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 void init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
58 bh->b_end_io = handler;
59 bh->b_private = private;
61 EXPORT_SYMBOL(init_buffer);
63 inline void touch_buffer(struct buffer_head *bh)
65 trace_block_touch_buffer(bh);
66 mark_page_accessed(bh->b_page);
68 EXPORT_SYMBOL(touch_buffer);
70 void __lock_buffer(struct buffer_head *bh)
72 wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
74 EXPORT_SYMBOL(__lock_buffer);
76 void unlock_buffer(struct buffer_head *bh)
78 clear_bit_unlock(BH_Lock, &bh->b_state);
79 smp_mb__after_atomic();
80 wake_up_bit(&bh->b_state, BH_Lock);
82 EXPORT_SYMBOL(unlock_buffer);
85 * Returns if the page has dirty or writeback buffers. If all the buffers
86 * are unlocked and clean then the PageDirty information is stale. If
87 * any of the pages are locked, it is assumed they are locked for IO.
89 void buffer_check_dirty_writeback(struct page *page,
90 bool *dirty, bool *writeback)
92 struct buffer_head *head, *bh;
96 BUG_ON(!PageLocked(page));
98 if (!page_has_buffers(page))
101 if (PageWriteback(page))
104 head = page_buffers(page);
107 if (buffer_locked(bh))
110 if (buffer_dirty(bh))
113 bh = bh->b_this_page;
114 } while (bh != head);
116 EXPORT_SYMBOL(buffer_check_dirty_writeback);
119 * Block until a buffer comes unlocked. This doesn't stop it
120 * from becoming locked again - you have to lock it yourself
121 * if you want to preserve its state.
123 void __wait_on_buffer(struct buffer_head * bh)
125 wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
127 EXPORT_SYMBOL(__wait_on_buffer);
130 __clear_page_buffers(struct page *page)
132 ClearPagePrivate(page);
133 set_page_private(page, 0);
137 static void buffer_io_error(struct buffer_head *bh, char *msg)
139 if (!test_bit(BH_Quiet, &bh->b_state))
140 printk_ratelimited(KERN_ERR
141 "Buffer I/O error on dev %pg, logical block %llu%s\n",
142 bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
146 * End-of-IO handler helper function which does not touch the bh after
148 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
149 * a race there is benign: unlock_buffer() only use the bh's address for
150 * hashing after unlocking the buffer, so it doesn't actually touch the bh
153 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
156 set_buffer_uptodate(bh);
158 /* This happens, due to failed read-ahead attempts. */
159 clear_buffer_uptodate(bh);
165 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
166 * unlock the buffer. This is what ll_rw_block uses too.
168 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
170 __end_buffer_read_notouch(bh, uptodate);
173 EXPORT_SYMBOL(end_buffer_read_sync);
175 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
178 set_buffer_uptodate(bh);
180 buffer_io_error(bh, ", lost sync page write");
181 mark_buffer_write_io_error(bh);
182 clear_buffer_uptodate(bh);
187 EXPORT_SYMBOL(end_buffer_write_sync);
190 * Various filesystems appear to want __find_get_block to be non-blocking.
191 * But it's the page lock which protects the buffers. To get around this,
192 * we get exclusion from try_to_free_buffers with the blockdev mapping's
195 * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
196 * may be quite high. This code could TryLock the page, and if that
197 * succeeds, there is no need to take private_lock. (But if
198 * private_lock is contended then so is mapping->tree_lock).
200 static struct buffer_head *
201 __find_get_block_slow(struct block_device *bdev, sector_t block)
203 struct inode *bd_inode = bdev->bd_inode;
204 struct address_space *bd_mapping = bd_inode->i_mapping;
205 struct buffer_head *ret = NULL;
207 struct buffer_head *bh;
208 struct buffer_head *head;
211 static DEFINE_RATELIMIT_STATE(last_warned, HZ, 1);
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
239 ratelimit_set_flags(&last_warned, RATELIMIT_MSG_ON_RELEASE);
240 if (all_mapped && __ratelimit(&last_warned)) {
241 printk("__find_get_block_slow() failed. block=%llu, "
242 "b_blocknr=%llu, b_state=0x%08lx, b_size=%zu, "
243 "device %pg blocksize: %d\n",
244 (unsigned long long)block,
245 (unsigned long long)bh->b_blocknr,
246 bh->b_state, bh->b_size, 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 mark_buffer_write_io_error(bh);
357 clear_buffer_uptodate(bh);
361 first = page_buffers(page);
362 local_irq_save(flags);
363 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
365 clear_buffer_async_write(bh);
367 tmp = bh->b_this_page;
369 if (buffer_async_write(tmp)) {
370 BUG_ON(!buffer_locked(tmp));
373 tmp = tmp->b_this_page;
375 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
376 local_irq_restore(flags);
377 end_page_writeback(page);
381 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
382 local_irq_restore(flags);
385 EXPORT_SYMBOL(end_buffer_async_write);
388 * If a page's buffers are under async readin (end_buffer_async_read
389 * completion) then there is a possibility that another thread of
390 * control could lock one of the buffers after it has completed
391 * but while some of the other buffers have not completed. This
392 * locked buffer would confuse end_buffer_async_read() into not unlocking
393 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
394 * that this buffer is not under async I/O.
396 * The page comes unlocked when it has no locked buffer_async buffers
399 * PageLocked prevents anyone starting new async I/O reads any of
402 * PageWriteback is used to prevent simultaneous writeout of the same
405 * PageLocked prevents anyone from starting writeback of a page which is
406 * under read I/O (PageWriteback is only ever set against a locked page).
408 static void mark_buffer_async_read(struct buffer_head *bh)
410 bh->b_end_io = end_buffer_async_read;
411 set_buffer_async_read(bh);
414 static void mark_buffer_async_write_endio(struct buffer_head *bh,
415 bh_end_io_t *handler)
417 bh->b_end_io = handler;
418 set_buffer_async_write(bh);
421 void mark_buffer_async_write(struct buffer_head *bh)
423 mark_buffer_async_write_endio(bh, end_buffer_async_write);
425 EXPORT_SYMBOL(mark_buffer_async_write);
429 * fs/buffer.c contains helper functions for buffer-backed address space's
430 * fsync functions. A common requirement for buffer-based filesystems is
431 * that certain data from the backing blockdev needs to be written out for
432 * a successful fsync(). For example, ext2 indirect blocks need to be
433 * written back and waited upon before fsync() returns.
435 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
436 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
437 * management of a list of dependent buffers at ->i_mapping->private_list.
439 * Locking is a little subtle: try_to_free_buffers() will remove buffers
440 * from their controlling inode's queue when they are being freed. But
441 * try_to_free_buffers() will be operating against the *blockdev* mapping
442 * at the time, not against the S_ISREG file which depends on those buffers.
443 * So the locking for private_list is via the private_lock in the address_space
444 * which backs the buffers. Which is different from the address_space
445 * against which the buffers are listed. So for a particular address_space,
446 * mapping->private_lock does *not* protect mapping->private_list! In fact,
447 * mapping->private_list will always be protected by the backing blockdev's
450 * Which introduces a requirement: all buffers on an address_space's
451 * ->private_list must be from the same address_space: the blockdev's.
453 * address_spaces which do not place buffers at ->private_list via these
454 * utility functions are free to use private_lock and private_list for
455 * whatever they want. The only requirement is that list_empty(private_list)
456 * be true at clear_inode() time.
458 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
459 * filesystems should do that. invalidate_inode_buffers() should just go
460 * BUG_ON(!list_empty).
462 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
463 * take an address_space, not an inode. And it should be called
464 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
467 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
468 * list if it is already on a list. Because if the buffer is on a list,
469 * it *must* already be on the right one. If not, the filesystem is being
470 * silly. This will save a ton of locking. But first we have to ensure
471 * that buffers are taken *off* the old inode's list when they are freed
472 * (presumably in truncate). That requires careful auditing of all
473 * filesystems (do it inside bforget()). It could also be done by bringing
478 * The buffer's backing address_space's private_lock must be held
480 static void __remove_assoc_queue(struct buffer_head *bh)
482 list_del_init(&bh->b_assoc_buffers);
483 WARN_ON(!bh->b_assoc_map);
484 bh->b_assoc_map = NULL;
487 int inode_has_buffers(struct inode *inode)
489 return !list_empty(&inode->i_data.private_list);
493 * osync is designed to support O_SYNC io. It waits synchronously for
494 * all already-submitted IO to complete, but does not queue any new
495 * writes to the disk.
497 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
498 * you dirty the buffers, and then use osync_inode_buffers to wait for
499 * completion. Any other dirty buffers which are not yet queued for
500 * write will not be flushed to disk by the osync.
502 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
504 struct buffer_head *bh;
510 list_for_each_prev(p, list) {
512 if (buffer_locked(bh)) {
516 if (!buffer_uptodate(bh))
527 static void do_thaw_one(struct super_block *sb, void *unused)
529 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
530 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
533 static void do_thaw_all(struct work_struct *work)
535 iterate_supers(do_thaw_one, NULL);
537 printk(KERN_WARNING "Emergency Thaw complete\n");
541 * emergency_thaw_all -- forcibly thaw every frozen filesystem
543 * Used for emergency unfreeze of all filesystems via SysRq
545 void emergency_thaw_all(void)
547 struct work_struct *work;
549 work = kmalloc(sizeof(*work), GFP_ATOMIC);
551 INIT_WORK(work, do_thaw_all);
557 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
558 * @mapping: the mapping which wants those buffers written
560 * Starts I/O against the buffers at mapping->private_list, and waits upon
563 * Basically, this is a convenience function for fsync().
564 * @mapping is a file or directory which needs those buffers to be written for
565 * a successful fsync().
567 int sync_mapping_buffers(struct address_space *mapping)
569 struct address_space *buffer_mapping = mapping->private_data;
571 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
574 return fsync_buffers_list(&buffer_mapping->private_lock,
575 &mapping->private_list);
577 EXPORT_SYMBOL(sync_mapping_buffers);
580 * Called when we've recently written block `bblock', and it is known that
581 * `bblock' was for a buffer_boundary() buffer. This means that the block at
582 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
583 * dirty, schedule it for IO. So that indirects merge nicely with their data.
