we want to make sure that all
679 * initially dirty buffers get waited on, but that any subsequently
680 * dirtied buffers don't. After all, we don't want fsync to last
681 * forever if somebody is actively writing to the file.
683 * Do this in two main stages: first we copy dirty buffers to a
684 * temporary inode list, queueing the writes as we go. Then we clean
685 * up, waiting for those writes to complete.
687 * During this second stage, any subsequent updates to the file may end
688 * up refiling the buffer on the original inode's dirty list again, so
689 * there is a chance we will end up with a buffer queued for write but
690 * not yet completed on that list. So, as a final cleanup we go through
691 * the osync code to catch these locked, dirty buffers without requeuing
692 * any newly dirty buffers for write.
694 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
696 struct buffer_head *bh;
697 struct list_head tmp;
698 struct address_space *mapping;
700 struct blk_plug plug;
702 INIT_LIST_HEAD(&tmp);
703 blk_start_plug(&plug);
706 while (!list_empty(list)) {
707 bh = BH_ENTRY(list->next);
708 mapping = bh->b_assoc_map;
709 __remove_assoc_queue(bh);
710 /* Avoid race with mark_buffer_dirty_inode() which does
711 * a lockless check and we rely on seeing the dirty bit */
713 if (buffer_dirty(bh) || buffer_locked(bh)) {
714 list_add(&bh->b_assoc_buffers, &tmp);
715 bh->b_assoc_map = mapping;
716 if (buffer_dirty(bh)) {
720 * Ensure any pending I/O completes so that
721 * write_dirty_buffer() actually writes the
722 * current contents - it is a noop if I/O is
723 * still in flight on potentially older
726 write_dirty_buffer(bh, REQ_SYNC);
729 * Kick off IO for the previous mapping. Note
730 * that we will not run the very last mapping,
731 * wait_on_buffer() will do that for us
732 * through sync_buffer().
741 blk_finish_plug(&plug);
744 while (!list_empty(&tmp)) {
745 bh = BH_ENTRY(tmp.prev);
747 mapping = bh->b_assoc_map;
748 __remove_assoc_queue(bh);
749 /* Avoid race with mark_buffer_dirty_inode() which does
750 * a lockless check and we rely on seeing the dirty bit */
752 if (buffer_dirty(bh)) {
753 list_add(&bh->b_assoc_buffers,
754 &mapping->private_list);
755 bh->b_assoc_map = mapping;
759 if (!buffer_uptodate(bh))
766 err2 = osync_buffers_list(lock, list);
774 * Invalidate any and all dirty buffers on a given inode. We are
775 * probably unmounting the fs, but that doesn't mean we have already
776 * done a sync(). Just drop the buffers from the inode list.
778 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
779 * assumes that all the buffers are against the blockdev. Not true
782 void invalidate_inode_buffers(struct inode *inode)
784 if (inode_has_buffers(inode)) {
785 struct address_space *mapping = &inode->i_data;
786 struct list_head *list = &mapping->private_list;
787 struct address_space *buffer_mapping = mapping->private_data;
789 spin_lock(&buffer_mapping->private_lock);
790 while (!list_empty(list))
791 __remove_assoc_queue(BH_ENTRY(list->next));
792 spin_unlock(&buffer_mapping->private_lock);
795 EXPORT_SYMBOL(invalidate_inode_buffers);
798 * Remove any clean buffers from the inode's buffer list. This is called
799 * when we're trying to free the inode itself. Those buffers can pin it.
801 * Returns true if all buffers were removed.
803 int remove_inode_buffers(struct inode *inode)
807 if (inode_has_buffers(inode)) {
808 struct address_space *mapping = &inode->i_data;
809 struct list_head *list = &mapping->private_list;
810 struct address_space *buffer_mapping = mapping->private_data;
812 spin_lock(&buffer_mapping->private_lock);
813 while (!list_empty(list)) {
814 struct buffer_head *bh = BH_ENTRY(list->next);
815 if (buffer_dirty(bh)) {
819 __remove_assoc_queue(bh);
821 spin_unlock(&buffer_mapping->private_lock);
827 * Create the appropriate buffers when given a page for data area and
828 * the size of each buffer.. Use the bh->b_this_page linked list to
829 * follow the buffers created. Return NULL if unable to create more
832 * The retry flag is used to differentiate async IO (paging, swapping)
833 * which may not fail from ordinary buffer allocations.
835 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
838 struct buffer_head *bh, *head;
839 gfp_t gfp = GFP_NOFS;
847 while ((offset -= size) >= 0) {
848 bh = alloc_buffer_head(gfp);
852 bh->b_this_page = head;
858 /* Link the buffer to its page */
859 set_bh_page(bh, page, offset);
863 * In case anything failed, we just free everything we got.
869 head = head->b_this_page;
870 free_buffer_head(bh);
876 EXPORT_SYMBOL_GPL(alloc_page_buffers);
879 link_dev_buffers(struct page *page, struct buffer_head *head)
881 struct buffer_head *bh, *tail;
886 bh = bh->b_this_page;
888 tail->b_this_page = head;
889 attach_page_buffers(page, head);
892 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
894 sector_t retval = ~((sector_t)0);
895 loff_t sz = i_size_read(bdev->bd_inode);
898 unsigned int sizebits = blksize_bits(size);
899 retval = (sz >> sizebits);
905 * Initialise the state of a blockdev page's buffers.
908 init_page_buffers(struct page *page, struct block_device *bdev,
909 sector_t block, int size)
911 struct buffer_head *head = page_buffers(page);
912 struct buffer_head *bh = head;
913 int uptodate = PageUptodate(page);
914 sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
917 if (!buffer_mapped(bh)) {
919 bh->b_private = NULL;
921 bh->b_blocknr = block;
923 set_buffer_uptodate(bh);
924 if (block < end_block)
925 set_buffer_mapped(bh);
928 bh = bh->b_this_page;
929 } while (bh != head);
932 * Caller needs to validate requested block against end of device.
938 * Create the page-cache page that contains the requested block.
940 * This is used purely for blockdev mappings.
943 grow_dev_page(struct block_device *bdev, sector_t block,
944 pgoff_t index, int size, int sizebits, gfp_t gfp)
946 struct inode *inode = bdev->bd_inode;
948 struct buffer_head *bh;
950 int ret = 0; /* Will call free_more_memory() */
953 gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
956 * XXX: __getblk_slow() can not really deal with failure and
957 * will endlessly loop on improvised global reclaim. Prefer
958 * looping in the allocator rather than here, at least that
959 * code knows what it's doing.
