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