585 void write_boundary_block(struct block_device *bdev,
586 sector_t bblock, unsigned blocksize)
588 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
590 if (buffer_dirty(bh))
591 ll_rw_block(REQ_OP_WRITE, 0, 1, &bh);
596 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
598 struct address_space *mapping = inode->i_mapping;
599 struct address_space *buffer_mapping = bh->b_page->mapping;
601 mark_buffer_dirty(bh);
602 if (!mapping->private_data) {
603 mapping->private_data = buffer_mapping;
605 BUG_ON(mapping->private_data != buffer_mapping);
607 if (!bh->b_assoc_map) {
608 spin_lock(&buffer_mapping->private_lock);
609 list_move_tail(&bh->b_assoc_buffers,
610 &mapping->private_list);
611 bh->b_assoc_map = mapping;
612 spin_unlock(&buffer_mapping->private_lock);
615 EXPORT_SYMBOL(mark_buffer_dirty_inode);
618 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
621 * If warn is true, then emit a warning if the page is not uptodate and has
622 * not been truncated.
624 * The caller must hold lock_page_memcg().
626 static void __set_page_dirty(struct page *page, struct address_space *mapping,
631 spin_lock_irqsave(&mapping->tree_lock, flags);
632 if (page->mapping) { /* Race with truncate? */
633 WARN_ON_ONCE(warn && !PageUptodate(page));
634 account_page_dirtied(page, mapping);
635 radix_tree_tag_set(&mapping->page_tree,
636 page_index(page), PAGECACHE_TAG_DIRTY);
638 spin_unlock_irqrestore(&mapping->tree_lock, flags);
642 * Add a page to the dirty page list.
644 * It is a sad fact of life that this function is called from several places
645 * deeply under spinlocking. It may not sleep.
647 * If the page has buffers, the uptodate buffers are set dirty, to preserve
648 * dirty-state coherency between the page and the buffers. It the page does
649 * not have buffers then when they are later attached they will all be set
652 * The buffers are dirtied before the page is dirtied. There's a small race
653 * window in which a writepage caller may see the page cleanness but not the
654 * buffer dirtiness. That's fine. If this code were to set the page dirty
655 * before the buffers, a concurrent writepage caller could clear the page dirty
656 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
657 * page on the dirty page list.
659 * We use private_lock to lock against try_to_free_buffers while using the
660 * page's buffer list. Also use this to protect against clean buffers being
661 * added to the page after it was set dirty.
663 * FIXME: may need to call ->reservepage here as well. That's rather up to the
664 * address_space though.
666 int __set_page_dirty_buffers(struct page *page)
669 struct address_space *mapping = page_mapping(page);
671 if (unlikely(!mapping))
672 return !TestSetPageDirty(page);
674 spin_lock(&mapping->private_lock);
675 if (page_has_buffers(page)) {
676 struct buffer_head *head = page_buffers(page);
677 struct buffer_head *bh = head;
680 set_buffer_dirty(bh);
681 bh = bh->b_this_page;
682 } while (bh != head);
685 * Lock out page->mem_cgroup migration to keep PageDirty
686 * synchronized with per-memcg dirty page counters.
688 lock_page_memcg(page);
689 newly_dirty = !TestSetPageDirty(page);
690 spin_unlock(&mapping->private_lock);
693 __set_page_dirty(page, mapping, 1);
695 unlock_page_memcg(page);
698 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
702 EXPORT_SYMBOL(__set_page_dirty_buffers);
705 * Write out and wait upon a list of buffers.
707 * We have conflicting pressures: we want to make sure that all
708 * initially dirty buffers get waited on, but that any subsequently
709 * dirtied buffers don't. After all, we don't want fsync to last
710 * forever if somebody is actively writing to the file.
712 * Do this in two main stages: first we copy dirty buffers to a
713 * temporary inode list, queueing the writes as we go. Then we clean
714 * up, waiting for those writes to complete.
716 * During this second stage, any subsequent updates to the file may end
717 * up refiling the buffer on the original inode's dirty list again, so
718 * there is a chance we will end up with a buffer queued for write but
719 * not yet completed on that list. So, as a final cleanup we go through
720 * the osync code to catch these locked, dirty buffers without requeuing
721 * any newly dirty buffers for write.
723 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
725 struct buffer_head *bh;
726 struct list_head tmp;
727 struct address_space *mapping;
729 struct blk_plug plug;
731 INIT_LIST_HEAD(&tmp);
732 blk_start_plug(&plug);
735 while (!list_empty(list)) {
736 bh = BH_ENTRY(list->next);
737 mapping = bh->b_assoc_map;
738 __remove_assoc_queue(bh);
739 /* Avoid race with mark_buffer_dirty_inode() which does
740 * a lockless check and we rely on seeing the dirty bit */
742 if (buffer_dirty(bh) || buffer_locked(bh)) {
743 list_add(&bh->b_assoc_buffers, &tmp);
744 bh->b_assoc_map = mapping;
745 if (buffer_dirty(bh)) {
749 * Ensure any pending I/O completes so that
750 * write_dirty_buffer() actually writes the
751 * current contents - it is a noop if I/O is
752 * still in flight on potentially older
755 write_dirty_buffer(bh, REQ_SYNC);
758 * Kick off IO for the previous mapping. Note
759 * that we will not run the very last mapping,
760 * wait_on_buffer() will do that for us
761 * through sync_buffer().
770 blk_finish_plug(&plug);
773 while (!list_empty(&tmp)) {
774 bh = BH_ENTRY(tmp.prev);
776 mapping = bh->b_assoc_map;
777 __remove_assoc_queue(bh);
778 /* Avoid race with mark_buffer_dirty_inode() which does
779 * a lockless check and we rely on seeing the dirty bit */
781 if (buffer_dirty(bh)) {
782 list_add(&bh->b_assoc_buffers,
783 &mapping->private_list);
784 bh->b_assoc_map = mapping;
788 if (!buffer_uptodate(bh))
795 err2 = osync_buffers_list(lock, list);
803 * Invalidate any and all dirty buffers on a given inode. We are
804 * probably unmounting the fs, but that doesn't mean we have already
805 * done a sync(). Just drop the buffers from the inode list.
807 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
808 * assumes that all the buffers are against the blockdev. Not true
811 void invalidate_inode_buffers(struct inode *inode)
813 if (inode_has_buffers(inode)) {
814 struct address_space *mapping = &inode->i_data;
815 struct list_head *list = &mapping->private_list;
816 struct address_space *buffer_mapping = mapping->private_data;
818 spin_lock(&buffer_mapping->private_lock);
819 while (!list_empty(list))
820 __remove_assoc_queue(BH_ENTRY(list->next));
821 spin_unlock(&buffer_mapping->private_lock);
824 EXPORT_SYMBOL(invalidate_inode_buffers);
827 * Remove any clean buffers from the inode's buffer list. This is called
828 * when we're trying to free the inode itself. Those buffers can pin it.
830 * Returns true if all buffers were removed.
832 int remove_inode_buffers(struct inode *inode)
836 if (inode_has_buffers(inode)) {
837 struct address_space *mapping = &inode->i_data;
838 struct list_head *list = &mapping->private_list;
839 struct address_space *buffer_mapping = mapping->private_data;
841 spin_lock(&buffer_mapping->private_lock);
842 while (!list_empty(list)) {
843 struct buffer_head *bh = BH_ENTRY(list->next);
844 if (buffer_dirty(bh)) {
848 __remove_assoc_queue(bh);
850 spin_unlock(&buffer_mapping->private_lock);
856 * Create the appropriate buffers when given a page for data area and
857 * the size of each buffer.. Use the bh->b_this_page linked list to
858 * follow the buffers created. Return NULL if unable to create more
861 * The retry flag is used to differentiate async IO (paging, swapping)
862 * which may not fail from ordinary buffer allocations.
864 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
867 struct buffer_head *bh, *head;
873 while ((offset -= size) >= 0) {
874 bh = alloc_buffer_head(GFP_NOFS);
878 bh->b_this_page = head;
884 /* Link the buffer to its page */
885 set_bh_page(bh, page, offset);
889 * In case anything failed, we just free everything we got.
895 head = head->b_this_page;
896 free_buffer_head(bh);
901 * Return failure for non-async IO requests. Async IO requests
902 * are not allowed to fail, so we have to wait until buffer heads
903 * become available. But we don't want tasks sleeping with
904 * partially complete buffers, so all were released above.
909 /* We're _really_ low on memory. Now we just
910 * wait for old buffer heads to become free due to
911 * finishing IO. Since this is an async request and
912 * the reserve list is empty, we're sure there are
913 * async buffer heads in use.
918 EXPORT_SYMBOL_GPL(alloc_page_buffers);
921 link_dev_buffers(struct page *page, struct buffer_head *head)
923 struct buffer_head *bh, *tail;
928 bh = bh->b_this_page;
930 tail->b_this_page = head;
931 attach_page_buffers(page, head);
934 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
936 sector_t retval = ~((sector_t)0);
937 loff_t sz = i_size_read(bdev->bd_inode);
940 unsigned int sizebits = blksize_bits(size);
941 retval = (sz >> sizebits);
947 * Initialise the state of a blockdev page's buffers.
950 init_page_buffers(struct page *page, struct block_device *bdev,
951 sector_t block, int size)
953 struct buffer_head *head = page_buffers(page);
954 struct buffer_head *bh = head;
955 int uptodate = PageUptodate(page);
956 sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
959 if (!buffer_mapped(bh)) {
960 init_buffer(bh, NULL, NULL);
962 bh->b_blocknr = block;
964 set_buffer_uptodate(bh);
965 if (block < end_block)
966 set_buffer_mapped(bh);
969 bh = bh->b_this_page;
970 } while (bh != head);
973 * Caller needs to validate requested block against end of device.
979 * Create the page-cache page that contains the requested block.
981 * This is used purely for blockdev mappings.
984 grow_dev_page(struct block_device *bdev, sector_t block,
985 pgoff_t index, int size, int sizebits, gfp_t gfp)
987 struct inode *inode = bdev->bd_inode;
989 struct buffer_head *bh;
991 int ret = 0; /* Will call free_more_memory() */
994 gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
997 * XXX: __getblk_slow() can not really deal with failure and
998 * will endlessly loop on improvised global reclaim. Prefer
999 * looping in the allocator rather than here, at least that
1000 * code knows what it's doing.