961 gfp_mask |= __GFP_NOFAIL;
963 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
965 BUG_ON(!PageLocked(page));
967 if (page_has_buffers(page)) {
968 bh = page_buffers(page);
969 if (bh->b_size == size) {
970 end_block = init_page_buffers(page, bdev,
971 (sector_t)index << sizebits,
975 if (!try_to_free_buffers(page))
980 * Allocate some buffers for this page
982 bh = alloc_page_buffers(page, size, true);
985 * Link the page to the buffers and initialise them. Take the
986 * lock to be atomic wrt __find_get_block(), which does not
987 * run under the page lock.
989 spin_lock(&inode->i_mapping->private_lock);
990 link_dev_buffers(page, bh);
991 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
993 spin_unlock(&inode->i_mapping->private_lock);
995 ret = (block < end_block) ? 1 : -ENXIO;
1003 * Create buffers for the specified block device block's page. If
1004 * that page was dirty, the buffers are set dirty also.
1007 grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
1015 } while ((size << sizebits) < PAGE_SIZE);
1017 index = block >> sizebits;
1020 * Check for a block which wants to lie outside our maximum possible
1021 * pagecache index. (this comparison is done using sector_t types).
1023 if (unlikely(index != block >> sizebits)) {
1024 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1026 __func__, (unsigned long long)block,
1031 /* Create a page with the proper size buffers.. */
1032 return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1035 static struct buffer_head *
1036 __getblk_slow(struct block_device *bdev, sector_t block,
1037 unsigned size, gfp_t gfp)
1039 /* Size must be multiple of hard sectorsize */
1040 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1041 (size < 512 || size > PAGE_SIZE))) {
1042 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1044 printk(KERN_ERR "logical block size: %d\n",
1045 bdev_logical_block_size(bdev));
1052 struct buffer_head *bh;
1055 bh = __find_get_block(bdev, block, size);
1059 ret = grow_buffers(bdev, block, size, gfp);
1066 * The relationship between dirty buffers and dirty pages:
1068 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1069 * the page is tagged dirty in its radix tree.
1071 * At all times, the dirtiness of the buffers represents the dirtiness of
1072 * subsections of the page. If the page has buffers, the page dirty bit is
1073 * merely a hint about the true dirty state.
1075 * When a page is set dirty in its entirety, all its buffers are marked dirty
1076 * (if the page has buffers).
1078 * When a buffer is marked dirty, its page is dirtied, but the page's other
1081 * Also. When blockdev buffers are explicitly read with bread(), they
1082 * individually become uptodate. But their backing page remains not
1083 * uptodate - even if all of its buffers are uptodate. A subsequent
1084 * block_read_full_page() against that page will discover all the uptodate
1085 * buffers, will set the page uptodate and will perform no I/O.
1089 * mark_buffer_dirty - mark a buffer_head as needing writeout
1090 * @bh: the buffer_head to mark dirty
1092 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1093 * backing page dirty, then tag the page as dirty in its address_space's radix
1094 * tree and then attach the address_space's inode to its superblock's dirty
1097 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1098 * i_pages lock and mapping->host->i_lock.
1100 void mark_buffer_dirty(struct buffer_head *bh)
1102 WARN_ON_ONCE(!buffer_uptodate(bh));
1104 trace_block_dirty_buffer(bh);
1107 * Very *carefully* optimize the it-is-already-dirty case.
1109 * Don't let the final "is it dirty" escape to before we
1110 * perhaps modified the buffer.
1112 if (buffer_dirty(bh)) {
1114 if (buffer_dirty(bh))
1118 if (!test_set_buffer_dirty(bh)) {
1119 struct page *page = bh->b_page;
1120 struct address_space *mapping = NULL;
1122 lock_page_memcg(page);
1123 if (!TestSetPageDirty(page)) {
1124 mapping = page_mapping(page);
1126 __set_page_dirty(page, mapping, 0);
1128 unlock_page_memcg(page);
1130 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1133 EXPORT_SYMBOL(mark_buffer_dirty);
1135 void mark_buffer_write_io_error(struct buffer_head *bh)
1137 set_buffer_write_io_error(bh);
1138 /* FIXME: do we need to set this in both places? */
1139 if (bh->b_page && bh->b_page->mapping)
1140 mapping_set_error(bh->b_page->mapping, -EIO);
1141 if (bh->b_assoc_map)
1142 mapping_set_error(bh->b_assoc_map, -EIO);
1144 EXPORT_SYMBOL(mark_buffer_write_io_error);
1147 * Decrement a buffer_head's reference count. If all buffers against a page
1148 * have zero reference count, are clean and unlocked, and if the page is clean
1149 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1150 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1151 * a page but it ends up not being freed, and buffers may later be reattached).
1153 void __brelse(struct buffer_head * buf)
1155 if (atomic_read(&buf->b_count)) {
1159 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1161 EXPORT_SYMBOL(__brelse);
1164 * bforget() is like brelse(), except it discards any
1165 * potentially dirty data.
1167 void __bforget(struct buffer_head *bh)
1169 clear_buffer_dirty(bh);
1170 if (bh->b_assoc_map) {
1171 struct address_space *buffer_mapping = bh->b_page->mapping;
1173 spin_lock(&buffer_mapping->private_lock);
1174 list_del_init(&bh->b_assoc_buffers);
1175 bh->b_assoc_map = NULL;
1176 spin_unlock(&buffer_mapping->private_lock);
1180 EXPORT_SYMBOL(__bforget);
1182 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1185 if (buffer_uptodate(bh)) {
1190 bh->b_end_io = end_buffer_read_sync;
1191 submit_bh(REQ_OP_READ, 0, bh);
1193 if (buffer_uptodate(bh))
1201 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1202 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1203 * refcount elevated by one when they're in an LRU. A buffer can only appear
1204 * once in a particular CPU's LRU. A single buffer can be present in multiple
1205 * CPU's LRUs at the same time.
1207 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1208 * sb_find_get_block().
1210 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1211 * a local interrupt disable for that.
