Loading...
1/*
2 * linux/fs/buffer.c
3 *
4 * Copyright (C) 1991, 1992, 2002 Linus Torvalds
5 */
6
7/*
8 * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9 *
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
12 *
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
15 *
16 * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17 *
18 * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19 */
20
21#include <linux/kernel.h>
22#include <linux/syscalls.h>
23#include <linux/fs.h>
24#include <linux/iomap.h>
25#include <linux/mm.h>
26#include <linux/percpu.h>
27#include <linux/slab.h>
28#include <linux/capability.h>
29#include <linux/blkdev.h>
30#include <linux/file.h>
31#include <linux/quotaops.h>
32#include <linux/highmem.h>
33#include <linux/export.h>
34#include <linux/backing-dev.h>
35#include <linux/writeback.h>
36#include <linux/hash.h>
37#include <linux/suspend.h>
38#include <linux/buffer_head.h>
39#include <linux/task_io_accounting_ops.h>
40#include <linux/bio.h>
41#include <linux/notifier.h>
42#include <linux/cpu.h>
43#include <linux/bitops.h>
44#include <linux/mpage.h>
45#include <linux/bit_spinlock.h>
46#include <linux/pagevec.h>
47#include <trace/events/block.h>
48
49static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
50static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
51 unsigned long bio_flags,
52 struct writeback_control *wbc);
53
54#define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
55
56void init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
57{
58 bh->b_end_io = handler;
59 bh->b_private = private;
60}
61EXPORT_SYMBOL(init_buffer);
62
63inline void touch_buffer(struct buffer_head *bh)
64{
65 trace_block_touch_buffer(bh);
66 mark_page_accessed(bh->b_page);
67}
68EXPORT_SYMBOL(touch_buffer);
69
70void __lock_buffer(struct buffer_head *bh)
71{
72 wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
73}
74EXPORT_SYMBOL(__lock_buffer);
75
76void unlock_buffer(struct buffer_head *bh)
77{
78 clear_bit_unlock(BH_Lock, &bh->b_state);
79 smp_mb__after_atomic();
80 wake_up_bit(&bh->b_state, BH_Lock);
81}
82EXPORT_SYMBOL(unlock_buffer);
83
84/*
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.
88 */
89void buffer_check_dirty_writeback(struct page *page,
90 bool *dirty, bool *writeback)
91{
92 struct buffer_head *head, *bh;
93 *dirty = false;
94 *writeback = false;
95
96 BUG_ON(!PageLocked(page));
97
98 if (!page_has_buffers(page))
99 return;
100
101 if (PageWriteback(page))
102 *writeback = true;
103
104 head = page_buffers(page);
105 bh = head;
106 do {
107 if (buffer_locked(bh))
108 *writeback = true;
109
110 if (buffer_dirty(bh))
111 *dirty = true;
112
113 bh = bh->b_this_page;
114 } while (bh != head);
115}
116EXPORT_SYMBOL(buffer_check_dirty_writeback);
117
118/*
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.
122 */
123void __wait_on_buffer(struct buffer_head * bh)
124{
125 wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
126}
127EXPORT_SYMBOL(__wait_on_buffer);
128
129static void
130__clear_page_buffers(struct page *page)
131{
132 ClearPagePrivate(page);
133 set_page_private(page, 0);
134 put_page(page);
135}
136
137static void buffer_io_error(struct buffer_head *bh, char *msg)
138{
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);
143}
144
145/*
146 * End-of-IO handler helper function which does not touch the bh after
147 * unlocking it.
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
151 * itself.
152 */
153static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
154{
155 if (uptodate) {
156 set_buffer_uptodate(bh);
157 } else {
158 /* This happens, due to failed read-ahead attempts. */
159 clear_buffer_uptodate(bh);
160 }
161 unlock_buffer(bh);
162}
163
164/*
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.
167 */
168void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
169{
170 __end_buffer_read_notouch(bh, uptodate);
171 put_bh(bh);
172}
173EXPORT_SYMBOL(end_buffer_read_sync);
174
175void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
176{
177 if (uptodate) {
178 set_buffer_uptodate(bh);
179 } else {
180 buffer_io_error(bh, ", lost sync page write");
181 set_buffer_write_io_error(bh);
182 clear_buffer_uptodate(bh);
183 }
184 unlock_buffer(bh);
185 put_bh(bh);
186}
187EXPORT_SYMBOL(end_buffer_write_sync);
188
189/*
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
193 * private_lock.
194 *
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).
199 */
200static struct buffer_head *
201__find_get_block_slow(struct block_device *bdev, sector_t block)
202{
203 struct inode *bd_inode = bdev->bd_inode;
204 struct address_space *bd_mapping = bd_inode->i_mapping;
205 struct buffer_head *ret = NULL;
206 pgoff_t index;
207 struct buffer_head *bh;
208 struct buffer_head *head;
209 struct page *page;
210 int all_mapped = 1;
211
212 index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
213 page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
214 if (!page)
215 goto out;
216
217 spin_lock(&bd_mapping->private_lock);
218 if (!page_has_buffers(page))
219 goto out_unlock;
220 head = page_buffers(page);
221 bh = head;
222 do {
223 if (!buffer_mapped(bh))
224 all_mapped = 0;
225 else if (bh->b_blocknr == block) {
226 ret = bh;
227 get_bh(bh);
228 goto out_unlock;
229 }
230 bh = bh->b_this_page;
231 } while (bh != head);
232
233 /* we might be here because some of the buffers on this page are
234 * not mapped. This is due to various races between
235 * file io on the block device and getblk. It gets dealt with
236 * elsewhere, don't buffer_error if we had some unmapped buffers
237 */
238 if (all_mapped) {
239 printk("__find_get_block_slow() failed. "
240 "block=%llu, b_blocknr=%llu\n",
241 (unsigned long long)block,
242 (unsigned long long)bh->b_blocknr);
243 printk("b_state=0x%08lx, b_size=%zu\n",
244 bh->b_state, bh->b_size);
245 printk("device %pg blocksize: %d\n", bdev,
246 1 << bd_inode->i_blkbits);
247 }
248out_unlock:
249 spin_unlock(&bd_mapping->private_lock);
250 put_page(page);
251out:
252 return ret;
253}
254
255/*
256 * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
257 */
258static void free_more_memory(void)
259{
260 struct zoneref *z;
261 int nid;
262
263 wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
264 yield();
265
266 for_each_online_node(nid) {
267
268 z = first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
269 gfp_zone(GFP_NOFS), NULL);
270 if (z->zone)
271 try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
272 GFP_NOFS, NULL);
273 }
274}
275
276/*
277 * I/O completion handler for block_read_full_page() - pages
278 * which come unlocked at the end of I/O.
279 */
280static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
281{
282 unsigned long flags;
283 struct buffer_head *first;
284 struct buffer_head *tmp;
285 struct page *page;
286 int page_uptodate = 1;
287
288 BUG_ON(!buffer_async_read(bh));
289
290 page = bh->b_page;
291 if (uptodate) {
292 set_buffer_uptodate(bh);
293 } else {
294 clear_buffer_uptodate(bh);
295 buffer_io_error(bh, ", async page read");
296 SetPageError(page);
297 }
298
299 /*
300 * Be _very_ careful from here on. Bad things can happen if
301 * two buffer heads end IO at almost the same time and both
302 * decide that the page is now completely done.
303 */
304 first = page_buffers(page);
305 local_irq_save(flags);
306 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
307 clear_buffer_async_read(bh);
308 unlock_buffer(bh);
309 tmp = bh;
310 do {
311 if (!buffer_uptodate(tmp))
312 page_uptodate = 0;
313 if (buffer_async_read(tmp)) {
314 BUG_ON(!buffer_locked(tmp));
315 goto still_busy;
316 }
317 tmp = tmp->b_this_page;
318 } while (tmp != bh);
319 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
320 local_irq_restore(flags);
321
322 /*
323 * If none of the buffers had errors and they are all
324 * uptodate then we can set the page uptodate.
325 */
326 if (page_uptodate && !PageError(page))
327 SetPageUptodate(page);
328 unlock_page(page);
329 return;
330
331still_busy:
332 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
333 local_irq_restore(flags);
334 return;
335}
336
337/*
338 * Completion handler for block_write_full_page() - pages which are unlocked
339 * during I/O, and which have PageWriteback cleared upon I/O completion.
340 */
341void end_buffer_async_write(struct buffer_head *bh, int uptodate)
342{
343 unsigned long flags;
344 struct buffer_head *first;
345 struct buffer_head *tmp;
346 struct page *page;
347
348 BUG_ON(!buffer_async_write(bh));
349
350 page = bh->b_page;
351 if (uptodate) {
352 set_buffer_uptodate(bh);
353 } else {
354 buffer_io_error(bh, ", lost async page write");
355 mapping_set_error(page->mapping, -EIO);
356 set_buffer_write_io_error(bh);
357 clear_buffer_uptodate(bh);
358 SetPageError(page);
359 }
360
361 first = page_buffers(page);
362 local_irq_save(flags);
363 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
364
365 clear_buffer_async_write(bh);
366 unlock_buffer(bh);
367 tmp = bh->b_this_page;
368 while (tmp != bh) {
369 if (buffer_async_write(tmp)) {
370 BUG_ON(!buffer_locked(tmp));
371 goto still_busy;
372 }
373 tmp = tmp->b_this_page;
374 }
375 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
376 local_irq_restore(flags);
377 end_page_writeback(page);
378 return;
379
380still_busy:
381 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
382 local_irq_restore(flags);
383 return;
384}
385EXPORT_SYMBOL(end_buffer_async_write);
386
387/*
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.
395 *
396 * The page comes unlocked when it has no locked buffer_async buffers
397 * left.
398 *
399 * PageLocked prevents anyone starting new async I/O reads any of
400 * the buffers.
401 *
402 * PageWriteback is used to prevent simultaneous writeout of the same
403 * page.
404 *
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).
407 */
408static void mark_buffer_async_read(struct buffer_head *bh)
409{
410 bh->b_end_io = end_buffer_async_read;
411 set_buffer_async_read(bh);
412}
413
414static void mark_buffer_async_write_endio(struct buffer_head *bh,
415 bh_end_io_t *handler)
416{
417 bh->b_end_io = handler;
418 set_buffer_async_write(bh);
419}
420
421void mark_buffer_async_write(struct buffer_head *bh)
422{
423 mark_buffer_async_write_endio(bh, end_buffer_async_write);
424}
425EXPORT_SYMBOL(mark_buffer_async_write);
426
427
428/*
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.
434 *
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.
438 *
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
448 * ->private_lock.
449 *
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.
452 *
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.
457 *
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).
461 *
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
465 * queued up.
466 *
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
474 * b_inode back.
475 */
476
477/*
478 * The buffer's backing address_space's private_lock must be held
479 */
480static void __remove_assoc_queue(struct buffer_head *bh)
481{
482 list_del_init(&bh->b_assoc_buffers);
483 WARN_ON(!bh->b_assoc_map);
484 if (buffer_write_io_error(bh))
485 set_bit(AS_EIO, &bh->b_assoc_map->flags);
486 bh->b_assoc_map = NULL;
487}
488
489int inode_has_buffers(struct inode *inode)
490{
491 return !list_empty(&inode->i_data.private_list);
492}
493
494/*
495 * osync is designed to support O_SYNC io. It waits synchronously for
496 * all already-submitted IO to complete, but does not queue any new
497 * writes to the disk.
498 *
499 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
500 * you dirty the buffers, and then use osync_inode_buffers to wait for
501 * completion. Any other dirty buffers which are not yet queued for
502 * write will not be flushed to disk by the osync.
503 */
504static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
505{
506 struct buffer_head *bh;
507 struct list_head *p;
508 int err = 0;
509
510 spin_lock(lock);
511repeat:
512 list_for_each_prev(p, list) {
513 bh = BH_ENTRY(p);
514 if (buffer_locked(bh)) {
515 get_bh(bh);
516 spin_unlock(lock);
517 wait_on_buffer(bh);
518 if (!buffer_uptodate(bh))
519 err = -EIO;
520 brelse(bh);
521 spin_lock(lock);
522 goto repeat;
523 }
524 }
525 spin_unlock(lock);
526 return err;
527}
528
529static void do_thaw_one(struct super_block *sb, void *unused)
530{
531 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
532 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
533}
534
535static void do_thaw_all(struct work_struct *work)
536{
537 iterate_supers(do_thaw_one, NULL);
538 kfree(work);
539 printk(KERN_WARNING "Emergency Thaw complete\n");
540}
541
542/**
543 * emergency_thaw_all -- forcibly thaw every frozen filesystem
544 *
545 * Used for emergency unfreeze of all filesystems via SysRq
546 */
547void emergency_thaw_all(void)
548{
549 struct work_struct *work;
550
551 work = kmalloc(sizeof(*work), GFP_ATOMIC);
552 if (work) {
553 INIT_WORK(work, do_thaw_all);
554 schedule_work(work);
555 }
556}
557
558/**
559 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
560 * @mapping: the mapping which wants those buffers written
561 *
562 * Starts I/O against the buffers at mapping->private_list, and waits upon
563 * that I/O.
564 *
565 * Basically, this is a convenience function for fsync().
566 * @mapping is a file or directory which needs those buffers to be written for
567 * a successful fsync().
568 */
569int sync_mapping_buffers(struct address_space *mapping)
570{
571 struct address_space *buffer_mapping = mapping->private_data;
572
573 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
574 return 0;
575
576 return fsync_buffers_list(&buffer_mapping->private_lock,
577 &mapping->private_list);
578}
579EXPORT_SYMBOL(sync_mapping_buffers);
580
581/*
582 * Called when we've recently written block `bblock', and it is known that
583 * `bblock' was for a buffer_boundary() buffer. This means that the block at
584 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
585 * dirty, schedule it for IO. So that indirects merge nicely with their data.
586 */
587void write_boundary_block(struct block_device *bdev,
588 sector_t bblock, unsigned blocksize)
589{
590 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
591 if (bh) {
592 if (buffer_dirty(bh))
593 ll_rw_block(REQ_OP_WRITE, 0, 1, &bh);
594 put_bh(bh);
595 }
596}
597
598void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
599{
600 struct address_space *mapping = inode->i_mapping;
601 struct address_space *buffer_mapping = bh->b_page->mapping;
602
603 mark_buffer_dirty(bh);
604 if (!mapping->private_data) {
605 mapping->private_data = buffer_mapping;
606 } else {
607 BUG_ON(mapping->private_data != buffer_mapping);
608 }
609 if (!bh->b_assoc_map) {
610 spin_lock(&buffer_mapping->private_lock);
611 list_move_tail(&bh->b_assoc_buffers,
612 &mapping->private_list);
613 bh->b_assoc_map = mapping;
614 spin_unlock(&buffer_mapping->private_lock);
615 }
616}
617EXPORT_SYMBOL(mark_buffer_dirty_inode);
618
619/*
620 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
621 * dirty.
622 *
623 * If warn is true, then emit a warning if the page is not uptodate and has
624 * not been truncated.
625 *
626 * The caller must hold lock_page_memcg().
627 */
628static void __set_page_dirty(struct page *page, struct address_space *mapping,
629 int warn)
630{
631 unsigned long flags;
632
633 spin_lock_irqsave(&mapping->tree_lock, flags);
634 if (page->mapping) { /* Race with truncate? */
635 WARN_ON_ONCE(warn && !PageUptodate(page));
636 account_page_dirtied(page, mapping);
637 radix_tree_tag_set(&mapping->page_tree,
638 page_index(page), PAGECACHE_TAG_DIRTY);
639 }
640 spin_unlock_irqrestore(&mapping->tree_lock, flags);
641}
642
643/*
644 * Add a page to the dirty page list.
645 *
646 * It is a sad fact of life that this function is called from several places
647 * deeply under spinlocking. It may not sleep.
648 *
649 * If the page has buffers, the uptodate buffers are set dirty, to preserve
650 * dirty-state coherency between the page and the buffers. It the page does
651 * not have buffers then when they are later attached they will all be set
652 * dirty.
653 *
654 * The buffers are dirtied before the page is dirtied. There's a small race
655 * window in which a writepage caller may see the page cleanness but not the
656 * buffer dirtiness. That's fine. If this code were to set the page dirty
657 * before the buffers, a concurrent writepage caller could clear the page dirty
658 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
659 * page on the dirty page list.
660 *
661 * We use private_lock to lock against try_to_free_buffers while using the
662 * page's buffer list. Also use this to protect against clean buffers being
663 * added to the page after it was set dirty.
664 *
665 * FIXME: may need to call ->reservepage here as well. That's rather up to the
666 * address_space though.
667 */
668int __set_page_dirty_buffers(struct page *page)
669{
670 int newly_dirty;
671 struct address_space *mapping = page_mapping(page);
672
673 if (unlikely(!mapping))
674 return !TestSetPageDirty(page);
675
676 spin_lock(&mapping->private_lock);
677 if (page_has_buffers(page)) {
678 struct buffer_head *head = page_buffers(page);
679 struct buffer_head *bh = head;
680
681 do {
682 set_buffer_dirty(bh);
683 bh = bh->b_this_page;
684 } while (bh != head);
685 }
686 /*
687 * Lock out page->mem_cgroup migration to keep PageDirty
688 * synchronized with per-memcg dirty page counters.
689 */
690 lock_page_memcg(page);
691 newly_dirty = !TestSetPageDirty(page);
692 spin_unlock(&mapping->private_lock);
693
694 if (newly_dirty)
695 __set_page_dirty(page, mapping, 1);
696
697 unlock_page_memcg(page);
698
699 if (newly_dirty)
700 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
701
702 return newly_dirty;
703}
704EXPORT_SYMBOL(__set_page_dirty_buffers);
705
706/*
707 * Write out and wait upon a list of buffers.
708 *
709 * We have conflicting pressures: we want to make sure that all
710 * initially dirty buffers get waited on, but that any subsequently
711 * dirtied buffers don't. After all, we don't want fsync to last
712 * forever if somebody is actively writing to the file.
713 *
714 * Do this in two main stages: first we copy dirty buffers to a
715 * temporary inode list, queueing the writes as we go. Then we clean
716 * up, waiting for those writes to complete.
717 *
718 * During this second stage, any subsequent updates to the file may end
719 * up refiling the buffer on the original inode's dirty list again, so
720 * there is a chance we will end up with a buffer queued for write but
721 * not yet completed on that list. So, as a final cleanup we go through
722 * the osync code to catch these locked, dirty buffers without requeuing
723 * any newly dirty buffers for write.
724 */
725static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
726{
727 struct buffer_head *bh;
728 struct list_head tmp;
729 struct address_space *mapping;
730 int err = 0, err2;
731 struct blk_plug plug;
732
733 INIT_LIST_HEAD(&tmp);
734 blk_start_plug(&plug);
735
736 spin_lock(lock);
737 while (!list_empty(list)) {
738 bh = BH_ENTRY(list->next);
739 mapping = bh->b_assoc_map;
740 __remove_assoc_queue(bh);
741 /* Avoid race with mark_buffer_dirty_inode() which does
742 * a lockless check and we rely on seeing the dirty bit */
743 smp_mb();
744 if (buffer_dirty(bh) || buffer_locked(bh)) {
745 list_add(&bh->b_assoc_buffers, &tmp);
746 bh->b_assoc_map = mapping;
747 if (buffer_dirty(bh)) {
748 get_bh(bh);
749 spin_unlock(lock);
750 /*
751 * Ensure any pending I/O completes so that
752 * write_dirty_buffer() actually writes the
753 * current contents - it is a noop if I/O is
754 * still in flight on potentially older
755 * contents.
756 */
757 write_dirty_buffer(bh, REQ_SYNC);
758
759 /*
760 * Kick off IO for the previous mapping. Note
761 * that we will not run the very last mapping,
762 * wait_on_buffer() will do that for us
763 * through sync_buffer().
764 */
765 brelse(bh);
766 spin_lock(lock);
767 }
768 }
769 }
770
771 spin_unlock(lock);
772 blk_finish_plug(&plug);
773 spin_lock(lock);
774
775 while (!list_empty(&tmp)) {
776 bh = BH_ENTRY(tmp.prev);
777 get_bh(bh);
778 mapping = bh->b_assoc_map;
779 __remove_assoc_queue(bh);
780 /* Avoid race with mark_buffer_dirty_inode() which does
781 * a lockless check and we rely on seeing the dirty bit */
782 smp_mb();
783 if (buffer_dirty(bh)) {
784 list_add(&bh->b_assoc_buffers,
785 &mapping->private_list);
786 bh->b_assoc_map = mapping;
787 }
788 spin_unlock(lock);
789 wait_on_buffer(bh);
790 if (!buffer_uptodate(bh))
791 err = -EIO;
792 brelse(bh);
793 spin_lock(lock);
794 }
795
796 spin_unlock(lock);
797 err2 = osync_buffers_list(lock, list);
798 if (err)
799 return err;
800 else
801 return err2;
802}
803
804/*
805 * Invalidate any and all dirty buffers on a given inode. We are
806 * probably unmounting the fs, but that doesn't mean we have already
807 * done a sync(). Just drop the buffers from the inode list.
808 *
809 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
810 * assumes that all the buffers are against the blockdev. Not true
811 * for reiserfs.
812 */
813void invalidate_inode_buffers(struct inode *inode)
814{
815 if (inode_has_buffers(inode)) {
816 struct address_space *mapping = &inode->i_data;
817 struct list_head *list = &mapping->private_list;
818 struct address_space *buffer_mapping = mapping->private_data;
819
820 spin_lock(&buffer_mapping->private_lock);
821 while (!list_empty(list))
822 __remove_assoc_queue(BH_ENTRY(list->next));
823 spin_unlock(&buffer_mapping->private_lock);
824 }
825}
826EXPORT_SYMBOL(invalidate_inode_buffers);
827
828/*
829 * Remove any clean buffers from the inode's buffer list. This is called
830 * when we're trying to free the inode itself. Those buffers can pin it.
831 *
832 * Returns true if all buffers were removed.
833 */
834int remove_inode_buffers(struct inode *inode)
835{
836 int ret = 1;
837
838 if (inode_has_buffers(inode)) {
839 struct address_space *mapping = &inode->i_data;
840 struct list_head *list = &mapping->private_list;
841 struct address_space *buffer_mapping = mapping->private_data;
842
843 spin_lock(&buffer_mapping->private_lock);
844 while (!list_empty(list)) {
845 struct buffer_head *bh = BH_ENTRY(list->next);
846 if (buffer_dirty(bh)) {
847 ret = 0;
848 break;
849 }
850 __remove_assoc_queue(bh);
851 }
852 spin_unlock(&buffer_mapping->private_lock);
853 }
854 return ret;
855}
856
857/*
858 * Create the appropriate buffers when given a page for data area and
859 * the size of each buffer.. Use the bh->b_this_page linked list to
860 * follow the buffers created. Return NULL if unable to create more
861 * buffers.
862 *
863 * The retry flag is used to differentiate async IO (paging, swapping)
864 * which may not fail from ordinary buffer allocations.
865 */
866struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
867 int retry)
868{
869 struct buffer_head *bh, *head;
870 long offset;
871
872try_again:
873 head = NULL;
874 offset = PAGE_SIZE;
875 while ((offset -= size) >= 0) {
876 bh = alloc_buffer_head(GFP_NOFS);
877 if (!bh)
878 goto no_grow;
879
880 bh->b_this_page = head;
881 bh->b_blocknr = -1;
882 head = bh;
883
884 bh->b_size = size;
885
886 /* Link the buffer to its page */
887 set_bh_page(bh, page, offset);
888 }
889 return head;
890/*
891 * In case anything failed, we just free everything we got.
892 */
893no_grow:
894 if (head) {
895 do {
896 bh = head;
897 head = head->b_this_page;
898 free_buffer_head(bh);
899 } while (head);
900 }
901
902 /*
903 * Return failure for non-async IO requests. Async IO requests
904 * are not allowed to fail, so we have to wait until buffer heads
905 * become available. But we don't want tasks sleeping with
906 * partially complete buffers, so all were released above.
907 */
908 if (!retry)
909 return NULL;
910
911 /* We're _really_ low on memory. Now we just
912 * wait for old buffer heads to become free due to
913 * finishing IO. Since this is an async request and
914 * the reserve list is empty, we're sure there are
915 * async buffer heads in use.
916 */
917 free_more_memory();
918 goto try_again;
919}
920EXPORT_SYMBOL_GPL(alloc_page_buffers);
921
922static inline void
923link_dev_buffers(struct page *page, struct buffer_head *head)
924{
925 struct buffer_head *bh, *tail;
926
927 bh = head;
928 do {
929 tail = bh;
930 bh = bh->b_this_page;
931 } while (bh);
932 tail->b_this_page = head;
933 attach_page_buffers(page, head);
934}
935
936static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
937{
938 sector_t retval = ~((sector_t)0);
939 loff_t sz = i_size_read(bdev->bd_inode);
940
941 if (sz) {
942 unsigned int sizebits = blksize_bits(size);
943 retval = (sz >> sizebits);
944 }
945 return retval;
946}
947
948/*
949 * Initialise the state of a blockdev page's buffers.
950 */
951static sector_t
952init_page_buffers(struct page *page, struct block_device *bdev,
953 sector_t block, int size)
954{
955 struct buffer_head *head = page_buffers(page);
956 struct buffer_head *bh = head;
957 int uptodate = PageUptodate(page);
958 sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
959
960 do {
961 if (!buffer_mapped(bh)) {
962 init_buffer(bh, NULL, NULL);
963 bh->b_bdev = bdev;
964 bh->b_blocknr = block;
965 if (uptodate)
966 set_buffer_uptodate(bh);
967 if (block < end_block)
968 set_buffer_mapped(bh);
969 }
970 block++;
971 bh = bh->b_this_page;
972 } while (bh != head);
973
974 /*
975 * Caller needs to validate requested block against end of device.
976 */
977 return end_block;
978}
979
980/*
981 * Create the page-cache page that contains the requested block.
982 *
983 * This is used purely for blockdev mappings.
