Merge tag 'xtensa-20190510' of git://github.com/jcmvbkbc/linux-xtensa
[sfrench/cifs-2.6.git] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/sched/signal.h>
23 #include <linux/syscalls.h>
24 #include <linux/fs.h>
25 #include <linux/iomap.h>
26 #include <linux/mm.h>
27 #include <linux/percpu.h>
28 #include <linux/slab.h>
29 #include <linux/capability.h>
30 #include <linux/blkdev.h>
31 #include <linux/file.h>
32 #include <linux/quotaops.h>
33 #include <linux/highmem.h>
34 #include <linux/export.h>
35 #include <linux/backing-dev.h>
36 #include <linux/writeback.h>
37 #include <linux/hash.h>
38 #include <linux/suspend.h>
39 #include <linux/buffer_head.h>
40 #include <linux/task_io_accounting_ops.h>
41 #include <linux/bio.h>
42 #include <linux/cpu.h>
43 #include <linux/bitops.h>
44 #include <linux/mpage.h>
45 #include <linux/bit_spinlock.h>
46 #include <linux/pagevec.h>
47 #include <linux/sched/mm.h>
48 #include <trace/events/block.h>
49
50 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
51 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
52                          enum rw_hint hint, struct writeback_control *wbc);
53
54 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
55
56 inline void touch_buffer(struct buffer_head *bh)
57 {
58         trace_block_touch_buffer(bh);
59         mark_page_accessed(bh->b_page);
60 }
61 EXPORT_SYMBOL(touch_buffer);
62
63 void __lock_buffer(struct buffer_head *bh)
64 {
65         wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
66 }
67 EXPORT_SYMBOL(__lock_buffer);
68
69 void unlock_buffer(struct buffer_head *bh)
70 {
71         clear_bit_unlock(BH_Lock, &bh->b_state);
72         smp_mb__after_atomic();
73         wake_up_bit(&bh->b_state, BH_Lock);
74 }
75 EXPORT_SYMBOL(unlock_buffer);
76
77 /*
78  * Returns if the page has dirty or writeback buffers. If all the buffers
79  * are unlocked and clean then the PageDirty information is stale. If
80  * any of the pages are locked, it is assumed they are locked for IO.
81  */
82 void buffer_check_dirty_writeback(struct page *page,
83                                      bool *dirty, bool *writeback)
84 {
85         struct buffer_head *head, *bh;
86         *dirty = false;
87         *writeback = false;
88
89         BUG_ON(!PageLocked(page));
90
91         if (!page_has_buffers(page))
92                 return;
93
94         if (PageWriteback(page))
95                 *writeback = true;
96
97         head = page_buffers(page);
98         bh = head;
99         do {
100                 if (buffer_locked(bh))
101                         *writeback = true;
102
103                 if (buffer_dirty(bh))
104                         *dirty = true;
105
106                 bh = bh->b_this_page;
107         } while (bh != head);
108 }
109 EXPORT_SYMBOL(buffer_check_dirty_writeback);
110
111 /*
112  * Block until a buffer comes unlocked.  This doesn't stop it
113  * from becoming locked again - you have to lock it yourself
114  * if you want to preserve its state.
115  */
116 void __wait_on_buffer(struct buffer_head * bh)
117 {
118         wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
119 }
120 EXPORT_SYMBOL(__wait_on_buffer);
121
122 static void
123 __clear_page_buffers(struct page *page)
124 {
125         ClearPagePrivate(page);
126         set_page_private(page, 0);
127         put_page(page);
128 }
129
130 static void buffer_io_error(struct buffer_head *bh, char *msg)
131 {
132         if (!test_bit(BH_Quiet, &bh->b_state))
133                 printk_ratelimited(KERN_ERR
134                         "Buffer I/O error on dev %pg, logical block %llu%s\n",
135                         bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
136 }
137
138 /*
139  * End-of-IO handler helper function which does not touch the bh after
140  * unlocking it.
141  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
142  * a race there is benign: unlock_buffer() only use the bh's address for
143  * hashing after unlocking the buffer, so it doesn't actually touch the bh
144  * itself.
145  */
146 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
147 {
148         if (uptodate) {
149                 set_buffer_uptodate(bh);
150         } else {
151                 /* This happens, due to failed read-ahead attempts. */
152                 clear_buffer_uptodate(bh);
153         }
154         unlock_buffer(bh);
155 }
156
157 /*
158  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
159  * unlock the buffer. This is what ll_rw_block uses too.