1214 #define BH_LRU_SIZE 16
1217 struct buffer_head *bhs[BH_LRU_SIZE];
1220 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1223 #define bh_lru_lock() local_irq_disable()
1224 #define bh_lru_unlock() local_irq_enable()
1226 #define bh_lru_lock() preempt_disable()
1227 #define bh_lru_unlock() preempt_enable()
1230 static inline void check_irqs_on(void)
1232 #ifdef irqs_disabled
1233 BUG_ON(irqs_disabled());
1238 * Install a buffer_head into this cpu's LRU. If not already in the LRU, it is
1239 * inserted at the front, and the buffer_head at the back if any is evicted.
1240 * Or, if already in the LRU it is moved to the front.
1242 static void bh_lru_install(struct buffer_head *bh)
1244 struct buffer_head *evictee = bh;
1251 b = this_cpu_ptr(&bh_lrus);
1252 for (i = 0; i < BH_LRU_SIZE; i++) {
1253 swap(evictee, b->bhs[i]);
1254 if (evictee == bh) {
1266 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1268 static struct buffer_head *
1269 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1271 struct buffer_head *ret = NULL;
1276 for (i = 0; i < BH_LRU_SIZE; i++) {
1277 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1279 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1280 bh->b_size == size) {
1283 __this_cpu_write(bh_lrus.bhs[i],
1284 __this_cpu_read(bh_lrus.bhs[i - 1]));
1287 __this_cpu_write(bh_lrus.bhs[0], bh);
1299 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1300 * it in the LRU and mark it as accessed. If it is not present then return
1303 struct buffer_head *
1304 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1306 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1309 /* __find_get_block_slow will mark the page accessed */
1310 bh = __find_get_block_slow(bdev, block);
1318 EXPORT_SYMBOL(__find_get_block);
1321 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1322 * which corresponds to the passed block_device, block and size. The
1323 * returned buffer has its reference count incremented.
1325 * __getblk_gfp() will lock up the machine if grow_dev_page's
1326 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1328 struct buffer_head *
1329 __getblk_gfp(struct block_device *bdev, sector_t block,
1330 unsigned size, gfp_t gfp)
1332 struct buffer_head *bh = __find_get_block(bdev, block, size);
1336 bh = __getblk_slow(bdev, block, size, gfp);
1339 EXPORT_SYMBOL(__getblk_gfp);
1342 * Do async read-ahead on a buffer..
1344 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1346 struct buffer_head *bh = __getblk(bdev, block, size);
1348 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1352 EXPORT_SYMBOL(__breadahead);
1355 * __bread_gfp() - reads a specified block and returns the bh
1356 * @bdev: the block_device to read from
1357 * @block: number of block
1358 * @size: size (in bytes) to read
1359 * @gfp: page allocation flag
1361 * Reads a specified block, and returns buffer head that contains it.
1362 * The page cache can be allocated from non-movable area
1363 * not to prevent page migration if you set gfp to zero.
1364 * It returns NULL if the block was unreadable.
1366 struct buffer_head *
1367 __bread_gfp(struct block_device *bdev, sector_t block,
1368 unsigned size, gfp_t gfp)
1370 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1372 if (likely(bh) && !buffer_uptodate(bh))
1373 bh = __bread_slow(bh);
1376 EXPORT_SYMBOL(__bread_gfp);
1379 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1380 * This doesn't race because it runs in each cpu either in irq
1381 * or with preempt disabled.
1383 static void invalidate_bh_lru(void *arg)
1385 struct bh_lru *b = &get_cpu_var(bh_lrus);
1388 for (i = 0; i < BH_LRU_SIZE; i++) {
1392 put_cpu_var(bh_lrus);
1395 static bool has_bh_in_lru(int cpu, void *dummy)
1397 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1400 for (i = 0; i < BH_LRU_SIZE; i++) {
1408 void invalidate_bh_lrus(void)
1410 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1412 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1414 void set_bh_page(struct buffer_head *bh,
1415 struct page *page, unsigned long offset)
1418 BUG_ON(offset >= PAGE_SIZE);
1419 if (PageHighMem(page))
1421 * This catches illegal uses and preserves the offset:
1423 bh->b_data = (char *)(0 + offset);
1425 bh->b_data = page_address(page) + offset;
1427 EXPORT_SYMBOL(set_bh_page);
1430 * Called when truncating a buffer on a page completely.
1433 /* Bits that are cleared during an invalidate */
1434 #define BUFFER_FLAGS_DISCARD \
1435 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1436 1 << BH_Delay | 1 << BH_Unwritten)
1438 static void discard_buffer(struct buffer_head * bh)
1440 unsigned long b_state, b_state_old;
1443 clear_buffer_dirty(bh);
1445 b_state = bh->b_state;
1447 b_state_old = cmpxchg(&bh->b_state, b_state,
1448 (b_state & ~BUFFER_FLAGS_DISCARD));
1449 if (b_state_old == b_state)
1451 b_state = b_state_old;
1457 * block_invalidatepage - invalidate part or all of a buffer-backed page
1459 * @page: the page which is affected
1460 * @offset: start of the range to invalidate
1461 * @length: length of the range to invalidate
1463 * block_invalidatepage() is called when all or part of the page has become
1464 * invalidated by a truncate operation.
1466 * block_invalidatepage() does not have to release all buffers, but it must
1467 * ensure that no dirty buffer is left outside @offset and that no I/O
1468 * is underway against any of the blocks which are outside the truncation
1469 * point. Because the caller is about to free (and possibly reuse) those
1472 void block_invalidatepage(struct page *page, unsigned int offset,
1473 unsigned int length)
1475 struct buffer_head *head, *bh, *next;
1476 unsigned int curr_off = 0;
1477 unsigned int stop = length + offset;
1479 BUG_ON(!PageLocked(page));
1480 if (!page_has_buffers(page))
1484 * Check for overflow
1486 BUG_ON(stop > PAGE_SIZE || stop < length);
1488 head = page_buffers(page);
1491 unsigned int next_off = curr_off + bh->b_size;
1492 next = bh->b_this_page;
1495 * Are we still fully in range ?
1497 if (next_off > stop)
1501 * is this block fully invalidated?
1503 if (offset <= curr_off)
1505 curr_off = next_off;
1507 } while (bh != head);
1510 * We release buffers only if the entire page is being invalidated.
