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 unsigned long bio_flags,
53 struct writeback_control *wbc);
55 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
57 void init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
59 bh->b_end_io = handler;
60 bh->b_private = private;
62 EXPORT_SYMBOL(init_buffer);
64 inline void touch_buffer(struct buffer_head *bh)
66 trace_block_touch_buffer(bh);
67 mark_page_accessed(bh->b_page);
69 EXPORT_SYMBOL(touch_buffer);
71 void __lock_buffer(struct buffer_head *bh)
73 wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
75 EXPORT_SYMBOL(__lock_buffer);
77 void unlock_buffer(struct buffer_head *bh)
79 clear_bit_unlock(BH_Lock, &bh->b_state);
80 smp_mb__after_atomic();
81 wake_up_bit(&bh->b_state, BH_Lock);
83 EXPORT_SYMBOL(unlock_buffer);
86 * Returns if the page has dirty or writeback buffers. If all the buffers
87 * are unlocked and clean then the PageDirty information is stale. If
88 * any of the pages are locked, it is assumed they are locked for IO.
90 void buffer_check_dirty_writeback(struct page *page,
91 bool *dirty, bool *writeback)
93 struct buffer_head *head, *bh;
97 BUG_ON(!PageLocked(page));
99 if (!page_has_buffers(page))
102 if (PageWriteback(page))
105 head = page_buffers(page);
108 if (buffer_locked(bh))
111 if (buffer_dirty(bh))
114 bh = bh->b_this_page;
115 } while (bh != head);
117 EXPORT_SYMBOL(buffer_check_dirty_writeback);
120 * Block until a buffer comes unlocked. This doesn't stop it
121 * from becoming locked again - you have to lock it yourself
122 * if you want to preserve its state.
124 void __wait_on_buffer(struct buffer_head * bh)
126 wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
128 EXPORT_SYMBOL(__wait_on_buffer);
131 __clear_page_buffers(struct page *page)
133 ClearPagePrivate(page);
134 set_page_private(page, 0);
138 static void buffer_io_error(struct buffer_head *bh, char *msg)
140 if (!test_bit(BH_Quiet, &bh->b_state))
141 printk_ratelimited(KERN_ERR
142 "Buffer I/O error on dev %pg, logical block %llu%s\n",
143 bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
147 * End-of-IO handler helper function which does not touch the bh after
149 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
150 * a race there is benign: unlock_buffer() only use the bh's address for
151 * hashing after unlocking the buffer, so it doesn't actually touch the bh
154 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
157 set_buffer_uptodate(bh);
159 /* This happens, due to failed read-ahead attempts. */
160 clear_buffer_uptodate(bh);
166 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
167 * unlock the buffer. This is what ll_rw_block uses too.
169 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
171 __end_buffer_read_notouch(bh, uptodate);
174 EXPORT_SYMBOL(end_buffer_read_sync);
176 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
179 set_buffer_uptodate(bh);
181 buffer_io_error(bh, ", lost sync page write");
182 set_buffer_write_io_error(bh);
183 clear_buffer_uptodate(bh);
188 EXPORT_SYMBOL(end_buffer_write_sync);
191 * Various filesystems appear to want __find_get_block to be non-blocking.
192 * But it's the page lock which protects the buffers. To get around this,
193 * we get exclusion from try_to_free_buffers with the blockdev mapping's
196 * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
197 * may be quite high. This code could TryLock the page, and if that
198 * succeeds, there is no need to take private_lock. (But if
199 * private_lock is contended then so is mapping->tree_lock).
201 static struct buffer_head *
202 __find_get_block_slow(struct block_device *bdev, sector_t block)
204 struct inode *bd_inode = bdev->bd_inode;
205 struct address_space *bd_mapping = bd_inode->i_mapping;
206 struct buffer_head *ret = NULL;
208 struct buffer_head *bh;
209 struct buffer_head *head;
213 index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
214 page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
218 spin_lock(&bd_mapping->private_lock);
219 if (!page_has_buffers(page))
221 head = page_buffers(page);
224 if (!buffer_mapped(bh))
226 else if (bh->b_blocknr == block) {
231 bh = bh->b_this_page;
232 } while (bh != head);
234 /* we might be here because some of the buffers on this page are
235 * not mapped. This is due to various races between
236 * file io on the block device and getblk. It gets dealt with
237 * elsewhere, don't buffer_error if we had some unmapped buffers
240 printk("__find_get_block_slow() failed. "
241 "block=%llu, b_blocknr=%llu\n",
242 (unsigned long long)block,
243 (unsigned long long)bh->b_blocknr);
244 printk("b_state=0x%08lx, b_size=%zu\n",
245 bh->b_state, bh->b_size);
246 printk("device %pg blocksize: %d\n", bdev,
247 1 << bd_inode->i_blkbits);
250 spin_unlock(&bd_mapping->private_lock);
257 * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
259 static void free_more_memory(void)
264 wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
267 for_each_online_node(nid) {
269 z = first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
270 gfp_zone(GFP_NOFS), NULL);
272 try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
278 * I/O completion handler for block_read_full_page() - pages
279 * which come unlocked at the end of I/O.
281 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
284 struct buffer_head *first;
285 struct buffer_head *tmp;
287 int page_uptodate = 1;
289 BUG_ON(!buffer_async_read(bh));
293 set_buffer_uptodate(bh);
295 clear_buffer_uptodate(bh);
296 buffer_io_error(bh, ", async page read");
301 * Be _very_ careful from here on. Bad things can happen if
302 * two buffer heads end IO at almost the same time and both
303 * decide that the page is now completely done.
305 first = page_buffers(page);
306 local_irq_save(flags);
307 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
308 clear_buffer_async_read(bh);
312 if (!buffer_uptodate(tmp))
314 if (buffer_async_read(tmp)) {
315 BUG_ON(!buffer_locked(tmp));
318 tmp = tmp->b_this_page;
320 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
321 local_irq_restore(flags);
324 * If none of the buffers had errors and they are all
325 * uptodate then we can set the page uptodate.
327 if (page_uptodate && !PageError(page))
328 SetPageUptodate(page);
333 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
334 local_irq_restore(flags);
339 * Completion handler for block_write_full_page() - pages which are unlocked
340 * during I/O, and which have PageWriteback cleared upon I/O completion.
342 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
345 struct buffer_head *first;
346 struct buffer_head *tmp;
349 BUG_ON(!buffer_async_write(bh));
353 set_buffer_uptodate(bh);
355 buffer_io_error(bh, ", lost async page write");
356 mapping_set_error(page->mapping, -EIO);
357 set_buffer_write_io_error(bh);
358 clear_buffer_uptodate(bh);
362 first = page_buffers(page);
363 local_irq_save(flags);
364 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
366 clear_buffer_async_write(bh);
368 tmp = bh->b_this_page;
370 if (buffer_async_write(tmp)) {
371 BUG_ON(!buffer_locked(tmp));
374 tmp = tmp->b_this_page;
376 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
377 local_irq_restore(flags);
378 end_page_writeback(page);
382 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
383 local_irq_restore(flags);
386 EXPORT_SYMBOL(end_buffer_async_write);
389 * If a page's buffers are under async readin (end_buffer_async_read
390 * completion) then there is a possibility that another thread of
391 * control could lock one of the buffers after it has completed
392 * but while some of the other buffers have not completed. This
393 * locked buffer would confuse end_buffer_async_read() into not unlocking
394 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
395 * that this buffer is not under async I/O.
397 * The page comes unlocked when it has no locked buffer_async buffers
400 * PageLocked prevents anyone starting new async I/O reads any of
403 * PageWriteback is used to prevent simultaneous writeout of the same
406 * PageLocked prevents anyone from starting writeback of a page which is
407 * under read I/O (PageWriteback is only ever set against a locked page).
409 static void mark_buffer_async_read(struct buffer_head *bh)
411 bh->b_end_io = end_buffer_async_read;
412 set_buffer_async_read(bh);
415 static void mark_buffer_async_write_endio(struct buffer_head *bh,
416 bh_end_io_t *handler)
418 bh->b_end_io = handler;
419 set_buffer_async_write(bh);
422 void mark_buffer_async_write(struct buffer_head *bh)
424 mark_buffer_async_write_endio(bh, end_buffer_async_write);
426 EXPORT_SYMBOL(mark_buffer_async_write);
430 * fs/buffer.c contains helper functions for buffer-backed address space's
431 * fsync functions. A common requirement for buffer-based filesystems is
432 * that certain data from the backing blockdev needs to be written out for
433 * a successful fsync(). For example, ext2 indirect blocks need to be
434 * written back and waited upon before fsync() returns.
436 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
437 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
438 * management of a list of dependent buffers at ->i_mapping->private_list.
440 * Locking is a little subtle: try_to_free_buffers() will remove buffers
441 * from their controlling inode's queue when they are being freed. But
442 * try_to_free_buffers() will be operating against the *blockdev* mapping
443 * at the time, not against the S_ISREG file which depends on those buffers.
444 * So the locking for private_list is via the private_lock in the address_space
445 * which backs the buffers. Which is different from the address_space
446 * against which the buffers are listed. So for a particular address_space,
447 * mapping->private_lock does *not* protect mapping->private_list! In fact,
448 * mapping->private_list will always be protected by the backing blockdev's
451 * Which introduces a requirement: all buffers on an address_space's
452 * ->private_list must be from the same address_space: the blockdev's.
454 * address_spaces which do not place buffers at ->private_list via these
455 * utility functions are free to use private_lock and private_list for
456 * whatever they want. The only requirement is that list_empty(private_list)
457 * be true at clear_inode() time.
459 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
460 * filesystems should do that. invalidate_inode_buffers() should just go
461 * BUG_ON(!list_empty).
463 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
464 * take an address_space, not an inode. And it should be called
465 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
468 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
469 * list if it is already on a list. Because if the buffer is on a list,
470 * it *must* already be on the right one. If not, the filesystem is being
471 * silly. This will save a ton of locking. But first we have to ensure
472 * that buffers are taken *off* the old inode's list when they are freed
473 * (presumably in truncate). That requires careful auditing of all
474 * filesystems (do it inside bforget()). It could also be done by bringing
479 * The buffer's backing address_space's private_lock must be held
481 static void __remove_assoc_queue(struct buffer_head *bh)
483 list_del_init(&bh->b_assoc_buffers);
484 WARN_ON(!bh->b_assoc_map);
485 if (buffer_write_io_error(bh))
486 set_bit(AS_EIO, &bh->b_assoc_map->flags);
487 bh->b_assoc_map = NULL;
490 int inode_has_buffers(struct inode *inode)
492 return !list_empty(&inode->i_data.private_list);
496 * osync is designed to support O_SYNC io. It waits synchronously for
497 * all already-submitted IO to complete, but does not queue any new
498 * writes to the disk.