160  */
161 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
162 {
163         __end_buffer_read_notouch(bh, uptodate);
164         put_bh(bh);
165 }
166 EXPORT_SYMBOL(end_buffer_read_sync);
167
168 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
169 {
170         if (uptodate) {
171                 set_buffer_uptodate(bh);
172         } else {
173                 buffer_io_error(bh, ", lost sync page write");
174                 mark_buffer_write_io_error(bh);
175                 clear_buffer_uptodate(bh);
176         }
177         unlock_buffer(bh);
178         put_bh(bh);
179 }
180 EXPORT_SYMBOL(end_buffer_write_sync);
181
182 /*
183  * Various filesystems appear to want __find_get_block to be non-blocking.
184  * But it's the page lock which protects the buffers.  To get around this,
185  * we get exclusion from try_to_free_buffers with the blockdev mapping's
186  * private_lock.
187  *
188  * Hack idea: for the blockdev mapping, private_lock contention
189  * may be quite high.  This code could TryLock the page, and if that
190  * succeeds, there is no need to take private_lock.
191  */
192 static struct buffer_head *
193 __find_get_block_slow(struct block_device *bdev, sector_t block)
194 {
195         struct inode *bd_inode = bdev->bd_inode;
196         struct address_space *bd_mapping = bd_inode->i_mapping;
197         struct buffer_head *ret = NULL;
198         pgoff_t index;
199         struct buffer_head *bh;
200         struct buffer_head *head;
201         struct page *page;
202         int all_mapped = 1;
203         static DEFINE_RATELIMIT_STATE(last_warned, HZ, 1);
204
205         index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
206         page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
207         if (!page)
208                 goto out;
209
210         spin_lock(&bd_mapping->private_lock);
211         if (!page_has_buffers(page))
212                 goto out_unlock;
213         head = page_buffers(page);
214         bh = head;
215         do {
216                 if (!buffer_mapped(bh))
217                         all_mapped = 0;
218                 else if (bh->b_blocknr == block) {
219                         ret = bh;
220                         get_bh(bh);
221                         goto out_unlock;
222                 }
223                 bh = bh->b_this_page;
224         } while (bh != head);
225
226         /* we might be here because some of the buffers on this page are
227          * not mapped.  This is due to various races between
228          * file io on the block device and getblk.  It gets dealt with
229          * elsewhere, don't buffer_error if we had some unmapped buffers
230          */
231         ratelimit_set_flags(&last_warned, RATELIMIT_MSG_ON_RELEASE);
232         if (all_mapped && __ratelimit(&last_warned)) {
233                 printk("__find_get_block_slow() failed. block=%llu, "
234                        "b_blocknr=%llu, b_state=0x%08lx, b_size=%zu, "
235                        "device %pg blocksize: %d\n",
236                        (unsigned long long)block,
237                        (unsigned long long)bh->b_blocknr,
238                        bh->b_state, bh->b_size, bdev,
239                        1 << bd_inode->i_blkbits);
240         }
241 out_unlock:
242         spin_unlock(&bd_mapping->private_lock);
243         put_page(page);
244 out:
245         return ret;
246 }
247
248 /*
249  * I/O completion handler for block_read_full_page() - pages
250  * which come unlocked at the end of I/O.
251  */
252 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
253 {
254         unsigned long flags;
255         struct buffer_head *first;
256         struct buffer_head *tmp;
257         struct page *page;
258         int page_uptodate = 1;
259
260         BUG_ON(!buffer_async_read(bh));
261
262         page = bh->b_page;
263         if (uptodate) {
264                 set_buffer_uptodate(bh);
265         } else {
266                 clear_buffer_uptodate(bh);
267                 buffer_io_error(bh, ", async page read");
268                 SetPageError(page);
269         }
270
271         /*
272          * Be _very_ careful from here on. Bad things can happen if
273          * two buffer heads end IO at almost the same time and both
274          * decide that the page is now completely done.
275          */
276         first = page_buffers(page);
277         local_irq_save(flags);
278         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
279         clear_buffer_async_read(bh);
280         unlock_buffer(bh);
281         tmp = bh;
282         do {
283                 if (!buffer_uptodate(tmp))
284                         page_uptodate = 0;
285                 if (buffer_async_read(tmp)) {
286                         BUG_ON(!buffer_locked(tmp));
287                         goto still_busy;
288                 }
289                 tmp = tmp->b_this_page;
290         } while (tmp != bh);
291         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
292         local_irq_restore(flags);
293
294         /*
295          * If none of the buffers had errors and they are all
296          * uptodate then we can set the page uptodate.