984 */
985static int
986grow_dev_page(struct block_device *bdev, sector_t block,
987 pgoff_t index, int size, int sizebits, gfp_t gfp)
988{
989 struct inode *inode = bdev->bd_inode;
990 struct page *page;
991 struct buffer_head *bh;
992 sector_t end_block;
993 int ret = 0; /* Will call free_more_memory() */
994 gfp_t gfp_mask;
995
996 gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
997
998 /*
999 * XXX: __getblk_slow() can not really deal with failure and
1000 * will endlessly loop on improvised global reclaim. Prefer
1001 * looping in the allocator rather than here, at least that
1002 * code knows what it's doing.
1003 */
1004 gfp_mask |= __GFP_NOFAIL;
1005
1006 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
1007 if (!page)
1008 return ret;
1009
1010 BUG_ON(!PageLocked(page));
1011
1012 if (page_has_buffers(page)) {
1013 bh = page_buffers(page);
1014 if (bh->b_size == size) {
1015 end_block = init_page_buffers(page, bdev,
1016 (sector_t)index << sizebits,
1017 size);
1018 goto done;
1019 }
1020 if (!try_to_free_buffers(page))
1021 goto failed;
1022 }
1023
1024 /*
1025 * Allocate some buffers for this page
1026 */
1027 bh = alloc_page_buffers(page, size, 0);
1028 if (!bh)
1029 goto failed;
1030
1031 /*
1032 * Link the page to the buffers and initialise them. Take the
1033 * lock to be atomic wrt __find_get_block(), which does not
1034 * run under the page lock.
1035 */
1036 spin_lock(&inode->i_mapping->private_lock);
1037 link_dev_buffers(page, bh);
1038 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1039 size);
1040 spin_unlock(&inode->i_mapping->private_lock);
1041done:
1042 ret = (block < end_block) ? 1 : -ENXIO;
1043failed:
1044 unlock_page(page);
1045 put_page(page);
1046 return ret;
1047}
1048
1049/*
1050 * Create buffers for the specified block device block's page. If
1051 * that page was dirty, the buffers are set dirty also.
1052 */
1053static int
1054grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
1055{
1056 pgoff_t index;
1057 int sizebits;
1058
1059 sizebits = -1;
1060 do {
1061 sizebits++;
1062 } while ((size << sizebits) < PAGE_SIZE);
1063
1064 index = block >> sizebits;
1065
1066 /*
1067 * Check for a block which wants to lie outside our maximum possible
1068 * pagecache index. (this comparison is done using sector_t types).
1069 */
1070 if (unlikely(index != block >> sizebits)) {
1071 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1072 "device %pg\n",
1073 __func__, (unsigned long long)block,
1074 bdev);
1075 return -EIO;
1076 }
1077
1078 /* Create a page with the proper size buffers.. */
1079 return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1080}
1081
1082static struct buffer_head *
1083__getblk_slow(struct block_device *bdev, sector_t block,
1084 unsigned size, gfp_t gfp)
1085{
1086 /* Size must be multiple of hard sectorsize */
1087 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1088 (size < 512 || size > PAGE_SIZE))) {
1089 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1090 size);
1091 printk(KERN_ERR "logical block size: %d\n",
1092 bdev_logical_block_size(bdev));
1093
1094 dump_stack();
1095 return NULL;
1096 }
1097
1098 for (;;) {
1099 struct buffer_head *bh;
1100 int ret;
1101
1102 bh = __find_get_block(bdev, block, size);
1103 if (bh)
1104 return bh;
1105
1106 ret = grow_buffers(bdev, block, size, gfp);
1107 if (ret < 0)
1108 return NULL;
1109 if (ret == 0)
1110 free_more_memory();
1111 }
1112}
1113
1114/*
1115 * The relationship between dirty buffers and dirty pages:
1116 *
1117 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1118 * the page is tagged dirty in its radix tree.
1119 *
1120 * At all times, the dirtiness of the buffers represents the dirtiness of
1121 * subsections of the page. If the page has buffers, the page dirty bit is
1122 * merely a hint about the true dirty state.
1123 *
1124 * When a page is set dirty in its entirety, all its buffers are marked dirty
1125 * (if the page has buffers).
1126 *
1127 * When a buffer is marked dirty, its page is dirtied, but the page's other
1128 * buffers are not.
1129 *
1130 * Also. When blockdev buffers are explicitly read with bread(), they
1131 * individually become uptodate. But their backing page remains not
1132 * uptodate - even if all of its buffers are uptodate. A subsequent
1133 * block_read_full_page() against that page will discover all the uptodate
1134 * buffers, will set the page uptodate and will perform no I/O.
1135 */
1136
1137/**
1138 * mark_buffer_dirty - mark a buffer_head as needing writeout
1139 * @bh: the buffer_head to mark dirty
1140 *
1141 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1142 * backing page dirty, then tag the page as dirty in its address_space's radix
1143 * tree and then attach the address_space's inode to its superblock's dirty
1144 * inode list.
1145 *
1146 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1147 * mapping->tree_lock and mapping->host->i_lock.
1148 */
1149void mark_buffer_dirty(struct buffer_head *bh)
1150{
1151 WARN_ON_ONCE(!buffer_uptodate(bh));
1152
1153 trace_block_dirty_buffer(bh);
1154
1155 /*
1156 * Very *carefully* optimize the it-is-already-dirty case.
1157 *
1158 * Don't let the final "is it dirty" escape to before we
1159 * perhaps modified the buffer.
1160 */
1161 if (buffer_dirty(bh)) {
1162 smp_mb();
1163 if (buffer_dirty(bh))
1164 return;
1165 }
1166
1167 if (!test_set_buffer_dirty(bh)) {
1168 struct page *page = bh->b_page;
1169 struct address_space *mapping = NULL;
1170
1171 lock_page_memcg(page);
1172 if (!TestSetPageDirty(page)) {
1173 mapping = page_mapping(page);
1174 if (mapping)
1175 __set_page_dirty(page, mapping, 0);
1176 }
1177 unlock_page_memcg(page);
1178 if (mapping)
1179 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1180 }
1181}
1182EXPORT_SYMBOL(mark_buffer_dirty);
1183
1184/*
1185 * Decrement a buffer_head's reference count. If all buffers against a page
1186 * have zero reference count, are clean and unlocked, and if the page is clean
1187 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1188 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1189 * a page but it ends up not being freed, and buffers may later be reattached).
1190 */
1191void __brelse(struct buffer_head * buf)
1192{
1193 if (atomic_read(&buf->b_count)) {
1194 put_bh(buf);
1195 return;
1196 }
1197 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1198}
1199EXPORT_SYMBOL(__brelse);
1200
1201/*
1202 * bforget() is like brelse(), except it discards any
1203 * potentially dirty data.
1204 */
1205void __bforget(struct buffer_head *bh)
1206{
1207 clear_buffer_dirty(bh);
1208 if (bh->b_assoc_map) {
1209 struct address_space *buffer_mapping = bh->b_page->mapping;
1210
1211 spin_lock(&buffer_mapping->private_lock);
1212 list_del_init(&bh->b_assoc_buffers);
1213 bh->b_assoc_map = NULL;
1214 spin_unlock(&buffer_mapping->private_lock);
1215 }
1216 __brelse(bh);
1217}
1218EXPORT_SYMBOL(__bforget);
1219
1220static struct buffer_head *__bread_slow(struct buffer_head *bh)
1221{
1222 lock_buffer(bh);
1223 if (buffer_uptodate(bh)) {
1224 unlock_buffer(bh);
1225 return bh;
1226 } else {
1227 get_bh(bh);
1228 bh->b_end_io = end_buffer_read_sync;
1229 submit_bh(REQ_OP_READ, 0, bh);
1230 wait_on_buffer(bh);
1231 if (buffer_uptodate(bh))
1232 return bh;
1233 }
1234 brelse(bh);
1235 return NULL;
1236}
1237
1238/*
1239 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1240 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1241 * refcount elevated by one when they're in an LRU. A buffer can only appear
1242 * once in a particular CPU's LRU. A single buffer can be present in multiple
1243 * CPU's LRUs at the same time.
1244 *
1245 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1246 * sb_find_get_block().
1247 *
1248 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1249 * a local interrupt disable for that.
1250 */
1251
1252#define BH_LRU_SIZE 16
1253
1254struct bh_lru {
1255 struct buffer_head *bhs[BH_LRU_SIZE];
1256};
1257
1258static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1259
1260#ifdef CONFIG_SMP
1261#define bh_lru_lock() local_irq_disable()
1262#define bh_lru_unlock() local_irq_enable()
1263#else
1264#define bh_lru_lock() preempt_disable()
1265#define bh_lru_unlock() preempt_enable()
1266#endif
1267
1268static inline void check_irqs_on(void)
1269{
1270#ifdef irqs_disabled
1271 BUG_ON(irqs_disabled());
1272#endif
1273}
1274
1275/*
1276 * The LRU management algorithm is dopey-but-simple. Sorry.
1277 */
1278static void bh_lru_install(struct buffer_head *bh)
1279{
1280 struct buffer_head *evictee = NULL;
1281
1282 check_irqs_on();
1283 bh_lru_lock();
1284 if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1285 struct buffer_head *bhs[BH_LRU_SIZE];
1286 int in;
1287 int out = 0;
1288
1289 get_bh(bh);
1290 bhs[out++] = bh;
1291 for (in = 0; in < BH_LRU_SIZE; in++) {
1292 struct buffer_head *bh2 =
1293 __this_cpu_read(bh_lrus.bhs[in]);
1294
1295 if (bh2 == bh) {
1296 __brelse(bh2);
1297 } else {
1298 if (out >= BH_LRU_SIZE) {
1299 BUG_ON(evictee != NULL);
1300 evictee = bh2;
1301 } else {
1302 bhs[out++] = bh2;
1303 }
1304 }
1305 }
1306 while (out < BH_LRU_SIZE)
1307 bhs[out++] = NULL;
1308 memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1309 }
1310 bh_lru_unlock();
1311
1312 if (evictee)
1313 __brelse(evictee);
1314}
1315
1316/*
1317 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1318 */
1319static struct buffer_head *
1320lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1321{
1322 struct buffer_head *ret = NULL;
1323 unsigned int i;
1324
1325 check_irqs_on();
1326 bh_lru_lock();
1327 for (i = 0; i < BH_LRU_SIZE; i++) {
1328 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1329
1330 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1331 bh->b_size == size) {
1332 if (i) {
1333 while (i) {
1334 __this_cpu_write(bh_lrus.bhs[i],
1335 __this_cpu_read(bh_lrus.bhs[i - 1]));
1336 i--;
1337 }
1338 __this_cpu_write(bh_lrus.bhs[0], bh);
1339 }
1340 get_bh(bh);
1341 ret = bh;
1342 break;
1343 }
1344 }
1345 bh_lru_unlock();
1346 return ret;
1347}
1348
1349/*
1350 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1351 * it in the LRU and mark it as accessed. If it is not present then return
1352 * NULL
1353 */
1354struct buffer_head *
1355__find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1356{
1357 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1358
1359 if (bh == NULL) {
1360 /* __find_get_block_slow will mark the page accessed */
1361 bh = __find_get_block_slow(bdev, block);
1362 if (bh)
1363 bh_lru_install(bh);
1364 } else
1365 touch_buffer(bh);
1366
1367 return bh;
1368}
1369EXPORT_SYMBOL(__find_get_block);
1370
1371/*
1372 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1373 * which corresponds to the passed block_device, block and size. The
1374 * returned buffer has its reference count incremented.
1375 *
1376 * __getblk_gfp() will lock up the machine if grow_dev_page's
1377 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1378 */
1379struct buffer_head *
1380__getblk_gfp(struct block_device *bdev, sector_t block,
1381 unsigned size, gfp_t gfp)
1382{
1383 struct buffer_head *bh = __find_get_block(bdev, block, size);
1384
1385 might_sleep();
1386 if (bh == NULL)
1387 bh = __getblk_slow(bdev, block, size, gfp);
1388 return bh;
1389}
1390EXPORT_SYMBOL(__getblk_gfp);
1391
1392/*
1393 * Do async read-ahead on a buffer..
1394 */
1395void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1396{
1397 struct buffer_head *bh = __getblk(bdev, block, size);
1398 if (likely(bh)) {
1399 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1400 brelse(bh);
1401 }
1402}
1403EXPORT_SYMBOL(__breadahead);
1404
1405/**
1406 * __bread_gfp() - reads a specified block and returns the bh
1407 * @bdev: the block_device to read from
1408 * @block: number of block
1409 * @size: size (in bytes) to read
1410 * @gfp: page allocation flag
1411 *
1412 * Reads a specified block, and returns buffer head that contains it.
1413 * The page cache can be allocated from non-movable area
1414 * not to prevent page migration if you set gfp to zero.
1415 * It returns NULL if the block was unreadable.
1416 */
1417struct buffer_head *
1418__bread_gfp(struct block_device *bdev, sector_t block,
1419 unsigned size, gfp_t gfp)
1420{
1421 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1422
1423 if (likely(bh) && !buffer_uptodate(bh))
1424 bh = __bread_slow(bh);
1425 return bh;
1426}
1427EXPORT_SYMBOL(__bread_gfp);
1428
1429/*
1430 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1431 * This doesn't race because it runs in each cpu either in irq
1432 * or with preempt disabled.
1433 */
1434static void invalidate_bh_lru(void *arg)
1435{
1436 struct bh_lru *b = &get_cpu_var(bh_lrus);
1437 int i;
1438
1439 for (i = 0; i < BH_LRU_SIZE; i++) {
1440 brelse(b->bhs[i]);
1441 b->bhs[i] = NULL;
1442 }
1443 put_cpu_var(bh_lrus);
1444}
1445
1446static bool has_bh_in_lru(int cpu, void *dummy)
1447{
1448 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1449 int i;
1450
1451 for (i = 0; i < BH_LRU_SIZE; i++) {
1452 if (b->bhs[i])
1453 return 1;
1454 }
1455
1456 return 0;
1457}
1458
1459void invalidate_bh_lrus(void)
1460{
1461 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1462}
1463EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1464
1465void set_bh_page(struct buffer_head *bh,
1466 struct page *page, unsigned long offset)
1467{
1468 bh->b_page = page;
1469 BUG_ON(offset >= PAGE_SIZE);
1470 if (PageHighMem(page))
1471 /*
1472 * This catches illegal uses and preserves the offset:
1473 */
1474 bh->b_data = (char *)(0 + offset);
1475 else
1476 bh->b_data = page_address(page) + offset;
1477}
1478EXPORT_SYMBOL(set_bh_page);
1479
1480/*
1481 * Called when truncating a buffer on a page completely.
1482 */
1483
1484/* Bits that are cleared during an invalidate */
1485#define BUFFER_FLAGS_DISCARD \
1486 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1487 1 << BH_Delay | 1 << BH_Unwritten)
1488
1489static void discard_buffer(struct buffer_head * bh)
1490{
1491 unsigned long b_state, b_state_old;
1492
1493 lock_buffer(bh);
1494 clear_buffer_dirty(bh);
1495 bh->b_bdev = NULL;
1496 b_state = bh->b_state;
1497 for (;;) {
1498 b_state_old = cmpxchg(&bh->b_state, b_state,
1499 (b_state & ~BUFFER_FLAGS_DISCARD));
1500 if (b_state_old == b_state)
1501 break;
1502 b_state = b_state_old;
1503 }
1504 unlock_buffer(bh);
1505}
1506
1507/**
1508 * block_invalidatepage - invalidate part or all of a buffer-backed page
1509 *
1510 * @page: the page which is affected
1511 * @offset: start of the range to invalidate
1512 * @length: length of the range to invalidate
1513 *
1514 * block_invalidatepage() is called when all or part of the page has become
1515 * invalidated by a truncate operation.
1516 *
1517 * block_invalidatepage() does not have to release all buffers, but it must
1518 * ensure that no dirty buffer is left outside @offset and that no I/O
1519 * is underway against any of the blocks which are outside the truncation
1520 * point. Because the caller is about to free (and possibly reuse) those
1521 * blocks on-disk.
1522 */
1523void block_invalidatepage(struct page *page, unsigned int offset,
1524 unsigned int length)
1525{
1526 struct buffer_head *head, *bh, *next;
1527 unsigned int curr_off = 0;
1528 unsigned int stop = length + offset;
1529
1530 BUG_ON(!PageLocked(page));
1531 if (!page_has_buffers(page))
1532 goto out;
1533
1534 /*
1535 * Check for overflow
1536 */
1537 BUG_ON(stop > PAGE_SIZE || stop < length);
1538
1539 head = page_buffers(page);
1540 bh = head;
1541 do {
1542 unsigned int next_off = curr_off + bh->b_size;
1543 next = bh->b_this_page;
1544
1545 /*
1546 * Are we still fully in range ?
1547 */
1548 if (next_off > stop)
1549 goto out;
1550
1551 /*
1552 * is this block fully invalidated?
1553 */
1554 if (offset <= curr_off)
1555 discard_buffer(bh);
1556 curr_off = next_off;
1557 bh = next;
1558 } while (bh != head);
1559
1560 /*
1561 * We release buffers only if the entire page is being invalidated.
1562 * The get_block cached value has been unconditionally invalidated,
1563 * so real IO is not possible anymore.
1564 */
1565 if (offset == 0)
1566 try_to_release_page(page, 0);
1567out:
1568 return;
1569}
1570EXPORT_SYMBOL(block_invalidatepage);
1571
1572
1573/*
1574 * We attach and possibly dirty the buffers atomically wrt
1575 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1576 * is already excluded via the page lock.
1577 */
1578void create_empty_buffers(struct page *page,
1579 unsigned long blocksize, unsigned long b_state)
1580{
1581 struct buffer_head *bh, *head, *tail;
1582
1583 head = alloc_page_buffers(page, blocksize, 1);
1584 bh = head;
1585 do {
1586 bh->b_state |= b_state;
1587 tail = bh;
1588 bh = bh->b_this_page;
1589 } while (bh);
1590 tail->b_this_page = head;
1591
1592 spin_lock(&page->mapping->private_lock);
1593 if (PageUptodate(page) || PageDirty(page)) {
1594 bh = head;
1595 do {
1596 if (PageDirty(page))
1597 set_buffer_dirty(bh);
1598 if (PageUptodate(page))
1599 set_buffer_uptodate(bh);
1600 bh = bh->b_this_page;
1601 } while (bh != head);
1602 }
1603 attach_page_buffers(page, head);
1604 spin_unlock(&page->mapping->private_lock);
1605}
1606EXPORT_SYMBOL(create_empty_buffers);
1607
1608/**
1609 * clean_bdev_aliases: clean a range of buffers in block device
1610 * @bdev: Block device to clean buffers in
1611 * @block: Start of a range of blocks to clean
1612 * @len: Number of blocks to clean
1613 *
1614 * We are taking a range of blocks for data and we don't want writeback of any
1615 * buffer-cache aliases starting from return from this function and until the
1616 * moment when something will explicitly mark the buffer dirty (hopefully that
1617 * will not happen until we will free that block ;-) We don't even need to mark
1618 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1619 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1620 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1621 * would confuse anyone who might pick it with bread() afterwards...
1622 *
1623 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1624 * writeout I/O going on against recently-freed buffers. We don't wait on that
1625 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1626 * need to. That happens here.
1627 */
1628void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1629{
1630 struct inode *bd_inode = bdev->bd_inode;
1631 struct address_space *bd_mapping = bd_inode->i_mapping;
1632 struct pagevec pvec;
1633 pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1634 pgoff_t end;
1635 int i;
1636 struct buffer_head *bh;
1637 struct buffer_head *head;
1638
1639 end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1640 pagevec_init(&pvec, 0);
1641 while (index <= end && pagevec_lookup(&pvec, bd_mapping, index,
1642 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
1643 for (i = 0; i < pagevec_count(&pvec); i++) {
1644 struct page *page = pvec.pages[i];
1645
1646 index = page->index;
1647 if (index > end)
1648 break;
1649 if (!page_has_buffers(page))
1650 continue;
1651 /*
1652 * We use page lock instead of bd_mapping->private_lock
1653 * to pin buffers here since we can afford to sleep and
1654 * it scales better than a global spinlock lock.
1655 */
1656 lock_page(page);
1657 /* Recheck when the page is locked which pins bhs */
1658 if (!page_has_buffers(page))
1659 goto unlock_page;
1660 head = page_buffers(page);
1661 bh = head;
1662 do {
1663 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1664 goto next;
1665 if (bh->b_blocknr >= block + len)
1666 break;
1667 clear_buffer_dirty(bh);
1668 wait_on_buffer(bh);
1669 clear_buffer_req(bh);
1670next:
1671 bh = bh->b_this_page;
1672 } while (bh != head);
1673unlock_page:
1674 unlock_page(page);
1675 }
1676 pagevec_release(&pvec);
1677 cond_resched();
1678 index++;
1679 }
1680}
1681EXPORT_SYMBOL(clean_bdev_aliases);
1682
1683/*
1684 * Size is a power-of-two in the range 512..PAGE_SIZE,
1685 * and the case we care about most is PAGE_SIZE.
1686 *
1687 * So this *could* possibly be written with those
1688 * constraints in mind (relevant mostly if some
1689 * architecture has a slow bit-scan instruction)
1690 */
1691static inline int block_size_bits(unsigned int blocksize)
1692{
1693 return ilog2(blocksize);
1694}
1695
1696static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1697{
1698 BUG_ON(!PageLocked(page));
1699
1700 if (!page_has_buffers(page))
1701 create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1702 return page_buffers(page);
1703}
1704
1705/*
1706 * NOTE! All mapped/uptodate combinations are valid:
1707 *
1708 * Mapped Uptodate Meaning
1709 *
1710 * No No "unknown" - must do get_block()
1711 * No Yes "hole" - zero-filled
1712 * Yes No "allocated" - allocated on disk, not read in
1713 * Yes Yes "valid" - allocated and up-to-date in memory.
1714 *
1715 * "Dirty" is valid only with the last case (mapped+uptodate).
1716 */
1717
1718/*
1719 * While block_write_full_page is writing back the dirty buffers under
1720 * the page lock, whoever dirtied the buffers may decide to clean them
1721 * again at any time. We handle that by only looking at the buffer
1722 * state inside lock_buffer().
1723 *
1724 * If block_write_full_page() is called for regular writeback
1725 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1726 * locked buffer. This only can happen if someone has written the buffer
1727 * directly, with submit_bh(). At the address_space level PageWriteback
1728 * prevents this contention from occurring.
1729 *
1730 * If block_write_full_page() is called with wbc->sync_mode ==
1731 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1732 * causes the writes to be flagged as synchronous writes.
1733 */
1734int __block_write_full_page(struct inode *inode, struct page *page,
1735 get_block_t *get_block, struct writeback_control *wbc,
1736 bh_end_io_t *handler)
1737{
1738 int err;
1739 sector_t block;
1740 sector_t last_block;
1741 struct buffer_head *bh, *head;
1742 unsigned int blocksize, bbits;
1743 int nr_underway = 0;
1744 int write_flags = wbc_to_write_flags(wbc);
1745
1746 head = create_page_buffers(page, inode,
1747 (1 << BH_Dirty)|(1 << BH_Uptodate));
1748
1749 /*
1750 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1751 * here, and the (potentially unmapped) buffers may become dirty at
1752 * any time. If a buffer becomes dirty here after we've inspected it
1753 * then we just miss that fact, and the page stays dirty.
1754 *
1755 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1756 * handle that here by just cleaning them.
1757 */
1758
1759 bh = head;
1760 blocksize = bh->b_size;
1761 bbits = block_size_bits(blocksize);
1762
1763 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1764 last_block = (i_size_read(inode) - 1) >> bbits;
1765
1766 /*
1767 * Get all the dirty buffers mapped to disk addresses and
1768 * handle any aliases from the underlying blockdev's mapping.
1769 */
1770 do {
1771 if (block > last_block) {
1772 /*
1773 * mapped buffers outside i_size will occur, because
1774 * this page can be outside i_size when there is a
1775 * truncate in progress.
1776 */
1777 /*
1778 * The buffer was zeroed by block_write_full_page()
1779 */
1780 clear_buffer_dirty(bh);
1781 set_buffer_uptodate(bh);
1782 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1783 buffer_dirty(bh)) {
1784 WARN_ON(bh->b_size != blocksize);
1785 err = get_block(inode, block, bh, 1);
1786 if (err)
1787 goto recover;
1788 clear_buffer_delay(bh);
1789 if (buffer_new(bh)) {
1790 /* blockdev mappings never come here */
1791 clear_buffer_new(bh);
1792 clean_bdev_bh_alias(bh);
1793 }
1794 }
1795 bh = bh->b_this_page;
1796 block++;
1797 } while (bh != head);
1798
1799 do {
1800 if (!buffer_mapped(bh))
1801 continue;
1802 /*
1803 * If it's a fully non-blocking write attempt and we cannot
1804 * lock the buffer then redirty the page. Note that this can
1805 * potentially cause a busy-wait loop from writeback threads
1806 * and kswapd activity, but those code paths have their own
1807 * higher-level throttling.
1808 */
1809 if (wbc->sync_mode != WB_SYNC_NONE) {
1810 lock_buffer(bh);
1811 } else if (!trylock_buffer(bh)) {
1812 redirty_page_for_writepage(wbc, page);
1813 continue;
1814 }
1815 if (test_clear_buffer_dirty(bh)) {
1816 mark_buffer_async_write_endio(bh, handler);
1817 } else {
1818 unlock_buffer(bh);
1819 }
1820 } while ((bh = bh->b_this_page) != head);
1821
1822 /*
1823 * The page and its buffers are protected by PageWriteback(), so we can
1824 * drop the bh refcounts early.
1825 */
1826 BUG_ON(PageWriteback(page));
1827 set_page_writeback(page);
1828
1829 do {
1830 struct buffer_head *next = bh->b_this_page;
1831 if (buffer_async_write(bh)) {
1832 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh, 0, wbc);
1833 nr_underway++;
1834 }
1835 bh = next;
1836 } while (bh != head);
1837 unlock_page(page);
1838
1839 err = 0;
1840done:
1841 if (nr_underway == 0) {
1842 /*
1843 * The page was marked dirty, but the buffers were
1844 * clean. Someone wrote them back by hand with
1845 * ll_rw_block/submit_bh. A rare case.