500 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
501 * you dirty the buffers, and then use osync_inode_buffers to wait for
502 * completion. Any other dirty buffers which are not yet queued for
503 * write will not be flushed to disk by the osync.
505 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
507 struct buffer_head *bh;
513 list_for_each_prev(p, list) {
515 if (buffer_locked(bh)) {
519 if (!buffer_uptodate(bh))
530 static void do_thaw_one(struct super_block *sb, void *unused)
532 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
533 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
536 static void do_thaw_all(struct work_struct *work)
538 iterate_supers(do_thaw_one, NULL);
540 printk(KERN_WARNING "Emergency Thaw complete\n");
544 * emergency_thaw_all -- forcibly thaw every frozen filesystem
546 * Used for emergency unfreeze of all filesystems via SysRq
548 void emergency_thaw_all(void)
550 struct work_struct *work;
552 work = kmalloc(sizeof(*work), GFP_ATOMIC);
554 INIT_WORK(work, do_thaw_all);
560 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
561 * @mapping: the mapping which wants those buffers written
563 * Starts I/O against the buffers at mapping->private_list, and waits upon
566 * Basically, this is a convenience function for fsync().
567 * @mapping is a file or directory which needs those buffers to be written for
568 * a successful fsync().
570 int sync_mapping_buffers(struct address_space *mapping)
572 struct address_space *buffer_mapping = mapping->private_data;
574 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
577 return fsync_buffers_list(&buffer_mapping->private_lock,
578 &mapping->private_list);
580 EXPORT_SYMBOL(sync_mapping_buffers);
583 * Called when we've recently written block `bblock', and it is known that
584 * `bblock' was for a buffer_boundary() buffer. This means that the block at
585 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
586 * dirty, schedule it for IO. So that indirects merge nicely with their data.
588 void write_boundary_block(struct block_device *bdev,
589 sector_t bblock, unsigned blocksize)
591 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
593 if (buffer_dirty(bh))
594 ll_rw_block(REQ_OP_WRITE, 0, 1, &bh);
599 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
601 struct address_space *mapping = inode->i_mapping;
602 struct address_space *buffer_mapping = bh->b_page->mapping;
604 mark_buffer_dirty(bh);
605 if (!mapping->private_data) {
606 mapping->private_data = buffer_mapping;
608 BUG_ON(mapping->private_data != buffer_mapping);
610 if (!bh->b_assoc_map) {
611 spin_lock(&buffer_mapping->private_lock);
612 list_move_tail(&bh->b_assoc_buffers,
613 &mapping->private_list);
614 bh->b_assoc_map = mapping;
615 spin_unlock(&buffer_mapping->private_lock);
618 EXPORT_SYMBOL(mark_buffer_dirty_inode);
621 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
624 * If warn is true, then emit a warning if the page is not uptodate and has
625 * not been truncated.
627 * The caller must hold lock_page_memcg().
629 static void __set_page_dirty(struct page *page, struct address_space *mapping,
634 spin_lock_irqsave(&mapping->tree_lock, flags);
635 if (page->mapping) { /* Race with truncate? */
636 WARN_ON_ONCE(warn && !PageUptodate(page));
637 account_page_dirtied(page, mapping);
638 radix_tree_tag_set(&mapping->page_tree,
639 page_index(page), PAGECACHE_TAG_DIRTY);
641 spin_unlock_irqrestore(&mapping->tree_lock, flags);
645 * Add a page to the dirty page list.
647 * It is a sad fact of life that this function is called from several places
648 * deeply under spinlocking. It may not sleep.
650 * If the page has buffers, the uptodate buffers are set dirty, to preserve
651 * dirty-state coherency between the page and the buffers. It the page does
652 * not have buffers then when they are later attached they will all be set
655 * The buffers are dirtied before the page is dirtied. There's a small race
656 * window in which a writepage caller may see the page cleanness but not the
657 * buffer dirtiness. That's fine. If this code were to set the page dirty
658 * before the buffers, a concurrent writepage caller could clear the page dirty
659 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
660 * page on the dirty page list.
662 * We use private_lock to lock against try_to_free_buffers while using the
663 * page's buffer list. Also use this to protect against clean buffers being
664 * added to the page after it was set dirty.
666 * FIXME: may need to call ->reservepage here as well. That's rather up to the
667 * address_space though.
669 int __set_page_dirty_buffers(struct page *page)
672 struct address_space *mapping = page_mapping(page);
674 if (unlikely(!mapping))
675 return !TestSetPageDirty(page);
677 spin_lock(&mapping->private_lock);
678 if (page_has_buffers(page)) {
679 struct buffer_head *head = page_buffers(page);
680 struct buffer_head *bh = head;
683 set_buffer_dirty(bh);
684 bh = bh->b_this_page;
685 } while (bh != head);
688 * Lock out page->mem_cgroup migration to keep PageDirty
689 * synchronized with per-memcg dirty page counters.
691 lock_page_memcg(page);
692 newly_dirty = !TestSetPageDirty(page);
693 spin_unlock(&mapping->private_lock);
696 __set_page_dirty(page, mapping, 1);
698 unlock_page_memcg(page);
701 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
705 EXPORT_SYMBOL(__set_page_dirty_buffers);
708 * Write out and wait upon a list of buffers.
710 * We have conflicting pressures: we want to make sure that all
711 * initially dirty buffers get waited on, but that any subsequently
712 * dirtied buffers don't. After all, we don't want fsync to last
713 * forever if somebody is actively writing to the file.
715 * Do this in two main stages: first we copy dirty buffers to a
716 * temporary inode list, queueing the writes as we go. Then we clean
717 * up, waiting for those writes to complete.
719 * During this second stage, any subsequent updates to the file may end
720 * up refiling the buffer on the original inode's dirty list again, so
721 * there is a chance we will end up with a buffer queued for write but
722 * not yet completed on that list. So, as a final cleanup we go through
723 * the osync code to catch these locked, dirty buffers without requeuing
724 * any newly dirty buffers for write.
726 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
728 struct buffer_head *bh;
729 struct list_head tmp;
730 struct address_space *mapping;
732 struct blk_plug plug;
734 INIT_LIST_HEAD(&tmp);
735 blk_start_plug(&plug);
738 while (!list_empty(list)) {
739 bh = BH_ENTRY(list->next);
740 mapping = bh->b_assoc_map;
741 __remove_assoc_queue(bh);
742 /* Avoid race with mark_buffer_dirty_inode() which does
743 * a lockless check and we rely on seeing the dirty bit */
745 if (buffer_dirty(bh) || buffer_locked(bh)) {
746 list_add(&bh->b_assoc_buffers, &tmp);
747 bh->b_assoc_map = mapping;
748 if (buffer_dirty(bh)) {
752 * Ensure any pending I/O completes so that
753 * write_dirty_buffer() actually writes the
754 * current contents - it is a noop if I/O is
755 * still in flight on potentially older
758 write_dirty_buffer(bh, REQ_SYNC);
761 * Kick off IO for the previous mapping. Note
762 * that we will not run the very last mapping,
763 * wait_on_buffer() will do that for us
764 * through sync_buffer().
773 blk_finish_plug(&plug);
776 while (!list_empty(&tmp)) {
777 bh = BH_ENTRY(tmp.prev);
779 mapping = bh->b_assoc_map;
780 __remove_assoc_queue(bh);
781 /* Avoid race with mark_buffer_dirty_inode() which does
782 * a lockless check and we rely on seeing the dirty bit */
784 if (buffer_dirty(bh)) {
785 list_add(&bh->b_assoc_buffers,
786 &mapping->private_list);
787 bh->b_assoc_map = mapping;
791 if (!buffer_uptodate(bh))
798 err2 = osync_buffers_list(lock, list);
806 * Invalidate any and all dirty buffers on a given inode. We are
807 * probably unmounting the fs, but that doesn't mean we have already
808 * done a sync(). Just drop the buffers from the inode list.
810 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
811 * assumes that all the buffers are against the blockdev. Not true
814 void invalidate_inode_buffers(struct inode *inode)
816 if (inode_has_buffers(inode)) {
817 struct address_space *mapping = &inode->i_data;
818 struct list_head *list = &mapping->private_list;
819 struct address_space *buffer_mapping = mapping->private_data;
821 spin_lock(&buffer_mapping->private_lock);
822 while (!list_empty(list))
823 __remove_assoc_queue(BH_ENTRY(list->next));
824 spin_unlock(&buffer_mapping->private_lock);
827 EXPORT_SYMBOL(invalidate_inode_buffers);
830 * Remove any clean buffers from the inode's buffer list. This is called
831 * when we're trying to free the inode itself. Those buffers can pin it.
833 * Returns true if all buffers were removed.
835 int remove_inode_buffers(struct inode *inode)
839 if (inode_has_buffers(inode)) {
840 struct address_space *mapping = &inode->i_data;
841 struct list_head *list = &mapping->private_list;
842 struct address_space *buffer_mapping = mapping->private_data;
844 spin_lock(&buffer_mapping->private_lock);
845 while (!list_empty(list)) {
846 struct buffer_head *bh = BH_ENTRY(list->next);
847 if (buffer_dirty(bh)) {
851 __remove_assoc_queue(bh);
853 spin_unlock(&buffer_mapping->private_lock);
859 * Create the appropriate buffers when given a page for data area and
860 * the size of each buffer.. Use the bh->b_this_page linked list to
861 * follow the buffers created. Return NULL if unable to create more
864 * The retry flag is used to differentiate async IO (paging, swapping)
865 * which may not fail from ordinary buffer allocations.
867 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
870 struct buffer_head *bh, *head;
876 while ((offset -= size) >= 0) {
877 bh = alloc_buffer_head(GFP_NOFS);
881 bh->b_this_page = head;
887 /* Link the buffer to its page */
888 set_bh_page(bh, page, offset);
892 * In case anything failed, we just free everything we got.