1002 gfp_mask |= __GFP_NOFAIL;
1004 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
1008 BUG_ON(!PageLocked(page));
1010 if (page_has_buffers(page)) {
1011 bh = page_buffers(page);
1012 if (bh->b_size == size) {
1013 end_block = init_page_buffers(page, bdev,
1014 (sector_t)index << sizebits,
1018 if (!try_to_free_buffers(page))
1023 * Allocate some buffers for this page
1025 bh = alloc_page_buffers(page, size, 0);
1030 * Link the page to the buffers and initialise them. Take the
1031 * lock to be atomic wrt __find_get_block(), which does not
1032 * run under the page lock.
1034 spin_lock(&inode->i_mapping->private_lock);
1035 link_dev_buffers(page, bh);
1036 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1038 spin_unlock(&inode->i_mapping->private_lock);
1040 ret = (block < end_block) ? 1 : -ENXIO;
1048 * Create buffers for the specified block device block's page. If
1049 * that page was dirty, the buffers are set dirty also.
1052 grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
1060 } while ((size << sizebits) < PAGE_SIZE);
1062 index = block >> sizebits;
1065 * Check for a block which wants to lie outside our maximum possible
1066 * pagecache index. (this comparison is done using sector_t types).
1068 if (unlikely(index != block >> sizebits)) {
1069 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1071 __func__, (unsigned long long)block,
1076 /* Create a page with the proper size buffers.. */
1077 return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1080 static struct buffer_head *
1081 __getblk_slow(struct block_device *bdev, sector_t block,
1082 unsigned size, gfp_t gfp)
1084 /* Size must be multiple of hard sectorsize */
1085 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1086 (size < 512 || size > PAGE_SIZE))) {
1087 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1089 printk(KERN_ERR "logical block size: %d\n",
1090 bdev_logical_block_size(bdev));
1097 struct buffer_head *bh;
1100 bh = __find_get_block(bdev, block, size);
1104 ret = grow_buffers(bdev, block, size, gfp);
1113 * The relationship between dirty buffers and dirty pages:
1115 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1116 * the page is tagged dirty in its radix tree.
1118 * At all times, the dirtiness of the buffers represents the dirtiness of
1119 * subsections of the page. If the page has buffers, the page dirty bit is
1120 * merely a hint about the true dirty state.
1122 * When a page is set dirty in its entirety, all its buffers are marked dirty
1123 * (if the page has buffers).
1125 * When a buffer is marked dirty, its page is dirtied, but the page's other
1128 * Also. When blockdev buffers are explicitly read with bread(), they
1129 * individually become uptodate. But their backing page remains not
1130 * uptodate - even if all of its buffers are uptodate. A subsequent
1131 * block_read_full_page() against that page will discover all the uptodate
1132 * buffers, will set the page uptodate and will perform no I/O.
1136 * mark_buffer_dirty - mark a buffer_head as needing writeout
1137 * @bh: the buffer_head to mark dirty
1139 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1140 * backing page dirty, then tag the page as dirty in its address_space's radix
1141 * tree and then attach the address_space's inode to its superblock's dirty
1144 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1145 * mapping->tree_lock and mapping->host->i_lock.
1147 void mark_buffer_dirty(struct buffer_head *bh)
1149 WARN_ON_ONCE(!buffer_uptodate(bh));
1151 trace_block_dirty_buffer(bh);
1154 * Very *carefully* optimize the it-is-already-dirty case.
1156 * Don't let the final "is it dirty" escape to before we
1157 * perhaps modified the buffer.
1159 if (buffer_dirty(bh)) {
1161 if (buffer_dirty(bh))
1165 if (!test_set_buffer_dirty(bh)) {
1166 struct page *page = bh->b_page;
1167 struct address_space *mapping = NULL;
1169 lock_page_memcg(page);
1170 if (!TestSetPageDirty(page)) {
1171 mapping = page_mapping(page);
1173 __set_page_dirty(page, mapping, 0);
1175 unlock_page_memcg(page);
1177 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1180 EXPORT_SYMBOL(mark_buffer_dirty);
1182 void mark_buffer_write_io_error(struct buffer_head *bh)
1184 set_buffer_write_io_error(bh);
1185 /* FIXME: do we need to set this in both places? */
1186 if (bh->b_page && bh->b_page->mapping)
1187 mapping_set_error(bh->b_page->mapping, -EIO);
1188 if (bh->b_assoc_map)
1189 mapping_set_error(bh->b_assoc_map, -EIO);
1191 EXPORT_SYMBOL(mark_buffer_write_io_error);
1194 * Decrement a buffer_head's reference count. If all buffers against a page
1195 * have zero reference count, are clean and unlocked, and if the page is clean
1196 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1197 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1198 * a page but it ends up not being freed, and buffers may later be reattached).
1200 void __brelse(struct buffer_head * buf)
1202 if (atomic_read(&buf->b_count)) {
1206 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1208 EXPORT_SYMBOL(__brelse);
1211 * bforget() is like brelse(), except it discards any
1212 * potentially dirty data.
1214 void __bforget(struct buffer_head *bh)
1216 clear_buffer_dirty(bh);
1217 if (bh->b_assoc_map) {
1218 struct address_space *buffer_mapping = bh->b_page->mapping;
1220 spin_lock(&buffer_mapping->private_lock);
1221 list_del_init(&bh->b_assoc_buffers);
1222 bh->b_assoc_map = NULL;
1223 spin_unlock(&buffer_mapping->private_lock);
1227 EXPORT_SYMBOL(__bforget);
1229 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1232 if (buffer_uptodate(bh)) {
1237 bh->b_end_io = end_buffer_read_sync;
1238 submit_bh(REQ_OP_READ, 0, bh);
1240 if (buffer_uptodate(bh))
1248 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1249 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1250 * refcount elevated by one when they're in an LRU. A buffer can only appear
1251 * once in a particular CPU's LRU. A single buffer can be present in multiple
1252 * CPU's LRUs at the same time.
1254 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1255 * sb_find_get_block().
1257 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1258 * a local interrupt disable for that.
1261 #define BH_LRU_SIZE 16
1264 struct buffer_head *bhs[BH_LRU_SIZE];
1267 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1270 #define bh_lru_lock() local_irq_disable()
1271 #define bh_lru_unlock() local_irq_enable()
1273 #define bh_lru_lock() preempt_disable()
1274 #define bh_lru_unlock() preempt_enable()
1277 static inline void check_irqs_on(void)
1279 #ifdef irqs_disabled
1280 BUG_ON(irqs_disabled());
1285 * Install a buffer_head into this cpu's LRU. If not already in the LRU, it is
1286 * inserted at the front, and the buffer_head at the back if any is evicted.
1287 * Or, if already in the LRU it is moved to the front.
1289 static void bh_lru_install(struct buffer_head *bh)
1291 struct buffer_head *evictee = bh;
1298 b = this_cpu_ptr(&bh_lrus);
1299 for (i = 0; i < BH_LRU_SIZE; i++) {
1300 swap(evictee, b->bhs[i]);
1301 if (evictee == bh) {
1313 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1315 static struct buffer_head *
1316 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1318 struct buffer_head *ret = NULL;
1323 for (i = 0; i < BH_LRU_SIZE; i++) {
1324 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1326 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1327 bh->b_size == size) {
1330 __this_cpu_write(bh_lrus.bhs[i],
1331 __this_cpu_read(bh_lrus.bhs[i - 1]));
1334 __this_cpu_write(bh_lrus.bhs[0], bh);
1346 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1347 * it in the LRU and mark it as accessed. If it is not present then return
1350 struct buffer_head *
1351 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1353 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1356 /* __find_get_block_slow will mark the page accessed */
1357 bh = __find_get_block_slow(bdev, block);
1365 EXPORT_SYMBOL(__find_get_block);
1368 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1369 * which corresponds to the passed block_device, block and size. The
1370 * returned buffer has its reference count incremented.
1372 * __getblk_gfp() will lock up the machine if grow_dev_page's
1373 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1375 struct buffer_head *
1376 __getblk_gfp(struct block_device *bdev, sector_t block,
1377 unsigned size, gfp_t gfp)
1379 struct buffer_head *bh = __find_get_block(bdev, block, size);
1383 bh = __getblk_slow(bdev, block, size, gfp);
1386 EXPORT_SYMBOL(__getblk_gfp);
1389 * Do async read-ahead on a buffer..
1391 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1393 struct buffer_head *bh = __getblk(bdev, block, size);
1395 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1399 EXPORT_SYMBOL(__breadahead);
1401 void __breadahead_gfp(struct block_device *bdev, sector_t block, unsigned size,
1404 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1406 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1410 EXPORT_SYMBOL(__breadahead_gfp);
1413 * __bread_gfp() - reads a specified block and returns the bh
1414 * @bdev: the block_device to read from
1415 * @block: number of block
1416 * @size: size (in bytes) to read
1417 * @gfp: page allocation flag
1419 * Reads a specified block, and returns buffer head that contains it.
1420 * The page cache can be allocated from non-movable area
1421 * not to prevent page migration if you set gfp to zero.
1422 * It returns NULL if the block was unreadable.
1424 struct buffer_head *
1425 __bread_gfp(struct block_device *bdev, sector_t block,
1426 unsigned size, gfp_t gfp)
1428 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1430 if (likely(bh) && !buffer_uptodate(bh))
1431 bh = __bread_slow(bh);
1434 EXPORT_SYMBOL(__bread_gfp);
1437 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1438 * This doesn't race because it runs in each cpu either in irq
1439 * or with preempt disabled.
1441 static void invalidate_bh_lru(void *arg)
1443 struct bh_lru *b = &get_cpu_var(bh_lrus);
1446 for (i = 0; i < BH_LRU_SIZE; i++) {
1450 put_cpu_var(bh_lrus);
1453 static bool has_bh_in_lru(int cpu, void *dummy)
1455 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1458 for (i = 0; i < BH_LRU_SIZE; i++) {
1466 void invalidate_bh_lrus(void)
1468 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1470 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1472 void set_bh_page(struct buffer_head *bh,
1473 struct page *page, unsigned long offset)
1476 BUG_ON(offset >= PAGE_SIZE);
1477 if (PageHighMem(page))
1479 * This catches illegal uses and preserves the offset:
1481 bh->b_data = (char *)(0 + offset);
1483 bh->b_data = page_address(page) + offset;
1485 EXPORT_SYMBOL(set_bh_page);
1488 * Called when truncating a buffer on a page completely.