1511 * The get_block cached value has been unconditionally invalidated,
1512 * so real IO is not possible anymore.
1514 if (length == PAGE_SIZE)
1515 try_to_release_page(page, 0);
1519 EXPORT_SYMBOL(block_invalidatepage);
1523 * We attach and possibly dirty the buffers atomically wrt
1524 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1525 * is already excluded via the page lock.
1527 void create_empty_buffers(struct page *page,
1528 unsigned long blocksize, unsigned long b_state)
1530 struct buffer_head *bh, *head, *tail;
1532 head = alloc_page_buffers(page, blocksize, true);
1535 bh->b_state |= b_state;
1537 bh = bh->b_this_page;
1539 tail->b_this_page = head;
1541 spin_lock(&page->mapping->private_lock);
1542 if (PageUptodate(page) || PageDirty(page)) {
1545 if (PageDirty(page))
1546 set_buffer_dirty(bh);
1547 if (PageUptodate(page))
1548 set_buffer_uptodate(bh);
1549 bh = bh->b_this_page;
1550 } while (bh != head);
1552 attach_page_buffers(page, head);
1553 spin_unlock(&page->mapping->private_lock);
1555 EXPORT_SYMBOL(create_empty_buffers);
1558 * clean_bdev_aliases: clean a range of buffers in block device
1559 * @bdev: Block device to clean buffers in
1560 * @block: Start of a range of blocks to clean
1561 * @len: Number of blocks to clean
1563 * We are taking a range of blocks for data and we don't want writeback of any
1564 * buffer-cache aliases starting from return from this function and until the
1565 * moment when something will explicitly mark the buffer dirty (hopefully that
1566 * will not happen until we will free that block ;-) We don't even need to mark
1567 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1568 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1569 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1570 * would confuse anyone who might pick it with bread() afterwards...
1572 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1573 * writeout I/O going on against recently-freed buffers. We don't wait on that
1574 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1575 * need to. That happens here.
1577 void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1579 struct inode *bd_inode = bdev->bd_inode;
1580 struct address_space *bd_mapping = bd_inode->i_mapping;
1581 struct pagevec pvec;
1582 pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1585 struct buffer_head *bh;
1586 struct buffer_head *head;
1588 end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1589 pagevec_init(&pvec);
1590 while (pagevec_lookup_range(&pvec, bd_mapping, &index, end)) {
1591 count = pagevec_count(&pvec);
1592 for (i = 0; i < count; i++) {
1593 struct page *page = pvec.pages[i];
1595 if (!page_has_buffers(page))
1598 * We use page lock instead of bd_mapping->private_lock
1599 * to pin buffers here since we can afford to sleep and
1600 * it scales better than a global spinlock lock.
1603 /* Recheck when the page is locked which pins bhs */
1604 if (!page_has_buffers(page))
1606 head = page_buffers(page);
1609 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1611 if (bh->b_blocknr >= block + len)
1613 clear_buffer_dirty(bh);
1615 clear_buffer_req(bh);
1617 bh = bh->b_this_page;
1618 } while (bh != head);
1622 pagevec_release(&pvec);
1624 /* End of range already reached? */
1625 if (index > end || !index)
1629 EXPORT_SYMBOL(clean_bdev_aliases);
1632 * Size is a power-of-two in the range 512..PAGE_SIZE,
1633 * and the case we care about most is PAGE_SIZE.
1635 * So this *could* possibly be written with those
1636 * constraints in mind (relevant mostly if some
1637 * architecture has a slow bit-scan instruction)
1639 static inline int block_size_bits(unsigned int blocksize)
1641 return ilog2(blocksize);
1644 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1646 BUG_ON(!PageLocked(page));
1648 if (!page_has_buffers(page))
1649 create_empty_buffers(page, 1 << READ_ONCE(inode->i_blkbits),
1651 return page_buffers(page);
1655 * NOTE! All mapped/uptodate combinations are valid:
1657 * Mapped Uptodate Meaning
1659 * No No "unknown" - must do get_block()
1660 * No Yes "hole" - zero-filled
1661 * Yes No "allocated" - allocated on disk, not read in
1662 * Yes Yes "valid" - allocated and up-to-date in memory.
1664 * "Dirty" is valid only with the last case (mapped+uptodate).
1668 * While block_write_full_page is writing back the dirty buffers under
1669 * the page lock, whoever dirtied the buffers may decide to clean them
1670 * again at any time. We handle that by only looking at the buffer
1671 * state inside lock_buffer().
1673 * If block_write_full_page() is called for regular writeback
1674 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1675 * locked buffer. This only can happen if someone has written the buffer
1676 * directly, with submit_bh(). At the address_space level PageWriteback
1677 * prevents this contention from occurring.
1679 * If block_write_full_page() is called with wbc->sync_mode ==
1680 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1681 * causes the writes to be flagged as synchronous writes.
1683 int __block_write_full_page(struct inode *inode, struct page *page,
1684 get_block_t *get_block, struct writeback_control *wbc,
1685 bh_end_io_t *handler)
1689 sector_t last_block;
1690 struct buffer_head *bh, *head;
1691 unsigned int blocksize, bbits;
1692 int nr_underway = 0;
1693 int write_flags = wbc_to_write_flags(wbc);
1695 head = create_page_buffers(page, inode,
1696 (1 << BH_Dirty)|(1 << BH_Uptodate));
1699 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1700 * here, and the (potentially unmapped) buffers may become dirty at
1701 * any time. If a buffer becomes dirty here after we've inspected it
1702 * then we just miss that fact, and the page stays dirty.
1704 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1705 * handle that here by just cleaning them.
1709 blocksize = bh->b_size;
1710 bbits = block_size_bits(blocksize);
1712 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1713 last_block = (i_size_read(inode) - 1) >> bbits;
1716 * Get all the dirty buffers mapped to disk addresses and
1717 * handle any aliases from the underlying blockdev's mapping.
1720 if (block > last_block) {
1722 * mapped buffers outside i_size will occur, because
1723 * this page can be outside i_size when there is a
1724 * truncate in progress.