898 head = head->b_this_page;
899 free_buffer_head(bh);
904 * Return failure for non-async IO requests. Async IO requests
905 * are not allowed to fail, so we have to wait until buffer heads
906 * become available. But we don't want tasks sleeping with
907 * partially complete buffers, so all were released above.
912 /* We're _really_ low on memory. Now we just
913 * wait for old buffer heads to become free due to
914 * finishing IO. Since this is an async request and
915 * the reserve list is empty, we're sure there are
916 * async buffer heads in use.
921 EXPORT_SYMBOL_GPL(alloc_page_buffers);
924 link_dev_buffers(struct page *page, struct buffer_head *head)
926 struct buffer_head *bh, *tail;
931 bh = bh->b_this_page;
933 tail->b_this_page = head;
934 attach_page_buffers(page, head);
937 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
939 sector_t retval = ~((sector_t)0);
940 loff_t sz = i_size_read(bdev->bd_inode);
943 unsigned int sizebits = blksize_bits(size);
944 retval = (sz >> sizebits);
950 * Initialise the state of a blockdev page's buffers.
953 init_page_buffers(struct page *page, struct block_device *bdev,
954 sector_t block, int size)
956 struct buffer_head *head = page_buffers(page);
957 struct buffer_head *bh = head;
958 int uptodate = PageUptodate(page);
959 sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
962 if (!buffer_mapped(bh)) {
963 init_buffer(bh, NULL, NULL);
965 bh->b_blocknr = block;
967 set_buffer_uptodate(bh);
968 if (block < end_block)
969 set_buffer_mapped(bh);
972 bh = bh->b_this_page;
973 } while (bh != head);
976 * Caller needs to validate requested block against end of device.
982 * Create the page-cache page that contains the requested block.
984 * This is used purely for blockdev mappings.
987 grow_dev_page(struct block_device *bdev, sector_t block,
988 pgoff_t index, int size, int sizebits, gfp_t gfp)
990 struct inode *inode = bdev->bd_inode;
992 struct buffer_head *bh;
994 int ret = 0; /* Will call free_more_memory() */
997 gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
1000 * XXX: __getblk_slow() can not really deal with failure and
1001 * will endlessly loop on improvised global reclaim. Prefer
1002 * looping in the allocator rather than here, at least that
1003 * code knows what it's doing.
1005 gfp_mask |= __GFP_NOFAIL;
1007 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
1011 BUG_ON(!PageLocked(page));
1013 if (page_has_buffers(page)) {
1014 bh = page_buffers(page);
1015 if (bh->b_size == size) {
1016 end_block = init_page_buffers(page, bdev,
1017 (sector_t)index << sizebits,
1021 if (!try_to_free_buffers(page))
1026 * Allocate some buffers for this page
1028 bh = alloc_page_buffers(page, size, 0);
1033 * Link the page to the buffers and initialise them. Take the
1034 * lock to be atomic wrt __find_get_block(), which does not
1035 * run under the page lock.
1037 spin_lock(&inode->i_mapping->private_lock);
1038 link_dev_buffers(page, bh);
1039 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1041 spin_unlock(&inode->i_mapping->private_lock);
1043 ret = (block < end_block) ? 1 : -ENXIO;
1051 * Create buffers for the specified block device block's page. If
1052 * that page was dirty, the buffers are set dirty also.
1055 grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
1063 } while ((size << sizebits) < PAGE_SIZE);
1065 index = block >> sizebits;
1068 * Check for a block which wants to lie outside our maximum possible
1069 * pagecache index. (this comparison is done using sector_t types).
1071 if (unlikely(index != block >> sizebits)) {
1072 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1074 __func__, (unsigned long long)block,
1079 /* Create a page with the proper size buffers.. */
1080 return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1083 static struct buffer_head *
1084 __getblk_slow(struct block_device *bdev, sector_t block,
1085 unsigned size, gfp_t gfp)
1087 /* Size must be multiple of hard sectorsize */
1088 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1089 (size < 512 || size > PAGE_SIZE))) {
1090 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1092 printk(KERN_ERR "logical block size: %d\n",
1093 bdev_logical_block_size(bdev));
1100 struct buffer_head *bh;
1103 bh = __find_get_block(bdev, block, size);
1107 ret = grow_buffers(bdev, block, size, gfp);
1116 * The relationship between dirty buffers and dirty pages:
1118 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1119 * the page is tagged dirty in its radix tree.
1121 * At all times, the dirtiness of the buffers represents the dirtiness of
1122 * subsections of the page. If the page has buffers, the page dirty bit is
1123 * merely a hint about the true dirty state.
1125 * When a page is set dirty in its entirety, all its buffers are marked dirty
1126 * (if the page has buffers).
1128 * When a buffer is marked dirty, its page is dirtied, but the page's other
1131 * Also. When blockdev buffers are explicitly read with bread(), they
1132 * individually become uptodate. But their backing page remains not
1133 * uptodate - even if all of its buffers are uptodate. A subsequent
1134 * block_read_full_page() against that page will discover all the uptodate
1135 * buffers, will set the page uptodate and will perform no I/O.
1139 * mark_buffer_dirty - mark a buffer_head as needing writeout
1140 * @bh: the buffer_head to mark dirty
1142 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1143 * backing page dirty, then tag the page as dirty in its address_space's radix
1144 * tree and then attach the address_space's inode to its superblock's dirty
1147 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1148 * mapping->tree_lock and mapping->host->i_lock.
1150 void mark_buffer_dirty(struct buffer_head *bh)
1152 WARN_ON_ONCE(!buffer_uptodate(bh));
1154 trace_block_dirty_buffer(bh);
1157 * Very *carefully* optimize the it-is-already-dirty case.
1159 * Don't let the final "is it dirty" escape to before we
1160 * perhaps modified the buffer.
1162 if (buffer_dirty(bh)) {
1164 if (buffer_dirty(bh))
1168 if (!test_set_buffer_dirty(bh)) {
1169 struct page *page = bh->b_page;
1170 struct address_space *mapping = NULL;
1172 lock_page_memcg(page);
1173 if (!TestSetPageDirty(page)) {
1174 mapping = page_mapping(page);
1176 __set_page_dirty(page, mapping, 0);
1178 unlock_page_memcg(page);
1180 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1183 EXPORT_SYMBOL(mark_buffer_dirty);
1186 * Decrement a buffer_head's reference count. If all buffers against a page
1187 * have zero reference count, are clean and unlocked, and if the page is clean
1188 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1189 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1190 * a page but it ends up not being freed, and buffers may later be reattached).
1192 void __brelse(struct buffer_head * buf)
1194 if (atomic_read(&buf->b_count)) {
1198 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1200 EXPORT_SYMBOL(__brelse);
1203 * bforget() is like brelse(), except it discards any
1204 * potentially dirty data.
1206 void __bforget(struct buffer_head *bh)
1208 clear_buffer_dirty(bh);
1209 if (bh->b_assoc_map) {
1210 struct address_space *buffer_mapping = bh->b_page->mapping;
1212 spin_lock(&buffer_mapping->private_lock);
1213 list_del_init(&bh->b_assoc_buffers);
1214 bh->b_assoc_map = NULL;
1215 spin_unlock(&buffer_mapping->private_lock);
1219 EXPORT_SYMBOL(__bforget);
1221 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1224 if (buffer_uptodate(bh)) {
1229 bh->b_end_io = end_buffer_read_sync;
1230 submit_bh(REQ_OP_READ, 0, bh);
1232 if (buffer_uptodate(bh))
1240 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1241 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1242 * refcount elevated by one when they're in an LRU. A buffer can only appear
1243 * once in a particular CPU's LRU. A single buffer can be present in multiple
1244 * CPU's LRUs at the same time.
1246 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1247 * sb_find_get_block().
1249 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1250 * a local interrupt disable for that.
1253 #define BH_LRU_SIZE 16
1256 struct buffer_head *bhs[BH_LRU_SIZE];
1259 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1262 #define bh_lru_lock() local_irq_disable()
1263 #define bh_lru_unlock() local_irq_enable()
1265 #define bh_lru_lock() preempt_disable()
1266 #define bh_lru_unlock() preempt_enable()
1269 static inline void check_irqs_on(void)
1271 #ifdef irqs_disabled
1272 BUG_ON(irqs_disabled());
1277 * The LRU management algorithm is dopey-but-simple. Sorry.
1279 static void bh_lru_install(struct buffer_head *bh)
1281 struct buffer_head *evictee = NULL;
1285 if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1286 struct buffer_head *bhs[BH_LRU_SIZE];
1292 for (in = 0; in < BH_LRU_SIZE; in++) {
1293 struct buffer_head *bh2 =
1294 __this_cpu_read(bh_lrus.bhs[in]);
1299 if (out >= BH_LRU_SIZE) {
1300 BUG_ON(evictee != NULL);
1307 while (out < BH_LRU_SIZE)
1309 memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1318 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1320 static struct buffer_head *
1321 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1323 struct buffer_head *ret = NULL;
1328 for (i = 0; i < BH_LRU_SIZE; i++) {
1329 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1331 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1332 bh->b_size == size) {
1335 __this_cpu_write(bh_lrus.bhs[i],
1336 __this_cpu_read(bh_lrus.bhs[i - 1]));
1339 __this_cpu_write(bh_lrus.bhs[0], bh);
1351 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1352 * it in the LRU and mark it as accessed. If it is not present then return
1355 struct buffer_head *
1356 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1358 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1361 /* __find_get_block_slow will mark the page accessed */
1362 bh = __find_get_block_slow(bdev, block);
1370 EXPORT_SYMBOL(__find_get_block);
1373 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1374 * which corresponds to the passed block_device, block and size. The
1375 * returned buffer has its reference count incremented.
1377 * __getblk_gfp() will lock up the machine if grow_dev_page's
1378 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1380 struct buffer_head *
1381 __getblk_gfp(struct block_device *bdev, sector_t block,
1382 unsigned size, gfp_t gfp)
1384 struct buffer_head *bh = __find_get_block(bdev, block, size);
1388 bh = __getblk_slow(bdev, block, size, gfp);
1391 EXPORT_SYMBOL(__getblk_gfp);
1394 * Do async read-ahead on a buffer..