1491 /* Bits that are cleared during an invalidate */
1492 #define BUFFER_FLAGS_DISCARD \
1493 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1494 1 << BH_Delay | 1 << BH_Unwritten)
1496 static void discard_buffer(struct buffer_head * bh)
1498 unsigned long b_state, b_state_old;
1501 clear_buffer_dirty(bh);
1503 b_state = bh->b_state;
1505 b_state_old = cmpxchg(&bh->b_state, b_state,
1506 (b_state & ~BUFFER_FLAGS_DISCARD));
1507 if (b_state_old == b_state)
1509 b_state = b_state_old;
1515 * block_invalidatepage - invalidate part or all of a buffer-backed page
1517 * @page: the page which is affected
1518 * @offset: start of the range to invalidate
1519 * @length: length of the range to invalidate
1521 * block_invalidatepage() is called when all or part of the page has become
1522 * invalidated by a truncate operation.
1524 * block_invalidatepage() does not have to release all buffers, but it must
1525 * ensure that no dirty buffer is left outside @offset and that no I/O
1526 * is underway against any of the blocks which are outside the truncation
1527 * point. Because the caller is about to free (and possibly reuse) those
1530 void block_invalidatepage(struct page *page, unsigned int offset,
1531 unsigned int length)
1533 struct buffer_head *head, *bh, *next;
1534 unsigned int curr_off = 0;
1535 unsigned int stop = length + offset;
1537 BUG_ON(!PageLocked(page));
1538 if (!page_has_buffers(page))
1542 * Check for overflow
1544 BUG_ON(stop > PAGE_SIZE || stop < length);
1546 head = page_buffers(page);
1549 unsigned int next_off = curr_off + bh->b_size;
1550 next = bh->b_this_page;
1553 * Are we still fully in range ?
1555 if (next_off > stop)
1559 * is this block fully invalidated?
1561 if (offset <= curr_off)
1563 curr_off = next_off;
1565 } while (bh != head);
1568 * We release buffers only if the entire page is being invalidated.
1569 * The get_block cached value has been unconditionally invalidated,
1570 * so real IO is not possible anymore.
1573 try_to_release_page(page, 0);
1577 EXPORT_SYMBOL(block_invalidatepage);
1581 * We attach and possibly dirty the buffers atomically wrt
1582 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1583 * is already excluded via the page lock.
1585 void create_empty_buffers(struct page *page,
1586 unsigned long blocksize, unsigned long b_state)
1588 struct buffer_head *bh, *head, *tail;
1590 head = alloc_page_buffers(page, blocksize, 1);
1593 bh->b_state |= b_state;
1595 bh = bh->b_this_page;
1597 tail->b_this_page = head;
1599 spin_lock(&page->mapping->private_lock);
1600 if (PageUptodate(page) || PageDirty(page)) {
1603 if (PageDirty(page))
1604 set_buffer_dirty(bh);
1605 if (PageUptodate(page))
1606 set_buffer_uptodate(bh);
1607 bh = bh->b_this_page;
1608 } while (bh != head);
1610 attach_page_buffers(page, head);
1611 spin_unlock(&page->mapping->private_lock);
1613 EXPORT_SYMBOL(create_empty_buffers);
1616 * clean_bdev_aliases: clean a range of buffers in block device
1617 * @bdev: Block device to clean buffers in
1618 * @block: Start of a range of blocks to clean
1619 * @len: Number of blocks to clean
1621 * We are taking a range of blocks for data and we don't want writeback of any
1622 * buffer-cache aliases starting from return from this function and until the
1623 * moment when something will explicitly mark the buffer dirty (hopefully that
1624 * will not happen until we will free that block ;-) We don't even need to mark
1625 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1626 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1627 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1628 * would confuse anyone who might pick it with bread() afterwards...
1630 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1631 * writeout I/O going on against recently-freed buffers. We don't wait on that
1632 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1633 * need to. That happens here.
1635 void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1637 struct inode *bd_inode = bdev->bd_inode;
1638 struct address_space *bd_mapping = bd_inode->i_mapping;
1639 struct pagevec pvec;
1640 pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1643 struct buffer_head *bh;
1644 struct buffer_head *head;
1646 end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1647 pagevec_init(&pvec, 0);
1648 while (pagevec_lookup_range(&pvec, bd_mapping, &index, end)) {
1649 count = pagevec_count(&pvec);
1650 for (i = 0; i < count; i++) {
1651 struct page *page = pvec.pages[i];
1653 if (!page_has_buffers(page))
1656 * We use page lock instead of bd_mapping->private_lock
1657 * to pin buffers here since we can afford to sleep and
1658 * it scales better than a global spinlock lock.
1661 /* Recheck when the page is locked which pins bhs */
1662 if (!page_has_buffers(page))
1664 head = page_buffers(page);
1667 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1669 if (bh->b_blocknr >= block + len)
1671 clear_buffer_dirty(bh);
1673 clear_buffer_req(bh);
1675 bh = bh->b_this_page;
1676 } while (bh != head);
1680 pagevec_release(&pvec);
1682 /* End of range already reached? */
1683 if (index > end || !index)
1687 EXPORT_SYMBOL(clean_bdev_aliases);
1690 * Size is a power-of-two in the range 512..PAGE_SIZE,
1691 * and the case we care about most is PAGE_SIZE.
1693 * So this *could* possibly be written with those
1694 * constraints in mind (relevant mostly if some
1695 * architecture has a slow bit-scan instruction)
1697 static inline int block_size_bits(unsigned int blocksize)
1699 return ilog2(blocksize);
1702 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1704 BUG_ON(!PageLocked(page));
1706 if (!page_has_buffers(page))
1707 create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1708 return page_buffers(page);
1712 * NOTE! All mapped/uptodate combinations are valid:
1714 * Mapped Uptodate Meaning
1716 * No No "unknown" - must do get_block()
1717 * No Yes "hole" - zero-filled
1718 * Yes No "allocated" - allocated on disk, not read in
1719 * Yes Yes "valid" - allocated and up-to-date in memory.
1721 * "Dirty" is valid only with the last case (mapped+uptodate).
1725 * While block_write_full_page is writing back the dirty buffers under
1726 * the page lock, whoever dirtied the buffers may decide to clean them
1727 * again at any time. We handle that by only looking at the buffer
1728 * state inside lock_buffer().
1730 * If block_write_full_page() is called for regular writeback
1731 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1732 * locked buffer. This only can happen if someone has written the buffer
1733 * directly, with submit_bh(). At the address_space level PageWriteback
1734 * prevents this contention from occurring.
1736 * If block_write_full_page() is called with wbc->sync_mode ==
1737 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1738 * causes the writes to be flagged as synchronous writes.
1740 int __block_write_full_page(struct inode *inode, struct page *page,
1741 get_block_t *get_block, struct writeback_control *wbc,
1742 bh_end_io_t *handler)
1746 sector_t last_block;
1747 struct buffer_head *bh, *head;
1748 unsigned int blocksize, bbits;
1749 int nr_underway = 0;
1750 int write_flags = wbc_to_write_flags(wbc);
1752 head = create_page_buffers(page, inode,
1753 (1 << BH_Dirty)|(1 << BH_Uptodate));
1756 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1757 * here, and the (potentially unmapped) buffers may become dirty at
1758 * any time. If a buffer becomes dirty here after we've inspected it
1759 * then we just miss that fact, and the page stays dirty.
1761 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1762 * handle that here by just cleaning them.
1766 blocksize = bh->b_size;
1767 bbits = block_size_bits(blocksize);
1769 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1770 last_block = (i_size_read(inode) - 1) >> bbits;
1773 * Get all the dirty buffers mapped to disk addresses and
1774 * handle any aliases from the underlying blockdev's mapping.
1777 if (block > last_block) {
1779 * mapped buffers outside i_size will occur, because
1780 * this page can be outside i_size when there is a
1781 * truncate in progress.
1784 * The buffer was zeroed by block_write_full_page()
1786 clear_buffer_dirty(bh);
1787 set_buffer_uptodate(bh);
1788 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1790 WARN_ON(bh->b_size != blocksize);
1791 err = get_block(inode, block, bh, 1);
1794 clear_buffer_delay(bh);
1795 if (buffer_new(bh)) {
1796 /* blockdev mappings never come here */
1797 clear_buffer_new(bh);
1798 clean_bdev_bh_alias(bh);
1801 bh = bh->b_this_page;
1803 } while (bh != head);
1806 if (!buffer_mapped(bh))
1809 * If it's a fully non-blocking write attempt and we cannot
1810 * lock the buffer then redirty the page. Note that this can
1811 * potentially cause a busy-wait loop from writeback threads
1812 * and kswapd activity, but those code paths have their own
1813 * higher-level throttling.
1815 if (wbc->sync_mode != WB_SYNC_NONE) {
1817 } else if (!trylock_buffer(bh)) {
1818 redirty_page_for_writepage(wbc, page);
1821 if (test_clear_buffer_dirty(bh)) {
1822 mark_buffer_async_write_endio(bh, handler);
1826 } while ((bh = bh->b_this_page) != head);
1829 * The page and its buffers are protected by PageWriteback(), so we can
1830 * drop the bh refcounts early.
1832 BUG_ON(PageWriteback(page));
1833 set_page_writeback(page);
1836 struct buffer_head *next = bh->b_this_page;
1837 if (buffer_async_write(bh)) {
1838 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1839 inode->i_write_hint, wbc);
1843 } while (bh != head);
1848 if (nr_underway == 0) {
1850 * The page was marked dirty, but the buffers were
1851 * clean. Someone wrote them back by hand with
1852 * ll_rw_block/submit_bh. A rare case.
1854 end_page_writeback(page);
1857 * The page and buffer_heads can be released at any time from
1865 * ENOSPC, or some other error. We may already have added some
1866 * blocks to the file, so we need to write these out to avoid
1867 * exposing stale data.
1868 * The page is currently locked and not marked for writeback
1871 /* Recovery: lock and submit the mapped buffers */
1873 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1874 !buffer_delay(bh)) {
1876 mark_buffer_async_write_endio(bh, handler);
1879 * The buffer may have been set dirty during
1880 * attachment to a dirty page.