1727 * The buffer was zeroed by block_write_full_page()
1729 clear_buffer_dirty(bh);
1730 set_buffer_uptodate(bh);
1731 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1733 WARN_ON(bh->b_size != blocksize);
1734 err = get_block(inode, block, bh, 1);
1737 clear_buffer_delay(bh);
1738 if (buffer_new(bh)) {
1739 /* blockdev mappings never come here */
1740 clear_buffer_new(bh);
1741 clean_bdev_bh_alias(bh);
1744 bh = bh->b_this_page;
1746 } while (bh != head);
1749 if (!buffer_mapped(bh))
1752 * If it's a fully non-blocking write attempt and we cannot
1753 * lock the buffer then redirty the page. Note that this can
1754 * potentially cause a busy-wait loop from writeback threads
1755 * and kswapd activity, but those code paths have their own
1756 * higher-level throttling.
1758 if (wbc->sync_mode != WB_SYNC_NONE) {
1760 } else if (!trylock_buffer(bh)) {
1761 redirty_page_for_writepage(wbc, page);
1764 if (test_clear_buffer_dirty(bh)) {
1765 mark_buffer_async_write_endio(bh, handler);
1769 } while ((bh = bh->b_this_page) != head);
1772 * The page and its buffers are protected by PageWriteback(), so we can
1773 * drop the bh refcounts early.
1775 BUG_ON(PageWriteback(page));
1776 set_page_writeback(page);
1779 struct buffer_head *next = bh->b_this_page;
1780 if (buffer_async_write(bh)) {
1781 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1782 inode->i_write_hint, wbc);
1786 } while (bh != head);
1791 if (nr_underway == 0) {
1793 * The page was marked dirty, but the buffers were
1794 * clean. Someone wrote them back by hand with
1795 * ll_rw_block/submit_bh. A rare case.
1797 end_page_writeback(page);
1800 * The page and buffer_heads can be released at any time from
1808 * ENOSPC, or some other error. We may already have added some
1809 * blocks to the file, so we need to write these out to avoid
1810 * exposing stale data.
1811 * The page is currently locked and not marked for writeback
1814 /* Recovery: lock and submit the mapped buffers */
1816 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1817 !buffer_delay(bh)) {
1819 mark_buffer_async_write_endio(bh, handler);
1822 * The buffer may have been set dirty during
1823 * attachment to a dirty page.
1825 clear_buffer_dirty(bh);
1827 } while ((bh = bh->b_this_page) != head);
1829 BUG_ON(PageWriteback(page));
1830 mapping_set_error(page->mapping, err);
1831 set_page_writeback(page);
1833 struct buffer_head *next = bh->b_this_page;
1834 if (buffer_async_write(bh)) {
1835 clear_buffer_dirty(bh);
1836 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1837 inode->i_write_hint, wbc);
1841 } while (bh != head);
1845 EXPORT_SYMBOL(__block_write_full_page);
1848 * If a page has any new buffers, zero them out here, and mark them uptodate
1849 * and dirty so they'll be written out (in order to prevent uninitialised
1850 * block data from leaking). And clear the new bit.
1852 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1854 unsigned int block_start, block_end;
1855 struct buffer_head *head, *bh;
1857 BUG_ON(!PageLocked(page));
1858 if (!page_has_buffers(page))
1861 bh = head = page_buffers(page);
1864 block_end = block_start + bh->b_size;
1866 if (buffer_new(bh)) {
1867 if (block_end > from && block_start < to) {
1868 if (!PageUptodate(page)) {
1869 unsigned start, size;
1871 start = max(from, block_start);
1872 size = min(to, block_end) - start;
1874 zero_user(page, start, size);
1875 set_buffer_uptodate(bh);
1878 clear_buffer_new(bh);
1879 mark_buffer_dirty(bh);
1883 block_start = block_end;
1884 bh = bh->b_this_page;
1885 } while (bh != head);
1887 EXPORT_SYMBOL(page_zero_new_buffers);
1890 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1891 struct iomap *iomap)
1893 loff_t offset = block << inode->i_blkbits;
1895 bh->b_bdev = iomap->bdev;
1898 * Block points to offset in file we need to map, iomap contains
1899 * the offset at which the map starts. If the map ends before the
1900 * current block, then do not map the buffer and let the caller
1903 BUG_ON(offset >= iomap->offset + iomap->length);
1905 switch (iomap->type) {
1908 * If the buffer is not up to date or beyond the current EOF,
1909 * we need to mark it as new to ensure sub-block zeroing is
1910 * executed if necessary.
1912 if (!buffer_uptodate(bh) ||
1913 (offset >= i_size_read(inode)))
1916 case IOMAP_DELALLOC:
1917 if (!buffer_uptodate(bh) ||
1918 (offset >= i_size_read(inode)))
1920 set_buffer_uptodate(bh);
1921 set_buffer_mapped(bh);
1922 set_buffer_delay(bh);
1924 case IOMAP_UNWRITTEN:
1926 * For unwritten regions, we always need to ensure that
1927 * sub-block writes cause the regions in the block we are not
1928 * writing to are zeroed. Set the buffer as new to ensure this.
1931 set_buffer_unwritten(bh);
1934 if (offset >= i_size_read(inode))
1936 bh->b_blocknr = (iomap->addr + offset - iomap->offset) >>
1938 set_buffer_mapped(bh);
1943 int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
1944 get_block_t *get_block, struct iomap *iomap)
1946 unsigned from = pos & (PAGE_SIZE - 1);
1947 unsigned to = from + len;
1948 struct inode *inode = page->mapping->host;
1949 unsigned block_start, block_end;
1952 unsigned blocksize, bbits;
1953 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1955 BUG_ON(!PageLocked(page));
1956 BUG_ON(from > PAGE_SIZE);
1957 BUG_ON(to > PAGE_SIZE);
1960 head = create_page_buffers(page, inode, 0);
1961 blocksize = head->b_size;
1962 bbits = block_size_bits(blocksize);
1964 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1966 for(bh = head, block_start = 0; bh != head || !block_start;
1967 block++, block_start=block_end, bh = bh->b_this_page) {
1968 block_end = block_start + blocksize;
1969 if (block_end <= from || block_start >= to) {
1970 if (PageUptodate(page)) {
1971 if (!buffer_uptodate(bh))
1972 set_buffer_uptodate(bh);
1977 clear_buffer_new(bh);
1978 if (!buffer_mapped(bh)) {
1979 WARN_ON(bh->b_size != blocksize);
1981 err = get_block(inode, block, bh, 1);
1985 iomap_to_bh(inode, block, bh, iomap);
1988 if (buffer_new(bh)) {
1989 clean_bdev_bh_alias(bh);
1990 if (PageUptodate(page)) {
1991 clear_buffer_new(bh);
1992 set_buffer_uptodate(bh);
1993 mark_buffer_dirty(bh);
1996 if (block_end > to || block_start < from)
1997 zero_user_segments(page,
2003 if (PageUptodate(page)) {
2004 if (!buffer_uptodate(bh))
2005 set_buffer_uptodate(bh);
2008 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2009 !buffer_unwritten(bh) &&
2010 (block_start < from || block_end > to)) {
2011 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2016 * If we issued read requests - let them complete.