1396 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1398 struct buffer_head *bh = __getblk(bdev, block, size);
1400 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1404 EXPORT_SYMBOL(__breadahead);
1407 * __bread_gfp() - reads a specified block and returns the bh
1408 * @bdev: the block_device to read from
1409 * @block: number of block
1410 * @size: size (in bytes) to read
1411 * @gfp: page allocation flag
1413 * Reads a specified block, and returns buffer head that contains it.
1414 * The page cache can be allocated from non-movable area
1415 * not to prevent page migration if you set gfp to zero.
1416 * It returns NULL if the block was unreadable.
1418 struct buffer_head *
1419 __bread_gfp(struct block_device *bdev, sector_t block,
1420 unsigned size, gfp_t gfp)
1422 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1424 if (likely(bh) && !buffer_uptodate(bh))
1425 bh = __bread_slow(bh);
1428 EXPORT_SYMBOL(__bread_gfp);
1431 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1432 * This doesn't race because it runs in each cpu either in irq
1433 * or with preempt disabled.
1435 static void invalidate_bh_lru(void *arg)
1437 struct bh_lru *b = &get_cpu_var(bh_lrus);
1440 for (i = 0; i < BH_LRU_SIZE; i++) {
1444 put_cpu_var(bh_lrus);
1447 static bool has_bh_in_lru(int cpu, void *dummy)
1449 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1452 for (i = 0; i < BH_LRU_SIZE; i++) {
1460 void invalidate_bh_lrus(void)
1462 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1464 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1466 void set_bh_page(struct buffer_head *bh,
1467 struct page *page, unsigned long offset)
1470 BUG_ON(offset >= PAGE_SIZE);
1471 if (PageHighMem(page))
1473 * This catches illegal uses and preserves the offset:
1475 bh->b_data = (char *)(0 + offset);
1477 bh->b_data = page_address(page) + offset;
1479 EXPORT_SYMBOL(set_bh_page);
1482 * Called when truncating a buffer on a page completely.
1485 /* Bits that are cleared during an invalidate */
1486 #define BUFFER_FLAGS_DISCARD \
1487 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1488 1 << BH_Delay | 1 << BH_Unwritten)
1490 static void discard_buffer(struct buffer_head * bh)
1492 unsigned long b_state, b_state_old;
1495 clear_buffer_dirty(bh);
1497 b_state = bh->b_state;
1499 b_state_old = cmpxchg(&bh->b_state, b_state,
1500 (b_state & ~BUFFER_FLAGS_DISCARD));
1501 if (b_state_old == b_state)
1503 b_state = b_state_old;
1509 * block_invalidatepage - invalidate part or all of a buffer-backed page
1511 * @page: the page which is affected
1512 * @offset: start of the range to invalidate
1513 * @length: length of the range to invalidate
1515 * block_invalidatepage() is called when all or part of the page has become
1516 * invalidated by a truncate operation.
1518 * block_invalidatepage() does not have to release all buffers, but it must
1519 * ensure that no dirty buffer is left outside @offset and that no I/O
1520 * is underway against any of the blocks which are outside the truncation
1521 * point. Because the caller is about to free (and possibly reuse) those
1524 void block_invalidatepage(struct page *page, unsigned int offset,
1525 unsigned int length)
1527 struct buffer_head *head, *bh, *next;
1528 unsigned int curr_off = 0;
1529 unsigned int stop = length + offset;
1531 BUG_ON(!PageLocked(page));
1532 if (!page_has_buffers(page))
1536 * Check for overflow
1538 BUG_ON(stop > PAGE_SIZE || stop < length);
1540 head = page_buffers(page);
1543 unsigned int next_off = curr_off + bh->b_size;
1544 next = bh->b_this_page;
1547 * Are we still fully in range ?
1549 if (next_off > stop)
1553 * is this block fully invalidated?
1555 if (offset <= curr_off)
1557 curr_off = next_off;
1559 } while (bh != head);
1562 * We release buffers only if the entire page is being invalidated.
1563 * The get_block cached value has been unconditionally invalidated,
1564 * so real IO is not possible anymore.
1567 try_to_release_page(page, 0);
1571 EXPORT_SYMBOL(block_invalidatepage);
1575 * We attach and possibly dirty the buffers atomically wrt
1576 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1577 * is already excluded via the page lock.
1579 void create_empty_buffers(struct page *page,
1580 unsigned long blocksize, unsigned long b_state)
1582 struct buffer_head *bh, *head, *tail;
1584 head = alloc_page_buffers(page, blocksize, 1);
1587 bh->b_state |= b_state;
1589 bh = bh->b_this_page;
1591 tail->b_this_page = head;
1593 spin_lock(&page->mapping->private_lock);
1594 if (PageUptodate(page) || PageDirty(page)) {
1597 if (PageDirty(page))
1598 set_buffer_dirty(bh);
1599 if (PageUptodate(page))
1600 set_buffer_uptodate(bh);
1601 bh = bh->b_this_page;
1602 } while (bh != head);
1604 attach_page_buffers(page, head);
1605 spin_unlock(&page->mapping->private_lock);
1607 EXPORT_SYMBOL(create_empty_buffers);
1610 * clean_bdev_aliases: clean a range of buffers in block device
1611 * @bdev: Block device to clean buffers in
1612 * @block: Start of a range of blocks to clean
1613 * @len: Number of blocks to clean
1615 * We are taking a range of blocks for data and we don't want writeback of any
1616 * buffer-cache aliases starting from return from this function and until the
1617 * moment when something will explicitly mark the buffer dirty (hopefully that
1618 * will not happen until we will free that block ;-) We don't even need to mark
1619 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1620 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1621 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1622 * would confuse anyone who might pick it with bread() afterwards...
1624 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1625 * writeout I/O going on against recently-freed buffers. We don't wait on that
1626 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1627 * need to. That happens here.
1629 void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1631 struct inode *bd_inode = bdev->bd_inode;
1632 struct address_space *bd_mapping = bd_inode->i_mapping;
1633 struct pagevec pvec;
1634 pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1637 struct buffer_head *bh;
1638 struct buffer_head *head;
1640 end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1641 pagevec_init(&pvec, 0);
1642 while (index <= end && pagevec_lookup(&pvec, bd_mapping, index,
1643 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
1644 for (i = 0; i < pagevec_count(&pvec); i++) {
1645 struct page *page = pvec.pages[i];
1647 index = page->index;
1650 if (!page_has_buffers(page))
1653 * We use page lock instead of bd_mapping->private_lock
1654 * to pin buffers here since we can afford to sleep and
1655 * it scales better than a global spinlock lock.
1658 /* Recheck when the page is locked which pins bhs */
1659 if (!page_has_buffers(page))
1661 head = page_buffers(page);
1664 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1666 if (bh->b_blocknr >= block + len)
1668 clear_buffer_dirty(bh);
1670 clear_buffer_req(bh);
1672 bh = bh->b_this_page;
1673 } while (bh != head);
1677 pagevec_release(&pvec);
1682 EXPORT_SYMBOL(clean_bdev_aliases);
1685 * Size is a power-of-two in the range 512..PAGE_SIZE,
1686 * and the case we care about most is PAGE_SIZE.
1688 * So this *could* possibly be written with those
1689 * constraints in mind (relevant mostly if some
1690 * architecture has a slow bit-scan instruction)
1692 static inline int block_size_bits(unsigned int blocksize)
1694 return ilog2(blocksize);
1697 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1699 BUG_ON(!PageLocked(page));
1701 if (!page_has_buffers(page))
1702 create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1703 return page_buffers(page);
1707 * NOTE! All mapped/uptodate combinations are valid:
1709 * Mapped Uptodate Meaning
1711 * No No "unknown" - must do get_block()
1712 * No Yes "hole" - zero-filled
1713 * Yes No "allocated" - allocated on disk, not read in
1714 * Yes Yes "valid" - allocated and up-to-date in memory.
1716 * "Dirty" is valid only with the last case (mapped+uptodate).
1720 * While block_write_full_page is writing back the dirty buffers under
1721 * the page lock, whoever dirtied the buffers may decide to clean them
1722 * again at any time. We handle that by only looking at the buffer
1723 * state inside lock_buffer().
1725 * If block_write_full_page() is called for regular writeback
1726 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1727 * locked buffer. This only can happen if someone has written the buffer
1728 * directly, with submit_bh(). At the address_space level PageWriteback
1729 * prevents this contention from occurring.
1731 * If block_write_full_page() is called with wbc->sync_mode ==
1732 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1733 * causes the writes to be flagged as synchronous writes.
1735 int __block_write_full_page(struct inode *inode, struct page *page,
1736 get_block_t *get_block, struct writeback_control *wbc,
1737 bh_end_io_t *handler)
1741 sector_t last_block;
1742 struct buffer_head *bh, *head;
1743 unsigned int blocksize, bbits;
1744 int nr_underway = 0;
1745 int write_flags = wbc_to_write_flags(wbc);
1747 head = create_page_buffers(page, inode,
1748 (1 << BH_Dirty)|(1 << BH_Uptodate));
1751 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1752 * here, and the (potentially unmapped) buffers may become dirty at
1753 * any time. If a buffer becomes dirty here after we've inspected it
1754 * then we just miss that fact, and the page stays dirty.
1756 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1757 * handle that here by just cleaning them.
1761 blocksize = bh->b_size;
1762 bbits = block_size_bits(blocksize);
1764 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1765 last_block = (i_size_read(inode) - 1) >> bbits;
1768 * Get all the dirty buffers mapped to disk addresses and
1769 * handle any aliases from the underlying blockdev's mapping.
1772 if (block > last_block) {
1774 * mapped buffers outside i_size will occur, because
1775 * this page can be outside i_size when there is a
1776 * truncate in progress.
1779 * The buffer was zeroed by block_write_full_page()
1781 clear_buffer_dirty(bh);
1782 set_buffer_uptodate(bh);
1783 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1785 WARN_ON(bh->b_size != blocksize);
1786 err = get_block(inode, block, bh, 1);
1789 clear_buffer_delay(bh);
1790 if (buffer_new(bh)) {
1791 /* blockdev mappings never come here */
1792 clear_buffer_new(bh);
1793 clean_bdev_bh_alias(bh);
1796 bh = bh->b_this_page;
1798 } while (bh != head);
1801 if (!buffer_mapped(bh))
1804 * If it's a fully non-blocking write attempt and we cannot
1805 * lock the buffer then redirty the page. Note that this can
1806 * potentially cause a busy-wait loop from writeback threads
1807 * and kswapd activity, but those code paths have their own
1808 * higher-level throttling.