1882 clear_buffer_dirty(bh);
1884 } while ((bh = bh->b_this_page) != head);
1886 BUG_ON(PageWriteback(page));
1887 mapping_set_error(page->mapping, err);
1888 set_page_writeback(page);
1890 struct buffer_head *next = bh->b_this_page;
1891 if (buffer_async_write(bh)) {
1892 clear_buffer_dirty(bh);
1893 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1894 inode->i_write_hint, wbc);
1898 } while (bh != head);
1902 EXPORT_SYMBOL(__block_write_full_page);
1905 * If a page has any new buffers, zero them out here, and mark them uptodate
1906 * and dirty so they'll be written out (in order to prevent uninitialised
1907 * block data from leaking). And clear the new bit.
1909 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1911 unsigned int block_start, block_end;
1912 struct buffer_head *head, *bh;
1914 BUG_ON(!PageLocked(page));
1915 if (!page_has_buffers(page))
1918 bh = head = page_buffers(page);
1921 block_end = block_start + bh->b_size;
1923 if (buffer_new(bh)) {
1924 if (block_end > from && block_start < to) {
1925 if (!PageUptodate(page)) {
1926 unsigned start, size;
1928 start = max(from, block_start);
1929 size = min(to, block_end) - start;
1931 zero_user(page, start, size);
1932 set_buffer_uptodate(bh);
1935 clear_buffer_new(bh);
1936 mark_buffer_dirty(bh);
1940 block_start = block_end;
1941 bh = bh->b_this_page;
1942 } while (bh != head);
1944 EXPORT_SYMBOL(page_zero_new_buffers);
1947 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1948 struct iomap *iomap)
1950 loff_t offset = block << inode->i_blkbits;
1952 bh->b_bdev = iomap->bdev;
1955 * Block points to offset in file we need to map, iomap contains
1956 * the offset at which the map starts. If the map ends before the
1957 * current block, then do not map the buffer and let the caller
1960 BUG_ON(offset >= iomap->offset + iomap->length);
1962 switch (iomap->type) {
1965 * If the buffer is not up to date or beyond the current EOF,
1966 * we need to mark it as new to ensure sub-block zeroing is
1967 * executed if necessary.
1969 if (!buffer_uptodate(bh) ||
1970 (offset >= i_size_read(inode)))
1973 case IOMAP_DELALLOC:
1974 if (!buffer_uptodate(bh) ||
1975 (offset >= i_size_read(inode)))
1977 set_buffer_uptodate(bh);
1978 set_buffer_mapped(bh);
1979 set_buffer_delay(bh);
1981 case IOMAP_UNWRITTEN:
1983 * For unwritten regions, we always need to ensure that
1984 * sub-block writes cause the regions in the block we are not
1985 * writing to are zeroed. Set the buffer as new to ensure this.
1988 set_buffer_unwritten(bh);
1991 if (offset >= i_size_read(inode))
1993 bh->b_blocknr = (iomap->blkno >> (inode->i_blkbits - 9)) +
1994 ((offset - iomap->offset) >> inode->i_blkbits);
1995 set_buffer_mapped(bh);
2000 int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
2001 get_block_t *get_block, struct iomap *iomap)
2003 unsigned from = pos & (PAGE_SIZE - 1);
2004 unsigned to = from + len;
2005 struct inode *inode = page->mapping->host;
2006 unsigned block_start, block_end;
2009 unsigned blocksize, bbits;
2010 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
2012 BUG_ON(!PageLocked(page));
2013 BUG_ON(from > PAGE_SIZE);
2014 BUG_ON(to > PAGE_SIZE);
2017 head = create_page_buffers(page, inode, 0);
2018 blocksize = head->b_size;
2019 bbits = block_size_bits(blocksize);
2021 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
2023 for(bh = head, block_start = 0; bh != head || !block_start;
2024 block++, block_start=block_end, bh = bh->b_this_page) {
2025 block_end = block_start + blocksize;
2026 if (block_end <= from || block_start >= to) {
2027 if (PageUptodate(page)) {
2028 if (!buffer_uptodate(bh))
2029 set_buffer_uptodate(bh);
2034 clear_buffer_new(bh);
2035 if (!buffer_mapped(bh)) {
2036 WARN_ON(bh->b_size != blocksize);
2038 err = get_block(inode, block, bh, 1);
2042 iomap_to_bh(inode, block, bh, iomap);
2045 if (buffer_new(bh)) {
2046 clean_bdev_bh_alias(bh);
2047 if (PageUptodate(page)) {
2048 clear_buffer_new(bh);
2049 set_buffer_uptodate(bh);
2050 mark_buffer_dirty(bh);
2053 if (block_end > to || block_start < from)
2054 zero_user_segments(page,
2060 if (PageUptodate(page)) {
2061 if (!buffer_uptodate(bh))
2062 set_buffer_uptodate(bh);
2065 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2066 !buffer_unwritten(bh) &&
2067 (block_start < from || block_end > to)) {
2068 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2073 * If we issued read requests - let them complete.
2075 while(wait_bh > wait) {
2076 wait_on_buffer(*--wait_bh);
2077 if (!buffer_uptodate(*wait_bh))
2081 page_zero_new_buffers(page, from, to);
2085 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2086 get_block_t *get_block)
2088 return __block_write_begin_int(page, pos, len, get_block, NULL);
2090 EXPORT_SYMBOL(__block_write_begin);
2092 static int __block_commit_write(struct inode *inode, struct page *page,
2093 unsigned from, unsigned to)
2095 unsigned block_start, block_end;
2098 struct buffer_head *bh, *head;
2100 bh = head = page_buffers(page);
2101 blocksize = bh->b_size;
2105 block_end = block_start + blocksize;
2106 if (block_end <= from || block_start >= to) {
2107 if (!buffer_uptodate(bh))
2110 set_buffer_uptodate(bh);
2111 mark_buffer_dirty(bh);
2113 clear_buffer_new(bh);
2115 block_start = block_end;
2116 bh = bh->b_this_page;
2117 } while (bh != head);
2120 * If this is a partial write which happened to make all buffers
2121 * uptodate then we can optimize away a bogus readpage() for
2122 * the next read(). Here we 'discover' whether the page went
2123 * uptodate as a result of this (potentially partial) write.
2126 SetPageUptodate(page);
2131 * block_write_begin takes care of the basic task of block allocation and
2132 * bringing partial write blocks uptodate first.
2134 * The filesystem needs to handle block truncation upon failure.
2136 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2137 unsigned flags, struct page **pagep, get_block_t *get_block)
2139 pgoff_t index = pos >> PAGE_SHIFT;
2143 page = grab_cache_page_write_begin(mapping, index, flags);
2147 status = __block_write_begin(page, pos, len, get_block);
2148 if (unlikely(status)) {
2157 EXPORT_SYMBOL(block_write_begin);
2159 int block_write_end(struct file *file, struct address_space *mapping,
2160 loff_t pos, unsigned len, unsigned copied,
2161 struct page *page, void *fsdata)
2163 struct inode *inode = mapping->host;
2166 start = pos & (PAGE_SIZE - 1);
2168 if (unlikely(copied < len)) {
2170 * The buffers that were written will now be uptodate, so we
2171 * don't have to worry about a readpage reading them and
2172 * overwriting a partial write. However if we have encountered
2173 * a short write and only partially written into a buffer, it
2174 * will not be marked uptodate, so a readpage might come in and
2175 * destroy our partial write.
2177 * Do the simplest thing, and just treat any short write to a
2178 * non uptodate page as a zero-length write, and force the
2179 * caller to redo the whole thing.
2181 if (!PageUptodate(page))
2184 page_zero_new_buffers(page, start+copied, start+len);
2186 flush_dcache_page(page);
2188 /* This could be a short (even 0-length) commit */
2189 __block_commit_write(inode, page, start, start+copied);
2193 EXPORT_SYMBOL(block_write_end);
2195 int generic_write_end(struct file *file, struct address_space *mapping,
2196 loff_t pos, unsigned len, unsigned copied,
2197 struct page *page, void *fsdata)
2199 struct inode *inode = mapping->host;
2200 loff_t old_size = inode->i_size;
2201 int i_size_changed = 0;
2203 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2206 * No need to use i_size_read() here, the i_size
2207 * cannot change under us because we hold i_mutex.
2209 * But it's important to update i_size while still holding page lock:
2210 * page writeout could otherwise come in and zero beyond i_size.
2212 if (pos+copied > inode->i_size) {
2213 i_size_write(inode, pos+copied);
2221 pagecache_isize_extended(inode, old_size, pos);
2223 * Don't mark the inode dirty under page lock. First, it unnecessarily
2224 * makes the holding time of page lock longer. Second, it forces lock
2225 * ordering of page lock and transaction start for journaling
2229 mark_inode_dirty(inode);
2233 EXPORT_SYMBOL(generic_write_end);
2236 * block_is_partially_uptodate checks whether buffers within a page are
2239 * Returns true if all buffers which correspond to a file portion
2240 * we want to read are uptodate.
2242 int block_is_partially_uptodate(struct page *page, unsigned long from,
2243 unsigned long count)
2245 unsigned block_start, block_end, blocksize;
2247 struct buffer_head *bh, *head;
2250 if (!page_has_buffers(page))
2253 head = page_buffers(page);
2254 blocksize = head->b_size;
2255 to = min_t(unsigned, PAGE_SIZE - from, count);
2257 if (from < blocksize && to > PAGE_SIZE - blocksize)
2263 block_end = block_start + blocksize;
2264 if (block_end > from && block_start < to) {
2265 if (!buffer_uptodate(bh)) {
2269 if (block_end >= to)
2272 block_start = block_end;
2273 bh = bh->b_this_page;
2274 } while (bh != head);
2278 EXPORT_SYMBOL(block_is_partially_uptodate);
2281 * Generic "read page" function for block devices that have the normal
2282 * get_block functionality. This is most of the block device filesystems.
2283 * Reads the page asynchronously --- the unlock_buffer() and
2284 * set/clear_buffer_uptodate() functions propagate buffer state into the
2285 * page struct once IO has completed.
2287 int block_read_full_page(struct page *page, get_block_t *get_block)
2289 struct inode *inode = page->mapping->host;
2290 sector_t iblock, lblock;
2291 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2292 unsigned int blocksize, bbits;
2294 int fully_mapped = 1;
2296 head = create_page_buffers(page, inode, 0);
2297 blocksize = head->b_size;
2298 bbits = block_size_bits(blocksize);
2300 iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2301 lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2307 if (buffer_uptodate(bh))
2310 if (!buffer_mapped(bh)) {
2314 if (iblock < lblock) {
2315 WARN_ON(bh->b_size != blocksize);
2316 err = get_block(inode, iblock, bh, 0);
2320 if (!buffer_mapped(bh)) {
2321 zero_user(page, i * blocksize, blocksize);
2323 set_buffer_uptodate(bh);
2327 * get_block() might have updated the buffer
2330 if (buffer_uptodate(bh))
2334 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2337 SetPageMappedToDisk(page);
2341 * All buffers are uptodate - we can set the page uptodate
2342 * as well. But not if get_block() returned an error.