1846 */
1847 end_page_writeback(page);
1848
1849 /*
1850 * The page and buffer_heads can be released at any time from
1851 * here on.
1852 */
1853 }
1854 return err;
1855
1856recover:
1857 /*
1858 * ENOSPC, or some other error. We may already have added some
1859 * blocks to the file, so we need to write these out to avoid
1860 * exposing stale data.
1861 * The page is currently locked and not marked for writeback
1862 */
1863 bh = head;
1864 /* Recovery: lock and submit the mapped buffers */
1865 do {
1866 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1867 !buffer_delay(bh)) {
1868 lock_buffer(bh);
1869 mark_buffer_async_write_endio(bh, handler);
1870 } else {
1871 /*
1872 * The buffer may have been set dirty during
1873 * attachment to a dirty page.
1874 */
1875 clear_buffer_dirty(bh);
1876 }
1877 } while ((bh = bh->b_this_page) != head);
1878 SetPageError(page);
1879 BUG_ON(PageWriteback(page));
1880 mapping_set_error(page->mapping, err);
1881 set_page_writeback(page);
1882 do {
1883 struct buffer_head *next = bh->b_this_page;
1884 if (buffer_async_write(bh)) {
1885 clear_buffer_dirty(bh);
1886 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh, 0, wbc);
1887 nr_underway++;
1888 }
1889 bh = next;
1890 } while (bh != head);
1891 unlock_page(page);
1892 goto done;
1893}
1894EXPORT_SYMBOL(__block_write_full_page);
1895
1896/*
1897 * If a page has any new buffers, zero them out here, and mark them uptodate
1898 * and dirty so they'll be written out (in order to prevent uninitialised
1899 * block data from leaking). And clear the new bit.
1900 */
1901void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1902{
1903 unsigned int block_start, block_end;
1904 struct buffer_head *head, *bh;
1905
1906 BUG_ON(!PageLocked(page));
1907 if (!page_has_buffers(page))
1908 return;
1909
1910 bh = head = page_buffers(page);
1911 block_start = 0;
1912 do {
1913 block_end = block_start + bh->b_size;
1914
1915 if (buffer_new(bh)) {
1916 if (block_end > from && block_start < to) {
1917 if (!PageUptodate(page)) {
1918 unsigned start, size;
1919
1920 start = max(from, block_start);
1921 size = min(to, block_end) - start;
1922
1923 zero_user(page, start, size);
1924 set_buffer_uptodate(bh);
1925 }
1926
1927 clear_buffer_new(bh);
1928 mark_buffer_dirty(bh);
1929 }
1930 }
1931
1932 block_start = block_end;
1933 bh = bh->b_this_page;
1934 } while (bh != head);
1935}
1936EXPORT_SYMBOL(page_zero_new_buffers);
1937
1938static void
1939iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1940 struct iomap *iomap)
1941{
1942 loff_t offset = block << inode->i_blkbits;
1943
1944 bh->b_bdev = iomap->bdev;
1945
1946 /*
1947 * Block points to offset in file we need to map, iomap contains
1948 * the offset at which the map starts. If the map ends before the
1949 * current block, then do not map the buffer and let the caller
1950 * handle it.
1951 */
1952 BUG_ON(offset >= iomap->offset + iomap->length);
1953
1954 switch (iomap->type) {
1955 case IOMAP_HOLE:
1956 /*
1957 * If the buffer is not up to date or beyond the current EOF,
1958 * we need to mark it as new to ensure sub-block zeroing is
1959 * executed if necessary.
1960 */
1961 if (!buffer_uptodate(bh) ||
1962 (offset >= i_size_read(inode)))
1963 set_buffer_new(bh);
1964 break;
1965 case IOMAP_DELALLOC:
1966 if (!buffer_uptodate(bh) ||
1967 (offset >= i_size_read(inode)))
1968 set_buffer_new(bh);
1969 set_buffer_uptodate(bh);
1970 set_buffer_mapped(bh);
1971 set_buffer_delay(bh);
1972 break;
1973 case IOMAP_UNWRITTEN:
1974 /*
1975 * For unwritten regions, we always need to ensure that
1976 * sub-block writes cause the regions in the block we are not
1977 * writing to are zeroed. Set the buffer as new to ensure this.
1978 */
1979 set_buffer_new(bh);
1980 set_buffer_unwritten(bh);
1981 /* FALLTHRU */
1982 case IOMAP_MAPPED:
1983 if (offset >= i_size_read(inode))
1984 set_buffer_new(bh);
1985 bh->b_blocknr = (iomap->blkno >> (inode->i_blkbits - 9)) +
1986 ((offset - iomap->offset) >> inode->i_blkbits);
1987 set_buffer_mapped(bh);
1988 break;
1989 }
1990}
1991
1992int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
1993 get_block_t *get_block, struct iomap *iomap)
1994{
1995 unsigned from = pos & (PAGE_SIZE - 1);
1996 unsigned to = from + len;
1997 struct inode *inode = page->mapping->host;
1998 unsigned block_start, block_end;
1999 sector_t block;
2000 int err = 0;
2001 unsigned blocksize, bbits;
2002 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
2003
2004 BUG_ON(!PageLocked(page));
2005 BUG_ON(from > PAGE_SIZE);
2006 BUG_ON(to > PAGE_SIZE);
2007 BUG_ON(from > to);
2008
2009 head = create_page_buffers(page, inode, 0);
2010 blocksize = head->b_size;
2011 bbits = block_size_bits(blocksize);
2012
2013 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
2014
2015 for(bh = head, block_start = 0; bh != head || !block_start;
2016 block++, block_start=block_end, bh = bh->b_this_page) {
2017 block_end = block_start + blocksize;
2018 if (block_end <= from || block_start >= to) {
2019 if (PageUptodate(page)) {
2020 if (!buffer_uptodate(bh))
2021 set_buffer_uptodate(bh);
2022 }
2023 continue;
2024 }
2025 if (buffer_new(bh))
2026 clear_buffer_new(bh);
2027 if (!buffer_mapped(bh)) {
2028 WARN_ON(bh->b_size != blocksize);
2029 if (get_block) {
2030 err = get_block(inode, block, bh, 1);
2031 if (err)
2032 break;
2033 } else {
2034 iomap_to_bh(inode, block, bh, iomap);
2035 }
2036
2037 if (buffer_new(bh)) {
2038 clean_bdev_bh_alias(bh);
2039 if (PageUptodate(page)) {
2040 clear_buffer_new(bh);
2041 set_buffer_uptodate(bh);
2042 mark_buffer_dirty(bh);
2043 continue;
2044 }
2045 if (block_end > to || block_start < from)
2046 zero_user_segments(page,
2047 to, block_end,
2048 block_start, from);
2049 continue;
2050 }
2051 }
2052 if (PageUptodate(page)) {
2053 if (!buffer_uptodate(bh))
2054 set_buffer_uptodate(bh);
2055 continue;
2056 }
2057 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2058 !buffer_unwritten(bh) &&
2059 (block_start < from || block_end > to)) {
2060 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2061 *wait_bh++=bh;
2062 }
2063 }
2064 /*
2065 * If we issued read requests - let them complete.
2066 */
2067 while(wait_bh > wait) {
2068 wait_on_buffer(*--wait_bh);
2069 if (!buffer_uptodate(*wait_bh))
2070 err = -EIO;
2071 }
2072 if (unlikely(err))
2073 page_zero_new_buffers(page, from, to);
2074 return err;
2075}
2076
2077int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2078 get_block_t *get_block)
2079{
2080 return __block_write_begin_int(page, pos, len, get_block, NULL);
2081}
2082EXPORT_SYMBOL(__block_write_begin);
2083
2084static int __block_commit_write(struct inode *inode, struct page *page,
2085 unsigned from, unsigned to)
2086{
2087 unsigned block_start, block_end;
2088 int partial = 0;
2089 unsigned blocksize;
2090 struct buffer_head *bh, *head;
2091
2092 bh = head = page_buffers(page);
2093 blocksize = bh->b_size;
2094
2095 block_start = 0;
2096 do {
2097 block_end = block_start + blocksize;
2098 if (block_end <= from || block_start >= to) {
2099 if (!buffer_uptodate(bh))
2100 partial = 1;
2101 } else {
2102 set_buffer_uptodate(bh);
2103 mark_buffer_dirty(bh);
2104 }
2105 clear_buffer_new(bh);
2106
2107 block_start = block_end;
2108 bh = bh->b_this_page;
2109 } while (bh != head);
2110
2111 /*
2112 * If this is a partial write which happened to make all buffers
2113 * uptodate then we can optimize away a bogus readpage() for
2114 * the next read(). Here we 'discover' whether the page went
2115 * uptodate as a result of this (potentially partial) write.
2116 */
2117 if (!partial)
2118 SetPageUptodate(page);
2119 return 0;
2120}
2121
2122/*
2123 * block_write_begin takes care of the basic task of block allocation and
2124 * bringing partial write blocks uptodate first.
2125 *
2126 * The filesystem needs to handle block truncation upon failure.
2127 */
2128int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2129 unsigned flags, struct page **pagep, get_block_t *get_block)
2130{
2131 pgoff_t index = pos >> PAGE_SHIFT;
2132 struct page *page;
2133 int status;
2134
2135 page = grab_cache_page_write_begin(mapping, index, flags);
2136 if (!page)
2137 return -ENOMEM;
2138
2139 status = __block_write_begin(page, pos, len, get_block);
2140 if (unlikely(status)) {
2141 unlock_page(page);
2142 put_page(page);
2143 page = NULL;
2144 }
2145
2146 *pagep = page;
2147 return status;
2148}
2149EXPORT_SYMBOL(block_write_begin);
2150
2151int block_write_end(struct file *file, struct address_space *mapping,
2152 loff_t pos, unsigned len, unsigned copied,
2153 struct page *page, void *fsdata)
2154{
2155 struct inode *inode = mapping->host;
2156 unsigned start;
2157
2158 start = pos & (PAGE_SIZE - 1);
2159
2160 if (unlikely(copied < len)) {
2161 /*
2162 * The buffers that were written will now be uptodate, so we
2163 * don't have to worry about a readpage reading them and
2164 * overwriting a partial write. However if we have encountered
2165 * a short write and only partially written into a buffer, it
2166 * will not be marked uptodate, so a readpage might come in and
2167 * destroy our partial write.
2168 *
2169 * Do the simplest thing, and just treat any short write to a
2170 * non uptodate page as a zero-length write, and force the
2171 * caller to redo the whole thing.
2172 */
2173 if (!PageUptodate(page))
2174 copied = 0;
2175
2176 page_zero_new_buffers(page, start+copied, start+len);
2177 }
2178 flush_dcache_page(page);
2179
2180 /* This could be a short (even 0-length) commit */
2181 __block_commit_write(inode, page, start, start+copied);
2182
2183 return copied;
2184}
2185EXPORT_SYMBOL(block_write_end);
2186
2187int generic_write_end(struct file *file, struct address_space *mapping,
2188 loff_t pos, unsigned len, unsigned copied,
2189 struct page *page, void *fsdata)
2190{
2191 struct inode *inode = mapping->host;
2192 loff_t old_size = inode->i_size;
2193 int i_size_changed = 0;
2194
2195 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2196
2197 /*
2198 * No need to use i_size_read() here, the i_size
2199 * cannot change under us because we hold i_mutex.
2200 *
2201 * But it's important to update i_size while still holding page lock:
2202 * page writeout could otherwise come in and zero beyond i_size.
2203 */
2204 if (pos+copied > inode->i_size) {
2205 i_size_write(inode, pos+copied);
2206 i_size_changed = 1;
2207 }
2208
2209 unlock_page(page);
2210 put_page(page);
2211
2212 if (old_size < pos)
2213 pagecache_isize_extended(inode, old_size, pos);
2214 /*
2215 * Don't mark the inode dirty under page lock. First, it unnecessarily
2216 * makes the holding time of page lock longer. Second, it forces lock
2217 * ordering of page lock and transaction start for journaling
2218 * filesystems.
2219 */
2220 if (i_size_changed)
2221 mark_inode_dirty(inode);
2222
2223 return copied;
2224}
2225EXPORT_SYMBOL(generic_write_end);
2226
2227/*
2228 * block_is_partially_uptodate checks whether buffers within a page are
2229 * uptodate or not.
2230 *
2231 * Returns true if all buffers which correspond to a file portion
2232 * we want to read are uptodate.
2233 */
2234int block_is_partially_uptodate(struct page *page, unsigned long from,
2235 unsigned long count)
2236{
2237 unsigned block_start, block_end, blocksize;
2238 unsigned to;
2239 struct buffer_head *bh, *head;
2240 int ret = 1;
2241
2242 if (!page_has_buffers(page))
2243 return 0;
2244
2245 head = page_buffers(page);
2246 blocksize = head->b_size;
2247 to = min_t(unsigned, PAGE_SIZE - from, count);
2248 to = from + to;
2249 if (from < blocksize && to > PAGE_SIZE - blocksize)
2250 return 0;
2251
2252 bh = head;
2253 block_start = 0;
2254 do {
2255 block_end = block_start + blocksize;
2256 if (block_end > from && block_start < to) {
2257 if (!buffer_uptodate(bh)) {
2258 ret = 0;
2259 break;
2260 }
2261 if (block_end >= to)
2262 break;
2263 }
2264 block_start = block_end;
2265 bh = bh->b_this_page;
2266 } while (bh != head);
2267
2268 return ret;
2269}
2270EXPORT_SYMBOL(block_is_partially_uptodate);
2271
2272/*
2273 * Generic "read page" function for block devices that have the normal
2274 * get_block functionality. This is most of the block device filesystems.
2275 * Reads the page asynchronously --- the unlock_buffer() and
2276 * set/clear_buffer_uptodate() functions propagate buffer state into the
2277 * page struct once IO has completed.
2278 */
2279int block_read_full_page(struct page *page, get_block_t *get_block)
2280{
2281 struct inode *inode = page->mapping->host;
2282 sector_t iblock, lblock;
2283 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2284 unsigned int blocksize, bbits;
2285 int nr, i;
2286 int fully_mapped = 1;
2287
2288 head = create_page_buffers(page, inode, 0);
2289 blocksize = head->b_size;
2290 bbits = block_size_bits(blocksize);
2291
2292 iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2293 lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2294 bh = head;
2295 nr = 0;
2296 i = 0;
2297
2298 do {
2299 if (buffer_uptodate(bh))
2300 continue;
2301
2302 if (!buffer_mapped(bh)) {
2303 int err = 0;
2304
2305 fully_mapped = 0;
2306 if (iblock < lblock) {
2307 WARN_ON(bh->b_size != blocksize);
2308 err = get_block(inode, iblock, bh, 0);
2309 if (err)
2310 SetPageError(page);
2311 }
2312 if (!buffer_mapped(bh)) {
2313 zero_user(page, i * blocksize, blocksize);
2314 if (!err)
2315 set_buffer_uptodate(bh);
2316 continue;
2317 }
2318 /*
2319 * get_block() might have updated the buffer
2320 * synchronously
2321 */
2322 if (buffer_uptodate(bh))
2323 continue;
2324 }
2325 arr[nr++] = bh;
2326 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2327
2328 if (fully_mapped)
2329 SetPageMappedToDisk(page);
2330
2331 if (!nr) {
2332 /*
2333 * All buffers are uptodate - we can set the page uptodate
2334 * as well. But not if get_block() returned an error.
2335 */
2336 if (!PageError(page))
2337 SetPageUptodate(page);
2338 unlock_page(page);
2339 return 0;
2340 }
2341
2342 /* Stage two: lock the buffers */
2343 for (i = 0; i < nr; i++) {
2344 bh = arr[i];
2345 lock_buffer(bh);
2346 mark_buffer_async_read(bh);
2347 }
2348
2349 /*
2350 * Stage 3: start the IO. Check for uptodateness
2351 * inside the buffer lock in case another process reading
2352 * the underlying blockdev brought it uptodate (the sct fix).
2353 */
2354 for (i = 0; i < nr; i++) {
2355 bh = arr[i];
2356 if (buffer_uptodate(bh))
2357 end_buffer_async_read(bh, 1);
2358 else
2359 submit_bh(REQ_OP_READ, 0, bh);
2360 }
2361 return 0;
2362}
2363EXPORT_SYMBOL(block_read_full_page);
2364
2365/* utility function for filesystems that need to do work on expanding
2366 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2367 * deal with the hole.
2368 */
2369int generic_cont_expand_simple(struct inode *inode, loff_t size)
2370{
2371 struct address_space *mapping = inode->i_mapping;
2372 struct page *page;
2373 void *fsdata;
2374 int err;
2375
2376 err = inode_newsize_ok(inode, size);
2377 if (err)
2378 goto out;
2379
2380 err = pagecache_write_begin(NULL, mapping, size, 0,
2381 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2382 &page, &fsdata);
2383 if (err)
2384 goto out;
2385
2386 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2387 BUG_ON(err > 0);
2388
2389out:
2390 return err;
2391}
2392EXPORT_SYMBOL(generic_cont_expand_simple);
2393
2394static int cont_expand_zero(struct file *file, struct address_space *mapping,
2395 loff_t pos, loff_t *bytes)
2396{
2397 struct inode *inode = mapping->host;
2398 unsigned blocksize = 1 << inode->i_blkbits;
2399 struct page *page;
2400 void *fsdata;
2401 pgoff_t index, curidx;
2402 loff_t curpos;
2403 unsigned zerofrom, offset, len;
2404 int err = 0;
2405
2406 index = pos >> PAGE_SHIFT;
2407 offset = pos & ~PAGE_MASK;
2408
2409 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2410 zerofrom = curpos & ~PAGE_MASK;
2411 if (zerofrom & (blocksize-1)) {
2412 *bytes |= (blocksize-1);
2413 (*bytes)++;
2414 }
2415 len = PAGE_SIZE - zerofrom;
2416
2417 err = pagecache_write_begin(file, mapping, curpos, len,
2418 AOP_FLAG_UNINTERRUPTIBLE,
2419 &page, &fsdata);
2420 if (err)
2421 goto out;
2422 zero_user(page, zerofrom, len);
2423 err = pagecache_write_end(file, mapping, curpos, len, len,
2424 page, fsdata);
2425 if (err < 0)
2426 goto out;
2427 BUG_ON(err != len);
2428 err = 0;
2429
2430 balance_dirty_pages_ratelimited(mapping);
2431
2432 if (unlikely(fatal_signal_pending(current))) {
2433 err = -EINTR;
2434 goto out;
2435 }
2436 }
2437
2438 /* page covers the boundary, find the boundary offset */
2439 if (index == curidx) {
2440 zerofrom = curpos & ~PAGE_MASK;
2441 /* if we will expand the thing last block will be filled */
2442 if (offset <= zerofrom) {
2443 goto out;
2444 }
2445 if (zerofrom & (blocksize-1)) {
2446 *bytes |= (blocksize-1);
2447 (*bytes)++;
2448 }
2449 len = offset - zerofrom;
2450
2451 err = pagecache_write_begin(file, mapping, curpos, len,
2452 AOP_FLAG_UNINTERRUPTIBLE,
2453 &page, &fsdata);
2454 if (err)
2455 goto out;
2456 zero_user(page, zerofrom, len);
2457 err = pagecache_write_end(file, mapping, curpos, len, len,
2458 page, fsdata);
2459 if (err < 0)
2460 goto out;
2461 BUG_ON(err != len);
2462 err = 0;
2463 }
2464out:
2465 return err;
2466}
2467
2468/*
2469 * For moronic filesystems that do not allow holes in file.
2470 * We may have to extend the file.
2471 */
2472int cont_write_begin(struct file *file, struct address_space *mapping,
2473 loff_t pos, unsigned len, unsigned flags,
2474 struct page **pagep, void **fsdata,
2475 get_block_t *get_block, loff_t *bytes)
2476{
2477 struct inode *inode = mapping->host;
2478 unsigned blocksize = 1 << inode->i_blkbits;
2479 unsigned zerofrom;
2480 int err;
2481
2482 err = cont_expand_zero(file, mapping, pos, bytes);
2483 if (err)
2484 return err;
2485
2486 zerofrom = *bytes & ~PAGE_MASK;
2487 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2488 *bytes |= (blocksize-1);
2489 (*bytes)++;
2490 }
2491
2492 return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2493}
2494EXPORT_SYMBOL(cont_write_begin);
2495
2496int block_commit_write(struct page *page, unsigned from, unsigned to)
2497{
2498 struct inode *inode = page->mapping->host;
2499 __block_commit_write(inode,page,from,to);
2500 return 0;
2501}
2502EXPORT_SYMBOL(block_commit_write);
2503
2504/*
2505 * block_page_mkwrite() is not allowed to change the file size as it gets
2506 * called from a page fault handler when a page is first dirtied. Hence we must
2507 * be careful to check for EOF conditions here. We set the page up correctly
2508 * for a written page which means we get ENOSPC checking when writing into
2509 * holes and correct delalloc and unwritten extent mapping on filesystems that
2510 * support these features.
2511 *
2512 * We are not allowed to take the i_mutex here so we have to play games to
2513 * protect against truncate races as the page could now be beyond EOF. Because
2514 * truncate writes the inode size before removing pages, once we have the
2515 * page lock we can determine safely if the page is beyond EOF. If it is not
2516 * beyond EOF, then the page is guaranteed safe against truncation until we
2517 * unlock the page.
2518 *
2519 * Direct callers of this function should protect against filesystem freezing
2520 * using sb_start_pagefault() - sb_end_pagefault() functions.
2521 */
2522int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2523 get_block_t get_block)
2524{
2525 struct page *page = vmf->page;
2526 struct inode *inode = file_inode(vma->vm_file);
2527 unsigned long end;
2528 loff_t size;
2529 int ret;
2530
2531 lock_page(page);
2532 size = i_size_read(inode);
2533 if ((page->mapping != inode->i_mapping) ||
2534 (page_offset(page) > size)) {
2535 /* We overload EFAULT to mean page got truncated */
2536 ret = -EFAULT;
2537 goto out_unlock;
2538 }
2539
2540 /* page is wholly or partially inside EOF */
2541 if (((page->index + 1) << PAGE_SHIFT) > size)
2542 end = size & ~PAGE_MASK;
2543 else
2544 end = PAGE_SIZE;
2545
2546 ret = __block_write_begin(page, 0, end, get_block);
2547 if (!ret)
2548 ret = block_commit_write(page, 0, end);
2549
2550 if (unlikely(ret < 0))
2551 goto out_unlock;
2552 set_page_dirty(page);
2553 wait_for_stable_page(page);
2554 return 0;
2555out_unlock:
2556 unlock_page(page);
2557 return ret;
2558}
2559EXPORT_SYMBOL(block_page_mkwrite);
2560
2561/*
2562 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2563 * immediately, while under the page lock. So it needs a special end_io
2564 * handler which does not touch the bh after unlocking it.
2565 */
2566static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2567{
2568 __end_buffer_read_notouch(bh, uptodate);
2569}
2570
2571/*
2572 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2573 * the page (converting it to circular linked list and taking care of page
2574 * dirty races).
2575 */
2576static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2577{
2578 struct buffer_head *bh;
2579
2580 BUG_ON(!PageLocked(page));
2581
2582 spin_lock(&page->mapping->private_lock);
2583 bh = head;
2584 do {
2585 if (PageDirty(page))
2586 set_buffer_dirty(bh);
2587 if (!bh->b_this_page)
2588 bh->b_this_page = head;
2589 bh = bh->b_this_page;
2590 } while (bh != head);
2591 attach_page_buffers(page, head);
2592 spin_unlock(&page->mapping->private_lock);
2593}
2594
2595/*
2596 * On entry, the page is fully not uptodate.
2597 * On exit the page is fully uptodate in the areas outside (from,to)
2598 * The filesystem needs to handle block truncation upon failure.
2599 */
2600int nobh_write_begin(struct address_space *mapping,
2601 loff_t pos, unsigned len, unsigned flags,
2602 struct page **pagep, void **fsdata,
2603 get_block_t *get_block)
2604{
2605 struct inode *inode = mapping->host;
2606 const unsigned blkbits = inode->i_blkbits;
2607 const unsigned blocksize = 1 << blkbits;
2608 struct buffer_head *head, *bh;
2609 struct page *page;
2610 pgoff_t index;
2611 unsigned from, to;
2612 unsigned block_in_page;
2613 unsigned block_start, block_end;
2614 sector_t block_in_file;
2615 int nr_reads = 0;
2616 int ret = 0;
2617 int is_mapped_to_disk = 1;
2618
2619 index = pos >> PAGE_SHIFT;
2620 from = pos & (PAGE_SIZE - 1);
2621 to = from + len;
2622
2623 page = grab_cache_page_write_begin(mapping, index, flags);
2624 if (!page)
2625 return -ENOMEM;
2626 *pagep = page;
2627 *fsdata = NULL;
2628
2629 if (page_has_buffers(page)) {
2630 ret = __block_write_begin(page, pos, len, get_block);
2631 if (unlikely(ret))
2632 goto out_release;
2633 return ret;
2634 }
2635
2636 if (PageMappedToDisk(page))
2637 return 0;
2638
2639 /*
2640 * Allocate buffers so that we can keep track of state, and potentially
2641 * attach them to the page if an error occurs. In the common case of
2642 * no error, they will just be freed again without ever being attached
2643 * to the page (which is all OK, because we're under the page lock).
2644 *
2645 * Be careful: the buffer linked list is a NULL terminated one, rather
2646 * than the circular one we're used to.
2647 */
2648 head = alloc_page_buffers(page, blocksize, 0);
2649 if (!head) {
2650 ret = -ENOMEM;
2651 goto out_release;
2652 }
2653
2654 block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2655
2656 /*
2657 * We loop across all blocks in the page, whether or not they are
2658 * part of the affected region. This is so we can discover if the
2659 * page is fully mapped-to-disk.