1810 if (wbc->sync_mode != WB_SYNC_NONE) {
1812 } else if (!trylock_buffer(bh)) {
1813 redirty_page_for_writepage(wbc, page);
1816 if (test_clear_buffer_dirty(bh)) {
1817 mark_buffer_async_write_endio(bh, handler);
1821 } while ((bh = bh->b_this_page) != head);
1824 * The page and its buffers are protected by PageWriteback(), so we can
1825 * drop the bh refcounts early.
1827 BUG_ON(PageWriteback(page));
1828 set_page_writeback(page);
1831 struct buffer_head *next = bh->b_this_page;
1832 if (buffer_async_write(bh)) {
1833 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh, 0, wbc);
1837 } while (bh != head);
1842 if (nr_underway == 0) {
1844 * The page was marked dirty, but the buffers were
1845 * clean. Someone wrote them back by hand with
1846 * ll_rw_block/submit_bh. A rare case.
1848 end_page_writeback(page);
1851 * The page and buffer_heads can be released at any time from
1859 * ENOSPC, or some other error. We may already have added some
1860 * blocks to the file, so we need to write these out to avoid
1861 * exposing stale data.
1862 * The page is currently locked and not marked for writeback
1865 /* Recovery: lock and submit the mapped buffers */
1867 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1868 !buffer_delay(bh)) {
1870 mark_buffer_async_write_endio(bh, handler);
1873 * The buffer may have been set dirty during
1874 * attachment to a dirty page.
1876 clear_buffer_dirty(bh);
1878 } while ((bh = bh->b_this_page) != head);
1880 BUG_ON(PageWriteback(page));
1881 mapping_set_error(page->mapping, err);
1882 set_page_writeback(page);
1884 struct buffer_head *next = bh->b_this_page;
1885 if (buffer_async_write(bh)) {
1886 clear_buffer_dirty(bh);
1887 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh, 0, wbc);
1891 } while (bh != head);
1895 EXPORT_SYMBOL(__block_write_full_page);
1898 * If a page has any new buffers, zero them out here, and mark them uptodate
1899 * and dirty so they'll be written out (in order to prevent uninitialised
1900 * block data from leaking). And clear the new bit.
1902 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1904 unsigned int block_start, block_end;
1905 struct buffer_head *head, *bh;
1907 BUG_ON(!PageLocked(page));
1908 if (!page_has_buffers(page))
1911 bh = head = page_buffers(page);
1914 block_end = block_start + bh->b_size;
1916 if (buffer_new(bh)) {
1917 if (block_end > from && block_start < to) {
1918 if (!PageUptodate(page)) {
1919 unsigned start, size;
1921 start = max(from, block_start);
1922 size = min(to, block_end) - start;
1924 zero_user(page, start, size);
1925 set_buffer_uptodate(bh);
1928 clear_buffer_new(bh);
1929 mark_buffer_dirty(bh);
1933 block_start = block_end;
1934 bh = bh->b_this_page;
1935 } while (bh != head);
1937 EXPORT_SYMBOL(page_zero_new_buffers);
1940 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1941 struct iomap *iomap)
1943 loff_t offset = block << inode->i_blkbits;
1945 bh->b_bdev = iomap->bdev;
1948 * Block points to offset in file we need to map, iomap contains
1949 * the offset at which the map starts. If the map ends before the
1950 * current block, then do not map the buffer and let the caller
1953 BUG_ON(offset >= iomap->offset + iomap->length);
1955 switch (iomap->type) {
1958 * If the buffer is not up to date or beyond the current EOF,
1959 * we need to mark it as new to ensure sub-block zeroing is
1960 * executed if necessary.
1962 if (!buffer_uptodate(bh) ||
1963 (offset >= i_size_read(inode)))
1966 case IOMAP_DELALLOC:
1967 if (!buffer_uptodate(bh) ||
1968 (offset >= i_size_read(inode)))
1970 set_buffer_uptodate(bh);
1971 set_buffer_mapped(bh);
1972 set_buffer_delay(bh);
1974 case IOMAP_UNWRITTEN:
1976 * For unwritten regions, we always need to ensure that
1977 * sub-block writes cause the regions in the block we are not
1978 * writing to are zeroed. Set the buffer as new to ensure this.
1981 set_buffer_unwritten(bh);
1984 if (offset >= i_size_read(inode))
1986 bh->b_blocknr = (iomap->blkno >> (inode->i_blkbits - 9)) +
1987 ((offset - iomap->offset) >> inode->i_blkbits);
1988 set_buffer_mapped(bh);
1993 int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
1994 get_block_t *get_block, struct iomap *iomap)
1996 unsigned from = pos & (PAGE_SIZE - 1);
1997 unsigned to = from + len;
1998 struct inode *inode = page->mapping->host;
1999 unsigned block_start, block_end;
2002 unsigned blocksize, bbits;
2003 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
2005 BUG_ON(!PageLocked(page));
2006 BUG_ON(from > PAGE_SIZE);
2007 BUG_ON(to > PAGE_SIZE);
2010 head = create_page_buffers(page, inode, 0);
2011 blocksize = head->b_size;
2012 bbits = block_size_bits(blocksize);
2014 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
2016 for(bh = head, block_start = 0; bh != head || !block_start;
2017 block++, block_start=block_end, bh = bh->b_this_page) {
2018 block_end = block_start + blocksize;
2019 if (block_end <= from || block_start >= to) {
2020 if (PageUptodate(page)) {
2021 if (!buffer_uptodate(bh))
2022 set_buffer_uptodate(bh);
2027 clear_buffer_new(bh);
2028 if (!buffer_mapped(bh)) {
2029 WARN_ON(bh->b_size != blocksize);
2031 err = get_block(inode, block, bh, 1);
2035 iomap_to_bh(inode, block, bh, iomap);
2038 if (buffer_new(bh)) {
2039 clean_bdev_bh_alias(bh);
2040 if (PageUptodate(page)) {
2041 clear_buffer_new(bh);
2042 set_buffer_uptodate(bh);
2043 mark_buffer_dirty(bh);
2046 if (block_end > to || block_start < from)
2047 zero_user_segments(page,
2053 if (PageUptodate(page)) {
2054 if (!buffer_uptodate(bh))
2055 set_buffer_uptodate(bh);
2058 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2059 !buffer_unwritten(bh) &&
2060 (block_start < from || block_end > to)) {
2061 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2066 * If we issued read requests - let them complete.
2068 while(wait_bh > wait) {
2069 wait_on_buffer(*--wait_bh);
2070 if (!buffer_uptodate(*wait_bh))
2074 page_zero_new_buffers(page, from, to);
2078 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2079 get_block_t *get_block)
2081 return __block_write_begin_int(page, pos, len, get_block, NULL);
2083 EXPORT_SYMBOL(__block_write_begin);
2085 static int __block_commit_write(struct inode *inode, struct page *page,
2086 unsigned from, unsigned to)
2088 unsigned block_start, block_end;
2091 struct buffer_head *bh, *head;
2093 bh = head = page_buffers(page);
2094 blocksize = bh->b_size;
2098 block_end = block_start + blocksize;
2099 if (block_end <= from || block_start >= to) {
2100 if (!buffer_uptodate(bh))
2103 set_buffer_uptodate(bh);
2104 mark_buffer_dirty(bh);
2106 clear_buffer_new(bh);
2108 block_start = block_end;
2109 bh = bh->b_this_page;
2110 } while (bh != head);
2113 * If this is a partial write which happened to make all buffers
2114 * uptodate then we can optimize away a bogus readpage() for
2115 * the next read(). Here we 'discover' whether the page went
2116 * uptodate as a result of this (potentially partial) write.
2119 SetPageUptodate(page);
2124 * block_write_begin takes care of the basic task of block allocation and
2125 * bringing partial write blocks uptodate first.
2127 * The filesystem needs to handle block truncation upon failure.
2129 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2130 unsigned flags, struct page **pagep, get_block_t *get_block)
2132 pgoff_t index = pos >> PAGE_SHIFT;
2136 page = grab_cache_page_write_begin(mapping, index, flags);
2140 status = __block_write_begin(page, pos, len, get_block);
2141 if (unlikely(status)) {
2150 EXPORT_SYMBOL(block_write_begin);
2152 int block_write_end(struct file *file, struct address_space *mapping,
2153 loff_t pos, unsigned len, unsigned copied,
2154 struct page *page, void *fsdata)
2156 struct inode *inode = mapping->host;
2159 start = pos & (PAGE_SIZE - 1);
2161 if (unlikely(copied < len)) {
2163 * The buffers that were written will now be uptodate, so we
2164 * don't have to worry about a readpage reading them and
2165 * overwriting a partial write. However if we have encountered
2166 * a short write and only partially written into a buffer, it
2167 * will not be marked uptodate, so a readpage might come in and
2168 * destroy our partial write.
2170 * Do the simplest thing, and just treat any short write to a
2171 * non uptodate page as a zero-length write, and force the
2172 * caller to redo the whole thing.
2174 if (!PageUptodate(page))
2177 page_zero_new_buffers(page, start+copied, start+len);
2179 flush_dcache_page(page);
2181 /* This could be a short (even 0-length) commit */
2182 __block_commit_write(inode, page, start, start+copied);
2186 EXPORT_SYMBOL(block_write_end);
2188 int generic_write_end(struct file *file, struct address_space *mapping,
2189 loff_t pos, unsigned len, unsigned copied,
2190 struct page *page, void *fsdata)
2192 struct inode *inode = mapping->host;
2193 loff_t old_size = inode->i_size;
2194 int i_size_changed = 0;
2196 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2199 * No need to use i_size_read() here, the i_size
2200 * cannot change under us because we hold i_mutex.
2202 * But it's important to update i_size while still holding page lock:
2203 * page writeout could otherwise come in and zero beyond i_size.