2018 while(wait_bh > wait) {
2019 wait_on_buffer(*--wait_bh);
2020 if (!buffer_uptodate(*wait_bh))
2024 page_zero_new_buffers(page, from, to);
2028 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2029 get_block_t *get_block)
2031 return __block_write_begin_int(page, pos, len, get_block, NULL);
2033 EXPORT_SYMBOL(__block_write_begin);
2035 static int __block_commit_write(struct inode *inode, struct page *page,
2036 unsigned from, unsigned to)
2038 unsigned block_start, block_end;
2041 struct buffer_head *bh, *head;
2043 bh = head = page_buffers(page);
2044 blocksize = bh->b_size;
2048 block_end = block_start + blocksize;
2049 if (block_end <= from || block_start >= to) {
2050 if (!buffer_uptodate(bh))
2053 set_buffer_uptodate(bh);
2054 mark_buffer_dirty(bh);
2056 clear_buffer_new(bh);
2058 block_start = block_end;
2059 bh = bh->b_this_page;
2060 } while (bh != head);
2063 * If this is a partial write which happened to make all buffers
2064 * uptodate then we can optimize away a bogus readpage() for
2065 * the next read(). Here we 'discover' whether the page went
2066 * uptodate as a result of this (potentially partial) write.
2069 SetPageUptodate(page);
2074 * block_write_begin takes care of the basic task of block allocation and
2075 * bringing partial write blocks uptodate first.
2077 * The filesystem needs to handle block truncation upon failure.
2079 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2080 unsigned flags, struct page **pagep, get_block_t *get_block)
2082 pgoff_t index = pos >> PAGE_SHIFT;
2086 page = grab_cache_page_write_begin(mapping, index, flags);
2090 status = __block_write_begin(page, pos, len, get_block);
2091 if (unlikely(status)) {
2100 EXPORT_SYMBOL(block_write_begin);
2102 int block_write_end(struct file *file, struct address_space *mapping,
2103 loff_t pos, unsigned len, unsigned copied,
2104 struct page *page, void *fsdata)
2106 struct inode *inode = mapping->host;
2109 start = pos & (PAGE_SIZE - 1);
2111 if (unlikely(copied < len)) {
2113 * The buffers that were written will now be uptodate, so we
2114 * don't have to worry about a readpage reading them and
2115 * overwriting a partial write. However if we have encountered
2116 * a short write and only partially written into a buffer, it
2117 * will not be marked uptodate, so a readpage might come in and
2118 * destroy our partial write.
2120 * Do the simplest thing, and just treat any short write to a
2121 * non uptodate page as a zero-length write, and force the
2122 * caller to redo the whole thing.
2124 if (!PageUptodate(page))
2127 page_zero_new_buffers(page, start+copied, start+len);
2129 flush_dcache_page(page);
2131 /* This could be a short (even 0-length) commit */
2132 __block_commit_write(inode, page, start, start+copied);
2136 EXPORT_SYMBOL(block_write_end);
2138 int generic_write_end(struct file *file, struct address_space *mapping,
2139 loff_t pos, unsigned len, unsigned copied,
2140 struct page *page, void *fsdata)
2142 struct inode *inode = mapping->host;
2143 loff_t old_size = inode->i_size;
2144 int i_size_changed = 0;
2146 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2149 * No need to use i_size_read() here, the i_size
2150 * cannot change under us because we hold i_mutex.
2152 * But it's important to update i_size while still holding page lock:
2153 * page writeout could otherwise come in and zero beyond i_size.
2155 if (pos+copied > inode->i_size) {
2156 i_size_write(inode, pos+copied);
2164 pagecache_isize_extended(inode, old_size, pos);
2166 * Don't mark the inode dirty under page lock. First, it unnecessarily
2167 * makes the holding time of page lock longer. Second, it forces lock
2168 * ordering of page lock and transaction start for journaling
2172 mark_inode_dirty(inode);
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 * The page may have dirty, unmapped buffers. For example,
2747 * they may have been added in ext3_writepage(). Make them
2748 * freeable here, so the page does not leak.
2751 /* Not really sure about this - do we need this ? */
2752 if (page->mapping->a_ops->invalidatepage)
2753 page->mapping->a_ops->invalidatepage(page, offset);
2756 return 0; /* don't care */
2760 * The page straddles i_size. It must be zeroed out on each and every
2761 * writepage invocation because it may be mmapped. "A file is mapped
2762 * in multiples of the page size. For a file that is not a multiple of
2763 * the page size, the remaining memory is zeroed when mapped, and
2764 * writes to that region are not written out to the file."