2660 */
2661 for (block_start = 0, block_in_page = 0, bh = head;
2662 block_start < PAGE_SIZE;
2663 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2664 int create;
2665
2666 block_end = block_start + blocksize;
2667 bh->b_state = 0;
2668 create = 1;
2669 if (block_start >= to)
2670 create = 0;
2671 ret = get_block(inode, block_in_file + block_in_page,
2672 bh, create);
2673 if (ret)
2674 goto failed;
2675 if (!buffer_mapped(bh))
2676 is_mapped_to_disk = 0;
2677 if (buffer_new(bh))
2678 clean_bdev_bh_alias(bh);
2679 if (PageUptodate(page)) {
2680 set_buffer_uptodate(bh);
2681 continue;
2682 }
2683 if (buffer_new(bh) || !buffer_mapped(bh)) {
2684 zero_user_segments(page, block_start, from,
2685 to, block_end);
2686 continue;
2687 }
2688 if (buffer_uptodate(bh))
2689 continue; /* reiserfs does this */
2690 if (block_start < from || block_end > to) {
2691 lock_buffer(bh);
2692 bh->b_end_io = end_buffer_read_nobh;
2693 submit_bh(REQ_OP_READ, 0, bh);
2694 nr_reads++;
2695 }
2696 }
2697
2698 if (nr_reads) {
2699 /*
2700 * The page is locked, so these buffers are protected from
2701 * any VM or truncate activity. Hence we don't need to care
2702 * for the buffer_head refcounts.
2703 */
2704 for (bh = head; bh; bh = bh->b_this_page) {
2705 wait_on_buffer(bh);
2706 if (!buffer_uptodate(bh))
2707 ret = -EIO;
2708 }
2709 if (ret)
2710 goto failed;
2711 }
2712
2713 if (is_mapped_to_disk)
2714 SetPageMappedToDisk(page);
2715
2716 *fsdata = head; /* to be released by nobh_write_end */
2717
2718 return 0;
2719
2720failed:
2721 BUG_ON(!ret);
2722 /*
2723 * Error recovery is a bit difficult. We need to zero out blocks that
2724 * were newly allocated, and dirty them to ensure they get written out.
2725 * Buffers need to be attached to the page at this point, otherwise
2726 * the handling of potential IO errors during writeout would be hard
2727 * (could try doing synchronous writeout, but what if that fails too?)
2728 */
2729 attach_nobh_buffers(page, head);
2730 page_zero_new_buffers(page, from, to);
2731
2732out_release:
2733 unlock_page(page);
2734 put_page(page);
2735 *pagep = NULL;
2736
2737 return ret;
2738}
2739EXPORT_SYMBOL(nobh_write_begin);
2740
2741int nobh_write_end(struct file *file, struct address_space *mapping,
2742 loff_t pos, unsigned len, unsigned copied,
2743 struct page *page, void *fsdata)
2744{
2745 struct inode *inode = page->mapping->host;
2746 struct buffer_head *head = fsdata;
2747 struct buffer_head *bh;
2748 BUG_ON(fsdata != NULL && page_has_buffers(page));
2749
2750 if (unlikely(copied < len) && head)
2751 attach_nobh_buffers(page, head);
2752 if (page_has_buffers(page))
2753 return generic_write_end(file, mapping, pos, len,
2754 copied, page, fsdata);
2755
2756 SetPageUptodate(page);
2757 set_page_dirty(page);
2758 if (pos+copied > inode->i_size) {
2759 i_size_write(inode, pos+copied);
2760 mark_inode_dirty(inode);
2761 }
2762
2763 unlock_page(page);
2764 put_page(page);
2765
2766 while (head) {
2767 bh = head;
2768 head = head->b_this_page;
2769 free_buffer_head(bh);
2770 }
2771
2772 return copied;
2773}
2774EXPORT_SYMBOL(nobh_write_end);
2775
2776/*
2777 * nobh_writepage() - based on block_full_write_page() except
2778 * that it tries to operate without attaching bufferheads to
2779 * the page.
2780 */
2781int nobh_writepage(struct page *page, get_block_t *get_block,
2782 struct writeback_control *wbc)
2783{
2784 struct inode * const inode = page->mapping->host;
2785 loff_t i_size = i_size_read(inode);
2786 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2787 unsigned offset;
2788 int ret;
2789
2790 /* Is the page fully inside i_size? */
2791 if (page->index < end_index)
2792 goto out;
2793
2794 /* Is the page fully outside i_size? (truncate in progress) */
2795 offset = i_size & (PAGE_SIZE-1);
2796 if (page->index >= end_index+1 || !offset) {
2797 /*
2798 * The page may have dirty, unmapped buffers. For example,
2799 * they may have been added in ext3_writepage(). Make them
2800 * freeable here, so the page does not leak.
2801 */
2802#if 0
2803 /* Not really sure about this - do we need this ? */
2804 if (page->mapping->a_ops->invalidatepage)
2805 page->mapping->a_ops->invalidatepage(page, offset);
2806#endif
2807 unlock_page(page);
2808 return 0; /* don't care */
2809 }
2810
2811 /*
2812 * The page straddles i_size. It must be zeroed out on each and every
2813 * writepage invocation because it may be mmapped. "A file is mapped
2814 * in multiples of the page size. For a file that is not a multiple of
2815 * the page size, the remaining memory is zeroed when mapped, and
2816 * writes to that region are not written out to the file."
2817 */
2818 zero_user_segment(page, offset, PAGE_SIZE);
2819out:
2820 ret = mpage_writepage(page, get_block, wbc);
2821 if (ret == -EAGAIN)
2822 ret = __block_write_full_page(inode, page, get_block, wbc,
2823 end_buffer_async_write);
2824 return ret;
2825}
2826EXPORT_SYMBOL(nobh_writepage);
2827
2828int nobh_truncate_page(struct address_space *mapping,
2829 loff_t from, get_block_t *get_block)
2830{
2831 pgoff_t index = from >> PAGE_SHIFT;
2832 unsigned offset = from & (PAGE_SIZE-1);
2833 unsigned blocksize;
2834 sector_t iblock;
2835 unsigned length, pos;
2836 struct inode *inode = mapping->host;
2837 struct page *page;
2838 struct buffer_head map_bh;
2839 int err;
2840
2841 blocksize = 1 << inode->i_blkbits;
2842 length = offset & (blocksize - 1);
2843
2844 /* Block boundary? Nothing to do */
2845 if (!length)
2846 return 0;
2847
2848 length = blocksize - length;
2849 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2850
2851 page = grab_cache_page(mapping, index);
2852 err = -ENOMEM;
2853 if (!page)
2854 goto out;
2855
2856 if (page_has_buffers(page)) {
2857has_buffers:
2858 unlock_page(page);
2859 put_page(page);
2860 return block_truncate_page(mapping, from, get_block);
2861 }
2862
2863 /* Find the buffer that contains "offset" */
2864 pos = blocksize;
2865 while (offset >= pos) {
2866 iblock++;
2867 pos += blocksize;
2868 }
2869
2870 map_bh.b_size = blocksize;
2871 map_bh.b_state = 0;
2872 err = get_block(inode, iblock, &map_bh, 0);
2873 if (err)
2874 goto unlock;
2875 /* unmapped? It's a hole - nothing to do */
2876 if (!buffer_mapped(&map_bh))
2877 goto unlock;
2878
2879 /* Ok, it's mapped. Make sure it's up-to-date */
2880 if (!PageUptodate(page)) {
2881 err = mapping->a_ops->readpage(NULL, page);
2882 if (err) {
2883 put_page(page);
2884 goto out;
2885 }
2886 lock_page(page);
2887 if (!PageUptodate(page)) {
2888 err = -EIO;
2889 goto unlock;
2890 }
2891 if (page_has_buffers(page))
2892 goto has_buffers;
2893 }
2894 zero_user(page, offset, length);
2895 set_page_dirty(page);
2896 err = 0;
2897
2898unlock:
2899 unlock_page(page);
2900 put_page(page);
2901out:
2902 return err;
2903}
2904EXPORT_SYMBOL(nobh_truncate_page);
2905
2906int block_truncate_page(struct address_space *mapping,
2907 loff_t from, get_block_t *get_block)
2908{
2909 pgoff_t index = from >> PAGE_SHIFT;
2910 unsigned offset = from & (PAGE_SIZE-1);
2911 unsigned blocksize;
2912 sector_t iblock;
2913 unsigned length, pos;
2914 struct inode *inode = mapping->host;
2915 struct page *page;
2916 struct buffer_head *bh;
2917 int err;
2918
2919 blocksize = 1 << inode->i_blkbits;
2920 length = offset & (blocksize - 1);
2921
2922 /* Block boundary? Nothing to do */
2923 if (!length)
2924 return 0;
2925
2926 length = blocksize - length;
2927 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2928
2929 page = grab_cache_page(mapping, index);
2930 err = -ENOMEM;
2931 if (!page)
2932 goto out;
2933
2934 if (!page_has_buffers(page))
2935 create_empty_buffers(page, blocksize, 0);
2936
2937 /* Find the buffer that contains "offset" */
2938 bh = page_buffers(page);
2939 pos = blocksize;
2940 while (offset >= pos) {
2941 bh = bh->b_this_page;
2942 iblock++;
2943 pos += blocksize;
2944 }
2945
2946 err = 0;
2947 if (!buffer_mapped(bh)) {
2948 WARN_ON(bh->b_size != blocksize);
2949 err = get_block(inode, iblock, bh, 0);
2950 if (err)
2951 goto unlock;
2952 /* unmapped? It's a hole - nothing to do */
2953 if (!buffer_mapped(bh))
2954 goto unlock;
2955 }
2956
2957 /* Ok, it's mapped. Make sure it's up-to-date */
2958 if (PageUptodate(page))
2959 set_buffer_uptodate(bh);
2960
2961 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2962 err = -EIO;
2963 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2964 wait_on_buffer(bh);
2965 /* Uhhuh. Read error. Complain and punt. */
2966 if (!buffer_uptodate(bh))
2967 goto unlock;
2968 }
2969
2970 zero_user(page, offset, length);
2971 mark_buffer_dirty(bh);
2972 err = 0;
2973
2974unlock:
2975 unlock_page(page);
2976 put_page(page);
2977out:
2978 return err;
2979}
2980EXPORT_SYMBOL(block_truncate_page);
2981
2982/*
2983 * The generic ->writepage function for buffer-backed address_spaces
2984 */
2985int block_write_full_page(struct page *page, get_block_t *get_block,
2986 struct writeback_control *wbc)
2987{
2988 struct inode * const inode = page->mapping->host;
2989 loff_t i_size = i_size_read(inode);
2990 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2991 unsigned offset;
2992
2993 /* Is the page fully inside i_size? */
2994 if (page->index < end_index)
2995 return __block_write_full_page(inode, page, get_block, wbc,
2996 end_buffer_async_write);
2997
2998 /* Is the page fully outside i_size? (truncate in progress) */
2999 offset = i_size & (PAGE_SIZE-1);
3000 if (page->index >= end_index+1 || !offset) {
3001 /*
3002 * The page may have dirty, unmapped buffers. For example,
3003 * they may have been added in ext3_writepage(). Make them
3004 * freeable here, so the page does not leak.
3005 */
3006 do_invalidatepage(page, 0, PAGE_SIZE);
3007 unlock_page(page);
3008 return 0; /* don't care */
3009 }
3010
3011 /*
3012 * The page straddles i_size. It must be zeroed out on each and every
3013 * writepage invocation because it may be mmapped. "A file is mapped
3014 * in multiples of the page size. For a file that is not a multiple of
3015 * the page size, the remaining memory is zeroed when mapped, and
3016 * writes to that region are not written out to the file."
3017 */
3018 zero_user_segment(page, offset, PAGE_SIZE);
3019 return __block_write_full_page(inode, page, get_block, wbc,
3020 end_buffer_async_write);
3021}
3022EXPORT_SYMBOL(block_write_full_page);
3023
3024sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
3025 get_block_t *get_block)
3026{
3027 struct buffer_head tmp;
3028 struct inode *inode = mapping->host;
3029 tmp.b_state = 0;
3030 tmp.b_blocknr = 0;
3031 tmp.b_size = 1 << inode->i_blkbits;
3032 get_block(inode, block, &tmp, 0);
3033 return tmp.b_blocknr;
3034}
3035EXPORT_SYMBOL(generic_block_bmap);
3036
3037static void end_bio_bh_io_sync(struct bio *bio)
3038{
3039 struct buffer_head *bh = bio->bi_private;
3040
3041 if (unlikely(bio_flagged(bio, BIO_QUIET)))
3042 set_bit(BH_Quiet, &bh->b_state);
3043
3044 bh->b_end_io(bh, !bio->bi_error);
3045 bio_put(bio);
3046}
3047
3048/*
3049 * This allows us to do IO even on the odd last sectors
3050 * of a device, even if the block size is some multiple
3051 * of the physical sector size.
3052 *
3053 * We'll just truncate the bio to the size of the device,
3054 * and clear the end of the buffer head manually.
3055 *
3056 * Truly out-of-range accesses will turn into actual IO
3057 * errors, this only handles the "we need to be able to
3058 * do IO at the final sector" case.
3059 */
3060void guard_bio_eod(int op, struct bio *bio)
3061{
3062 sector_t maxsector;
3063 struct bio_vec *bvec = &bio->bi_io_vec[bio->bi_vcnt - 1];
3064 unsigned truncated_bytes;
3065
3066 maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
3067 if (!maxsector)
3068 return;
3069
3070 /*
3071 * If the *whole* IO is past the end of the device,
3072 * let it through, and the IO layer will turn it into
3073 * an EIO.
3074 */
3075 if (unlikely(bio->bi_iter.bi_sector >= maxsector))
3076 return;
3077
3078 maxsector -= bio->bi_iter.bi_sector;
3079 if (likely((bio->bi_iter.bi_size >> 9) <= maxsector))
3080 return;
3081
3082 /* Uhhuh. We've got a bio that straddles the device size! */
3083 truncated_bytes = bio->bi_iter.bi_size - (maxsector << 9);
3084
3085 /* Truncate the bio.. */
3086 bio->bi_iter.bi_size -= truncated_bytes;
3087 bvec->bv_len -= truncated_bytes;
3088
3089 /* ..and clear the end of the buffer for reads */
3090 if (op == REQ_OP_READ) {
3091 zero_user(bvec->bv_page, bvec->bv_offset + bvec->bv_len,
3092 truncated_bytes);
3093 }
3094}
3095
3096static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3097 unsigned long bio_flags, struct writeback_control *wbc)
3098{
3099 struct bio *bio;
3100
3101 BUG_ON(!buffer_locked(bh));
3102 BUG_ON(!buffer_mapped(bh));
3103 BUG_ON(!bh->b_end_io);
3104 BUG_ON(buffer_delay(bh));
3105 BUG_ON(buffer_unwritten(bh));
3106
3107 /*
3108 * Only clear out a write error when rewriting
3109 */
3110 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3111 clear_buffer_write_io_error(bh);
3112
3113 /*
3114 * from here on down, it's all bio -- do the initial mapping,
3115 * submit_bio -> generic_make_request may further map this bio around
3116 */
3117 bio = bio_alloc(GFP_NOIO, 1);
3118
3119 if (wbc) {
3120 wbc_init_bio(wbc, bio);
3121 wbc_account_io(wbc, bh->b_page, bh->b_size);
3122 }
3123
3124 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3125 bio->bi_bdev = bh->b_bdev;
3126
3127 bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3128 BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3129
3130 bio->bi_end_io = end_bio_bh_io_sync;
3131 bio->bi_private = bh;
3132 bio->bi_flags |= bio_flags;
3133
3134 /* Take care of bh's that straddle the end of the device */
3135 guard_bio_eod(op, bio);
3136
3137 if (buffer_meta(bh))
3138 op_flags |= REQ_META;
3139 if (buffer_prio(bh))
3140 op_flags |= REQ_PRIO;
3141 bio_set_op_attrs(bio, op, op_flags);
3142
3143 submit_bio(bio);
3144 return 0;
3145}
3146
3147int _submit_bh(int op, int op_flags, struct buffer_head *bh,
3148 unsigned long bio_flags)
3149{
3150 return submit_bh_wbc(op, op_flags, bh, bio_flags, NULL);
3151}
3152EXPORT_SYMBOL_GPL(_submit_bh);
3153
3154int submit_bh(int op, int op_flags, struct buffer_head *bh)
3155{
3156 return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3157}
3158EXPORT_SYMBOL(submit_bh);
3159
3160/**
3161 * ll_rw_block: low-level access to block devices (DEPRECATED)
3162 * @op: whether to %READ or %WRITE
3163 * @op_flags: req_flag_bits
3164 * @nr: number of &struct buffer_heads in the array
3165 * @bhs: array of pointers to &struct buffer_head
3166 *
3167 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3168 * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3169 * @op_flags contains flags modifying the detailed I/O behavior, most notably
3170 * %REQ_RAHEAD.
3171 *
3172 * This function drops any buffer that it cannot get a lock on (with the
3173 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3174 * request, and any buffer that appears to be up-to-date when doing read
3175 * request. Further it marks as clean buffers that are processed for
3176 * writing (the buffer cache won't assume that they are actually clean
3177 * until the buffer gets unlocked).
3178 *
3179 * ll_rw_block sets b_end_io to simple completion handler that marks
3180 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3181 * any waiters.
3182 *
3183 * All of the buffers must be for the same device, and must also be a
3184 * multiple of the current approved size for the device.
3185 */
3186void ll_rw_block(int op, int op_flags, int nr, struct buffer_head *bhs[])
3187{
3188 int i;
3189
3190 for (i = 0; i < nr; i++) {
3191 struct buffer_head *bh = bhs[i];
3192
3193 if (!trylock_buffer(bh))
3194 continue;
3195 if (op == WRITE) {
3196 if (test_clear_buffer_dirty(bh)) {
3197 bh->b_end_io = end_buffer_write_sync;
3198 get_bh(bh);
3199 submit_bh(op, op_flags, bh);
3200 continue;
3201 }
3202 } else {
3203 if (!buffer_uptodate(bh)) {
3204 bh->b_end_io = end_buffer_read_sync;
3205 get_bh(bh);
3206 submit_bh(op, op_flags, bh);
3207 continue;
3208 }
3209 }
3210 unlock_buffer(bh);
3211 }
3212}
3213EXPORT_SYMBOL(ll_rw_block);
3214
3215void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3216{
3217 lock_buffer(bh);
3218 if (!test_clear_buffer_dirty(bh)) {
3219 unlock_buffer(bh);
3220 return;
3221 }
3222 bh->b_end_io = end_buffer_write_sync;
3223 get_bh(bh);
3224 submit_bh(REQ_OP_WRITE, op_flags, bh);
3225}
3226EXPORT_SYMBOL(write_dirty_buffer);
3227
3228/*
3229 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3230 * and then start new I/O and then wait upon it. The caller must have a ref on
3231 * the buffer_head.
3232 */
3233int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3234{
3235 int ret = 0;
3236
3237 WARN_ON(atomic_read(&bh->b_count) < 1);
3238 lock_buffer(bh);
3239 if (test_clear_buffer_dirty(bh)) {
3240 get_bh(bh);
3241 bh->b_end_io = end_buffer_write_sync;
3242 ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3243 wait_on_buffer(bh);
3244 if (!ret && !buffer_uptodate(bh))
3245 ret = -EIO;
3246 } else {
3247 unlock_buffer(bh);
3248 }
3249 return ret;
3250}
3251EXPORT_SYMBOL(__sync_dirty_buffer);
3252
3253int sync_dirty_buffer(struct buffer_head *bh)
3254{
3255 return __sync_dirty_buffer(bh, REQ_SYNC);
3256}
3257EXPORT_SYMBOL(sync_dirty_buffer);
3258
3259/*
3260 * try_to_free_buffers() checks if all the buffers on this particular page
3261 * are unused, and releases them if so.
3262 *
3263 * Exclusion against try_to_free_buffers may be obtained by either
3264 * locking the page or by holding its mapping's private_lock.
3265 *
3266 * If the page is dirty but all the buffers are clean then we need to
3267 * be sure to mark the page clean as well. This is because the page
3268 * may be against a block device, and a later reattachment of buffers
3269 * to a dirty page will set *all* buffers dirty. Which would corrupt
3270 * filesystem data on the same device.
3271 *
3272 * The same applies to regular filesystem pages: if all the buffers are
3273 * clean then we set the page clean and proceed. To do that, we require
3274 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3275 * private_lock.
3276 *
3277 * try_to_free_buffers() is non-blocking.
3278 */
3279static inline int buffer_busy(struct buffer_head *bh)
3280{
3281 return atomic_read(&bh->b_count) |
3282 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3283}
3284
3285static int
3286drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3287{
3288 struct buffer_head *head = page_buffers(page);
3289 struct buffer_head *bh;
3290
3291 bh = head;
3292 do {
3293 if (buffer_write_io_error(bh) && page->mapping)
3294 mapping_set_error(page->mapping, -EIO);
3295 if (buffer_busy(bh))
3296 goto failed;
3297 bh = bh->b_this_page;
3298 } while (bh != head);
3299
3300 do {
3301 struct buffer_head *next = bh->b_this_page;
3302
3303 if (bh->b_assoc_map)
3304 __remove_assoc_queue(bh);
3305 bh = next;
3306 } while (bh != head);
3307 *buffers_to_free = head;
3308 __clear_page_buffers(page);
3309 return 1;
3310failed:
3311 return 0;
3312}
3313
3314int try_to_free_buffers(struct page *page)
3315{
3316 struct address_space * const mapping = page->mapping;
3317 struct buffer_head *buffers_to_free = NULL;
3318 int ret = 0;
3319
3320 BUG_ON(!PageLocked(page));
3321 if (PageWriteback(page))
3322 return 0;
3323
3324 if (mapping == NULL) { /* can this still happen? */
3325 ret = drop_buffers(page, &buffers_to_free);
3326 goto out;
3327 }
3328
3329 spin_lock(&mapping->private_lock);
3330 ret = drop_buffers(page, &buffers_to_free);
3331
3332 /*
3333 * If the filesystem writes its buffers by hand (eg ext3)
3334 * then we can have clean buffers against a dirty page. We
3335 * clean the page here; otherwise the VM will never notice
3336 * that the filesystem did any IO at all.
3337 *
3338 * Also, during truncate, discard_buffer will have marked all
3339 * the page's buffers clean. We discover that here and clean
3340 * the page also.
3341 *
3342 * private_lock must be held over this entire operation in order
3343 * to synchronise against __set_page_dirty_buffers and prevent the
3344 * dirty bit from being lost.
3345 */
3346 if (ret)
3347 cancel_dirty_page(page);
3348 spin_unlock(&mapping->private_lock);
3349out:
3350 if (buffers_to_free) {
3351 struct buffer_head *bh = buffers_to_free;
3352
3353 do {
3354 struct buffer_head *next = bh->b_this_page;
3355 free_buffer_head(bh);
3356 bh = next;
3357 } while (bh != buffers_to_free);
3358 }
3359 return ret;
3360}
3361EXPORT_SYMBOL(try_to_free_buffers);
3362
3363/*
3364 * There are no bdflush tunables left. But distributions are
3365 * still running obsolete flush daemons, so we terminate them here.
3366 *
3367 * Use of bdflush() is deprecated and will be removed in a future kernel.
3368 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3369 */
3370SYSCALL_DEFINE2(bdflush, int, func, long, data)
3371{
3372 static int msg_count;
3373
3374 if (!capable(CAP_SYS_ADMIN))
3375 return -EPERM;
3376
3377 if (msg_count < 5) {
3378 msg_count++;
3379 printk(KERN_INFO
3380 "warning: process `%s' used the obsolete bdflush"
3381 " system call\n", current->comm);
3382 printk(KERN_INFO "Fix your initscripts?\n");
3383 }
3384
3385 if (func == 1)
3386 do_exit(0);
3387 return 0;
3388}
3389
3390/*
3391 * Buffer-head allocation
3392 */
3393static struct kmem_cache *bh_cachep __read_mostly;
3394
3395/*
3396 * Once the number of bh's in the machine exceeds this level, we start
3397 * stripping them in writeback.
3398 */
3399static unsigned long max_buffer_heads;
3400
3401int buffer_heads_over_limit;
3402
3403struct bh_accounting {
3404 int nr; /* Number of live bh's */
3405 int ratelimit; /* Limit cacheline bouncing */
3406};
3407
3408static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3409
3410static void recalc_bh_state(void)
3411{
3412 int i;
3413 int tot = 0;
3414
3415 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3416 return;
3417 __this_cpu_write(bh_accounting.ratelimit, 0);
3418 for_each_online_cpu(i)
3419 tot += per_cpu(bh_accounting, i).nr;
3420 buffer_heads_over_limit = (tot > max_buffer_heads);
3421}
3422
3423struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3424{
3425 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3426 if (ret) {
3427 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3428 preempt_disable();
3429 __this_cpu_inc(bh_accounting.nr);
3430 recalc_bh_state();
3431 preempt_enable();
3432 }
3433 return ret;
3434}
3435EXPORT_SYMBOL(alloc_buffer_head);
3436
3437void free_buffer_head(struct buffer_head *bh)
3438{
3439 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3440 kmem_cache_free(bh_cachep, bh);
3441 preempt_disable();
3442 __this_cpu_dec(bh_accounting.nr);
3443 recalc_bh_state();
3444 preempt_enable();
3445}
3446EXPORT_SYMBOL(free_buffer_head);
3447
3448static int buffer_exit_cpu_dead(unsigned int cpu)
3449{
3450 int i;
3451 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3452
3453 for (i = 0; i < BH_LRU_SIZE; i++) {
3454 brelse(b->bhs[i]);
3455 b->bhs[i] = NULL;
3456 }
3457 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3458 per_cpu(bh_accounting, cpu).nr = 0;
3459 return 0;
3460}
3461
3462/**
3463 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3464 * @bh: struct buffer_head
3465 *
3466 * Return true if the buffer is up-to-date and false,
3467 * with the buffer locked, if not.