2205 if (pos+copied > inode->i_size) {
2206 i_size_write(inode, pos+copied);
2214 pagecache_isize_extended(inode, old_size, pos);
2216 * Don't mark the inode dirty under page lock. First, it unnecessarily
2217 * makes the holding time of page lock longer. Second, it forces lock
2218 * ordering of page lock and transaction start for journaling
2222 mark_inode_dirty(inode);
2226 EXPORT_SYMBOL(generic_write_end);
2229 * block_is_partially_uptodate checks whether buffers within a page are
2232 * Returns true if all buffers which correspond to a file portion
2233 * we want to read are uptodate.
2235 int block_is_partially_uptodate(struct page *page, unsigned long from,
2236 unsigned long count)
2238 unsigned block_start, block_end, blocksize;
2240 struct buffer_head *bh, *head;
2243 if (!page_has_buffers(page))
2246 head = page_buffers(page);
2247 blocksize = head->b_size;
2248 to = min_t(unsigned, PAGE_SIZE - from, count);
2250 if (from < blocksize && to > PAGE_SIZE - blocksize)
2256 block_end = block_start + blocksize;
2257 if (block_end > from && block_start < to) {
2258 if (!buffer_uptodate(bh)) {
2262 if (block_end >= to)
2265 block_start = block_end;
2266 bh = bh->b_this_page;
2267 } while (bh != head);
2271 EXPORT_SYMBOL(block_is_partially_uptodate);
2274 * Generic "read page" function for block devices that have the normal
2275 * get_block functionality. This is most of the block device filesystems.
2276 * Reads the page asynchronously --- the unlock_buffer() and
2277 * set/clear_buffer_uptodate() functions propagate buffer state into the
2278 * page struct once IO has completed.
2280 int block_read_full_page(struct page *page, get_block_t *get_block)
2282 struct inode *inode = page->mapping->host;
2283 sector_t iblock, lblock;
2284 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2285 unsigned int blocksize, bbits;
2287 int fully_mapped = 1;
2289 head = create_page_buffers(page, inode, 0);
2290 blocksize = head->b_size;
2291 bbits = block_size_bits(blocksize);
2293 iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2294 lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2300 if (buffer_uptodate(bh))
2303 if (!buffer_mapped(bh)) {
2307 if (iblock < lblock) {
2308 WARN_ON(bh->b_size != blocksize);
2309 err = get_block(inode, iblock, bh, 0);
2313 if (!buffer_mapped(bh)) {
2314 zero_user(page, i * blocksize, blocksize);
2316 set_buffer_uptodate(bh);
2320 * get_block() might have updated the buffer
2323 if (buffer_uptodate(bh))
2327 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2330 SetPageMappedToDisk(page);
2334 * All buffers are uptodate - we can set the page uptodate
2335 * as well. But not if get_block() returned an error.
2337 if (!PageError(page))
2338 SetPageUptodate(page);
2343 /* Stage two: lock the buffers */
2344 for (i = 0; i < nr; i++) {
2347 mark_buffer_async_read(bh);
2351 * Stage 3: start the IO. Check for uptodateness
2352 * inside the buffer lock in case another process reading
2353 * the underlying blockdev brought it uptodate (the sct fix).
2355 for (i = 0; i < nr; i++) {
2357 if (buffer_uptodate(bh))
2358 end_buffer_async_read(bh, 1);
2360 submit_bh(REQ_OP_READ, 0, bh);
2364 EXPORT_SYMBOL(block_read_full_page);
2366 /* utility function for filesystems that need to do work on expanding
2367 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2368 * deal with the hole.
2370 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2372 struct address_space *mapping = inode->i_mapping;
2377 err = inode_newsize_ok(inode, size);
2381 err = pagecache_write_begin(NULL, mapping, size, 0,
2382 AOP_FLAG_CONT_EXPAND, &page, &fsdata);
2386 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2392 EXPORT_SYMBOL(generic_cont_expand_simple);
2394 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2395 loff_t pos, loff_t *bytes)
2397 struct inode *inode = mapping->host;
2398 unsigned int blocksize = i_blocksize(inode);
2401 pgoff_t index, curidx;
2403 unsigned zerofrom, offset, len;
2406 index = pos >> PAGE_SHIFT;
2407 offset = pos & ~PAGE_MASK;
2409 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2410 zerofrom = curpos & ~PAGE_MASK;
2411 if (zerofrom & (blocksize-1)) {
2412 *bytes |= (blocksize-1);
2415 len = PAGE_SIZE - zerofrom;
2417 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2421 zero_user(page, zerofrom, len);
2422 err = pagecache_write_end(file, mapping, curpos, len, len,
2429 balance_dirty_pages_ratelimited(mapping);
2431 if (unlikely(fatal_signal_pending(current))) {
2437 /* page covers the boundary, find the boundary offset */
2438 if (index == curidx) {
2439 zerofrom = curpos & ~PAGE_MASK;
2440 /* if we will expand the thing last block will be filled */
2441 if (offset <= zerofrom) {
2444 if (zerofrom & (blocksize-1)) {
2445 *bytes |= (blocksize-1);
2448 len = offset - zerofrom;
2450 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2454 zero_user(page, zerofrom, len);
2455 err = pagecache_write_end(file, mapping, curpos, len, len,
2467 * For moronic filesystems that do not allow holes in file.
2468 * We may have to extend the file.
2470 int cont_write_begin(struct file *file, struct address_space *mapping,
2471 loff_t pos, unsigned len, unsigned flags,
2472 struct page **pagep, void **fsdata,
2473 get_block_t *get_block, loff_t *bytes)
2475 struct inode *inode = mapping->host;
2476 unsigned int blocksize = i_blocksize(inode);
2477 unsigned int zerofrom;
2480 err = cont_expand_zero(file, mapping, pos, bytes);
2484 zerofrom = *bytes & ~PAGE_MASK;
2485 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2486 *bytes |= (blocksize-1);
2490 return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2492 EXPORT_SYMBOL(cont_write_begin);
2494 int block_commit_write(struct page *page, unsigned from, unsigned to)
2496 struct inode *inode = page->mapping->host;
2497 __block_commit_write(inode,page,from,to);
2500 EXPORT_SYMBOL(block_commit_write);
2503 * block_page_mkwrite() is not allowed to change the file size as it gets
2504 * called from a page fault handler when a page is first dirtied. Hence we must
2505 * be careful to check for EOF conditions here. We set the page up correctly
2506 * for a written page which means we get ENOSPC checking when writing into
2507 * holes and correct delalloc and unwritten extent mapping on filesystems that
2508 * support these features.
2510 * We are not allowed to take the i_mutex here so we have to play games to
2511 * protect against truncate races as the page could now be beyond EOF. Because
2512 * truncate writes the inode size before removing pages, once we have the
2513 * page lock we can determine safely if the page is beyond EOF. If it is not
2514 * beyond EOF, then the page is guaranteed safe against truncation until we
2517 * Direct callers of this function should protect against filesystem freezing
2518 * using sb_start_pagefault() - sb_end_pagefault() functions.
2520 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2521 get_block_t get_block)
2523 struct page *page = vmf->page;
2524 struct inode *inode = file_inode(vma->vm_file);
2530 size = i_size_read(inode);
2531 if ((page->mapping != inode->i_mapping) ||
2532 (page_offset(page) > size)) {
2533 /* We overload EFAULT to mean page got truncated */
2538 /* page is wholly or partially inside EOF */
2539 if (((page->index + 1) << PAGE_SHIFT) > size)
2540 end = size & ~PAGE_MASK;
2544 ret = __block_write_begin(page, 0, end, get_block);
2546 ret = block_commit_write(page, 0, end);
2548 if (unlikely(ret < 0))
2550 set_page_dirty(page);
2551 wait_for_stable_page(page);
2557 EXPORT_SYMBOL(block_page_mkwrite);
2560 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2561 * immediately, while under the page lock. So it needs a special end_io
2562 * handler which does not touch the bh after unlocking it.
2564 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2566 __end_buffer_read_notouch(bh, uptodate);
2570 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2571 * the page (converting it to circular linked list and taking care of page
2574 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2576 struct buffer_head *bh;
2578 BUG_ON(!PageLocked(page));
2580 spin_lock(&page->mapping->private_lock);
2583 if (PageDirty(page))
2584 set_buffer_dirty(bh);
2585 if (!bh->b_this_page)
2586 bh->b_this_page = head;
2587 bh = bh->b_this_page;
2588 } while (bh != head);
2589 attach_page_buffers(page, head);
2590 spin_unlock(&page->mapping->private_lock);
2594 * On entry, the page is fully not uptodate.
2595 * On exit the page is fully uptodate in the areas outside (from,to)
2596 * The filesystem needs to handle block truncation upon failure.
2598 int nobh_write_begin(struct address_space *mapping,
2599 loff_t pos, unsigned len, unsigned flags,
2600 struct page **pagep, void **fsdata,
2601 get_block_t *get_block)
2603 struct inode *inode = mapping->host;
2604 const unsigned blkbits = inode->i_blkbits;
2605 const unsigned blocksize = 1 << blkbits;
2606 struct buffer_head *head, *bh;
2610 unsigned block_in_page;
2611 unsigned block_start, block_end;
2612 sector_t block_in_file;
2615 int is_mapped_to_disk = 1;
2617 index = pos >> PAGE_SHIFT;
2618 from = pos & (PAGE_SIZE - 1);
2621 page = grab_cache_page_write_begin(mapping, index, flags);
2627 if (page_has_buffers(page)) {
2628 ret = __block_write_begin(page, pos, len, get_block);
2634 if (PageMappedToDisk(page))
2638 * Allocate buffers so that we can keep track of state, and potentially
2639 * attach them to the page if an error occurs. In the common case of
2640 * no error, they will just be freed again without ever being attached
2641 * to the page (which is all OK, because we're under the page lock).
2643 * Be careful: the buffer linked list is a NULL terminated one, rather
2644 * than the circular one we're used to.
2646 head = alloc_page_buffers(page, blocksize, 0);
2652 block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2655 * We loop across all blocks in the page, whether or not they are
2656 * part of the affected region. This is so we can discover if the
2657 * page is fully mapped-to-disk.