2766 zero_user_segment(page, offset, PAGE_SIZE);
2768 ret = mpage_writepage(page, get_block, wbc);
2770 ret = __block_write_full_page(inode, page, get_block, wbc,
2771 end_buffer_async_write);
2774 EXPORT_SYMBOL(nobh_writepage);
2776 int nobh_truncate_page(struct address_space *mapping,
2777 loff_t from, get_block_t *get_block)
2779 pgoff_t index = from >> PAGE_SHIFT;
2780 unsigned offset = from & (PAGE_SIZE-1);
2783 unsigned length, pos;
2784 struct inode *inode = mapping->host;
2786 struct buffer_head map_bh;
2789 blocksize = i_blocksize(inode);
2790 length = offset & (blocksize - 1);
2792 /* Block boundary? Nothing to do */
2796 length = blocksize - length;
2797 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2799 page = grab_cache_page(mapping, index);
2804 if (page_has_buffers(page)) {
2808 return block_truncate_page(mapping, from, get_block);
2811 /* Find the buffer that contains "offset" */
2813 while (offset >= pos) {
2818 map_bh.b_size = blocksize;
2820 err = get_block(inode, iblock, &map_bh, 0);
2823 /* unmapped? It's a hole - nothing to do */
2824 if (!buffer_mapped(&map_bh))
2827 /* Ok, it's mapped. Make sure it's up-to-date */
2828 if (!PageUptodate(page)) {
2829 err = mapping->a_ops->readpage(NULL, page);
2835 if (!PageUptodate(page)) {
2839 if (page_has_buffers(page))
2842 zero_user(page, offset, length);
2843 set_page_dirty(page);
2852 EXPORT_SYMBOL(nobh_truncate_page);
2854 int block_truncate_page(struct address_space *mapping,
2855 loff_t from, get_block_t *get_block)
2857 pgoff_t index = from >> PAGE_SHIFT;
2858 unsigned offset = from & (PAGE_SIZE-1);
2861 unsigned length, pos;
2862 struct inode *inode = mapping->host;
2864 struct buffer_head *bh;
2867 blocksize = i_blocksize(inode);
2868 length = offset & (blocksize - 1);
2870 /* Block boundary? Nothing to do */
2874 length = blocksize - length;
2875 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2877 page = grab_cache_page(mapping, index);
2882 if (!page_has_buffers(page))
2883 create_empty_buffers(page, blocksize, 0);
2885 /* Find the buffer that contains "offset" */
2886 bh = page_buffers(page);
2888 while (offset >= pos) {
2889 bh = bh->b_this_page;
2895 if (!buffer_mapped(bh)) {
2896 WARN_ON(bh->b_size != blocksize);
2897 err = get_block(inode, iblock, bh, 0);
2900 /* unmapped? It's a hole - nothing to do */
2901 if (!buffer_mapped(bh))
2905 /* Ok, it's mapped. Make sure it's up-to-date */
2906 if (PageUptodate(page))
2907 set_buffer_uptodate(bh);
2909 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2911 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2913 /* Uhhuh. Read error. Complain and punt. */
2914 if (!buffer_uptodate(bh))
2918 zero_user(page, offset, length);
2919 mark_buffer_dirty(bh);
2928 EXPORT_SYMBOL(block_truncate_page);
2931 * The generic ->writepage function for buffer-backed address_spaces
2933 int block_write_full_page(struct page *page, get_block_t *get_block,
2934 struct writeback_control *wbc)
2936 struct inode * const inode = page->mapping->host;
2937 loff_t i_size = i_size_read(inode);
2938 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2941 /* Is the page fully inside i_size? */
2942 if (page->index < end_index)
2943 return __block_write_full_page(inode, page, get_block, wbc,
2944 end_buffer_async_write);
2946 /* Is the page fully outside i_size? (truncate in progress) */
2947 offset = i_size & (PAGE_SIZE-1);
2948 if (page->index >= end_index+1 || !offset) {
2950 * The page may have dirty, unmapped buffers. For example,
2951 * they may have been added in ext3_writepage(). Make them
2952 * freeable here, so the page does not leak.
2954 do_invalidatepage(page, 0, PAGE_SIZE);
2956 return 0; /* don't care */
2960 * The page straddles i_size. It must be zeroed out on each and every
2961 * writepage invocation because it may be mmapped. "A file is mapped
2962 * in multiples of the page size. For a file that is not a multiple of
2963 * the page size, the remaining memory is zeroed when mapped, and
2964 * writes to that region are not written out to the file."
2966 zero_user_segment(page, offset, PAGE_SIZE);
2967 return __block_write_full_page(inode, page, get_block, wbc,
2968 end_buffer_async_write);
2970 EXPORT_SYMBOL(block_write_full_page);
2972 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2973 get_block_t *get_block)
2975 struct inode *inode = mapping->host;
2976 struct buffer_head tmp = {
2977 .b_size = i_blocksize(inode),
2980 get_block(inode, block, &tmp, 0);
2981 return tmp.b_blocknr;
2983 EXPORT_SYMBOL(generic_block_bmap);
2985 static void end_bio_bh_io_sync(struct bio *bio)
2987 struct buffer_head *bh = bio->bi_private;
2989 if (unlikely(bio_flagged(bio, BIO_QUIET)))
2990 set_bit(BH_Quiet, &bh->b_state);
2992 bh->b_end_io(bh, !bio->bi_status);
2997 * This allows us to do IO even on the odd last sectors
2998 * of a device, even if the block size is some multiple
2999 * of the physical sector size.
3001 * We'll just truncate the bio to the size of the device,
3002 * and clear the end of the buffer head manually.
3004 * Truly out-of-range accesses will turn into actual IO
3005 * errors, this only handles the "we need to be able to
3006 * do IO at the final sector" case.