3468 */
3469int bh_uptodate_or_lock(struct buffer_head *bh)
3470{
3471 if (!buffer_uptodate(bh)) {
3472 lock_buffer(bh);
3473 if (!buffer_uptodate(bh))
3474 return 0;
3475 unlock_buffer(bh);
3476 }
3477 return 1;
3478}
3479EXPORT_SYMBOL(bh_uptodate_or_lock);
3480
3481/**
3482 * bh_submit_read - Submit a locked buffer for reading
3483 * @bh: struct buffer_head
3484 *
3485 * Returns zero on success and -EIO on error.
3486 */
3487int bh_submit_read(struct buffer_head *bh)
3488{
3489 BUG_ON(!buffer_locked(bh));
3490
3491 if (buffer_uptodate(bh)) {
3492 unlock_buffer(bh);
3493 return 0;
3494 }
3495
3496 get_bh(bh);
3497 bh->b_end_io = end_buffer_read_sync;
3498 submit_bh(REQ_OP_READ, 0, bh);
3499 wait_on_buffer(bh);
3500 if (buffer_uptodate(bh))
3501 return 0;
3502 return -EIO;
3503}
3504EXPORT_SYMBOL(bh_submit_read);
3505
3506void __init buffer_init(void)
3507{
3508 unsigned long nrpages;
3509 int ret;
3510
3511 bh_cachep = kmem_cache_create("buffer_head",
3512 sizeof(struct buffer_head), 0,
3513 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3514 SLAB_MEM_SPREAD),
3515 NULL);
3516
3517 /*
3518 * Limit the bh occupancy to 10% of ZONE_NORMAL
3519 */
3520 nrpages = (nr_free_buffer_pages() * 10) / 100;
3521 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3522 ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3523 NULL, buffer_exit_cpu_dead);
3524 WARN_ON(ret < 0);
3525}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/fs/buffer.c
4 *
5 * Copyright (C) 1991, 1992, 2002 Linus Torvalds
6 */
7
8/*
9 * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
10 *
11 * Removed a lot of unnecessary code and simplified things now that
12 * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
13 *
14 * Speed up hash, lru, and free list operations. Use gfp() for allocating
15 * hash table, use SLAB cache for buffer heads. SMP threading. -DaveM
16 *
17 * Added 32k buffer block sizes - these are required older ARM systems. - RMK
18 *
19 * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
20 */
21
22#include <linux/kernel.h>
23#include <linux/sched/signal.h>
24#include <linux/syscalls.h>
25#include <linux/fs.h>
26#include <linux/iomap.h>
27#include <linux/mm.h>
28#include <linux/percpu.h>
29#include <linux/slab.h>
30#include <linux/capability.h>
31#include <linux/blkdev.h>
32#include <linux/file.h>
33#include <linux/quotaops.h>
34#include <linux/highmem.h>
35#include <linux/export.h>
36#include <linux/backing-dev.h>
37#include <linux/writeback.h>
38#include <linux/hash.h>
39#include <linux/suspend.h>
40#include <linux/buffer_head.h>
41#include <linux/task_io_accounting_ops.h>
42#include <linux/bio.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 <linux/sched/mm.h>
49#include <trace/events/block.h>
50#include <linux/fscrypt.h>
51#include <linux/fsverity.h>
52#include <linux/sched/isolation.h>
53
54#include "internal.h"
55
56static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
57static void submit_bh_wbc(blk_opf_t opf, struct buffer_head *bh,
58 enum rw_hint hint, struct writeback_control *wbc);
59
60#define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
61
62inline void touch_buffer(struct buffer_head *bh)
63{
64 trace_block_touch_buffer(bh);
65 folio_mark_accessed(bh->b_folio);
66}
67EXPORT_SYMBOL(touch_buffer);
68
69void __lock_buffer(struct buffer_head *bh)
70{
71 wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
72}
73EXPORT_SYMBOL(__lock_buffer);
74
75void unlock_buffer(struct buffer_head *bh)
76{
77 clear_bit_unlock(BH_Lock, &bh->b_state);
78 smp_mb__after_atomic();
79 wake_up_bit(&bh->b_state, BH_Lock);
80}
81EXPORT_SYMBOL(unlock_buffer);
82
83/*
84 * Returns if the folio has dirty or writeback buffers. If all the buffers
85 * are unlocked and clean then the folio_test_dirty information is stale. If
86 * any of the buffers are locked, it is assumed they are locked for IO.
87 */
88void buffer_check_dirty_writeback(struct folio *folio,
89 bool *dirty, bool *writeback)
90{
91 struct buffer_head *head, *bh;
92 *dirty = false;
93 *writeback = false;
94
95 BUG_ON(!folio_test_locked(folio));
96
97 head = folio_buffers(folio);
98 if (!head)
99 return;
100
101 if (folio_test_writeback(folio))
102 *writeback = true;
103
104 bh = head;
105 do {
106 if (buffer_locked(bh))
107 *writeback = true;
108
109 if (buffer_dirty(bh))
110 *dirty = true;
111
112 bh = bh->b_this_page;
113 } while (bh != head);
114}
115
116/*
117 * Block until a buffer comes unlocked. This doesn't stop it
118 * from becoming locked again - you have to lock it yourself
119 * if you want to preserve its state.
120 */
121void __wait_on_buffer(struct buffer_head * bh)
122{
123 wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
124}
125EXPORT_SYMBOL(__wait_on_buffer);
126
127static void buffer_io_error(struct buffer_head *bh, char *msg)
128{
129 if (!test_bit(BH_Quiet, &bh->b_state))
130 printk_ratelimited(KERN_ERR
131 "Buffer I/O error on dev %pg, logical block %llu%s\n",
132 bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
133}
134
135/*
136 * End-of-IO handler helper function which does not touch the bh after
137 * unlocking it.
138 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
139 * a race there is benign: unlock_buffer() only use the bh's address for
140 * hashing after unlocking the buffer, so it doesn't actually touch the bh
141 * itself.
142 */
143static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
144{
145 if (uptodate) {
146 set_buffer_uptodate(bh);
147 } else {
148 /* This happens, due to failed read-ahead attempts. */
149 clear_buffer_uptodate(bh);
150 }
151 unlock_buffer(bh);
152}
153
154/*
155 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
156 * unlock the buffer.
157 */
158void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
159{
160 __end_buffer_read_notouch(bh, uptodate);
161 put_bh(bh);
162}
163EXPORT_SYMBOL(end_buffer_read_sync);
164
165void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
166{
167 if (uptodate) {
168 set_buffer_uptodate(bh);
169 } else {
170 buffer_io_error(bh, ", lost sync page write");
171 mark_buffer_write_io_error(bh);
172 clear_buffer_uptodate(bh);
173 }
174 unlock_buffer(bh);
175 put_bh(bh);
176}
177EXPORT_SYMBOL(end_buffer_write_sync);
178
179/*
180 * Various filesystems appear to want __find_get_block to be non-blocking.
181 * But it's the page lock which protects the buffers. To get around this,
182 * we get exclusion from try_to_free_buffers with the blockdev mapping's
183 * i_private_lock.
184 *
185 * Hack idea: for the blockdev mapping, i_private_lock contention
186 * may be quite high. This code could TryLock the page, and if that
187 * succeeds, there is no need to take i_private_lock.
188 */
189static struct buffer_head *
190__find_get_block_slow(struct block_device *bdev, sector_t block)
191{
192 struct inode *bd_inode = bdev->bd_inode;
193 struct address_space *bd_mapping = bd_inode->i_mapping;
194 struct buffer_head *ret = NULL;
195 pgoff_t index;
196 struct buffer_head *bh;
197 struct buffer_head *head;
198 struct folio *folio;
199 int all_mapped = 1;
200 static DEFINE_RATELIMIT_STATE(last_warned, HZ, 1);
201
202 index = ((loff_t)block << bd_inode->i_blkbits) / PAGE_SIZE;
203 folio = __filemap_get_folio(bd_mapping, index, FGP_ACCESSED, 0);
204 if (IS_ERR(folio))
205 goto out;
206
207 spin_lock(&bd_mapping->i_private_lock);
208 head = folio_buffers(folio);
209 if (!head)
210 goto out_unlock;
211 bh = head;
212 do {
213 if (!buffer_mapped(bh))
214 all_mapped = 0;
215 else if (bh->b_blocknr == block) {
216 ret = bh;
217 get_bh(bh);
218 goto out_unlock;
219 }
220 bh = bh->b_this_page;
221 } while (bh != head);
222
223 /* we might be here because some of the buffers on this page are
224 * not mapped. This is due to various races between
225 * file io on the block device and getblk. It gets dealt with
226 * elsewhere, don't buffer_error if we had some unmapped buffers
227 */
228 ratelimit_set_flags(&last_warned, RATELIMIT_MSG_ON_RELEASE);
229 if (all_mapped && __ratelimit(&last_warned)) {
230 printk("__find_get_block_slow() failed. block=%llu, "
231 "b_blocknr=%llu, b_state=0x%08lx, b_size=%zu, "
232 "device %pg blocksize: %d\n",
233 (unsigned long long)block,
234 (unsigned long long)bh->b_blocknr,
235 bh->b_state, bh->b_size, bdev,
236 1 << bd_inode->i_blkbits);
237 }
238out_unlock:
239 spin_unlock(&bd_mapping->i_private_lock);
240 folio_put(folio);
241out:
242 return ret;
243}
244
245static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
246{
247 unsigned long flags;
248 struct buffer_head *first;
249 struct buffer_head *tmp;
250 struct folio *folio;
251 int folio_uptodate = 1;
252
253 BUG_ON(!buffer_async_read(bh));
254
255 folio = bh->b_folio;
256 if (uptodate) {
257 set_buffer_uptodate(bh);
258 } else {
259 clear_buffer_uptodate(bh);
260 buffer_io_error(bh, ", async page read");
261 folio_set_error(folio);
262 }
263
264 /*
265 * Be _very_ careful from here on. Bad things can happen if
266 * two buffer heads end IO at almost the same time and both
267 * decide that the page is now completely done.
268 */
269 first = folio_buffers(folio);
270 spin_lock_irqsave(&first->b_uptodate_lock, flags);
271 clear_buffer_async_read(bh);
272 unlock_buffer(bh);
273 tmp = bh;
274 do {
275 if (!buffer_uptodate(tmp))
276 folio_uptodate = 0;
277 if (buffer_async_read(tmp)) {
278 BUG_ON(!buffer_locked(tmp));
279 goto still_busy;
280 }
281 tmp = tmp->b_this_page;
282 } while (tmp != bh);
283 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
284
285 folio_end_read(folio, folio_uptodate);
286 return;
287
288still_busy:
289 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
290 return;
291}
292
293struct postprocess_bh_ctx {
294 struct work_struct work;
295 struct buffer_head *bh;
296};
297
298static void verify_bh(struct work_struct *work)
299{
300 struct postprocess_bh_ctx *ctx =
301 container_of(work, struct postprocess_bh_ctx, work);
302 struct buffer_head *bh = ctx->bh;
303 bool valid;
304
305 valid = fsverity_verify_blocks(bh->b_folio, bh->b_size, bh_offset(bh));
306 end_buffer_async_read(bh, valid);
307 kfree(ctx);
308}
309
310static bool need_fsverity(struct buffer_head *bh)
311{
312 struct folio *folio = bh->b_folio;
313 struct inode *inode = folio->mapping->host;
314
315 return fsverity_active(inode) &&
316 /* needed by ext4 */
317 folio->index < DIV_ROUND_UP(inode->i_size, PAGE_SIZE);
318}
319
320static void decrypt_bh(struct work_struct *work)
321{
322 struct postprocess_bh_ctx *ctx =
323 container_of(work, struct postprocess_bh_ctx, work);
324 struct buffer_head *bh = ctx->bh;
325 int err;
326
327 err = fscrypt_decrypt_pagecache_blocks(bh->b_folio, bh->b_size,
328 bh_offset(bh));
329 if (err == 0 && need_fsverity(bh)) {
330 /*
331 * We use different work queues for decryption and for verity
332 * because verity may require reading metadata pages that need
333 * decryption, and we shouldn't recurse to the same workqueue.
334 */
335 INIT_WORK(&ctx->work, verify_bh);
336 fsverity_enqueue_verify_work(&ctx->work);
337 return;
338 }
339 end_buffer_async_read(bh, err == 0);
340 kfree(ctx);
341}
342
343/*
344 * I/O completion handler for block_read_full_folio() - pages
345 * which come unlocked at the end of I/O.
346 */
347static void end_buffer_async_read_io(struct buffer_head *bh, int uptodate)
348{
349 struct inode *inode = bh->b_folio->mapping->host;
350 bool decrypt = fscrypt_inode_uses_fs_layer_crypto(inode);
351 bool verify = need_fsverity(bh);
352
353 /* Decrypt (with fscrypt) and/or verify (with fsverity) if needed. */
354 if (uptodate && (decrypt || verify)) {
355 struct postprocess_bh_ctx *ctx =
356 kmalloc(sizeof(*ctx), GFP_ATOMIC);
357
358 if (ctx) {
359 ctx->bh = bh;
360 if (decrypt) {
361 INIT_WORK(&ctx->work, decrypt_bh);
362 fscrypt_enqueue_decrypt_work(&ctx->work);
363 } else {
364 INIT_WORK(&ctx->work, verify_bh);
365 fsverity_enqueue_verify_work(&ctx->work);
366 }
367 return;
368 }
369 uptodate = 0;
370 }
371 end_buffer_async_read(bh, uptodate);
372}
373
374/*
375 * Completion handler for block_write_full_folio() - folios which are unlocked
376 * during I/O, and which have the writeback flag cleared upon I/O completion.
377 */
378static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
379{
380 unsigned long flags;
381 struct buffer_head *first;
382 struct buffer_head *tmp;
383 struct folio *folio;
384
385 BUG_ON(!buffer_async_write(bh));
386
387 folio = bh->b_folio;
388 if (uptodate) {
389 set_buffer_uptodate(bh);
390 } else {
391 buffer_io_error(bh, ", lost async page write");
392 mark_buffer_write_io_error(bh);
393 clear_buffer_uptodate(bh);
394 folio_set_error(folio);
395 }
396
397 first = folio_buffers(folio);
398 spin_lock_irqsave(&first->b_uptodate_lock, flags);
399
400 clear_buffer_async_write(bh);
401 unlock_buffer(bh);
402 tmp = bh->b_this_page;
403 while (tmp != bh) {
404 if (buffer_async_write(tmp)) {
405 BUG_ON(!buffer_locked(tmp));
406 goto still_busy;
407 }
408 tmp = tmp->b_this_page;
409 }
410 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
411 folio_end_writeback(folio);
412 return;
413
414still_busy:
415 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
416 return;
417}
418
419/*
420 * If a page's buffers are under async readin (end_buffer_async_read
421 * completion) then there is a possibility that another thread of
422 * control could lock one of the buffers after it has completed
423 * but while some of the other buffers have not completed. This
424 * locked buffer would confuse end_buffer_async_read() into not unlocking
425 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
426 * that this buffer is not under async I/O.
427 *
428 * The page comes unlocked when it has no locked buffer_async buffers
429 * left.
430 *
431 * PageLocked prevents anyone starting new async I/O reads any of
432 * the buffers.
433 *
434 * PageWriteback is used to prevent simultaneous writeout of the same
435 * page.
436 *
437 * PageLocked prevents anyone from starting writeback of a page which is
438 * under read I/O (PageWriteback is only ever set against a locked page).
439 */
440static void mark_buffer_async_read(struct buffer_head *bh)
441{
442 bh->b_end_io = end_buffer_async_read_io;
443 set_buffer_async_read(bh);
444}
445
446static void mark_buffer_async_write_endio(struct buffer_head *bh,
447 bh_end_io_t *handler)
448{
449 bh->b_end_io = handler;
450 set_buffer_async_write(bh);
451}
452
453void mark_buffer_async_write(struct buffer_head *bh)
454{
455 mark_buffer_async_write_endio(bh, end_buffer_async_write);
456}
457EXPORT_SYMBOL(mark_buffer_async_write);
458
459
460/*
461 * fs/buffer.c contains helper functions for buffer-backed address space's
462 * fsync functions. A common requirement for buffer-based filesystems is
463 * that certain data from the backing blockdev needs to be written out for
464 * a successful fsync(). For example, ext2 indirect blocks need to be
465 * written back and waited upon before fsync() returns.
466 *
467 * The functions mark_buffer_dirty_inode(), fsync_inode_buffers(),
468 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
469 * management of a list of dependent buffers at ->i_mapping->i_private_list.
470 *
471 * Locking is a little subtle: try_to_free_buffers() will remove buffers
472 * from their controlling inode's queue when they are being freed. But
473 * try_to_free_buffers() will be operating against the *blockdev* mapping
474 * at the time, not against the S_ISREG file which depends on those buffers.
475 * So the locking for i_private_list is via the i_private_lock in the address_space
476 * which backs the buffers. Which is different from the address_space
477 * against which the buffers are listed. So for a particular address_space,
478 * mapping->i_private_lock does *not* protect mapping->i_private_list! In fact,
479 * mapping->i_private_list will always be protected by the backing blockdev's
480 * ->i_private_lock.
481 *
482 * Which introduces a requirement: all buffers on an address_space's
483 * ->i_private_list must be from the same address_space: the blockdev's.
484 *
485 * address_spaces which do not place buffers at ->i_private_list via these
486 * utility functions are free to use i_private_lock and i_private_list for
487 * whatever they want. The only requirement is that list_empty(i_private_list)
488 * be true at clear_inode() time.
489 *
490 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
491 * filesystems should do that. invalidate_inode_buffers() should just go
492 * BUG_ON(!list_empty).
493 *
494 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
495 * take an address_space, not an inode. And it should be called
496 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
497 * queued up.
498 *
499 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
500 * list if it is already on a list. Because if the buffer is on a list,
501 * it *must* already be on the right one. If not, the filesystem is being
502 * silly. This will save a ton of locking. But first we have to ensure
503 * that buffers are taken *off* the old inode's list when they are freed
504 * (presumably in truncate). That requires careful auditing of all
505 * filesystems (do it inside bforget()). It could also be done by bringing
506 * b_inode back.
507 */
508
509/*
510 * The buffer's backing address_space's i_private_lock must be held
511 */
512static void __remove_assoc_queue(struct buffer_head *bh)
513{
514 list_del_init(&bh->b_assoc_buffers);
515 WARN_ON(!bh->b_assoc_map);
516 bh->b_assoc_map = NULL;
517}
518
519int inode_has_buffers(struct inode *inode)
520{
521 return !list_empty(&inode->i_data.i_private_list);
522}
523
524/*
525 * osync is designed to support O_SYNC io. It waits synchronously for
526 * all already-submitted IO to complete, but does not queue any new
527 * writes to the disk.
528 *
529 * To do O_SYNC writes, just queue the buffer writes with write_dirty_buffer
530 * as you dirty the buffers, and then use osync_inode_buffers to wait for
531 * completion. Any other dirty buffers which are not yet queued for
532 * write will not be flushed to disk by the osync.
533 */
534static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
535{
536 struct buffer_head *bh;
537 struct list_head *p;
538 int err = 0;
539
540 spin_lock(lock);
541repeat:
542 list_for_each_prev(p, list) {
543 bh = BH_ENTRY(p);
544 if (buffer_locked(bh)) {
545 get_bh(bh);
546 spin_unlock(lock);
547 wait_on_buffer(bh);
548 if (!buffer_uptodate(bh))
549 err = -EIO;
550 brelse(bh);
551 spin_lock(lock);
552 goto repeat;
553 }
554 }
555 spin_unlock(lock);
556 return err;
557}
558
559/**
560 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
561 * @mapping: the mapping which wants those buffers written
562 *
563 * Starts I/O against the buffers at mapping->i_private_list, and waits upon
564 * that I/O.
565 *
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().
569 */
570int sync_mapping_buffers(struct address_space *mapping)
571{
572 struct address_space *buffer_mapping = mapping->i_private_data;
573
574 if (buffer_mapping == NULL || list_empty(&mapping->i_private_list))
575 return 0;
576
577 return fsync_buffers_list(&buffer_mapping->i_private_lock,
578 &mapping->i_private_list);
579}
580EXPORT_SYMBOL(sync_mapping_buffers);
581
582/**
583 * generic_buffers_fsync_noflush - generic buffer fsync implementation
584 * for simple filesystems with no inode lock
585 *
586 * @file: file to synchronize
587 * @start: start offset in bytes
588 * @end: end offset in bytes (inclusive)
589 * @datasync: only synchronize essential metadata if true
590 *
591 * This is a generic implementation of the fsync method for simple
592 * filesystems which track all non-inode metadata in the buffers list
593 * hanging off the address_space structure.
594 */
595int generic_buffers_fsync_noflush(struct file *file, loff_t start, loff_t end,
596 bool datasync)
597{
598 struct inode *inode = file->f_mapping->host;
599 int err;
600 int ret;
601
602 err = file_write_and_wait_range(file, start, end);
603 if (err)
604 return err;
605
606 ret = sync_mapping_buffers(inode->i_mapping);
607 if (!(inode->i_state & I_DIRTY_ALL))
608 goto out;
609 if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
610 goto out;
611
612 err = sync_inode_metadata(inode, 1);
613 if (ret == 0)
614 ret = err;
615
616out:
617 /* check and advance again to catch errors after syncing out buffers */
618 err = file_check_and_advance_wb_err(file);
619 if (ret == 0)
620 ret = err;
621 return ret;
622}
623EXPORT_SYMBOL(generic_buffers_fsync_noflush);
624
625/**
626 * generic_buffers_fsync - generic buffer fsync implementation
627 * for simple filesystems with no inode lock
628 *
629 * @file: file to synchronize
630 * @start: start offset in bytes
631 * @end: end offset in bytes (inclusive)
632 * @datasync: only synchronize essential metadata if true
633 *
634 * This is a generic implementation of the fsync method for simple
635 * filesystems which track all non-inode metadata in the buffers list
636 * hanging off the address_space structure. This also makes sure that
637 * a device cache flush operation is called at the end.
638 */
639int generic_buffers_fsync(struct file *file, loff_t start, loff_t end,
640 bool datasync)
641{
642 struct inode *inode = file->f_mapping->host;
643 int ret;
644
645 ret = generic_buffers_fsync_noflush(file, start, end, datasync);
646 if (!ret)
647 ret = blkdev_issue_flush(inode->i_sb->s_bdev);
648 return ret;
649}
650EXPORT_SYMBOL(generic_buffers_fsync);
651
652/*
653 * Called when we've recently written block `bblock', and it is known that
654 * `bblock' was for a buffer_boundary() buffer. This means that the block at
655 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
656 * dirty, schedule it for IO. So that indirects merge nicely with their data.
657 */
658void write_boundary_block(struct block_device *bdev,
659 sector_t bblock, unsigned blocksize)
660{
661 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
662 if (bh) {
663 if (buffer_dirty(bh))
664 write_dirty_buffer(bh, 0);
665 put_bh(bh);
666 }
667}
668
669void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
670{
671 struct address_space *mapping = inode->i_mapping;
672 struct address_space *buffer_mapping = bh->b_folio->mapping;
673
674 mark_buffer_dirty(bh);
675 if (!mapping->i_private_data) {
676 mapping->i_private_data = buffer_mapping;
677 } else {
678 BUG_ON(mapping->i_private_data != buffer_mapping);
679 }
680 if (!bh->b_assoc_map) {
681 spin_lock(&buffer_mapping->i_private_lock);
682 list_move_tail(&bh->b_assoc_buffers,
683 &mapping->i_private_list);
684 bh->b_assoc_map = mapping;
685 spin_unlock(&buffer_mapping->i_private_lock);
686 }
687}
688EXPORT_SYMBOL(mark_buffer_dirty_inode);
689
690/*
691 * Add a page to the dirty page list.
692 *
693 * It is a sad fact of life that this function is called from several places
694 * deeply under spinlocking. It may not sleep.
695 *
696 * If the page has buffers, the uptodate buffers are set dirty, to preserve
697 * dirty-state coherency between the page and the buffers. It the page does
698 * not have buffers then when they are later attached they will all be set
699 * dirty.
700 *
701 * The buffers are dirtied before the page is dirtied. There's a small race
702 * window in which a writepage caller may see the page cleanness but not the
703 * buffer dirtiness. That's fine. If this code were to set the page dirty
704 * before the buffers, a concurrent writepage caller could clear the page dirty
705 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
706 * page on the dirty page list.
707 *
708 * We use i_private_lock to lock against try_to_free_buffers while using the
709 * page's buffer list. Also use this to protect against clean buffers being
710 * added to the page after it was set dirty.
711 *
712 * FIXME: may need to call ->reservepage here as well. That's rather up to the
713 * address_space though.
714 */
715bool block_dirty_folio(struct address_space *mapping, struct folio *folio)
716{
717 struct buffer_head *head;
718 bool newly_dirty;
719
720 spin_lock(&mapping->i_private_lock);
721 head = folio_buffers(folio);
722 if (head) {
723 struct buffer_head *bh = head;
724
725 do {
726 set_buffer_dirty(bh);
727 bh = bh->b_this_page;
728 } while (bh != head);
729 }
730 /*
731 * Lock out page's memcg migration to keep PageDirty
732 * synchronized with per-memcg dirty page counters.
733 */
734 folio_memcg_lock(folio);
735 newly_dirty = !folio_test_set_dirty(folio);
736 spin_unlock(&mapping->i_private_lock);
737
738 if (newly_dirty)
739 __folio_mark_dirty(folio, mapping, 1);
740
741 folio_memcg_unlock(folio);
742
743 if (newly_dirty)
744 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
745
746 return newly_dirty;
747}
748EXPORT_SYMBOL(block_dirty_folio);
749
750/*
751 * Write out and wait upon a list of buffers.
752 *
753 * We have conflicting pressures: we want to make sure that all
754 * initially dirty buffers get waited on, but that any subsequently
755 * dirtied buffers don't. After all, we don't want fsync to last
756 * forever if somebody is actively writing to the file.