2659 for (block_start = 0, block_in_page = 0, bh = head;
2660 block_start < PAGE_SIZE;
2661 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2664 block_end = block_start + blocksize;
2667 if (block_start >= to)
2669 ret = get_block(inode, block_in_file + block_in_page,
2673 if (!buffer_mapped(bh))
2674 is_mapped_to_disk = 0;
2676 clean_bdev_bh_alias(bh);
2677 if (PageUptodate(page)) {
2678 set_buffer_uptodate(bh);
2681 if (buffer_new(bh) || !buffer_mapped(bh)) {
2682 zero_user_segments(page, block_start, from,
2686 if (buffer_uptodate(bh))
2687 continue; /* reiserfs does this */
2688 if (block_start < from || block_end > to) {
2690 bh->b_end_io = end_buffer_read_nobh;
2691 submit_bh(REQ_OP_READ, 0, bh);
2698 * The page is locked, so these buffers are protected from
2699 * any VM or truncate activity. Hence we don't need to care
2700 * for the buffer_head refcounts.
2702 for (bh = head; bh; bh = bh->b_this_page) {
2704 if (!buffer_uptodate(bh))
2711 if (is_mapped_to_disk)
2712 SetPageMappedToDisk(page);
2714 *fsdata = head; /* to be released by nobh_write_end */
2721 * Error recovery is a bit difficult. We need to zero out blocks that
2722 * were newly allocated, and dirty them to ensure they get written out.
2723 * Buffers need to be attached to the page at this point, otherwise
2724 * the handling of potential IO errors during writeout would be hard
2725 * (could try doing synchronous writeout, but what if that fails too?)
2727 attach_nobh_buffers(page, head);
2728 page_zero_new_buffers(page, from, to);
2737 EXPORT_SYMBOL(nobh_write_begin);
2739 int nobh_write_end(struct file *file, struct address_space *mapping,
2740 loff_t pos, unsigned len, unsigned copied,
2741 struct page *page, void *fsdata)
2743 struct inode *inode = page->mapping->host;
2744 struct buffer_head *head = fsdata;
2745 struct buffer_head *bh;
2746 BUG_ON(fsdata != NULL && page_has_buffers(page));
2748 if (unlikely(copied < len) && head)
2749 attach_nobh_buffers(page, head);
2750 if (page_has_buffers(page))
2751 return generic_write_end(file, mapping, pos, len,
2752 copied, page, fsdata);
2754 SetPageUptodate(page);
2755 set_page_dirty(page);
2756 if (pos+copied > inode->i_size) {
2757 i_size_write(inode, pos+copied);
2758 mark_inode_dirty(inode);
2766 head = head->b_this_page;
2767 free_buffer_head(bh);
2772 EXPORT_SYMBOL(nobh_write_end);
2775 * nobh_writepage() - based on block_full_write_page() except
2776 * that it tries to operate without attaching bufferheads to
2779 int nobh_writepage(struct page *page, get_block_t *get_block,
2780 struct writeback_control *wbc)
2782 struct inode * const inode = page->mapping->host;
2783 loff_t i_size = i_size_read(inode);
2784 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2788 /* Is the page fully inside i_size? */
2789 if (page->index < end_index)
2792 /* Is the page fully outside i_size? (truncate in progress) */
2793 offset = i_size & (PAGE_SIZE-1);
2794 if (page->index >= end_index+1 || !offset) {
2796 * The page may have dirty, unmapped buffers. For example,
2797 * they may have been added in ext3_writepage(). Make them
2798 * freeable here, so the page does not leak.
2801 /* Not really sure about this - do we need this ? */
2802 if (page->mapping->a_ops->invalidatepage)
2803 page->mapping->a_ops->invalidatepage(page, offset);
2806 return 0; /* don't care */
2810 * The page straddles i_size. It must be zeroed out on each and every
2811 * writepage invocation because it may be mmapped. "A file is mapped
2812 * in multiples of the page size. For a file that is not a multiple of
2813 * the page size, the remaining memory is zeroed when mapped, and
2814 * writes to that region are not written out to the file."
2816 zero_user_segment(page, offset, PAGE_SIZE);
2818 ret = mpage_writepage(page, get_block, wbc);
2820 ret = __block_write_full_page(inode, page, get_block, wbc,
2821 end_buffer_async_write);
2824 EXPORT_SYMBOL(nobh_writepage);
2826 int nobh_truncate_page(struct address_space *mapping,
2827 loff_t from, get_block_t *get_block)
2829 pgoff_t index = from >> PAGE_SHIFT;
2830 unsigned offset = from & (PAGE_SIZE-1);
2833 unsigned length, pos;
2834 struct inode *inode = mapping->host;
2836 struct buffer_head map_bh;
2839 blocksize = i_blocksize(inode);
2840 length = offset & (blocksize - 1);
2842 /* Block boundary? Nothing to do */
2846 length = blocksize - length;
2847 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2849 page = grab_cache_page(mapping, index);
2854 if (page_has_buffers(page)) {
2858 return block_truncate_page(mapping, from, get_block);
2861 /* Find the buffer that contains "offset" */
2863 while (offset >= pos) {
2868 map_bh.b_size = blocksize;
2870 err = get_block(inode, iblock, &map_bh, 0);
2873 /* unmapped? It's a hole - nothing to do */
2874 if (!buffer_mapped(&map_bh))
2877 /* Ok, it's mapped. Make sure it's up-to-date */
2878 if (!PageUptodate(page)) {
2879 err = mapping->a_ops->readpage(NULL, page);
2885 if (!PageUptodate(page)) {
2889 if (page_has_buffers(page))
2892 zero_user(page, offset, length);
2893 set_page_dirty(page);
2902 EXPORT_SYMBOL(nobh_truncate_page);
2904 int block_truncate_page(struct address_space *mapping,
2905 loff_t from, get_block_t *get_block)
2907 pgoff_t index = from >> PAGE_SHIFT;
2908 unsigned offset = from & (PAGE_SIZE-1);
2911 unsigned length, pos;
2912 struct inode *inode = mapping->host;
2914 struct buffer_head *bh;
2917 blocksize = i_blocksize(inode);
2918 length = offset & (blocksize - 1);
2920 /* Block boundary? Nothing to do */
2924 length = blocksize - length;
2925 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2927 page = grab_cache_page(mapping, index);
2932 if (!page_has_buffers(page))
2933 create_empty_buffers(page, blocksize, 0);
2935 /* Find the buffer that contains "offset" */
2936 bh = page_buffers(page);
2938 while (offset >= pos) {
2939 bh = bh->b_this_page;
2945 if (!buffer_mapped(bh)) {
2946 WARN_ON(bh->b_size != blocksize);
2947 err = get_block(inode, iblock, bh, 0);
2950 /* unmapped? It's a hole - nothing to do */
2951 if (!buffer_mapped(bh))
2955 /* Ok, it's mapped. Make sure it's up-to-date */
2956 if (PageUptodate(page))
2957 set_buffer_uptodate(bh);
2959 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2961 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2963 /* Uhhuh. Read error. Complain and punt. */
2964 if (!buffer_uptodate(bh))
2968 zero_user(page, offset, length);
2969 mark_buffer_dirty(bh);
2978 EXPORT_SYMBOL(block_truncate_page);
2981 * The generic ->writepage function for buffer-backed address_spaces
2983 int block_write_full_page(struct page *page, get_block_t *get_block,
2984 struct writeback_control *wbc)
2986 struct inode * const inode = page->mapping->host;
2987 loff_t i_size = i_size_read(inode);
2988 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2991 /* Is the page fully inside i_size? */
2992 if (page->index < end_index)
2993 return __block_write_full_page(inode, page, get_block, wbc,
2994 end_buffer_async_write);
2996 /* Is the page fully outside i_size? (truncate in progress) */
2997 offset = i_size & (PAGE_SIZE-1);
2998 if (page->index >= end_index+1 || !offset) {
3000 * The page may have dirty, unmapped buffers. For example,
3001 * they may have been added in ext3_writepage(). Make them
3002 * freeable here, so the page does not leak.
3004 do_invalidatepage(page, 0, PAGE_SIZE);
3006 return 0; /* don't care */
3010 * The page straddles i_size. It must be zeroed out on each and every
3011 * writepage invocation because it may be mmapped. "A file is mapped
3012 * in multiples of the page size. For a file that is not a multiple of
3013 * the page size, the remaining memory is zeroed when mapped, and
3014 * writes to that region are not written out to the file."
3016 zero_user_segment(page, offset, PAGE_SIZE);
3017 return __block_write_full_page(inode, page, get_block, wbc,
3018 end_buffer_async_write);
3020 EXPORT_SYMBOL(block_write_full_page);
3022 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
3023 get_block_t *get_block)
3025 struct buffer_head tmp;
3026 struct inode *inode = mapping->host;
3029 tmp.b_size = i_blocksize(inode);
3030 get_block(inode, block, &tmp, 0);
3031 return tmp.b_blocknr;
3033 EXPORT_SYMBOL(generic_block_bmap);
3035 static void end_bio_bh_io_sync(struct bio *bio)
3037 struct buffer_head *bh = bio->bi_private;
3039 if (unlikely(bio_flagged(bio, BIO_QUIET)))
3040 set_bit(BH_Quiet, &bh->b_state);
3042 bh->b_end_io(bh, !bio->bi_error);
3047 * This allows us to do IO even on the odd last sectors
3048 * of a device, even if the block size is some multiple
3049 * of the physical sector size.
3051 * We'll just truncate the bio to the size of the device,
3052 * and clear the end of the buffer head manually.
3054 * Truly out-of-range accesses will turn into actual IO
3055 * errors, this only handles the "we need to be able to
3056 * do IO at the final sector" case.