2344 if (!PageError(page))
2345 SetPageUptodate(page);
2350 /* Stage two: lock the buffers */
2351 for (i = 0; i < nr; i++) {
2354 mark_buffer_async_read(bh);
2358 * Stage 3: start the IO. Check for uptodateness
2359 * inside the buffer lock in case another process reading
2360 * the underlying blockdev brought it uptodate (the sct fix).
2362 for (i = 0; i < nr; i++) {
2364 if (buffer_uptodate(bh))
2365 end_buffer_async_read(bh, 1);
2367 submit_bh(REQ_OP_READ, 0, bh);
2371 EXPORT_SYMBOL(block_read_full_page);
2373 /* utility function for filesystems that need to do work on expanding
2374 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2375 * deal with the hole.
2377 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2379 struct address_space *mapping = inode->i_mapping;
2381 void *fsdata = NULL;
2384 err = inode_newsize_ok(inode, size);
2388 err = pagecache_write_begin(NULL, mapping, size, 0,
2389 AOP_FLAG_CONT_EXPAND, &page, &fsdata);
2393 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2399 EXPORT_SYMBOL(generic_cont_expand_simple);
2401 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2402 loff_t pos, loff_t *bytes)
2404 struct inode *inode = mapping->host;
2405 unsigned int blocksize = i_blocksize(inode);
2407 void *fsdata = NULL;
2408 pgoff_t index, curidx;
2410 unsigned zerofrom, offset, len;
2413 index = pos >> PAGE_SHIFT;
2414 offset = pos & ~PAGE_MASK;
2416 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2417 zerofrom = curpos & ~PAGE_MASK;
2418 if (zerofrom & (blocksize-1)) {
2419 *bytes |= (blocksize-1);
2422 len = PAGE_SIZE - zerofrom;
2424 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2428 zero_user(page, zerofrom, len);
2429 err = pagecache_write_end(file, mapping, curpos, len, len,
2436 balance_dirty_pages_ratelimited(mapping);
2438 if (unlikely(fatal_signal_pending(current))) {
2444 /* page covers the boundary, find the boundary offset */
2445 if (index == curidx) {
2446 zerofrom = curpos & ~PAGE_MASK;
2447 /* if we will expand the thing last block will be filled */
2448 if (offset <= zerofrom) {
2451 if (zerofrom & (blocksize-1)) {
2452 *bytes |= (blocksize-1);
2455 len = offset - zerofrom;
2457 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2461 zero_user(page, zerofrom, len);
2462 err = pagecache_write_end(file, mapping, curpos, len, len,
2474 * For moronic filesystems that do not allow holes in file.
2475 * We may have to extend the file.
2477 int cont_write_begin(struct file *file, struct address_space *mapping,
2478 loff_t pos, unsigned len, unsigned flags,
2479 struct page **pagep, void **fsdata,
2480 get_block_t *get_block, loff_t *bytes)
2482 struct inode *inode = mapping->host;
2483 unsigned int blocksize = i_blocksize(inode);
2484 unsigned int zerofrom;
2487 err = cont_expand_zero(file, mapping, pos, bytes);
2491 zerofrom = *bytes & ~PAGE_MASK;
2492 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2493 *bytes |= (blocksize-1);
2497 return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2499 EXPORT_SYMBOL(cont_write_begin);
2501 int block_commit_write(struct page *page, unsigned from, unsigned to)
2503 struct inode *inode = page->mapping->host;
2504 __block_commit_write(inode,page,from,to);
2507 EXPORT_SYMBOL(block_commit_write);
2510 * block_page_mkwrite() is not allowed to change the file size as it gets
2511 * called from a page fault handler when a page is first dirtied. Hence we must
2512 * be careful to check for EOF conditions here. We set the page up correctly
2513 * for a written page which means we get ENOSPC checking when writing into
2514 * holes and correct delalloc and unwritten extent mapping on filesystems that
2515 * support these features.
2517 * We are not allowed to take the i_mutex here so we have to play games to
2518 * protect against truncate races as the page could now be beyond EOF. Because
2519 * truncate writes the inode size before removing pages, once we have the
2520 * page lock we can determine safely if the page is beyond EOF. If it is not
2521 * beyond EOF, then the page is guaranteed safe against truncation until we
2524 * Direct callers of this function should protect against filesystem freezing
2525 * using sb_start_pagefault() - sb_end_pagefault() functions.
2527 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2528 get_block_t get_block)
2530 struct page *page = vmf->page;
2531 struct inode *inode = file_inode(vma->vm_file);
2537 size = i_size_read(inode);
2538 if ((page->mapping != inode->i_mapping) ||
2539 (page_offset(page) > size)) {
2540 /* We overload EFAULT to mean page got truncated */
2545 /* page is wholly or partially inside EOF */
2546 if (((page->index + 1) << PAGE_SHIFT) > size)
2547 end = size & ~PAGE_MASK;
2551 ret = __block_write_begin(page, 0, end, get_block);
2553 ret = block_commit_write(page, 0, end);
2555 if (unlikely(ret < 0))
2557 set_page_dirty(page);
2558 wait_for_stable_page(page);
2564 EXPORT_SYMBOL(block_page_mkwrite);
2567 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2568 * immediately, while under the page lock. So it needs a special end_io
2569 * handler which does not touch the bh after unlocking it.
2571 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2573 __end_buffer_read_notouch(bh, uptodate);
2577 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2578 * the page (converting it to circular linked list and taking care of page
2581 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2583 struct buffer_head *bh;
2585 BUG_ON(!PageLocked(page));
2587 spin_lock(&page->mapping->private_lock);
2590 if (PageDirty(page))
2591 set_buffer_dirty(bh);
2592 if (!bh->b_this_page)
2593 bh->b_this_page = head;
2594 bh = bh->b_this_page;
2595 } while (bh != head);
2596 attach_page_buffers(page, head);
2597 spin_unlock(&page->mapping->private_lock);
2601 * On entry, the page is fully not uptodate.
2602 * On exit the page is fully uptodate in the areas outside (from,to)
2603 * The filesystem needs to handle block truncation upon failure.
2605 int nobh_write_begin(struct address_space *mapping,
2606 loff_t pos, unsigned len, unsigned flags,
2607 struct page **pagep, void **fsdata,
2608 get_block_t *get_block)
2610 struct inode *inode = mapping->host;
2611 const unsigned blkbits = inode->i_blkbits;
2612 const unsigned blocksize = 1 << blkbits;
2613 struct buffer_head *head, *bh;
2617 unsigned block_in_page;
2618 unsigned block_start, block_end;
2619 sector_t block_in_file;
2622 int is_mapped_to_disk = 1;
2624 index = pos >> PAGE_SHIFT;
2625 from = pos & (PAGE_SIZE - 1);
2628 page = grab_cache_page_write_begin(mapping, index, flags);
2634 if (page_has_buffers(page)) {
2635 ret = __block_write_begin(page, pos, len, get_block);
2641 if (PageMappedToDisk(page))
2645 * Allocate buffers so that we can keep track of state, and potentially
2646 * attach them to the page if an error occurs. In the common case of
2647 * no error, they will just be freed again without ever being attached
2648 * to the page (which is all OK, because we're under the page lock).
2650 * Be careful: the buffer linked list is a NULL terminated one, rather
2651 * than the circular one we're used to.
2653 head = alloc_page_buffers(page, blocksize, 0);
2659 block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2662 * We loop across all blocks in the page, whether or not they are
2663 * part of the affected region. This is so we can discover if the
2664 * page is fully mapped-to-disk.
2666 for (block_start = 0, block_in_page = 0, bh = head;
2667 block_start < PAGE_SIZE;
2668 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2671 block_end = block_start + blocksize;
2674 if (block_start >= to)
2676 ret = get_block(inode, block_in_file + block_in_page,
2680 if (!buffer_mapped(bh))
2681 is_mapped_to_disk = 0;
2683 clean_bdev_bh_alias(bh);
2684 if (PageUptodate(page)) {
2685 set_buffer_uptodate(bh);
2688 if (buffer_new(bh) || !buffer_mapped(bh)) {
2689 zero_user_segments(page, block_start, from,
2693 if (buffer_uptodate(bh))
2694 continue; /* reiserfs does this */
2695 if (block_start < from || block_end > to) {
2697 bh->b_end_io = end_buffer_read_nobh;
2698 submit_bh(REQ_OP_READ, 0, bh);
2705 * The page is locked, so these buffers are protected from
2706 * any VM or truncate activity. Hence we don't need to care
2707 * for the buffer_head refcounts.
2709 for (bh = head; bh; bh = bh->b_this_page) {
2711 if (!buffer_uptodate(bh))
2718 if (is_mapped_to_disk)
2719 SetPageMappedToDisk(page);
2721 *fsdata = head; /* to be released by nobh_write_end */
2728 * Error recovery is a bit difficult. We need to zero out blocks that
2729 * were newly allocated, and dirty them to ensure they get written out.
2730 * Buffers need to be attached to the page at this point, otherwise
2731 * the handling of potential IO errors during writeout would be hard
2732 * (could try doing synchronous writeout, but what if that fails too?)