757 *
758 * Do this in two main stages: first we copy dirty buffers to a
759 * temporary inode list, queueing the writes as we go. Then we clean
760 * up, waiting for those writes to complete.
761 *
762 * During this second stage, any subsequent updates to the file may end
763 * up refiling the buffer on the original inode's dirty list again, so
764 * there is a chance we will end up with a buffer queued for write but
765 * not yet completed on that list. So, as a final cleanup we go through
766 * the osync code to catch these locked, dirty buffers without requeuing
767 * any newly dirty buffers for write.
768 */
769static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
770{
771 struct buffer_head *bh;
772 struct list_head tmp;
773 struct address_space *mapping;
774 int err = 0, err2;
775 struct blk_plug plug;
776
777 INIT_LIST_HEAD(&tmp);
778 blk_start_plug(&plug);
779
780 spin_lock(lock);
781 while (!list_empty(list)) {
782 bh = BH_ENTRY(list->next);
783 mapping = bh->b_assoc_map;
784 __remove_assoc_queue(bh);
785 /* Avoid race with mark_buffer_dirty_inode() which does
786 * a lockless check and we rely on seeing the dirty bit */
787 smp_mb();
788 if (buffer_dirty(bh) || buffer_locked(bh)) {
789 list_add(&bh->b_assoc_buffers, &tmp);
790 bh->b_assoc_map = mapping;
791 if (buffer_dirty(bh)) {
792 get_bh(bh);
793 spin_unlock(lock);
794 /*
795 * Ensure any pending I/O completes so that
796 * write_dirty_buffer() actually writes the
797 * current contents - it is a noop if I/O is
798 * still in flight on potentially older
799 * contents.
800 */
801 write_dirty_buffer(bh, REQ_SYNC);
802
803 /*
804 * Kick off IO for the previous mapping. Note
805 * that we will not run the very last mapping,
806 * wait_on_buffer() will do that for us
807 * through sync_buffer().
808 */
809 brelse(bh);
810 spin_lock(lock);
811 }
812 }
813 }
814
815 spin_unlock(lock);
816 blk_finish_plug(&plug);
817 spin_lock(lock);
818
819 while (!list_empty(&tmp)) {
820 bh = BH_ENTRY(tmp.prev);
821 get_bh(bh);
822 mapping = bh->b_assoc_map;
823 __remove_assoc_queue(bh);
824 /* Avoid race with mark_buffer_dirty_inode() which does
825 * a lockless check and we rely on seeing the dirty bit */
826 smp_mb();
827 if (buffer_dirty(bh)) {
828 list_add(&bh->b_assoc_buffers,
829 &mapping->i_private_list);
830 bh->b_assoc_map = mapping;
831 }
832 spin_unlock(lock);
833 wait_on_buffer(bh);
834 if (!buffer_uptodate(bh))
835 err = -EIO;
836 brelse(bh);
837 spin_lock(lock);
838 }
839
840 spin_unlock(lock);
841 err2 = osync_buffers_list(lock, list);
842 if (err)
843 return err;
844 else
845 return err2;
846}
847
848/*
849 * Invalidate any and all dirty buffers on a given inode. We are
850 * probably unmounting the fs, but that doesn't mean we have already
851 * done a sync(). Just drop the buffers from the inode list.
852 *
853 * NOTE: we take the inode's blockdev's mapping's i_private_lock. Which
854 * assumes that all the buffers are against the blockdev. Not true
855 * for reiserfs.
856 */
857void invalidate_inode_buffers(struct inode *inode)
858{
859 if (inode_has_buffers(inode)) {
860 struct address_space *mapping = &inode->i_data;
861 struct list_head *list = &mapping->i_private_list;
862 struct address_space *buffer_mapping = mapping->i_private_data;
863
864 spin_lock(&buffer_mapping->i_private_lock);
865 while (!list_empty(list))
866 __remove_assoc_queue(BH_ENTRY(list->next));
867 spin_unlock(&buffer_mapping->i_private_lock);
868 }
869}
870EXPORT_SYMBOL(invalidate_inode_buffers);
871
872/*
873 * Remove any clean buffers from the inode's buffer list. This is called
874 * when we're trying to free the inode itself. Those buffers can pin it.
875 *
876 * Returns true if all buffers were removed.
877 */
878int remove_inode_buffers(struct inode *inode)
879{
880 int ret = 1;
881
882 if (inode_has_buffers(inode)) {
883 struct address_space *mapping = &inode->i_data;
884 struct list_head *list = &mapping->i_private_list;
885 struct address_space *buffer_mapping = mapping->i_private_data;
886
887 spin_lock(&buffer_mapping->i_private_lock);
888 while (!list_empty(list)) {
889 struct buffer_head *bh = BH_ENTRY(list->next);
890 if (buffer_dirty(bh)) {
891 ret = 0;
892 break;
893 }
894 __remove_assoc_queue(bh);
895 }
896 spin_unlock(&buffer_mapping->i_private_lock);
897 }
898 return ret;
899}
900
901/*
902 * Create the appropriate buffers when given a folio for data area and
903 * the size of each buffer.. Use the bh->b_this_page linked list to
904 * follow the buffers created. Return NULL if unable to create more
905 * buffers.
906 *
907 * The retry flag is used to differentiate async IO (paging, swapping)
908 * which may not fail from ordinary buffer allocations.
909 */
910struct buffer_head *folio_alloc_buffers(struct folio *folio, unsigned long size,
911 gfp_t gfp)
912{
913 struct buffer_head *bh, *head;
914 long offset;
915 struct mem_cgroup *memcg, *old_memcg;
916
917 /* The folio lock pins the memcg */
918 memcg = folio_memcg(folio);
919 old_memcg = set_active_memcg(memcg);
920
921 head = NULL;
922 offset = folio_size(folio);
923 while ((offset -= size) >= 0) {
924 bh = alloc_buffer_head(gfp);
925 if (!bh)
926 goto no_grow;
927
928 bh->b_this_page = head;
929 bh->b_blocknr = -1;
930 head = bh;
931
932 bh->b_size = size;
933
934 /* Link the buffer to its folio */
935 folio_set_bh(bh, folio, offset);
936 }
937out:
938 set_active_memcg(old_memcg);
939 return head;
940/*
941 * In case anything failed, we just free everything we got.
942 */
943no_grow:
944 if (head) {
945 do {
946 bh = head;
947 head = head->b_this_page;
948 free_buffer_head(bh);
949 } while (head);
950 }
951
952 goto out;
953}
954EXPORT_SYMBOL_GPL(folio_alloc_buffers);
955
956struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
957 bool retry)
958{
959 gfp_t gfp = GFP_NOFS | __GFP_ACCOUNT;
960 if (retry)
961 gfp |= __GFP_NOFAIL;
962
963 return folio_alloc_buffers(page_folio(page), size, gfp);
964}
965EXPORT_SYMBOL_GPL(alloc_page_buffers);
966
967static inline void link_dev_buffers(struct folio *folio,
968 struct buffer_head *head)
969{
970 struct buffer_head *bh, *tail;
971
972 bh = head;
973 do {
974 tail = bh;
975 bh = bh->b_this_page;
976 } while (bh);
977 tail->b_this_page = head;
978 folio_attach_private(folio, head);
979}
980
981static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
982{
983 sector_t retval = ~((sector_t)0);
984 loff_t sz = bdev_nr_bytes(bdev);
985
986 if (sz) {
987 unsigned int sizebits = blksize_bits(size);
988 retval = (sz >> sizebits);
989 }
990 return retval;
991}
992
993/*
994 * Initialise the state of a blockdev folio's buffers.
995 */
996static sector_t folio_init_buffers(struct folio *folio,
997 struct block_device *bdev, unsigned size)
998{
999 struct buffer_head *head = folio_buffers(folio);
1000 struct buffer_head *bh = head;
1001 bool uptodate = folio_test_uptodate(folio);
1002 sector_t block = div_u64(folio_pos(folio), size);
1003 sector_t end_block = blkdev_max_block(bdev, size);
1004
1005 do {
1006 if (!buffer_mapped(bh)) {
1007 bh->b_end_io = NULL;
1008 bh->b_private = NULL;
1009 bh->b_bdev = bdev;
1010 bh->b_blocknr = block;
1011 if (uptodate)
1012 set_buffer_uptodate(bh);
1013 if (block < end_block)
1014 set_buffer_mapped(bh);
1015 }
1016 block++;
1017 bh = bh->b_this_page;
1018 } while (bh != head);
1019
1020 /*
1021 * Caller needs to validate requested block against end of device.
1022 */
1023 return end_block;
1024}
1025
1026/*
1027 * Create the page-cache folio that contains the requested block.
1028 *
1029 * This is used purely for blockdev mappings.
1030 *
1031 * Returns false if we have a failure which cannot be cured by retrying
1032 * without sleeping. Returns true if we succeeded, or the caller should retry.
1033 */
1034static bool grow_dev_folio(struct block_device *bdev, sector_t block,
1035 pgoff_t index, unsigned size, gfp_t gfp)
1036{
1037 struct inode *inode = bdev->bd_inode;
1038 struct folio *folio;
1039 struct buffer_head *bh;
1040 sector_t end_block = 0;
1041
1042 folio = __filemap_get_folio(inode->i_mapping, index,
1043 FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp);
1044 if (IS_ERR(folio))
1045 return false;
1046
1047 bh = folio_buffers(folio);
1048 if (bh) {
1049 if (bh->b_size == size) {
1050 end_block = folio_init_buffers(folio, bdev, size);
1051 goto unlock;
1052 }
1053
1054 /*
1055 * Retrying may succeed; for example the folio may finish
1056 * writeback, or buffers may be cleaned. This should not
1057 * happen very often; maybe we have old buffers attached to
1058 * this blockdev's page cache and we're trying to change
1059 * the block size?
1060 */
1061 if (!try_to_free_buffers(folio)) {
1062 end_block = ~0ULL;
1063 goto unlock;
1064 }
1065 }
1066
1067 bh = folio_alloc_buffers(folio, size, gfp | __GFP_ACCOUNT);
1068 if (!bh)
1069 goto unlock;
1070
1071 /*
1072 * Link the folio to the buffers and initialise them. Take the
1073 * lock to be atomic wrt __find_get_block(), which does not
1074 * run under the folio lock.
1075 */
1076 spin_lock(&inode->i_mapping->i_private_lock);
1077 link_dev_buffers(folio, bh);
1078 end_block = folio_init_buffers(folio, bdev, size);
1079 spin_unlock(&inode->i_mapping->i_private_lock);
1080unlock:
1081 folio_unlock(folio);
1082 folio_put(folio);
1083 return block < end_block;
1084}
1085
1086/*
1087 * Create buffers for the specified block device block's folio. If
1088 * that folio was dirty, the buffers are set dirty also. Returns false
1089 * if we've hit a permanent error.
1090 */
1091static bool grow_buffers(struct block_device *bdev, sector_t block,
1092 unsigned size, gfp_t gfp)
1093{
1094 loff_t pos;
1095
1096 /*
1097 * Check for a block which lies outside our maximum possible
1098 * pagecache index.
1099 */
1100 if (check_mul_overflow(block, (sector_t)size, &pos) || pos > MAX_LFS_FILESIZE) {
1101 printk(KERN_ERR "%s: requested out-of-range block %llu for device %pg\n",
1102 __func__, (unsigned long long)block,
1103 bdev);
1104 return false;
1105 }
1106
1107 /* Create a folio with the proper size buffers */
1108 return grow_dev_folio(bdev, block, pos / PAGE_SIZE, size, gfp);
1109}
1110
1111static struct buffer_head *
1112__getblk_slow(struct block_device *bdev, sector_t block,
1113 unsigned size, gfp_t gfp)
1114{
1115 /* Size must be multiple of hard sectorsize */
1116 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1117 (size < 512 || size > PAGE_SIZE))) {
1118 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1119 size);
1120 printk(KERN_ERR "logical block size: %d\n",
1121 bdev_logical_block_size(bdev));
1122
1123 dump_stack();
1124 return NULL;
1125 }
1126
1127 for (;;) {
1128 struct buffer_head *bh;
1129
1130 bh = __find_get_block(bdev, block, size);
1131 if (bh)
1132 return bh;
1133
1134 if (!grow_buffers(bdev, block, size, gfp))
1135 return NULL;
1136 }
1137}
1138
1139/*
1140 * The relationship between dirty buffers and dirty pages:
1141 *
1142 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1143 * the page is tagged dirty in the page cache.
1144 *
1145 * At all times, the dirtiness of the buffers represents the dirtiness of
1146 * subsections of the page. If the page has buffers, the page dirty bit is
1147 * merely a hint about the true dirty state.
1148 *
1149 * When a page is set dirty in its entirety, all its buffers are marked dirty
1150 * (if the page has buffers).
1151 *
1152 * When a buffer is marked dirty, its page is dirtied, but the page's other
1153 * buffers are not.
1154 *
1155 * Also. When blockdev buffers are explicitly read with bread(), they
1156 * individually become uptodate. But their backing page remains not
1157 * uptodate - even if all of its buffers are uptodate. A subsequent
1158 * block_read_full_folio() against that folio will discover all the uptodate
1159 * buffers, will set the folio uptodate and will perform no I/O.
1160 */
1161
1162/**
1163 * mark_buffer_dirty - mark a buffer_head as needing writeout
1164 * @bh: the buffer_head to mark dirty
1165 *
1166 * mark_buffer_dirty() will set the dirty bit against the buffer, then set
1167 * its backing page dirty, then tag the page as dirty in the page cache
1168 * and then attach the address_space's inode to its superblock's dirty
1169 * inode list.
1170 *
1171 * mark_buffer_dirty() is atomic. It takes bh->b_folio->mapping->i_private_lock,
1172 * i_pages lock and mapping->host->i_lock.
1173 */
1174void mark_buffer_dirty(struct buffer_head *bh)
1175{
1176 WARN_ON_ONCE(!buffer_uptodate(bh));
1177
1178 trace_block_dirty_buffer(bh);
1179
1180 /*
1181 * Very *carefully* optimize the it-is-already-dirty case.
1182 *
1183 * Don't let the final "is it dirty" escape to before we
1184 * perhaps modified the buffer.
1185 */
1186 if (buffer_dirty(bh)) {
1187 smp_mb();
1188 if (buffer_dirty(bh))
1189 return;
1190 }
1191
1192 if (!test_set_buffer_dirty(bh)) {
1193 struct folio *folio = bh->b_folio;
1194 struct address_space *mapping = NULL;
1195
1196 folio_memcg_lock(folio);
1197 if (!folio_test_set_dirty(folio)) {
1198 mapping = folio->mapping;
1199 if (mapping)
1200 __folio_mark_dirty(folio, mapping, 0);
1201 }
1202 folio_memcg_unlock(folio);
1203 if (mapping)
1204 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1205 }
1206}
1207EXPORT_SYMBOL(mark_buffer_dirty);
1208
1209void mark_buffer_write_io_error(struct buffer_head *bh)
1210{
1211 set_buffer_write_io_error(bh);
1212 /* FIXME: do we need to set this in both places? */
1213 if (bh->b_folio && bh->b_folio->mapping)
1214 mapping_set_error(bh->b_folio->mapping, -EIO);
1215 if (bh->b_assoc_map) {
1216 mapping_set_error(bh->b_assoc_map, -EIO);
1217 errseq_set(&bh->b_assoc_map->host->i_sb->s_wb_err, -EIO);
1218 }
1219}
1220EXPORT_SYMBOL(mark_buffer_write_io_error);
1221
1222/*
1223 * Decrement a buffer_head's reference count. If all buffers against a page
1224 * have zero reference count, are clean and unlocked, and if the page is clean
1225 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1226 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1227 * a page but it ends up not being freed, and buffers may later be reattached).
1228 */
1229void __brelse(struct buffer_head * buf)
1230{
1231 if (atomic_read(&buf->b_count)) {
1232 put_bh(buf);
1233 return;
1234 }
1235 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1236}
1237EXPORT_SYMBOL(__brelse);
1238
1239/*
1240 * bforget() is like brelse(), except it discards any
1241 * potentially dirty data.
1242 */
1243void __bforget(struct buffer_head *bh)
1244{
1245 clear_buffer_dirty(bh);
1246 if (bh->b_assoc_map) {
1247 struct address_space *buffer_mapping = bh->b_folio->mapping;
1248
1249 spin_lock(&buffer_mapping->i_private_lock);
1250 list_del_init(&bh->b_assoc_buffers);
1251 bh->b_assoc_map = NULL;
1252 spin_unlock(&buffer_mapping->i_private_lock);
1253 }
1254 __brelse(bh);
1255}
1256EXPORT_SYMBOL(__bforget);
1257
1258static struct buffer_head *__bread_slow(struct buffer_head *bh)
1259{
1260 lock_buffer(bh);
1261 if (buffer_uptodate(bh)) {
1262 unlock_buffer(bh);
1263 return bh;
1264 } else {
1265 get_bh(bh);
1266 bh->b_end_io = end_buffer_read_sync;
1267 submit_bh(REQ_OP_READ, bh);
1268 wait_on_buffer(bh);
1269 if (buffer_uptodate(bh))
1270 return bh;
1271 }
1272 brelse(bh);
1273 return NULL;
1274}
1275
1276/*
1277 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1278 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1279 * refcount elevated by one when they're in an LRU. A buffer can only appear
1280 * once in a particular CPU's LRU. A single buffer can be present in multiple
1281 * CPU's LRUs at the same time.
1282 *
1283 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1284 * sb_find_get_block().
1285 *
1286 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1287 * a local interrupt disable for that.
1288 */
1289
1290#define BH_LRU_SIZE 16
1291
1292struct bh_lru {
1293 struct buffer_head *bhs[BH_LRU_SIZE];
1294};
1295
1296static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1297
1298#ifdef CONFIG_SMP
1299#define bh_lru_lock() local_irq_disable()
1300#define bh_lru_unlock() local_irq_enable()
1301#else
1302#define bh_lru_lock() preempt_disable()
1303#define bh_lru_unlock() preempt_enable()
1304#endif
1305
1306static inline void check_irqs_on(void)
1307{
1308#ifdef irqs_disabled
1309 BUG_ON(irqs_disabled());
1310#endif
1311}
1312
1313/*
1314 * Install a buffer_head into this cpu's LRU. If not already in the LRU, it is
1315 * inserted at the front, and the buffer_head at the back if any is evicted.
1316 * Or, if already in the LRU it is moved to the front.
1317 */
1318static void bh_lru_install(struct buffer_head *bh)
1319{
1320 struct buffer_head *evictee = bh;
1321 struct bh_lru *b;
1322 int i;
1323
1324 check_irqs_on();
1325 bh_lru_lock();
1326
1327 /*
1328 * the refcount of buffer_head in bh_lru prevents dropping the
1329 * attached page(i.e., try_to_free_buffers) so it could cause
1330 * failing page migration.
1331 * Skip putting upcoming bh into bh_lru until migration is done.
1332 */
1333 if (lru_cache_disabled() || cpu_is_isolated(smp_processor_id())) {
1334 bh_lru_unlock();
1335 return;
1336 }
1337
1338 b = this_cpu_ptr(&bh_lrus);
1339 for (i = 0; i < BH_LRU_SIZE; i++) {
1340 swap(evictee, b->bhs[i]);
1341 if (evictee == bh) {
1342 bh_lru_unlock();
1343 return;
1344 }
1345 }
1346
1347 get_bh(bh);
1348 bh_lru_unlock();
1349 brelse(evictee);
1350}
1351
1352/*
1353 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1354 */
1355static struct buffer_head *
1356lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1357{
1358 struct buffer_head *ret = NULL;
1359 unsigned int i;
1360
1361 check_irqs_on();
1362 bh_lru_lock();
1363 if (cpu_is_isolated(smp_processor_id())) {
1364 bh_lru_unlock();
1365 return NULL;
1366 }
1367 for (i = 0; i < BH_LRU_SIZE; i++) {
1368 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1369
1370 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1371 bh->b_size == size) {
1372 if (i) {
1373 while (i) {
1374 __this_cpu_write(bh_lrus.bhs[i],
1375 __this_cpu_read(bh_lrus.bhs[i - 1]));
1376 i--;
1377 }
1378 __this_cpu_write(bh_lrus.bhs[0], bh);
1379 }
1380 get_bh(bh);
1381 ret = bh;
1382 break;
1383 }
1384 }
1385 bh_lru_unlock();
1386 return ret;
1387}
1388
1389/*
1390 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1391 * it in the LRU and mark it as accessed. If it is not present then return
1392 * NULL
1393 */
1394struct buffer_head *
1395__find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1396{
1397 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1398
1399 if (bh == NULL) {
1400 /* __find_get_block_slow will mark the page accessed */
1401 bh = __find_get_block_slow(bdev, block);
1402 if (bh)
1403 bh_lru_install(bh);
1404 } else
1405 touch_buffer(bh);
1406
1407 return bh;
1408}
1409EXPORT_SYMBOL(__find_get_block);
1410
1411/**
1412 * bdev_getblk - Get a buffer_head in a block device's buffer cache.
1413 * @bdev: The block device.
1414 * @block: The block number.
1415 * @size: The size of buffer_heads for this @bdev.
1416 * @gfp: The memory allocation flags to use.
1417 *
1418 * Return: The buffer head, or NULL if memory could not be allocated.
1419 */
1420struct buffer_head *bdev_getblk(struct block_device *bdev, sector_t block,
1421 unsigned size, gfp_t gfp)
1422{
1423 struct buffer_head *bh = __find_get_block(bdev, block, size);
1424
1425 might_alloc(gfp);
1426 if (bh)
1427 return bh;
1428
1429 return __getblk_slow(bdev, block, size, gfp);
1430}
1431EXPORT_SYMBOL(bdev_getblk);
1432
1433/*
1434 * Do async read-ahead on a buffer..
1435 */
1436void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1437{
1438 struct buffer_head *bh = bdev_getblk(bdev, block, size,
1439 GFP_NOWAIT | __GFP_MOVABLE);
1440
1441 if (likely(bh)) {
1442 bh_readahead(bh, REQ_RAHEAD);
1443 brelse(bh);
1444 }
1445}
1446EXPORT_SYMBOL(__breadahead);
1447
1448/**
1449 * __bread_gfp() - reads a specified block and returns the bh
1450 * @bdev: the block_device to read from
1451 * @block: number of block
1452 * @size: size (in bytes) to read
1453 * @gfp: page allocation flag
1454 *
1455 * Reads a specified block, and returns buffer head that contains it.
1456 * The page cache can be allocated from non-movable area
1457 * not to prevent page migration if you set gfp to zero.
1458 * It returns NULL if the block was unreadable.
1459 */
1460struct buffer_head *
1461__bread_gfp(struct block_device *bdev, sector_t block,
1462 unsigned size, gfp_t gfp)
1463{
1464 struct buffer_head *bh;
1465
1466 gfp |= mapping_gfp_constraint(bdev->bd_inode->i_mapping, ~__GFP_FS);
1467
1468 /*
1469 * Prefer looping in the allocator rather than here, at least that
1470 * code knows what it's doing.
1471 */
1472 gfp |= __GFP_NOFAIL;
1473
1474 bh = bdev_getblk(bdev, block, size, gfp);
1475
1476 if (likely(bh) && !buffer_uptodate(bh))
1477 bh = __bread_slow(bh);
1478 return bh;
1479}
1480EXPORT_SYMBOL(__bread_gfp);
1481
1482static void __invalidate_bh_lrus(struct bh_lru *b)
1483{
1484 int i;
1485
1486 for (i = 0; i < BH_LRU_SIZE; i++) {
1487 brelse(b->bhs[i]);
1488 b->bhs[i] = NULL;
1489 }
1490}
1491/*
1492 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1493 * This doesn't race because it runs in each cpu either in irq
1494 * or with preempt disabled.
1495 */
1496static void invalidate_bh_lru(void *arg)
1497{
1498 struct bh_lru *b = &get_cpu_var(bh_lrus);
1499
1500 __invalidate_bh_lrus(b);
1501 put_cpu_var(bh_lrus);
1502}
1503
1504bool has_bh_in_lru(int cpu, void *dummy)
1505{
1506 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1507 int i;
1508
1509 for (i = 0; i < BH_LRU_SIZE; i++) {
1510 if (b->bhs[i])
1511 return true;
1512 }
1513
1514 return false;
1515}
1516
1517void invalidate_bh_lrus(void)
1518{
1519 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1);
1520}
1521EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1522
1523/*
1524 * It's called from workqueue context so we need a bh_lru_lock to close
1525 * the race with preemption/irq.
1526 */
1527void invalidate_bh_lrus_cpu(void)
1528{
1529 struct bh_lru *b;
1530
1531 bh_lru_lock();
1532 b = this_cpu_ptr(&bh_lrus);
1533 __invalidate_bh_lrus(b);
1534 bh_lru_unlock();
1535}
1536
1537void folio_set_bh(struct buffer_head *bh, struct folio *folio,
1538 unsigned long offset)
1539{
1540 bh->b_folio = folio;
1541 BUG_ON(offset >= folio_size(folio));
1542 if (folio_test_highmem(folio))
1543 /*
1544 * This catches illegal uses and preserves the offset:
1545 */
1546 bh->b_data = (char *)(0 + offset);
1547 else
1548 bh->b_data = folio_address(folio) + offset;
1549}
1550EXPORT_SYMBOL(folio_set_bh);
1551
1552/*
1553 * Called when truncating a buffer on a page completely.
1554 */
1555
1556/* Bits that are cleared during an invalidate */
1557#define BUFFER_FLAGS_DISCARD \
1558 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1559 1 << BH_Delay | 1 << BH_Unwritten)
1560
1561static void discard_buffer(struct buffer_head * bh)
1562{
1563 unsigned long b_state;
1564
1565 lock_buffer(bh);
1566 clear_buffer_dirty(bh);
1567 bh->b_bdev = NULL;
1568 b_state = READ_ONCE(bh->b_state);
1569 do {
1570 } while (!try_cmpxchg(&bh->b_state, &b_state,
1571 b_state & ~BUFFER_FLAGS_DISCARD));
1572 unlock_buffer(bh);
1573}
1574
1575/**
1576 * block_invalidate_folio - Invalidate part or all of a buffer-backed folio.