3058 void guard_bio_eod(int op, struct bio *bio)
3061 struct bio_vec *bvec = &bio->bi_io_vec[bio->bi_vcnt - 1];
3062 unsigned truncated_bytes;
3064 maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
3069 * If the *whole* IO is past the end of the device,
3070 * let it through, and the IO layer will turn it into
3073 if (unlikely(bio->bi_iter.bi_sector >= maxsector))
3076 maxsector -= bio->bi_iter.bi_sector;
3077 if (likely((bio->bi_iter.bi_size >> 9) <= maxsector))
3080 /* Uhhuh. We've got a bio that straddles the device size! */
3081 truncated_bytes = bio->bi_iter.bi_size - (maxsector << 9);
3083 /* Truncate the bio.. */
3084 bio->bi_iter.bi_size -= truncated_bytes;
3085 bvec->bv_len -= truncated_bytes;
3087 /* ..and clear the end of the buffer for reads */
3088 if (op == REQ_OP_READ) {
3089 zero_user(bvec->bv_page, bvec->bv_offset + bvec->bv_len,
3094 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3095 unsigned long bio_flags, struct writeback_control *wbc)
3099 BUG_ON(!buffer_locked(bh));
3100 BUG_ON(!buffer_mapped(bh));
3101 BUG_ON(!bh->b_end_io);
3102 BUG_ON(buffer_delay(bh));
3103 BUG_ON(buffer_unwritten(bh));
3106 * Only clear out a write error when rewriting
3108 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3109 clear_buffer_write_io_error(bh);
3112 * from here on down, it's all bio -- do the initial mapping,
3113 * submit_bio -> generic_make_request may further map this bio around
3115 bio = bio_alloc(GFP_NOIO, 1);
3118 wbc_init_bio(wbc, bio);
3119 wbc_account_io(wbc, bh->b_page, bh->b_size);
3122 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3123 bio->bi_bdev = bh->b_bdev;
3125 bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3126 BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3128 bio->bi_end_io = end_bio_bh_io_sync;
3129 bio->bi_private = bh;
3130 bio->bi_flags |= bio_flags;
3132 /* Take care of bh's that straddle the end of the device */
3133 guard_bio_eod(op, bio);
3135 if (buffer_meta(bh))
3136 op_flags |= REQ_META;
3137 if (buffer_prio(bh))
3138 op_flags |= REQ_PRIO;
3139 bio_set_op_attrs(bio, op, op_flags);
3145 int _submit_bh(int op, int op_flags, struct buffer_head *bh,
3146 unsigned long bio_flags)
3148 return submit_bh_wbc(op, op_flags, bh, bio_flags, NULL);
3150 EXPORT_SYMBOL_GPL(_submit_bh);
3152 int submit_bh(int op, int op_flags, struct buffer_head *bh)
3154 return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3156 EXPORT_SYMBOL(submit_bh);
3159 * ll_rw_block: low-level access to block devices (DEPRECATED)
3160 * @op: whether to %READ or %WRITE
3161 * @op_flags: req_flag_bits
3162 * @nr: number of &struct buffer_heads in the array
3163 * @bhs: array of pointers to &struct buffer_head
3165 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3166 * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3167 * @op_flags contains flags modifying the detailed I/O behavior, most notably
3170 * This function drops any buffer that it cannot get a lock on (with the
3171 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3172 * request, and any buffer that appears to be up-to-date when doing read
3173 * request. Further it marks as clean buffers that are processed for
3174 * writing (the buffer cache won't assume that they are actually clean
3175 * until the buffer gets unlocked).
3177 * ll_rw_block sets b_end_io to simple completion handler that marks
3178 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3181 * All of the buffers must be for the same device, and must also be a
3182 * multiple of the current approved size for the device.
3184 void ll_rw_block(int op, int op_flags, int nr, struct buffer_head *bhs[])
3188 for (i = 0; i < nr; i++) {
3189 struct buffer_head *bh = bhs[i];
3191 if (!trylock_buffer(bh))
3194 if (test_clear_buffer_dirty(bh)) {
3195 bh->b_end_io = end_buffer_write_sync;
3197 submit_bh(op, op_flags, bh);
3201 if (!buffer_uptodate(bh)) {
3202 bh->b_end_io = end_buffer_read_sync;
3204 submit_bh(op, op_flags, bh);
3211 EXPORT_SYMBOL(ll_rw_block);
3213 void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3216 if (!test_clear_buffer_dirty(bh)) {
3220 bh->b_end_io = end_buffer_write_sync;
3222 submit_bh(REQ_OP_WRITE, op_flags, bh);
3224 EXPORT_SYMBOL(write_dirty_buffer);
3227 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3228 * and then start new I/O and then wait upon it. The caller must have a ref on
3231 int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3235 WARN_ON(atomic_read(&bh->b_count) < 1);
3237 if (test_clear_buffer_dirty(bh)) {
3239 bh->b_end_io = end_buffer_write_sync;
3240 ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3242 if (!ret && !buffer_uptodate(bh))
3249 EXPORT_SYMBOL(__sync_dirty_buffer);
3251 int sync_dirty_buffer(struct buffer_head *bh)
3253 return __sync_dirty_buffer(bh, REQ_SYNC);
3255 EXPORT_SYMBOL(sync_dirty_buffer);
3258 * try_to_free_buffers() checks if all the buffers on this particular page
3259 * are unused, and releases them if so.
3261 * Exclusion against try_to_free_buffers may be obtained by either
3262 * locking the page or by holding its mapping's private_lock.
3264 * If the page is dirty but all the buffers are clean then we need to
3265 * be sure to mark the page clean as well. This is because the page
3266 * may be against a block device, and a later reattachment of buffers
3267 * to a dirty page will set *all* buffers dirty. Which would corrupt
3268 * filesystem data on the same device.
3270 * The same applies to regular filesystem pages: if all the buffers are
3271 * clean then we set the page clean and proceed. To do that, we require
3272 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3275 * try_to_free_buffers() is non-blocking.
3277 static inline int buffer_busy(struct buffer_head *bh)
3279 return atomic_read(&bh->b_count) |
3280 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3284 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3286 struct buffer_head *head = page_buffers(page);
3287 struct buffer_head *bh;
3291 if (buffer_write_io_error(bh) && page->mapping)
3292 mapping_set_error(page->mapping, -EIO);
3293 if (buffer_busy(bh))
3295 bh = bh->b_this_page;
3296 } while (bh != head);
3299 struct buffer_head *next = bh->b_this_page;
3301 if (bh->b_assoc_map)
3302 __remove_assoc_queue(bh);
3304 } while (bh != head);
3305 *buffers_to_free = head;
3306 __clear_page_buffers(page);
3312 int try_to_free_buffers(struct page *page)
3314 struct address_space * const mapping = page->mapping;
3315 struct buffer_head *buffers_to_free = NULL;
3318 BUG_ON(!PageLocked(page));
3319 if (PageWriteback(page))
3322 if (mapping == NULL) { /* can this still happen? */
3323 ret = drop_buffers(page, &buffers_to_free);
3327 spin_lock(&mapping->private_lock);
3328 ret = drop_buffers(page, &buffers_to_free);
3331 * If the filesystem writes its buffers by hand (eg ext3)
3332 * then we can have clean buffers against a dirty page. We
3333 * clean the page here; otherwise the VM will never notice
3334 * that the filesystem did any IO at all.
3336 * Also, during truncate, discard_buffer will have marked all
3337 * the page's buffers clean. We discover that here and clean
3340 * private_lock must be held over this entire operation in order
3341 * to synchronise against __set_page_dirty_buffers and prevent the
3342 * dirty bit from being lost.
3345 cancel_dirty_page(page);
3346 spin_unlock(&mapping->private_lock);
3348 if (buffers_to_free) {
3349 struct buffer_head *bh = buffers_to_free;
3352 struct buffer_head *next = bh->b_this_page;
3353 free_buffer_head(bh);
3355 } while (bh != buffers_to_free);
3359 EXPORT_SYMBOL(try_to_free_buffers);
3362 * There are no bdflush tunables left. But distributions are
3363 * still running obsolete flush daemons, so we terminate them here.
3365 * Use of bdflush() is deprecated and will be removed in a future kernel.
3366 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3368 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3370 static int msg_count;
3372 if (!capable(CAP_SYS_ADMIN))
3375 if (msg_count < 5) {
3378 "warning: process `%s' used the obsolete bdflush"
3379 " system call\n", current->comm);
3380 printk(KERN_INFO "Fix your initscripts?\n");
3389 * Buffer-head allocation
3391 static struct kmem_cache *bh_cachep __read_mostly;
3394 * Once the number of bh's in the machine exceeds this level, we start
3395 * stripping them in writeback.
3397 static unsigned long max_buffer_heads;
3399 int buffer_heads_over_limit;
3401 struct bh_accounting {
3402 int nr; /* Number of live bh's */
3403 int ratelimit; /* Limit cacheline bouncing */
3406 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3408 static void recalc_bh_state(void)
3413 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3415 __this_cpu_write(bh_accounting.ratelimit, 0);
3416 for_each_online_cpu(i)
3417 tot += per_cpu(bh_accounting, i).nr;
3418 buffer_heads_over_limit = (tot > max_buffer_heads);
3421 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3423 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3425 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3427 __this_cpu_inc(bh_accounting.nr);
3433 EXPORT_SYMBOL(alloc_buffer_head);
3435 void free_buffer_head(struct buffer_head *bh)
3437 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3438 kmem_cache_free(bh_cachep, bh);
3440 __this_cpu_dec(bh_accounting.nr);
3444 EXPORT_SYMBOL(free_buffer_head);
3446 static int buffer_exit_cpu_dead(unsigned int cpu)
3449 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3451 for (i = 0; i < BH_LRU_SIZE; i++) {
3455 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3456 per_cpu(bh_accounting, cpu).nr = 0;
3461 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3462 * @bh: struct buffer_head
3464 * Return true if the buffer is up-to-date and false,
3465 * with the buffer locked, if not.
3467 int bh_uptodate_or_lock(struct buffer_head *bh)
3469 if (!buffer_uptodate(bh)) {
3471 if (!buffer_uptodate(bh))
3477 EXPORT_SYMBOL(bh_uptodate_or_lock);
3480 * bh_submit_read - Submit a locked buffer for reading
3481 * @bh: struct buffer_head
3483 * Returns zero on success and -EIO on error.
3485 int bh_submit_read(struct buffer_head *bh)
3487 BUG_ON(!buffer_locked(bh));
3489 if (buffer_uptodate(bh)) {
3495 bh->b_end_io = end_buffer_read_sync;
3496 submit_bh(REQ_OP_READ, 0, bh);
3498 if (buffer_uptodate(bh))
3502 EXPORT_SYMBOL(bh_submit_read);
3504 void __init buffer_init(void)
3506 unsigned long nrpages;
3509 bh_cachep = kmem_cache_create("buffer_head",
3510 sizeof(struct buffer_head), 0,
3511 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3516 * Limit the bh occupancy to 10% of ZONE_NORMAL
3518 nrpages = (nr_free_buffer_pages() * 10) / 100;
3519 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3520 ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3521 NULL, buffer_exit_cpu_dead);