2734 attach_nobh_buffers(page, head);
2735 page_zero_new_buffers(page, from, to);
2744 EXPORT_SYMBOL(nobh_write_begin);
2746 int nobh_write_end(struct file *file, struct address_space *mapping,
2747 loff_t pos, unsigned len, unsigned copied,
2748 struct page *page, void *fsdata)
2750 struct inode *inode = page->mapping->host;
2751 struct buffer_head *head = fsdata;
2752 struct buffer_head *bh;
2753 BUG_ON(fsdata != NULL && page_has_buffers(page));
2755 if (unlikely(copied < len) && head)
2756 attach_nobh_buffers(page, head);
2757 if (page_has_buffers(page))
2758 return generic_write_end(file, mapping, pos, len,
2759 copied, page, fsdata);
2761 SetPageUptodate(page);
2762 set_page_dirty(page);
2763 if (pos+copied > inode->i_size) {
2764 i_size_write(inode, pos+copied);
2765 mark_inode_dirty(inode);
2773 head = head->b_this_page;
2774 free_buffer_head(bh);
2779 EXPORT_SYMBOL(nobh_write_end);
2782 * nobh_writepage() - based on block_full_write_page() except
2783 * that it tries to operate without attaching bufferheads to
2786 int nobh_writepage(struct page *page, get_block_t *get_block,
2787 struct writeback_control *wbc)
2789 struct inode * const inode = page->mapping->host;
2790 loff_t i_size = i_size_read(inode);
2791 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2795 /* Is the page fully inside i_size? */
2796 if (page->index < end_index)
2799 /* Is the page fully outside i_size? (truncate in progress) */
2800 offset = i_size & (PAGE_SIZE-1);
2801 if (page->index >= end_index+1 || !offset) {
2803 return 0; /* don't care */
2807 * The page straddles i_size. It must be zeroed out on each and every
2808 * writepage invocation because it may be mmapped. "A file is mapped
2809 * in multiples of the page size. For a file that is not a multiple of
2810 * the page size, the remaining memory is zeroed when mapped, and
2811 * writes to that region are not written out to the file."
2813 zero_user_segment(page, offset, PAGE_SIZE);
2815 ret = mpage_writepage(page, get_block, wbc);
2817 ret = __block_write_full_page(inode, page, get_block, wbc,
2818 end_buffer_async_write);
2821 EXPORT_SYMBOL(nobh_writepage);
2823 int nobh_truncate_page(struct address_space *mapping,
2824 loff_t from, get_block_t *get_block)
2826 pgoff_t index = from >> PAGE_SHIFT;
2827 unsigned offset = from & (PAGE_SIZE-1);
2830 unsigned length, pos;
2831 struct inode *inode = mapping->host;
2833 struct buffer_head map_bh;
2836 blocksize = i_blocksize(inode);
2837 length = offset & (blocksize - 1);
2839 /* Block boundary? Nothing to do */
2843 length = blocksize - length;
2844 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2846 page = grab_cache_page(mapping, index);
2851 if (page_has_buffers(page)) {
2855 return block_truncate_page(mapping, from, get_block);
2858 /* Find the buffer that contains "offset" */
2860 while (offset >= pos) {
2865 map_bh.b_size = blocksize;
2867 err = get_block(inode, iblock, &map_bh, 0);
2870 /* unmapped? It's a hole - nothing to do */
2871 if (!buffer_mapped(&map_bh))
2874 /* Ok, it's mapped. Make sure it's up-to-date */
2875 if (!PageUptodate(page)) {
2876 err = mapping->a_ops->readpage(NULL, page);
2882 if (!PageUptodate(page)) {
2886 if (page_has_buffers(page))
2889 zero_user(page, offset, length);
2890 set_page_dirty(page);
2899 EXPORT_SYMBOL(nobh_truncate_page);
2901 int block_truncate_page(struct address_space *mapping,
2902 loff_t from, get_block_t *get_block)
2904 pgoff_t index = from >> PAGE_SHIFT;
2905 unsigned offset = from & (PAGE_SIZE-1);
2908 unsigned length, pos;
2909 struct inode *inode = mapping->host;
2911 struct buffer_head *bh;
2914 blocksize = i_blocksize(inode);
2915 length = offset & (blocksize - 1);
2917 /* Block boundary? Nothing to do */
2921 length = blocksize - length;
2922 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2924 page = grab_cache_page(mapping, index);
2929 if (!page_has_buffers(page))
2930 create_empty_buffers(page, blocksize, 0);
2932 /* Find the buffer that contains "offset" */
2933 bh = page_buffers(page);
2935 while (offset >= pos) {
2936 bh = bh->b_this_page;
2942 if (!buffer_mapped(bh)) {
2943 WARN_ON(bh->b_size != blocksize);
2944 err = get_block(inode, iblock, bh, 0);
2947 /* unmapped? It's a hole - nothing to do */
2948 if (!buffer_mapped(bh))
2952 /* Ok, it's mapped. Make sure it's up-to-date */
2953 if (PageUptodate(page))
2954 set_buffer_uptodate(bh);
2956 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2958 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2960 /* Uhhuh. Read error. Complain and punt. */
2961 if (!buffer_uptodate(bh))
2965 zero_user(page, offset, length);
2966 mark_buffer_dirty(bh);
2975 EXPORT_SYMBOL(block_truncate_page);
2978 * The generic ->writepage function for buffer-backed address_spaces
2980 int block_write_full_page(struct page *page, get_block_t *get_block,
2981 struct writeback_control *wbc)
2983 struct inode * const inode = page->mapping->host;
2984 loff_t i_size = i_size_read(inode);
2985 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2988 /* Is the page fully inside i_size? */
2989 if (page->index < end_index)
2990 return __block_write_full_page(inode, page, get_block, wbc,
2991 end_buffer_async_write);
2993 /* Is the page fully outside i_size? (truncate in progress) */
2994 offset = i_size & (PAGE_SIZE-1);
2995 if (page->index >= end_index+1 || !offset) {
2997 return 0; /* don't care */
3001 * The page straddles i_size. It must be zeroed out on each and every
3002 * writepage invocation because it may be mmapped. "A file is mapped
3003 * in multiples of the page size. For a file that is not a multiple of
3004 * the page size, the remaining memory is zeroed when mapped, and
3005 * writes to that region are not written out to the file."
3007 zero_user_segment(page, offset, PAGE_SIZE);
3008 return __block_write_full_page(inode, page, get_block, wbc,
3009 end_buffer_async_write);
3011 EXPORT_SYMBOL(block_write_full_page);
3013 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
3014 get_block_t *get_block)
3016 struct inode *inode = mapping->host;
3017 struct buffer_head tmp = {
3018 .b_size = i_blocksize(inode),
3021 get_block(inode, block, &tmp, 0);
3022 return tmp.b_blocknr;
3024 EXPORT_SYMBOL(generic_block_bmap);
3026 static void end_bio_bh_io_sync(struct bio *bio)
3028 struct buffer_head *bh = bio->bi_private;
3030 if (unlikely(bio_flagged(bio, BIO_QUIET)))
3031 set_bit(BH_Quiet, &bh->b_state);
3033 bh->b_end_io(bh, !bio->bi_status);
3038 * This allows us to do IO even on the odd last sectors
3039 * of a device, even if the block size is some multiple
3040 * of the physical sector size.
3042 * We'll just truncate the bio to the size of the device,
3043 * and clear the end of the buffer head manually.
3045 * Truly out-of-range accesses will turn into actual IO
3046 * errors, this only handles the "we need to be able to
3047 * do IO at the final sector" case.
3049 void guard_bio_eod(int op, struct bio *bio)
3052 struct bio_vec *bvec = &bio->bi_io_vec[bio->bi_vcnt - 1];
3053 unsigned truncated_bytes;
3054 struct hd_struct *part;
3057 part = __disk_get_part(bio->bi_disk, bio->bi_partno);
3059 maxsector = part_nr_sects_read(part);
3061 maxsector = get_capacity(bio->bi_disk);
3068 * If the *whole* IO is past the end of the device,
3069 * let it through, and the IO layer will turn it into
3072 if (unlikely(bio->bi_iter.bi_sector >= maxsector))
3075 maxsector -= bio->bi_iter.bi_sector;
3076 if (likely((bio->bi_iter.bi_size >> 9) <= maxsector))
3079 /* Uhhuh. We've got a bio that straddles the device size! */
3080 truncated_bytes = bio->bi_iter.bi_size - (maxsector << 9);
3083 * The bio contains more than one segment which spans EOD, just return
3084 * and let IO layer turn it into an EIO
3086 if (truncated_bytes > bvec->bv_len)
3089 /* Truncate the bio.. */
3090 bio->bi_iter.bi_size -= truncated_bytes;
3091 bvec->bv_len -= truncated_bytes;
3093 /* ..and clear the end of the buffer for reads */
3094 if (op == REQ_OP_READ) {
3095 zero_user(bvec->bv_page, bvec->bv_offset + bvec->bv_len,
3100 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3101 enum rw_hint write_hint, struct writeback_control *wbc)
3105 BUG_ON(!buffer_locked(bh));
3106 BUG_ON(!buffer_mapped(bh));
3107 BUG_ON(!bh->b_end_io);
3108 BUG_ON(buffer_delay(bh));
3109 BUG_ON(buffer_unwritten(bh));
3112 * Only clear out a write error when rewriting
3114 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3115 clear_buffer_write_io_error(bh);
3118 * from here on down, it's all bio -- do the initial mapping,
3119 * submit_bio -> generic_make_request may further map this bio around
3121 bio = bio_alloc(GFP_NOIO, 1);
3124 wbc_init_bio(wbc, bio);
3125 wbc_account_io(wbc, bh->b_page, bh->b_size);
3128 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3129 bio_set_dev(bio, bh->b_bdev);
3130 bio->bi_write_hint = write_hint;
3132 bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3133 BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3135 bio->bi_end_io = end_bio_bh_io_sync;
3136 bio->bi_private = bh;
3138 /* Take care of bh's that straddle the end of the device */
3139 guard_bio_eod(op, bio);
3141 if (buffer_meta(bh))
3142 op_flags |= REQ_META;
3143 if (buffer_prio(bh))
3144 op_flags |= REQ_PRIO;
3145 bio_set_op_attrs(bio, op, op_flags);
3151 int submit_bh(int op, int op_flags, struct buffer_head *bh)
3153 return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3155 EXPORT_SYMBOL(submit_bh);
3158 * ll_rw_block: low-level access to block devices (DEPRECATED)
3159 * @op: whether to %READ or %WRITE
3160 * @op_flags: req_flag_bits
3161 * @nr: number of &struct buffer_heads in the array
3162 * @bhs: array of pointers to &struct buffer_head
3164 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3165 * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3166 * @op_flags contains flags modifying the detailed I/O behavior, most notably
3169 * This function drops any buffer that it cannot get a lock on (with the
3170 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3171 * request, and any buffer that appears to be up-to-date when doing read
3172 * request. Further it marks as clean buffers that are processed for
3173 * writing (the buffer cache won't assume that they are actually clean
3174 * until the buffer gets unlocked).
3176 * ll_rw_block sets b_end_io to simple completion handler that marks
3177 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3180 * All of the buffers must be for the same device, and must also be a
3181 * multiple of the current approved size for the device.