1577 * @folio: The folio which is affected.
1578 * @offset: start of the range to invalidate
1579 * @length: length of the range to invalidate
1580 *
1581 * block_invalidate_folio() is called when all or part of the folio has been
1582 * invalidated by a truncate operation.
1583 *
1584 * block_invalidate_folio() does not have to release all buffers, but it must
1585 * ensure that no dirty buffer is left outside @offset and that no I/O
1586 * is underway against any of the blocks which are outside the truncation
1587 * point. Because the caller is about to free (and possibly reuse) those
1588 * blocks on-disk.
1589 */
1590void block_invalidate_folio(struct folio *folio, size_t offset, size_t length)
1591{
1592 struct buffer_head *head, *bh, *next;
1593 size_t curr_off = 0;
1594 size_t stop = length + offset;
1595
1596 BUG_ON(!folio_test_locked(folio));
1597
1598 /*
1599 * Check for overflow
1600 */
1601 BUG_ON(stop > folio_size(folio) || stop < length);
1602
1603 head = folio_buffers(folio);
1604 if (!head)
1605 return;
1606
1607 bh = head;
1608 do {
1609 size_t next_off = curr_off + bh->b_size;
1610 next = bh->b_this_page;
1611
1612 /*
1613 * Are we still fully in range ?
1614 */
1615 if (next_off > stop)
1616 goto out;
1617
1618 /*
1619 * is this block fully invalidated?
1620 */
1621 if (offset <= curr_off)
1622 discard_buffer(bh);
1623 curr_off = next_off;
1624 bh = next;
1625 } while (bh != head);
1626
1627 /*
1628 * We release buffers only if the entire folio is being invalidated.
1629 * The get_block cached value has been unconditionally invalidated,
1630 * so real IO is not possible anymore.
1631 */
1632 if (length == folio_size(folio))
1633 filemap_release_folio(folio, 0);
1634out:
1635 return;
1636}
1637EXPORT_SYMBOL(block_invalidate_folio);
1638
1639/*
1640 * We attach and possibly dirty the buffers atomically wrt
1641 * block_dirty_folio() via i_private_lock. try_to_free_buffers
1642 * is already excluded via the folio lock.
1643 */
1644struct buffer_head *create_empty_buffers(struct folio *folio,
1645 unsigned long blocksize, unsigned long b_state)
1646{
1647 struct buffer_head *bh, *head, *tail;
1648 gfp_t gfp = GFP_NOFS | __GFP_ACCOUNT | __GFP_NOFAIL;
1649
1650 head = folio_alloc_buffers(folio, blocksize, gfp);
1651 bh = head;
1652 do {
1653 bh->b_state |= b_state;
1654 tail = bh;
1655 bh = bh->b_this_page;
1656 } while (bh);
1657 tail->b_this_page = head;
1658
1659 spin_lock(&folio->mapping->i_private_lock);
1660 if (folio_test_uptodate(folio) || folio_test_dirty(folio)) {
1661 bh = head;
1662 do {
1663 if (folio_test_dirty(folio))
1664 set_buffer_dirty(bh);
1665 if (folio_test_uptodate(folio))
1666 set_buffer_uptodate(bh);
1667 bh = bh->b_this_page;
1668 } while (bh != head);
1669 }
1670 folio_attach_private(folio, head);
1671 spin_unlock(&folio->mapping->i_private_lock);
1672
1673 return head;
1674}
1675EXPORT_SYMBOL(create_empty_buffers);
1676
1677/**
1678 * clean_bdev_aliases: clean a range of buffers in block device
1679 * @bdev: Block device to clean buffers in
1680 * @block: Start of a range of blocks to clean
1681 * @len: Number of blocks to clean
1682 *
1683 * We are taking a range of blocks for data and we don't want writeback of any
1684 * buffer-cache aliases starting from return from this function and until the
1685 * moment when something will explicitly mark the buffer dirty (hopefully that
1686 * will not happen until we will free that block ;-) We don't even need to mark
1687 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1688 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1689 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1690 * would confuse anyone who might pick it with bread() afterwards...
1691 *
1692 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1693 * writeout I/O going on against recently-freed buffers. We don't wait on that
1694 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1695 * need to. That happens here.
1696 */
1697void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1698{
1699 struct inode *bd_inode = bdev->bd_inode;
1700 struct address_space *bd_mapping = bd_inode->i_mapping;
1701 struct folio_batch fbatch;
1702 pgoff_t index = ((loff_t)block << bd_inode->i_blkbits) / PAGE_SIZE;
1703 pgoff_t end;
1704 int i, count;
1705 struct buffer_head *bh;
1706 struct buffer_head *head;
1707
1708 end = ((loff_t)(block + len - 1) << bd_inode->i_blkbits) / PAGE_SIZE;
1709 folio_batch_init(&fbatch);
1710 while (filemap_get_folios(bd_mapping, &index, end, &fbatch)) {
1711 count = folio_batch_count(&fbatch);
1712 for (i = 0; i < count; i++) {
1713 struct folio *folio = fbatch.folios[i];
1714
1715 if (!folio_buffers(folio))
1716 continue;
1717 /*
1718 * We use folio lock instead of bd_mapping->i_private_lock
1719 * to pin buffers here since we can afford to sleep and
1720 * it scales better than a global spinlock lock.
1721 */
1722 folio_lock(folio);
1723 /* Recheck when the folio is locked which pins bhs */
1724 head = folio_buffers(folio);
1725 if (!head)
1726 goto unlock_page;
1727 bh = head;
1728 do {
1729 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1730 goto next;
1731 if (bh->b_blocknr >= block + len)
1732 break;
1733 clear_buffer_dirty(bh);
1734 wait_on_buffer(bh);
1735 clear_buffer_req(bh);
1736next:
1737 bh = bh->b_this_page;
1738 } while (bh != head);
1739unlock_page:
1740 folio_unlock(folio);
1741 }
1742 folio_batch_release(&fbatch);
1743 cond_resched();
1744 /* End of range already reached? */
1745 if (index > end || !index)
1746 break;
1747 }
1748}
1749EXPORT_SYMBOL(clean_bdev_aliases);
1750
1751static struct buffer_head *folio_create_buffers(struct folio *folio,
1752 struct inode *inode,
1753 unsigned int b_state)
1754{
1755 struct buffer_head *bh;
1756
1757 BUG_ON(!folio_test_locked(folio));
1758
1759 bh = folio_buffers(folio);
1760 if (!bh)
1761 bh = create_empty_buffers(folio,
1762 1 << READ_ONCE(inode->i_blkbits), b_state);
1763 return bh;
1764}
1765
1766/*
1767 * NOTE! All mapped/uptodate combinations are valid:
1768 *
1769 * Mapped Uptodate Meaning
1770 *
1771 * No No "unknown" - must do get_block()
1772 * No Yes "hole" - zero-filled
1773 * Yes No "allocated" - allocated on disk, not read in
1774 * Yes Yes "valid" - allocated and up-to-date in memory.
1775 *
1776 * "Dirty" is valid only with the last case (mapped+uptodate).
1777 */
1778
1779/*
1780 * While block_write_full_folio is writing back the dirty buffers under
1781 * the page lock, whoever dirtied the buffers may decide to clean them
1782 * again at any time. We handle that by only looking at the buffer
1783 * state inside lock_buffer().
1784 *
1785 * If block_write_full_folio() is called for regular writeback
1786 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1787 * locked buffer. This only can happen if someone has written the buffer
1788 * directly, with submit_bh(). At the address_space level PageWriteback
1789 * prevents this contention from occurring.
1790 *
1791 * If block_write_full_folio() is called with wbc->sync_mode ==
1792 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1793 * causes the writes to be flagged as synchronous writes.
1794 */
1795int __block_write_full_folio(struct inode *inode, struct folio *folio,
1796 get_block_t *get_block, struct writeback_control *wbc)
1797{
1798 int err;
1799 sector_t block;
1800 sector_t last_block;
1801 struct buffer_head *bh, *head;
1802 size_t blocksize;
1803 int nr_underway = 0;
1804 blk_opf_t write_flags = wbc_to_write_flags(wbc);
1805
1806 head = folio_create_buffers(folio, inode,
1807 (1 << BH_Dirty) | (1 << BH_Uptodate));
1808
1809 /*
1810 * Be very careful. We have no exclusion from block_dirty_folio
1811 * here, and the (potentially unmapped) buffers may become dirty at
1812 * any time. If a buffer becomes dirty here after we've inspected it
1813 * then we just miss that fact, and the folio stays dirty.
1814 *
1815 * Buffers outside i_size may be dirtied by block_dirty_folio;
1816 * handle that here by just cleaning them.
1817 */
1818
1819 bh = head;
1820 blocksize = bh->b_size;
1821
1822 block = div_u64(folio_pos(folio), blocksize);
1823 last_block = div_u64(i_size_read(inode) - 1, blocksize);
1824
1825 /*
1826 * Get all the dirty buffers mapped to disk addresses and
1827 * handle any aliases from the underlying blockdev's mapping.
1828 */
1829 do {
1830 if (block > last_block) {
1831 /*
1832 * mapped buffers outside i_size will occur, because
1833 * this folio can be outside i_size when there is a
1834 * truncate in progress.
1835 */
1836 /*
1837 * The buffer was zeroed by block_write_full_folio()
1838 */
1839 clear_buffer_dirty(bh);
1840 set_buffer_uptodate(bh);
1841 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1842 buffer_dirty(bh)) {
1843 WARN_ON(bh->b_size != blocksize);
1844 err = get_block(inode, block, bh, 1);
1845 if (err)
1846 goto recover;
1847 clear_buffer_delay(bh);
1848 if (buffer_new(bh)) {
1849 /* blockdev mappings never come here */
1850 clear_buffer_new(bh);
1851 clean_bdev_bh_alias(bh);
1852 }
1853 }
1854 bh = bh->b_this_page;
1855 block++;
1856 } while (bh != head);
1857
1858 do {
1859 if (!buffer_mapped(bh))
1860 continue;
1861 /*
1862 * If it's a fully non-blocking write attempt and we cannot
1863 * lock the buffer then redirty the folio. Note that this can
1864 * potentially cause a busy-wait loop from writeback threads
1865 * and kswapd activity, but those code paths have their own
1866 * higher-level throttling.
1867 */
1868 if (wbc->sync_mode != WB_SYNC_NONE) {
1869 lock_buffer(bh);
1870 } else if (!trylock_buffer(bh)) {
1871 folio_redirty_for_writepage(wbc, folio);
1872 continue;
1873 }
1874 if (test_clear_buffer_dirty(bh)) {
1875 mark_buffer_async_write_endio(bh,
1876 end_buffer_async_write);
1877 } else {
1878 unlock_buffer(bh);
1879 }
1880 } while ((bh = bh->b_this_page) != head);
1881
1882 /*
1883 * The folio and its buffers are protected by the writeback flag,
1884 * so we can drop the bh refcounts early.
1885 */
1886 BUG_ON(folio_test_writeback(folio));
1887 folio_start_writeback(folio);
1888
1889 do {
1890 struct buffer_head *next = bh->b_this_page;
1891 if (buffer_async_write(bh)) {
1892 submit_bh_wbc(REQ_OP_WRITE | write_flags, bh,
1893 inode->i_write_hint, wbc);
1894 nr_underway++;
1895 }
1896 bh = next;
1897 } while (bh != head);
1898 folio_unlock(folio);
1899
1900 err = 0;
1901done:
1902 if (nr_underway == 0) {
1903 /*
1904 * The folio was marked dirty, but the buffers were
1905 * clean. Someone wrote them back by hand with
1906 * write_dirty_buffer/submit_bh. A rare case.
1907 */
1908 folio_end_writeback(folio);
1909
1910 /*
1911 * The folio and buffer_heads can be released at any time from
1912 * here on.
1913 */
1914 }
1915 return err;
1916
1917recover:
1918 /*
1919 * ENOSPC, or some other error. We may already have added some
1920 * blocks to the file, so we need to write these out to avoid
1921 * exposing stale data.
1922 * The folio is currently locked and not marked for writeback
1923 */
1924 bh = head;
1925 /* Recovery: lock and submit the mapped buffers */
1926 do {
1927 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1928 !buffer_delay(bh)) {
1929 lock_buffer(bh);
1930 mark_buffer_async_write_endio(bh,
1931 end_buffer_async_write);
1932 } else {
1933 /*
1934 * The buffer may have been set dirty during
1935 * attachment to a dirty folio.
1936 */
1937 clear_buffer_dirty(bh);
1938 }
1939 } while ((bh = bh->b_this_page) != head);
1940 folio_set_error(folio);
1941 BUG_ON(folio_test_writeback(folio));
1942 mapping_set_error(folio->mapping, err);
1943 folio_start_writeback(folio);
1944 do {
1945 struct buffer_head *next = bh->b_this_page;
1946 if (buffer_async_write(bh)) {
1947 clear_buffer_dirty(bh);
1948 submit_bh_wbc(REQ_OP_WRITE | write_flags, bh,
1949 inode->i_write_hint, wbc);
1950 nr_underway++;
1951 }
1952 bh = next;
1953 } while (bh != head);
1954 folio_unlock(folio);
1955 goto done;
1956}
1957EXPORT_SYMBOL(__block_write_full_folio);
1958
1959/*
1960 * If a folio has any new buffers, zero them out here, and mark them uptodate
1961 * and dirty so they'll be written out (in order to prevent uninitialised
1962 * block data from leaking). And clear the new bit.
1963 */
1964void folio_zero_new_buffers(struct folio *folio, size_t from, size_t to)
1965{
1966 size_t block_start, block_end;
1967 struct buffer_head *head, *bh;
1968
1969 BUG_ON(!folio_test_locked(folio));
1970 head = folio_buffers(folio);
1971 if (!head)
1972 return;
1973
1974 bh = head;
1975 block_start = 0;
1976 do {
1977 block_end = block_start + bh->b_size;
1978
1979 if (buffer_new(bh)) {
1980 if (block_end > from && block_start < to) {
1981 if (!folio_test_uptodate(folio)) {
1982 size_t start, xend;
1983
1984 start = max(from, block_start);
1985 xend = min(to, block_end);
1986
1987 folio_zero_segment(folio, start, xend);
1988 set_buffer_uptodate(bh);
1989 }
1990
1991 clear_buffer_new(bh);
1992 mark_buffer_dirty(bh);
1993 }
1994 }
1995
1996 block_start = block_end;
1997 bh = bh->b_this_page;
1998 } while (bh != head);
1999}
2000EXPORT_SYMBOL(folio_zero_new_buffers);
2001
2002static int
2003iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
2004 const struct iomap *iomap)
2005{
2006 loff_t offset = (loff_t)block << inode->i_blkbits;
2007
2008 bh->b_bdev = iomap->bdev;
2009
2010 /*
2011 * Block points to offset in file we need to map, iomap contains
2012 * the offset at which the map starts. If the map ends before the
2013 * current block, then do not map the buffer and let the caller
2014 * handle it.
2015 */
2016 if (offset >= iomap->offset + iomap->length)
2017 return -EIO;
2018
2019 switch (iomap->type) {
2020 case IOMAP_HOLE:
2021 /*
2022 * If the buffer is not up to date or beyond the current EOF,
2023 * we need to mark it as new to ensure sub-block zeroing is
2024 * executed if necessary.
2025 */
2026 if (!buffer_uptodate(bh) ||
2027 (offset >= i_size_read(inode)))
2028 set_buffer_new(bh);
2029 return 0;
2030 case IOMAP_DELALLOC:
2031 if (!buffer_uptodate(bh) ||
2032 (offset >= i_size_read(inode)))
2033 set_buffer_new(bh);
2034 set_buffer_uptodate(bh);
2035 set_buffer_mapped(bh);
2036 set_buffer_delay(bh);
2037 return 0;
2038 case IOMAP_UNWRITTEN:
2039 /*
2040 * For unwritten regions, we always need to ensure that regions
2041 * in the block we are not writing to are zeroed. Mark the
2042 * buffer as new to ensure this.
2043 */
2044 set_buffer_new(bh);
2045 set_buffer_unwritten(bh);
2046 fallthrough;
2047 case IOMAP_MAPPED:
2048 if ((iomap->flags & IOMAP_F_NEW) ||
2049 offset >= i_size_read(inode)) {
2050 /*
2051 * This can happen if truncating the block device races
2052 * with the check in the caller as i_size updates on
2053 * block devices aren't synchronized by i_rwsem for
2054 * block devices.
2055 */
2056 if (S_ISBLK(inode->i_mode))
2057 return -EIO;
2058 set_buffer_new(bh);
2059 }
2060 bh->b_blocknr = (iomap->addr + offset - iomap->offset) >>
2061 inode->i_blkbits;
2062 set_buffer_mapped(bh);
2063 return 0;
2064 default:
2065 WARN_ON_ONCE(1);
2066 return -EIO;
2067 }
2068}
2069
2070int __block_write_begin_int(struct folio *folio, loff_t pos, unsigned len,
2071 get_block_t *get_block, const struct iomap *iomap)
2072{
2073 size_t from = offset_in_folio(folio, pos);
2074 size_t to = from + len;
2075 struct inode *inode = folio->mapping->host;
2076 size_t block_start, block_end;
2077 sector_t block;
2078 int err = 0;
2079 size_t blocksize;
2080 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
2081
2082 BUG_ON(!folio_test_locked(folio));
2083 BUG_ON(to > folio_size(folio));
2084 BUG_ON(from > to);
2085
2086 head = folio_create_buffers(folio, inode, 0);
2087 blocksize = head->b_size;
2088 block = div_u64(folio_pos(folio), blocksize);
2089
2090 for (bh = head, block_start = 0; bh != head || !block_start;
2091 block++, block_start=block_end, bh = bh->b_this_page) {
2092 block_end = block_start + blocksize;
2093 if (block_end <= from || block_start >= to) {
2094 if (folio_test_uptodate(folio)) {
2095 if (!buffer_uptodate(bh))
2096 set_buffer_uptodate(bh);
2097 }
2098 continue;
2099 }
2100 if (buffer_new(bh))
2101 clear_buffer_new(bh);
2102 if (!buffer_mapped(bh)) {
2103 WARN_ON(bh->b_size != blocksize);
2104 if (get_block)
2105 err = get_block(inode, block, bh, 1);
2106 else
2107 err = iomap_to_bh(inode, block, bh, iomap);
2108 if (err)
2109 break;
2110
2111 if (buffer_new(bh)) {
2112 clean_bdev_bh_alias(bh);
2113 if (folio_test_uptodate(folio)) {
2114 clear_buffer_new(bh);
2115 set_buffer_uptodate(bh);
2116 mark_buffer_dirty(bh);
2117 continue;
2118 }
2119 if (block_end > to || block_start < from)
2120 folio_zero_segments(folio,
2121 to, block_end,
2122 block_start, from);
2123 continue;
2124 }
2125 }
2126 if (folio_test_uptodate(folio)) {
2127 if (!buffer_uptodate(bh))
2128 set_buffer_uptodate(bh);
2129 continue;
2130 }
2131 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2132 !buffer_unwritten(bh) &&
2133 (block_start < from || block_end > to)) {
2134 bh_read_nowait(bh, 0);
2135 *wait_bh++=bh;
2136 }
2137 }
2138 /*
2139 * If we issued read requests - let them complete.
2140 */
2141 while(wait_bh > wait) {
2142 wait_on_buffer(*--wait_bh);
2143 if (!buffer_uptodate(*wait_bh))
2144 err = -EIO;
2145 }
2146 if (unlikely(err))
2147 folio_zero_new_buffers(folio, from, to);
2148 return err;
2149}
2150
2151int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2152 get_block_t *get_block)
2153{
2154 return __block_write_begin_int(page_folio(page), pos, len, get_block,
2155 NULL);
2156}
2157EXPORT_SYMBOL(__block_write_begin);
2158
2159static void __block_commit_write(struct folio *folio, size_t from, size_t to)
2160{
2161 size_t block_start, block_end;
2162 bool partial = false;
2163 unsigned blocksize;
2164 struct buffer_head *bh, *head;
2165
2166 bh = head = folio_buffers(folio);
2167 blocksize = bh->b_size;
2168
2169 block_start = 0;
2170 do {
2171 block_end = block_start + blocksize;
2172 if (block_end <= from || block_start >= to) {
2173 if (!buffer_uptodate(bh))
2174 partial = true;
2175 } else {
2176 set_buffer_uptodate(bh);
2177 mark_buffer_dirty(bh);
2178 }
2179 if (buffer_new(bh))
2180 clear_buffer_new(bh);
2181
2182 block_start = block_end;
2183 bh = bh->b_this_page;
2184 } while (bh != head);
2185
2186 /*
2187 * If this is a partial write which happened to make all buffers
2188 * uptodate then we can optimize away a bogus read_folio() for
2189 * the next read(). Here we 'discover' whether the folio went
2190 * uptodate as a result of this (potentially partial) write.
2191 */
2192 if (!partial)
2193 folio_mark_uptodate(folio);
2194}
2195
2196/*
2197 * block_write_begin takes care of the basic task of block allocation and
2198 * bringing partial write blocks uptodate first.
2199 *
2200 * The filesystem needs to handle block truncation upon failure.
2201 */
2202int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2203 struct page **pagep, get_block_t *get_block)
2204{
2205 pgoff_t index = pos >> PAGE_SHIFT;
2206 struct page *page;
2207 int status;
2208
2209 page = grab_cache_page_write_begin(mapping, index);
2210 if (!page)
2211 return -ENOMEM;
2212
2213 status = __block_write_begin(page, pos, len, get_block);
2214 if (unlikely(status)) {
2215 unlock_page(page);
2216 put_page(page);
2217 page = NULL;
2218 }
2219
2220 *pagep = page;
2221 return status;
2222}
2223EXPORT_SYMBOL(block_write_begin);
2224
2225int block_write_end(struct file *file, struct address_space *mapping,
2226 loff_t pos, unsigned len, unsigned copied,
2227 struct page *page, void *fsdata)
2228{
2229 struct folio *folio = page_folio(page);
2230 size_t start = pos - folio_pos(folio);
2231
2232 if (unlikely(copied < len)) {
2233 /*
2234 * The buffers that were written will now be uptodate, so
2235 * we don't have to worry about a read_folio reading them
2236 * and overwriting a partial write. However if we have
2237 * encountered a short write and only partially written
2238 * into a buffer, it will not be marked uptodate, so a
2239 * read_folio might come in and destroy our partial write.
2240 *
2241 * Do the simplest thing, and just treat any short write to a
2242 * non uptodate folio as a zero-length write, and force the
2243 * caller to redo the whole thing.
2244 */
2245 if (!folio_test_uptodate(folio))
2246 copied = 0;
2247
2248 folio_zero_new_buffers(folio, start+copied, start+len);
2249 }
2250 flush_dcache_folio(folio);
2251
2252 /* This could be a short (even 0-length) commit */
2253 __block_commit_write(folio, start, start + copied);
2254
2255 return copied;
2256}
2257EXPORT_SYMBOL(block_write_end);
2258
2259int generic_write_end(struct file *file, struct address_space *mapping,
2260 loff_t pos, unsigned len, unsigned copied,
2261 struct page *page, void *fsdata)
2262{
2263 struct inode *inode = mapping->host;
2264 loff_t old_size = inode->i_size;
2265 bool i_size_changed = false;
2266
2267 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2268
2269 /*
2270 * No need to use i_size_read() here, the i_size cannot change under us
2271 * because we hold i_rwsem.
2272 *
2273 * But it's important to update i_size while still holding page lock:
2274 * page writeout could otherwise come in and zero beyond i_size.
2275 */
2276 if (pos + copied > inode->i_size) {
2277 i_size_write(inode, pos + copied);
2278 i_size_changed = true;
2279 }
2280
2281 unlock_page(page);
2282 put_page(page);
2283
2284 if (old_size < pos)
2285 pagecache_isize_extended(inode, old_size, pos);
2286 /*
2287 * Don't mark the inode dirty under page lock. First, it unnecessarily
2288 * makes the holding time of page lock longer. Second, it forces lock
2289 * ordering of page lock and transaction start for journaling
2290 * filesystems.
2291 */
2292 if (i_size_changed)
2293 mark_inode_dirty(inode);
2294 return copied;
2295}
2296EXPORT_SYMBOL(generic_write_end);
2297
2298/*
2299 * block_is_partially_uptodate checks whether buffers within a folio are
2300 * uptodate or not.
2301 *
2302 * Returns true if all buffers which correspond to the specified part
2303 * of the folio are uptodate.
2304 */
2305bool block_is_partially_uptodate(struct folio *folio, size_t from, size_t count)
2306{
2307 unsigned block_start, block_end, blocksize;
2308 unsigned to;
2309 struct buffer_head *bh, *head;
2310 bool ret = true;
2311
2312 head = folio_buffers(folio);
2313 if (!head)
2314 return false;
2315 blocksize = head->b_size;
2316 to = min_t(unsigned, folio_size(folio) - from, count);
2317 to = from + to;
2318 if (from < blocksize && to > folio_size(folio) - blocksize)
2319 return false;
2320
2321 bh = head;
2322 block_start = 0;
2323 do {
2324 block_end = block_start + blocksize;
2325 if (block_end > from && block_start < to) {
2326 if (!buffer_uptodate(bh)) {
2327 ret = false;
2328 break;
2329 }
2330 if (block_end >= to)
2331 break;
2332 }
2333 block_start = block_end;
2334 bh = bh->b_this_page;
2335 } while (bh != head);
2336
2337 return ret;
2338}
2339EXPORT_SYMBOL(block_is_partially_uptodate);
2340
2341/*
2342 * Generic "read_folio" function for block devices that have the normal
2343 * get_block functionality. This is most of the block device filesystems.
2344 * Reads the folio asynchronously --- the unlock_buffer() and
2345 * set/clear_buffer_uptodate() functions propagate buffer state into the
2346 * folio once IO has completed.
