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