297          */
298         if (page_uptodate && !PageError(page))
299                 SetPageUptodate(page);
300         unlock_page(page);
301         return;
302
303 still_busy:
304         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
305         local_irq_restore(flags);
306         return;
307 }
308
309 /*
310  * Completion handler for block_write_full_page() - pages which are unlocked
311  * during I/O, and which have PageWriteback cleared upon I/O completion.
312  */
313 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
314 {
315         unsigned long flags;
316         struct buffer_head *first;
317         struct buffer_head *tmp;
318         struct page *page;
319
320         BUG_ON(!buffer_async_write(bh));
321
322         page = bh->b_page;
323         if (uptodate) {
324                 set_buffer_uptodate(bh);
325         } else {
326                 buffer_io_error(bh, ", lost async page write");
327                 mark_buffer_write_io_error(bh);
328                 clear_buffer_uptodate(bh);
329                 SetPageError(page);
330         }
331
332         first = page_buffers(page);
333         local_irq_save(flags);
334         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
335
336         clear_buffer_async_write(bh);
337         unlock_buffer(bh);
338         tmp = bh->b_this_page;
339         while (tmp != bh) {
340                 if (buffer_async_write(tmp)) {
341                         BUG_ON(!buffer_locked(tmp));
342                         goto still_busy;
343                 }
344                 tmp = tmp->b_this_page;
345         }
346         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
347         local_irq_restore(flags);
348         end_page_writeback(page);
349         return;
350
351 still_busy:
352         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
353         local_irq_restore(flags);
354         return;
355 }
356 EXPORT_SYMBOL(end_buffer_async_write);
357
358 /*
359  * If a page's buffers are under async readin (end_buffer_async_read
360  * completion) then there is a possibility that another thread of
361  * control could lock one of the buffers after it has completed
362  * but while some of the other buffers have not completed.  This
363  * locked buffer would confuse end_buffer_async_read() into not unlocking
364  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
365  * that this buffer is not under async I/O.
366  *
367  * The page comes unlocked when it has no locked buffer_async buffers
368  * left.
369  *
370  * PageLocked prevents anyone starting new async I/O reads any of
371  * the buffers.
372  *
373  * PageWriteback is used to prevent simultaneous writeout of the same
374  * page.
375  *
376  * PageLocked prevents anyone from starting writeback of a page which is
377  * under read I/O (PageWriteback is only ever set against a locked page).
378  */
379 static void mark_buffer_async_read(struct buffer_head *bh)
380 {
381         bh->b_end_io = end_buffer_async_read;
382         set_buffer_async_read(bh);
383 }
384
385 static void mark_buffer_async_write_endio(struct buffer_head *bh,
386                                           bh_end_io_t *handler)
387 {
388         bh->b_end_io = handler;
389         set_buffer_async_write(bh);
390 }
391
392 void mark_buffer_async_write(struct buffer_head *bh)
393 {
394         mark_buffer_async_write_endio(bh, end_buffer_async_write);
395 }
396 EXPORT_SYMBOL(mark_buffer_async_write);
397
398
399 /*
400  * fs/buffer.c contains helper functions for buffer-backed address space's
401  * fsync functions.  A common requirement for buffer-based filesystems is
402  * that certain data from the backing blockdev needs to be written out for
403  * a successful fsync().  For example, ext2 indirect blocks need to be
404  * written back and waited upon before fsync() returns.
405  *
406  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
407  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
408  * management of a list of dependent buffers at ->i_mapping->private_list.
409  *
410  * Locking is a little subtle: try_to_free_buffers() will remove buffers
411  * from their controlling inode's queue when they are being freed.  But
412  * try_to_free_buffers() will be operating against the *blockdev* mapping
413  * at the time, not against the S_ISREG file which depends on those buffers.
414  * So the locking for private_list is via the private_lock in the address_space
415  * which backs the buffers.  Which is different from the address_space 
416  * against which the buffers are listed.  So for a particular address_space,
417  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
418  * mapping->private_list will always be protected by the backing blockdev's
419  * ->private_lock.
420  *
421  * Which introduces a requirement: all buffers on an address_space's
422  * ->private_list must be from the same address_space: the blockdev's.
423  *
424  * address_spaces which do not place buffers at ->private_list via these
425  * utility functions are free to use private_lock and private_list for
426  * whatever they want.  The only requirement is that list_empty(private_list)
427  * be true at clear_inode() time.