2347 */
2348int block_read_full_folio(struct folio *folio, get_block_t *get_block)
2349{
2350 struct inode *inode = folio->mapping->host;
2351 sector_t iblock, lblock;
2352 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2353 size_t blocksize;
2354 int nr, i;
2355 int fully_mapped = 1;
2356 bool page_error = false;
2357 loff_t limit = i_size_read(inode);
2358
2359 /* This is needed for ext4. */
2360 if (IS_ENABLED(CONFIG_FS_VERITY) && IS_VERITY(inode))
2361 limit = inode->i_sb->s_maxbytes;
2362
2363 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
2364
2365 head = folio_create_buffers(folio, inode, 0);
2366 blocksize = head->b_size;
2367
2368 iblock = div_u64(folio_pos(folio), blocksize);
2369 lblock = div_u64(limit + blocksize - 1, blocksize);
2370 bh = head;
2371 nr = 0;
2372 i = 0;
2373
2374 do {
2375 if (buffer_uptodate(bh))
2376 continue;
2377
2378 if (!buffer_mapped(bh)) {
2379 int err = 0;
2380
2381 fully_mapped = 0;
2382 if (iblock < lblock) {
2383 WARN_ON(bh->b_size != blocksize);
2384 err = get_block(inode, iblock, bh, 0);
2385 if (err) {
2386 folio_set_error(folio);
2387 page_error = true;
2388 }
2389 }
2390 if (!buffer_mapped(bh)) {
2391 folio_zero_range(folio, i * blocksize,
2392 blocksize);
2393 if (!err)
2394 set_buffer_uptodate(bh);
2395 continue;
2396 }
2397 /*
2398 * get_block() might have updated the buffer
2399 * synchronously
2400 */
2401 if (buffer_uptodate(bh))
2402 continue;
2403 }
2404 arr[nr++] = bh;
2405 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2406
2407 if (fully_mapped)
2408 folio_set_mappedtodisk(folio);
2409
2410 if (!nr) {
2411 /*
2412 * All buffers are uptodate or get_block() returned an
2413 * error when trying to map them - we can finish the read.
2414 */
2415 folio_end_read(folio, !page_error);
2416 return 0;
2417 }
2418
2419 /* Stage two: lock the buffers */
2420 for (i = 0; i < nr; i++) {
2421 bh = arr[i];
2422 lock_buffer(bh);
2423 mark_buffer_async_read(bh);
2424 }
2425
2426 /*
2427 * Stage 3: start the IO. Check for uptodateness
2428 * inside the buffer lock in case another process reading
2429 * the underlying blockdev brought it uptodate (the sct fix).
2430 */
2431 for (i = 0; i < nr; i++) {
2432 bh = arr[i];
2433 if (buffer_uptodate(bh))
2434 end_buffer_async_read(bh, 1);
2435 else
2436 submit_bh(REQ_OP_READ, bh);
2437 }
2438 return 0;
2439}
2440EXPORT_SYMBOL(block_read_full_folio);
2441
2442/* utility function for filesystems that need to do work on expanding
2443 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2444 * deal with the hole.
2445 */
2446int generic_cont_expand_simple(struct inode *inode, loff_t size)
2447{
2448 struct address_space *mapping = inode->i_mapping;
2449 const struct address_space_operations *aops = mapping->a_ops;
2450 struct page *page;
2451 void *fsdata = NULL;
2452 int err;
2453
2454 err = inode_newsize_ok(inode, size);
2455 if (err)
2456 goto out;
2457
2458 err = aops->write_begin(NULL, mapping, size, 0, &page, &fsdata);
2459 if (err)
2460 goto out;
2461
2462 err = aops->write_end(NULL, mapping, size, 0, 0, page, fsdata);
2463 BUG_ON(err > 0);
2464
2465out:
2466 return err;
2467}
2468EXPORT_SYMBOL(generic_cont_expand_simple);
2469
2470static int cont_expand_zero(struct file *file, struct address_space *mapping,
2471 loff_t pos, loff_t *bytes)
2472{
2473 struct inode *inode = mapping->host;
2474 const struct address_space_operations *aops = mapping->a_ops;
2475 unsigned int blocksize = i_blocksize(inode);
2476 struct page *page;
2477 void *fsdata = NULL;
2478 pgoff_t index, curidx;
2479 loff_t curpos;
2480 unsigned zerofrom, offset, len;
2481 int err = 0;
2482
2483 index = pos >> PAGE_SHIFT;
2484 offset = pos & ~PAGE_MASK;
2485
2486 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2487 zerofrom = curpos & ~PAGE_MASK;
2488 if (zerofrom & (blocksize-1)) {
2489 *bytes |= (blocksize-1);
2490 (*bytes)++;
2491 }
2492 len = PAGE_SIZE - zerofrom;
2493
2494 err = aops->write_begin(file, mapping, curpos, len,
2495 &page, &fsdata);
2496 if (err)
2497 goto out;
2498 zero_user(page, zerofrom, len);
2499 err = aops->write_end(file, mapping, curpos, len, len,
2500 page, fsdata);
2501 if (err < 0)
2502 goto out;
2503 BUG_ON(err != len);
2504 err = 0;
2505
2506 balance_dirty_pages_ratelimited(mapping);
2507
2508 if (fatal_signal_pending(current)) {
2509 err = -EINTR;
2510 goto out;
2511 }
2512 }
2513
2514 /* page covers the boundary, find the boundary offset */
2515 if (index == curidx) {
2516 zerofrom = curpos & ~PAGE_MASK;
2517 /* if we will expand the thing last block will be filled */
2518 if (offset <= zerofrom) {
2519 goto out;
2520 }
2521 if (zerofrom & (blocksize-1)) {
2522 *bytes |= (blocksize-1);
2523 (*bytes)++;
2524 }
2525 len = offset - zerofrom;
2526
2527 err = aops->write_begin(file, mapping, curpos, len,
2528 &page, &fsdata);
2529 if (err)
2530 goto out;
2531 zero_user(page, zerofrom, len);
2532 err = aops->write_end(file, mapping, curpos, len, len,
2533 page, fsdata);
2534 if (err < 0)
2535 goto out;
2536 BUG_ON(err != len);
2537 err = 0;
2538 }
2539out:
2540 return err;
2541}
2542
2543/*
2544 * For moronic filesystems that do not allow holes in file.
2545 * We may have to extend the file.
2546 */
2547int cont_write_begin(struct file *file, struct address_space *mapping,
2548 loff_t pos, unsigned len,
2549 struct page **pagep, void **fsdata,
2550 get_block_t *get_block, loff_t *bytes)
2551{
2552 struct inode *inode = mapping->host;
2553 unsigned int blocksize = i_blocksize(inode);
2554 unsigned int zerofrom;
2555 int err;
2556
2557 err = cont_expand_zero(file, mapping, pos, bytes);
2558 if (err)
2559 return err;
2560
2561 zerofrom = *bytes & ~PAGE_MASK;
2562 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2563 *bytes |= (blocksize-1);
2564 (*bytes)++;
2565 }
2566
2567 return block_write_begin(mapping, pos, len, pagep, get_block);
2568}
2569EXPORT_SYMBOL(cont_write_begin);
2570
2571void block_commit_write(struct page *page, unsigned from, unsigned to)
2572{
2573 struct folio *folio = page_folio(page);
2574 __block_commit_write(folio, from, to);
2575}
2576EXPORT_SYMBOL(block_commit_write);
2577
2578/*
2579 * block_page_mkwrite() is not allowed to change the file size as it gets
2580 * called from a page fault handler when a page is first dirtied. Hence we must
2581 * be careful to check for EOF conditions here. We set the page up correctly
2582 * for a written page which means we get ENOSPC checking when writing into
2583 * holes and correct delalloc and unwritten extent mapping on filesystems that
2584 * support these features.
2585 *
2586 * We are not allowed to take the i_mutex here so we have to play games to
2587 * protect against truncate races as the page could now be beyond EOF. Because
2588 * truncate writes the inode size before removing pages, once we have the
2589 * page lock we can determine safely if the page is beyond EOF. If it is not
2590 * beyond EOF, then the page is guaranteed safe against truncation until we
2591 * unlock the page.
2592 *
2593 * Direct callers of this function should protect against filesystem freezing
2594 * using sb_start_pagefault() - sb_end_pagefault() functions.
2595 */
2596int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2597 get_block_t get_block)
2598{
2599 struct folio *folio = page_folio(vmf->page);
2600 struct inode *inode = file_inode(vma->vm_file);
2601 unsigned long end;
2602 loff_t size;
2603 int ret;
2604
2605 folio_lock(folio);
2606 size = i_size_read(inode);
2607 if ((folio->mapping != inode->i_mapping) ||
2608 (folio_pos(folio) >= size)) {
2609 /* We overload EFAULT to mean page got truncated */
2610 ret = -EFAULT;
2611 goto out_unlock;
2612 }
2613
2614 end = folio_size(folio);
2615 /* folio is wholly or partially inside EOF */
2616 if (folio_pos(folio) + end > size)
2617 end = size - folio_pos(folio);
2618
2619 ret = __block_write_begin_int(folio, 0, end, get_block, NULL);
2620 if (unlikely(ret))
2621 goto out_unlock;
2622
2623 __block_commit_write(folio, 0, end);
2624
2625 folio_mark_dirty(folio);
2626 folio_wait_stable(folio);
2627 return 0;
2628out_unlock:
2629 folio_unlock(folio);
2630 return ret;
2631}
2632EXPORT_SYMBOL(block_page_mkwrite);
2633
2634int block_truncate_page(struct address_space *mapping,
2635 loff_t from, get_block_t *get_block)
2636{
2637 pgoff_t index = from >> PAGE_SHIFT;
2638 unsigned blocksize;
2639 sector_t iblock;
2640 size_t offset, length, pos;
2641 struct inode *inode = mapping->host;
2642 struct folio *folio;
2643 struct buffer_head *bh;
2644 int err = 0;
2645
2646 blocksize = i_blocksize(inode);
2647 length = from & (blocksize - 1);
2648
2649 /* Block boundary? Nothing to do */
2650 if (!length)
2651 return 0;
2652
2653 length = blocksize - length;
2654 iblock = ((loff_t)index * PAGE_SIZE) >> inode->i_blkbits;
2655
2656 folio = filemap_grab_folio(mapping, index);
2657 if (IS_ERR(folio))
2658 return PTR_ERR(folio);
2659
2660 bh = folio_buffers(folio);
2661 if (!bh)
2662 bh = create_empty_buffers(folio, blocksize, 0);
2663
2664 /* Find the buffer that contains "offset" */
2665 offset = offset_in_folio(folio, from);
2666 pos = blocksize;
2667 while (offset >= pos) {
2668 bh = bh->b_this_page;
2669 iblock++;
2670 pos += blocksize;
2671 }
2672
2673 if (!buffer_mapped(bh)) {
2674 WARN_ON(bh->b_size != blocksize);
2675 err = get_block(inode, iblock, bh, 0);
2676 if (err)
2677 goto unlock;
2678 /* unmapped? It's a hole - nothing to do */
2679 if (!buffer_mapped(bh))
2680 goto unlock;
2681 }
2682
2683 /* Ok, it's mapped. Make sure it's up-to-date */
2684 if (folio_test_uptodate(folio))
2685 set_buffer_uptodate(bh);
2686
2687 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2688 err = bh_read(bh, 0);
2689 /* Uhhuh. Read error. Complain and punt. */
2690 if (err < 0)
2691 goto unlock;
2692 }
2693
2694 folio_zero_range(folio, offset, length);
2695 mark_buffer_dirty(bh);
2696
2697unlock:
2698 folio_unlock(folio);
2699 folio_put(folio);
2700
2701 return err;
2702}
2703EXPORT_SYMBOL(block_truncate_page);
2704
2705/*
2706 * The generic ->writepage function for buffer-backed address_spaces
2707 */
2708int block_write_full_folio(struct folio *folio, struct writeback_control *wbc,
2709 void *get_block)
2710{
2711 struct inode * const inode = folio->mapping->host;
2712 loff_t i_size = i_size_read(inode);
2713
2714 /* Is the folio fully inside i_size? */
2715 if (folio_pos(folio) + folio_size(folio) <= i_size)
2716 return __block_write_full_folio(inode, folio, get_block, wbc);
2717
2718 /* Is the folio fully outside i_size? (truncate in progress) */
2719 if (folio_pos(folio) >= i_size) {
2720 folio_unlock(folio);
2721 return 0; /* don't care */
2722 }
2723
2724 /*
2725 * The folio straddles i_size. It must be zeroed out on each and every
2726 * writepage invocation because it may be mmapped. "A file is mapped
2727 * in multiples of the page size. For a file that is not a multiple of
2728 * the page size, the remaining memory is zeroed when mapped, and
2729 * writes to that region are not written out to the file."
2730 */
2731 folio_zero_segment(folio, offset_in_folio(folio, i_size),
2732 folio_size(folio));
2733 return __block_write_full_folio(inode, folio, get_block, wbc);
2734}
2735
2736sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2737 get_block_t *get_block)
2738{
2739 struct inode *inode = mapping->host;
2740 struct buffer_head tmp = {
2741 .b_size = i_blocksize(inode),
2742 };
2743
2744 get_block(inode, block, &tmp, 0);
2745 return tmp.b_blocknr;
2746}
2747EXPORT_SYMBOL(generic_block_bmap);
2748
2749static void end_bio_bh_io_sync(struct bio *bio)
2750{
2751 struct buffer_head *bh = bio->bi_private;
2752
2753 if (unlikely(bio_flagged(bio, BIO_QUIET)))
2754 set_bit(BH_Quiet, &bh->b_state);
2755
2756 bh->b_end_io(bh, !bio->bi_status);
2757 bio_put(bio);
2758}
2759
2760static void submit_bh_wbc(blk_opf_t opf, struct buffer_head *bh,
2761 enum rw_hint write_hint,
2762 struct writeback_control *wbc)
2763{
2764 const enum req_op op = opf & REQ_OP_MASK;
2765 struct bio *bio;
2766
2767 BUG_ON(!buffer_locked(bh));
2768 BUG_ON(!buffer_mapped(bh));
2769 BUG_ON(!bh->b_end_io);
2770 BUG_ON(buffer_delay(bh));
2771 BUG_ON(buffer_unwritten(bh));
2772
2773 /*
2774 * Only clear out a write error when rewriting
2775 */
2776 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
2777 clear_buffer_write_io_error(bh);
2778
2779 if (buffer_meta(bh))
2780 opf |= REQ_META;
2781 if (buffer_prio(bh))
2782 opf |= REQ_PRIO;
2783
2784 bio = bio_alloc(bh->b_bdev, 1, opf, GFP_NOIO);
2785
2786 fscrypt_set_bio_crypt_ctx_bh(bio, bh, GFP_NOIO);
2787
2788 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2789 bio->bi_write_hint = write_hint;
2790
2791 __bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
2792
2793 bio->bi_end_io = end_bio_bh_io_sync;
2794 bio->bi_private = bh;
2795
2796 /* Take care of bh's that straddle the end of the device */
2797 guard_bio_eod(bio);
2798
2799 if (wbc) {
2800 wbc_init_bio(wbc, bio);
2801 wbc_account_cgroup_owner(wbc, bh->b_page, bh->b_size);
2802 }
2803
2804 submit_bio(bio);
2805}
2806
2807void submit_bh(blk_opf_t opf, struct buffer_head *bh)
2808{
2809 submit_bh_wbc(opf, bh, WRITE_LIFE_NOT_SET, NULL);
2810}
2811EXPORT_SYMBOL(submit_bh);
2812
2813void write_dirty_buffer(struct buffer_head *bh, blk_opf_t op_flags)
2814{
2815 lock_buffer(bh);
2816 if (!test_clear_buffer_dirty(bh)) {
2817 unlock_buffer(bh);
2818 return;
2819 }
2820 bh->b_end_io = end_buffer_write_sync;
2821 get_bh(bh);
2822 submit_bh(REQ_OP_WRITE | op_flags, bh);
2823}
2824EXPORT_SYMBOL(write_dirty_buffer);
2825
2826/*
2827 * For a data-integrity writeout, we need to wait upon any in-progress I/O
2828 * and then start new I/O and then wait upon it. The caller must have a ref on
2829 * the buffer_head.
2830 */
2831int __sync_dirty_buffer(struct buffer_head *bh, blk_opf_t op_flags)
2832{
2833 WARN_ON(atomic_read(&bh->b_count) < 1);
2834 lock_buffer(bh);
2835 if (test_clear_buffer_dirty(bh)) {
2836 /*
2837 * The bh should be mapped, but it might not be if the
2838 * device was hot-removed. Not much we can do but fail the I/O.
2839 */
2840 if (!buffer_mapped(bh)) {
2841 unlock_buffer(bh);
2842 return -EIO;
2843 }
2844
2845 get_bh(bh);
2846 bh->b_end_io = end_buffer_write_sync;
2847 submit_bh(REQ_OP_WRITE | op_flags, bh);
2848 wait_on_buffer(bh);
2849 if (!buffer_uptodate(bh))
2850 return -EIO;
2851 } else {
2852 unlock_buffer(bh);
2853 }
2854 return 0;
2855}
2856EXPORT_SYMBOL(__sync_dirty_buffer);
2857
2858int sync_dirty_buffer(struct buffer_head *bh)
2859{
2860 return __sync_dirty_buffer(bh, REQ_SYNC);
2861}
2862EXPORT_SYMBOL(sync_dirty_buffer);
2863
2864/*
2865 * try_to_free_buffers() checks if all the buffers on this particular folio
2866 * are unused, and releases them if so.
2867 *
2868 * Exclusion against try_to_free_buffers may be obtained by either
2869 * locking the folio or by holding its mapping's i_private_lock.
2870 *
2871 * If the folio is dirty but all the buffers are clean then we need to
2872 * be sure to mark the folio clean as well. This is because the folio
2873 * may be against a block device, and a later reattachment of buffers
2874 * to a dirty folio will set *all* buffers dirty. Which would corrupt
2875 * filesystem data on the same device.
2876 *
2877 * The same applies to regular filesystem folios: if all the buffers are
2878 * clean then we set the folio clean and proceed. To do that, we require
2879 * total exclusion from block_dirty_folio(). That is obtained with
2880 * i_private_lock.
2881 *
2882 * try_to_free_buffers() is non-blocking.
2883 */
2884static inline int buffer_busy(struct buffer_head *bh)
2885{
2886 return atomic_read(&bh->b_count) |
2887 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
2888}
2889
2890static bool
2891drop_buffers(struct folio *folio, struct buffer_head **buffers_to_free)
2892{
2893 struct buffer_head *head = folio_buffers(folio);
2894 struct buffer_head *bh;
2895
2896 bh = head;
2897 do {
2898 if (buffer_busy(bh))
2899 goto failed;
2900 bh = bh->b_this_page;
2901 } while (bh != head);
2902
2903 do {
2904 struct buffer_head *next = bh->b_this_page;
2905
2906 if (bh->b_assoc_map)
2907 __remove_assoc_queue(bh);
2908 bh = next;
2909 } while (bh != head);
2910 *buffers_to_free = head;
2911 folio_detach_private(folio);
2912 return true;
2913failed:
2914 return false;
2915}
2916
2917bool try_to_free_buffers(struct folio *folio)
2918{
2919 struct address_space * const mapping = folio->mapping;
2920 struct buffer_head *buffers_to_free = NULL;
2921 bool ret = 0;
2922
2923 BUG_ON(!folio_test_locked(folio));
2924 if (folio_test_writeback(folio))
2925 return false;
2926
2927 if (mapping == NULL) { /* can this still happen? */
2928 ret = drop_buffers(folio, &buffers_to_free);
2929 goto out;
2930 }
2931
2932 spin_lock(&mapping->i_private_lock);
2933 ret = drop_buffers(folio, &buffers_to_free);
2934
2935 /*
2936 * If the filesystem writes its buffers by hand (eg ext3)
2937 * then we can have clean buffers against a dirty folio. We
2938 * clean the folio here; otherwise the VM will never notice
2939 * that the filesystem did any IO at all.
2940 *
2941 * Also, during truncate, discard_buffer will have marked all
2942 * the folio's buffers clean. We discover that here and clean
2943 * the folio also.
2944 *
2945 * i_private_lock must be held over this entire operation in order
2946 * to synchronise against block_dirty_folio and prevent the
2947 * dirty bit from being lost.
2948 */
2949 if (ret)
2950 folio_cancel_dirty(folio);
2951 spin_unlock(&mapping->i_private_lock);
2952out:
2953 if (buffers_to_free) {
2954 struct buffer_head *bh = buffers_to_free;
2955
2956 do {
2957 struct buffer_head *next = bh->b_this_page;
2958 free_buffer_head(bh);
2959 bh = next;
2960 } while (bh != buffers_to_free);
2961 }
2962 return ret;
2963}
2964EXPORT_SYMBOL(try_to_free_buffers);
2965
2966/*
2967 * Buffer-head allocation
2968 */
2969static struct kmem_cache *bh_cachep __ro_after_init;
2970
2971/*
2972 * Once the number of bh's in the machine exceeds this level, we start
2973 * stripping them in writeback.
2974 */
2975static unsigned long max_buffer_heads __ro_after_init;
2976
2977int buffer_heads_over_limit;
2978
2979struct bh_accounting {
2980 int nr; /* Number of live bh's */
2981 int ratelimit; /* Limit cacheline bouncing */
2982};
2983
2984static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
2985
2986static void recalc_bh_state(void)
2987{
2988 int i;
2989 int tot = 0;
2990
2991 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
2992 return;
2993 __this_cpu_write(bh_accounting.ratelimit, 0);
2994 for_each_online_cpu(i)
2995 tot += per_cpu(bh_accounting, i).nr;
2996 buffer_heads_over_limit = (tot > max_buffer_heads);
2997}
2998
2999struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3000{
3001 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3002 if (ret) {
3003 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3004 spin_lock_init(&ret->b_uptodate_lock);
3005 preempt_disable();
3006 __this_cpu_inc(bh_accounting.nr);
3007 recalc_bh_state();
3008 preempt_enable();
3009 }
3010 return ret;
3011}
3012EXPORT_SYMBOL(alloc_buffer_head);
3013
3014void free_buffer_head(struct buffer_head *bh)
3015{
3016 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3017 kmem_cache_free(bh_cachep, bh);
3018 preempt_disable();
3019 __this_cpu_dec(bh_accounting.nr);
3020 recalc_bh_state();
3021 preempt_enable();
3022}
3023EXPORT_SYMBOL(free_buffer_head);
3024
3025static int buffer_exit_cpu_dead(unsigned int cpu)
3026{
3027 int i;
3028 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3029
3030 for (i = 0; i < BH_LRU_SIZE; i++) {
3031 brelse(b->bhs[i]);
3032 b->bhs[i] = NULL;
3033 }
3034 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3035 per_cpu(bh_accounting, cpu).nr = 0;
3036 return 0;
3037}
3038
3039/**
3040 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3041 * @bh: struct buffer_head
3042 *
3043 * Return true if the buffer is up-to-date and false,
3044 * with the buffer locked, if not.
3045 */
3046int bh_uptodate_or_lock(struct buffer_head *bh)
3047{
3048 if (!buffer_uptodate(bh)) {
3049 lock_buffer(bh);
3050 if (!buffer_uptodate(bh))
3051 return 0;
3052 unlock_buffer(bh);
3053 }
3054 return 1;
3055}
3056EXPORT_SYMBOL(bh_uptodate_or_lock);
3057
3058/**
3059 * __bh_read - Submit read for a locked buffer
3060 * @bh: struct buffer_head
3061 * @op_flags: appending REQ_OP_* flags besides REQ_OP_READ
3062 * @wait: wait until reading finish
3063 *
3064 * Returns zero on success or don't wait, and -EIO on error.
3065 */
3066int __bh_read(struct buffer_head *bh, blk_opf_t op_flags, bool wait)
3067{
3068 int ret = 0;
3069
3070 BUG_ON(!buffer_locked(bh));
3071
3072 get_bh(bh);
3073 bh->b_end_io = end_buffer_read_sync;
3074 submit_bh(REQ_OP_READ | op_flags, bh);
3075 if (wait) {
3076 wait_on_buffer(bh);
3077 if (!buffer_uptodate(bh))
3078 ret = -EIO;
3079 }
3080 return ret;
3081}
3082EXPORT_SYMBOL(__bh_read);
3083
3084/**
3085 * __bh_read_batch - Submit read for a batch of unlocked buffers
3086 * @nr: entry number of the buffer batch
3087 * @bhs: a batch of struct buffer_head
3088 * @op_flags: appending REQ_OP_* flags besides REQ_OP_READ
3089 * @force_lock: force to get a lock on the buffer if set, otherwise drops any
3090 * buffer that cannot lock.
3091 *
3092 * Returns zero on success or don't wait, and -EIO on error.
3093 */
3094void __bh_read_batch(int nr, struct buffer_head *bhs[],
3095 blk_opf_t op_flags, bool force_lock)
3096{
3097 int i;
3098
3099 for (i = 0; i < nr; i++) {
3100 struct buffer_head *bh = bhs[i];
3101
3102 if (buffer_uptodate(bh))
3103 continue;
3104
3105 if (force_lock)
3106 lock_buffer(bh);
3107 else
3108 if (!trylock_buffer(bh))
3109 continue;
3110
3111 if (buffer_uptodate(bh)) {
3112 unlock_buffer(bh);
3113 continue;
3114 }
3115
3116 bh->b_end_io = end_buffer_read_sync;
3117 get_bh(bh);
3118 submit_bh(REQ_OP_READ | op_flags, bh);
3119 }
3120}
3121EXPORT_SYMBOL(__bh_read_batch);
3122
3123void __init buffer_init(void)
3124{
3125 unsigned long nrpages;
3126 int ret;
3127
3128 bh_cachep = KMEM_CACHE(buffer_head,
3129 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC);
3130 /*
3131 * Limit the bh occupancy to 10% of ZONE_NORMAL
3132 */
3133 nrpages = (nr_free_buffer_pages() * 10) / 100;
3134 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3135 ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3136 NULL, buffer_exit_cpu_dead);
3137 WARN_ON(ret < 0);
3138}