3008 void guard_bio_eod(int op, struct bio *bio)
3011 struct bio_vec *bvec = bio_last_bvec_all(bio);
3012 unsigned truncated_bytes;
3013 struct hd_struct *part;
3016 part = __disk_get_part(bio->bi_disk, bio->bi_partno);
3018 maxsector = part_nr_sects_read(part);
3020 maxsector = get_capacity(bio->bi_disk);
3027 * If the *whole* IO is past the end of the device,
3028 * let it through, and the IO layer will turn it into
3031 if (unlikely(bio->bi_iter.bi_sector >= maxsector))
3034 maxsector -= bio->bi_iter.bi_sector;
3035 if (likely((bio->bi_iter.bi_size >> 9) <= maxsector))
3038 /* Uhhuh. We've got a bio that straddles the device size! */
3039 truncated_bytes = bio->bi_iter.bi_size - (maxsector << 9);
3041 /* Truncate the bio.. */
3042 bio->bi_iter.bi_size -= truncated_bytes;
3043 bvec->bv_len -= truncated_bytes;
3045 /* ..and clear the end of the buffer for reads */
3046 if (op == REQ_OP_READ) {
3047 zero_user(bvec->bv_page, bvec->bv_offset + bvec->bv_len,
3052 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3053 enum rw_hint write_hint, struct writeback_control *wbc)
3057 BUG_ON(!buffer_locked(bh));
3058 BUG_ON(!buffer_mapped(bh));
3059 BUG_ON(!bh->b_end_io);
3060 BUG_ON(buffer_delay(bh));
3061 BUG_ON(buffer_unwritten(bh));
3064 * Only clear out a write error when rewriting
3066 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3067 clear_buffer_write_io_error(bh);
3070 * from here on down, it's all bio -- do the initial mapping,
3071 * submit_bio -> generic_make_request may further map this bio around
3073 bio = bio_alloc(GFP_NOIO, 1);
3076 wbc_init_bio(wbc, bio);
3077 wbc_account_io(wbc, bh->b_page, bh->b_size);
3080 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3081 bio_set_dev(bio, bh->b_bdev);
3082 bio->bi_write_hint = write_hint;
3084 bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3085 BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3087 bio->bi_end_io = end_bio_bh_io_sync;
3088 bio->bi_private = bh;
3090 /* Take care of bh's that straddle the end of the device */
3091 guard_bio_eod(op, bio);
3093 if (buffer_meta(bh))
3094 op_flags |= REQ_META;
3095 if (buffer_prio(bh))
3096 op_flags |= REQ_PRIO;
3097 bio_set_op_attrs(bio, op, op_flags);
3103 int submit_bh(int op, int op_flags, struct buffer_head *bh)
3105 return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3107 EXPORT_SYMBOL(submit_bh);
3110 * ll_rw_block: low-level access to block devices (DEPRECATED)
3111 * @op: whether to %READ or %WRITE
3112 * @op_flags: req_flag_bits
3113 * @nr: number of &struct buffer_heads in the array
3114 * @bhs: array of pointers to &struct buffer_head
3116 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3117 * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3118 * @op_flags contains flags modifying the detailed I/O behavior, most notably
3121 * This function drops any buffer that it cannot get a lock on (with the
3122 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3123 * request, and any buffer that appears to be up-to-date when doing read
3124 * request. Further it marks as clean buffers that are processed for
3125 * writing (the buffer cache won't assume that they are actually clean
3126 * until the buffer gets unlocked).
3128 * ll_rw_block sets b_end_io to simple completion handler that marks
3129 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3132 * All of the buffers must be for the same device, and must also be a
3133 * multiple of the current approved size for the device.
3135 void ll_rw_block(int op, int op_flags, int nr, struct buffer_head *bhs[])
3139 for (i = 0; i < nr; i++) {
3140 struct buffer_head *bh = bhs[i];
3142 if (!trylock_buffer(bh))
3145 if (test_clear_buffer_dirty(bh)) {
3146 bh->b_end_io = end_buffer_write_sync;
3148 submit_bh(op, op_flags, bh);
3152 if (!buffer_uptodate(bh)) {
3153 bh->b_end_io = end_buffer_read_sync;
3155 submit_bh(op, op_flags, bh);
3162 EXPORT_SYMBOL(ll_rw_block);
3164 void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3167 if (!test_clear_buffer_dirty(bh)) {
3171 bh->b_end_io = end_buffer_write_sync;
3173 submit_bh(REQ_OP_WRITE, op_flags, bh);
3175 EXPORT_SYMBOL(write_dirty_buffer);
3178 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3179 * and then start new I/O and then wait upon it. The caller must have a ref on
3182 int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3186 WARN_ON(atomic_read(&bh->b_count) < 1);
3188 if (test_clear_buffer_dirty(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);
3454 * Seek for SEEK_DATA / SEEK_HOLE within @page, starting at @lastoff.
3456 * Returns the offset within the file on success, and -ENOENT otherwise.
3459 page_seek_hole_data(struct page *page, loff_t lastoff, int whence)
3461 loff_t offset = page_offset(page);
3462 struct buffer_head *bh, *head;
3463 bool seek_data = whence == SEEK_DATA;
3465 if (lastoff < offset)
3468 bh = head = page_buffers(page);
3470 offset += bh->b_size;
3471 if (lastoff >= offset)
3475 * Unwritten extents that have data in the page cache covering
3476 * them can be identified by the BH_Unwritten state flag.
3477 * Pages with multiple buffers might have a mix of holes, data
3478 * and unwritten extents - any buffer with valid data in it
3479 * should have BH_Uptodate flag set on it.
3482 if ((buffer_unwritten(bh) || buffer_uptodate(bh)) == seek_data)
3486 } while ((bh = bh->b_this_page) != head);
3491 * Seek for SEEK_DATA / SEEK_HOLE in the page cache.
3493 * Within unwritten extents, the page cache determines which parts are holes
3494 * and which are data: unwritten and uptodate buffer heads count as data;
3495 * everything else counts as a hole.
3497 * Returns the resulting offset on successs, and -ENOENT otherwise.
3500 page_cache_seek_hole_data(struct inode *inode, loff_t offset, loff_t length,
3503 pgoff_t index = offset >> PAGE_SHIFT;
3504 pgoff_t end = DIV_ROUND_UP(offset + length, PAGE_SIZE);
3505 loff_t lastoff = offset;
3506 struct pagevec pvec;
3511 pagevec_init(&pvec);
3514 unsigned nr_pages, i;
3516 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping, &index,
3521 for (i = 0; i < nr_pages; i++) {
3522 struct page *page = pvec.pages[i];
3525 * At this point, the page may be truncated or
3526 * invalidated (changing page->mapping to NULL), or
3527 * even swizzled back from swapper_space to tmpfs file
3528 * mapping. However, page->index will not change
3529 * because we have a reference on the page.
3531 * If current page offset is beyond where we've ended,
3532 * we've found a hole.
3534 if (whence == SEEK_HOLE &&
3535 lastoff < page_offset(page))
3539 if (likely(page->mapping == inode->i_mapping) &&
3540 page_has_buffers(page)) {
3541 lastoff = page_seek_hole_data(page, lastoff, whence);
3548 lastoff = page_offset(page) + PAGE_SIZE;
3550 pagevec_release(&pvec);
3551 } while (index < end);
3553 /* When no page at lastoff and we are not done, we found a hole. */
3554 if (whence != SEEK_HOLE)
3558 if (lastoff < offset + length)
3563 pagevec_release(&pvec);
3567 void __init buffer_init(void)
3569 unsigned long nrpages;
3572 bh_cachep = kmem_cache_create("buffer_head",
3573 sizeof(struct buffer_head), 0,
3574 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3579 * Limit the bh occupancy to 10% of ZONE_NORMAL
3581 nrpages = (nr_free_buffer_pages() * 10) / 100;
3582 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3583 ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3584 NULL, buffer_exit_cpu_dead);