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