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