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