3183 void ll_rw_block(int op, int op_flags, int nr, struct buffer_head *bhs[])
3187 for (i = 0; i < nr; i++) {
3188 struct buffer_head *bh = bhs[i];
3190 if (!trylock_buffer(bh))
3193 if (test_clear_buffer_dirty(bh)) {
3194 bh->b_end_io = end_buffer_write_sync;
3196 submit_bh(op, op_flags, bh);
3200 if (!buffer_uptodate(bh)) {
3201 bh->b_end_io = end_buffer_read_sync;
3203 submit_bh(op, op_flags, bh);
3210 EXPORT_SYMBOL(ll_rw_block);
3212 void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3215 if (!test_clear_buffer_dirty(bh)) {
3219 bh->b_end_io = end_buffer_write_sync;
3221 submit_bh(REQ_OP_WRITE, op_flags, bh);
3223 EXPORT_SYMBOL(write_dirty_buffer);
3226 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3227 * and then start new I/O and then wait upon it. The caller must have a ref on
3230 int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3234 WARN_ON(atomic_read(&bh->b_count) < 1);
3236 if (test_clear_buffer_dirty(bh)) {
3238 * The bh should be mapped, but it might not be if the
3239 * device was hot-removed. Not much we can do but fail the I/O.
3241 if (!buffer_mapped(bh)) {
3247 bh->b_end_io = end_buffer_write_sync;
3248 ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3250 if (!ret && !buffer_uptodate(bh))
3257 EXPORT_SYMBOL(__sync_dirty_buffer);
3259 int sync_dirty_buffer(struct buffer_head *bh)
3261 return __sync_dirty_buffer(bh, REQ_SYNC);
3263 EXPORT_SYMBOL(sync_dirty_buffer);
3266 * try_to_free_buffers() checks if all the buffers on this particular page
3267 * are unused, and releases them if so.
3269 * Exclusion against try_to_free_buffers may be obtained by either
3270 * locking the page or by holding its mapping's private_lock.
3272 * If the page is dirty but all the buffers are clean then we need to
3273 * be sure to mark the page clean as well. This is because the page
3274 * may be against a block device, and a later reattachment of buffers
3275 * to a dirty page will set *all* buffers dirty. Which would corrupt
3276 * filesystem data on the same device.
3278 * The same applies to regular filesystem pages: if all the buffers are
3279 * clean then we set the page clean and proceed. To do that, we require
3280 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3283 * try_to_free_buffers() is non-blocking.
3285 static inline int buffer_busy(struct buffer_head *bh)
3287 return atomic_read(&bh->b_count) |
3288 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3292 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3294 struct buffer_head *head = page_buffers(page);
3295 struct buffer_head *bh;
3299 if (buffer_busy(bh))
3301 bh = bh->b_this_page;
3302 } while (bh != head);
3305 struct buffer_head *next = bh->b_this_page;
3307 if (bh->b_assoc_map)
3308 __remove_assoc_queue(bh);
3310 } while (bh != head);
3311 *buffers_to_free = head;
3312 __clear_page_buffers(page);
3318 int try_to_free_buffers(struct page *page)
3320 struct address_space * const mapping = page->mapping;
3321 struct buffer_head *buffers_to_free = NULL;
3324 BUG_ON(!PageLocked(page));
3325 if (PageWriteback(page))
3328 if (mapping == NULL) { /* can this still happen? */
3329 ret = drop_buffers(page, &buffers_to_free);
3333 spin_lock(&mapping->private_lock);
3334 ret = drop_buffers(page, &buffers_to_free);
3337 * If the filesystem writes its buffers by hand (eg ext3)
3338 * then we can have clean buffers against a dirty page. We
3339 * clean the page here; otherwise the VM will never notice
3340 * that the filesystem did any IO at all.
3342 * Also, during truncate, discard_buffer will have marked all
3343 * the page's buffers clean. We discover that here and clean
3346 * private_lock must be held over this entire operation in order
3347 * to synchronise against __set_page_dirty_buffers and prevent the
3348 * dirty bit from being lost.
3351 cancel_dirty_page(page);
3352 spin_unlock(&mapping->private_lock);
3354 if (buffers_to_free) {
3355 struct buffer_head *bh = buffers_to_free;
3358 struct buffer_head *next = bh->b_this_page;
3359 free_buffer_head(bh);
3361 } while (bh != buffers_to_free);
3365 EXPORT_SYMBOL(try_to_free_buffers);
3368 * There are no bdflush tunables left. But distributions are
3369 * still running obsolete flush daemons, so we terminate them here.
3371 * Use of bdflush() is deprecated and will be removed in a future kernel.
3372 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3374 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3376 static int msg_count;
3378 if (!capable(CAP_SYS_ADMIN))
3381 if (msg_count < 5) {
3384 "warning: process `%s' used the obsolete bdflush"
3385 " system call\n", current->comm);
3386 printk(KERN_INFO "Fix your initscripts?\n");
3395 * Buffer-head allocation
3397 static struct kmem_cache *bh_cachep __read_mostly;
3400 * Once the number of bh's in the machine exceeds this level, we start
3401 * stripping them in writeback.
3403 static unsigned long max_buffer_heads;
3405 int buffer_heads_over_limit;
3407 struct bh_accounting {
3408 int nr; /* Number of live bh's */
3409 int ratelimit; /* Limit cacheline bouncing */
3412 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3414 static void recalc_bh_state(void)
3419 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3421 __this_cpu_write(bh_accounting.ratelimit, 0);
3422 for_each_online_cpu(i)
3423 tot += per_cpu(bh_accounting, i).nr;
3424 buffer_heads_over_limit = (tot > max_buffer_heads);
3427 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3429 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3431 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3433 __this_cpu_inc(bh_accounting.nr);
3439 EXPORT_SYMBOL(alloc_buffer_head);
3441 void free_buffer_head(struct buffer_head *bh)
3443 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3444 kmem_cache_free(bh_cachep, bh);
3446 __this_cpu_dec(bh_accounting.nr);
3450 EXPORT_SYMBOL(free_buffer_head);
3452 static int buffer_exit_cpu_dead(unsigned int cpu)
3455 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3457 for (i = 0; i < BH_LRU_SIZE; i++) {
3461 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3462 per_cpu(bh_accounting, cpu).nr = 0;
3467 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3468 * @bh: struct buffer_head
3470 * Return true if the buffer is up-to-date and false,
3471 * with the buffer locked, if not.
3473 int bh_uptodate_or_lock(struct buffer_head *bh)
3475 if (!buffer_uptodate(bh)) {
3477 if (!buffer_uptodate(bh))
3483 EXPORT_SYMBOL(bh_uptodate_or_lock);
3486 * bh_submit_read - Submit a locked buffer for reading
3487 * @bh: struct buffer_head
3489 * Returns zero on success and -EIO on error.
3491 int bh_submit_read(struct buffer_head *bh)
3493 BUG_ON(!buffer_locked(bh));
3495 if (buffer_uptodate(bh)) {
3501 bh->b_end_io = end_buffer_read_sync;
3502 submit_bh(REQ_OP_READ, 0, bh);
3504 if (buffer_uptodate(bh))
3508 EXPORT_SYMBOL(bh_submit_read);
3511 * Seek for SEEK_DATA / SEEK_HOLE within @page, starting at @lastoff.
3513 * Returns the offset within the file on success, and -ENOENT otherwise.
3516 page_seek_hole_data(struct page *page, loff_t lastoff, int whence)
3518 loff_t offset = page_offset(page);
3519 struct buffer_head *bh, *head;
3520 bool seek_data = whence == SEEK_DATA;
3522 if (lastoff < offset)
3525 bh = head = page_buffers(page);
3527 offset += bh->b_size;
3528 if (lastoff >= offset)
3532 * Unwritten extents that have data in the page cache covering
3533 * them can be identified by the BH_Unwritten state flag.
3534 * Pages with multiple buffers might have a mix of holes, data
3535 * and unwritten extents - any buffer with valid data in it
3536 * should have BH_Uptodate flag set on it.
3539 if ((buffer_unwritten(bh) || buffer_uptodate(bh)) == seek_data)
3543 } while ((bh = bh->b_this_page) != head);
3548 * Seek for SEEK_DATA / SEEK_HOLE in the page cache.
3550 * Within unwritten extents, the page cache determines which parts are holes
3551 * and which are data: unwritten and uptodate buffer heads count as data;
3552 * everything else counts as a hole.
3554 * Returns the resulting offset on successs, and -ENOENT otherwise.
3557 page_cache_seek_hole_data(struct inode *inode, loff_t offset, loff_t length,
3560 pgoff_t index = offset >> PAGE_SHIFT;
3561 pgoff_t end = DIV_ROUND_UP(offset + length, PAGE_SIZE);
3562 loff_t lastoff = offset;
3563 struct pagevec pvec;
3568 pagevec_init(&pvec, 0);
3571 unsigned nr_pages, i;
3573 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping, &index,
3578 for (i = 0; i < nr_pages; i++) {
3579 struct page *page = pvec.pages[i];
3582 * At this point, the page may be truncated or
3583 * invalidated (changing page->mapping to NULL), or
3584 * even swizzled back from swapper_space to tmpfs file
3585 * mapping. However, page->index will not change
3586 * because we have a reference on the page.
3588 * If current page offset is beyond where we've ended,
3589 * we've found a hole.
3591 if (whence == SEEK_HOLE &&
3592 lastoff < page_offset(page))
3596 if (likely(page->mapping == inode->i_mapping) &&
3597 page_has_buffers(page)) {
3598 lastoff = page_seek_hole_data(page, lastoff, whence);
3605 lastoff = page_offset(page) + PAGE_SIZE;
3607 pagevec_release(&pvec);
3608 } while (index < end);
3610 /* When no page at lastoff and we are not done, we found a hole. */
3611 if (whence != SEEK_HOLE)
3615 if (lastoff < offset + length)
3620 pagevec_release(&pvec);
3624 void __init buffer_init(void)
3626 unsigned long nrpages;
3629 bh_cachep = kmem_cache_create("buffer_head",
3630 sizeof(struct buffer_head), 0,
3631 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3636 * Limit the bh occupancy to 10% of ZONE_NORMAL
3638 nrpages = (nr_free_buffer_pages() * 10) / 100;
3639 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3640 ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3641 NULL, buffer_exit_cpu_dead);