428  *
429  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
430  * filesystems should do that.  invalidate_inode_buffers() should just go
431  * BUG_ON(!list_empty).
432  *
433  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
434  * take an address_space, not an inode.  And it should be called
435  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
436  * queued up.
437  *
438  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
439  * list if it is already on a list.  Because if the buffer is on a list,
440  * it *must* already be on the right one.  If not, the filesystem is being
441  * silly.  This will save a ton of locking.  But first we have to ensure
442  * that buffers are taken *off* the old inode's list when they are freed
443  * (presumably in truncate).  That requires careful auditing of all
444  * filesystems (do it inside bforget()).  It could also be done by bringing
445  * b_inode back.
446  */
447
448 /*
449  * The buffer's backing address_space's private_lock must be held
450  */
451 static void __remove_assoc_queue(struct buffer_head *bh)
452 {
453         list_del_init(&bh->b_assoc_buffers);
454         WARN_ON(!bh->b_assoc_map);
455         bh->b_assoc_map = NULL;
456 }
457
458 int inode_has_buffers(struct inode *inode)
459 {
460         return !list_empty(&inode->i_data.private_list);
461 }
462
463 /*
464  * osync is designed to support O_SYNC io.  It waits synchronously for
465  * all already-submitted IO to complete, but does not queue any new
466  * writes to the disk.
467  *
468  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
469  * you dirty the buffers, and then use osync_inode_buffers to wait for
470  * completion.  Any other dirty buffers which are not yet queued for
471  * write will not be flushed to disk by the osync.
472  */
473 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
474 {
475         struct buffer_head *bh;
476         struct list_head *p;
477         int err = 0;
478
479         spin_lock(lock);
480 repeat:
481         list_for_each_prev(p, list) {
482                 bh = BH_ENTRY(p);
483                 if (buffer_locked(bh)) {
484                         get_bh(bh);
485                         spin_unlock(lock);
486                         wait_on_buffer(bh);
487                         if (!buffer_uptodate(bh))
488                                 err = -EIO;
489                         brelse(bh);
490                         spin_lock(lock);
491                         goto repeat;
492                 }
493         }
494         spin_unlock(lock);
495         return err;
496 }
497
498 void emergency_thaw_bdev(struct super_block *sb)
499 {
500         while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
501                 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
502 }
503
504 /**
505  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
506  * @mapping: the mapping which wants those buffers written
507  *
508  * Starts I/O against the buffers at mapping->private_list, and waits upon
509  * that I/O.
510  *
511  * Basically, this is a convenience function for fsync().
512  * @mapping is a file or directory which needs those buffers to be written for
513  * a successful fsync().
514  */
515 int sync_mapping_buffers(struct address_space *mapping)
516 {
517         struct address_space *buffer_mapping = mapping->private_data;
518
519         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
520                 return 0;
521
522         return fsync_buffers_list(&buffer_mapping->private_lock,
523                                         &mapping->private_list);
524 }
525 EXPORT_SYMBOL(sync_mapping_buffers);
526
527 /*
528  * Called when we've recently written block `bblock', and it is known that
529  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
530  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
531  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
532  */
533 void write_boundary_block(struct block_device *bdev,
534                         sector_t bblock, unsigned blocksize)
535 {
536         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
537         if (bh) {
538                 if (buffer_dirty(bh))
539                         ll_rw_block(REQ_OP_WRITE, 0, 1, &bh);
540                 put_bh(bh);
541         }
542 }
543
544 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
545 {
546         struct address_space *mapping = inode->i_mapping;
547         struct address_space *buffer_mapping = bh->b_page->mapping;
548
549         mark_buffer_dirty(bh);
550         if (!mapping->private_data) {
551                 mapping->private_data = buffer_mapping;
552         } else {
553                 BUG_ON(mapping->private_data != buffer_mapping);
554         }
555         if (!bh->b_assoc_map) {
556                 spin_lock(&buffer_mapping->private_lock);
557                 list_move_tail(&bh->b_assoc_buffers,
558                                 &mapping->private_list);
559                 bh->b_assoc_map = mapping;
560                 spin_unlock(&buffer_mapping->private_lock);
561         }
562 }
563 EXPORT_SYMBOL(mark_buffer_dirty_inode);
564
565 /*
566  * Mark the page dirty, and set it dirty in the page cache, and mark the inode
567  * dirty.
568  *
569  * If warn is true, then emit a warning if the page is not uptodate and has
570  * not been truncated.
571  *
572  * The caller must hold lock_page_memcg().
573  */
574 void __set_page_dirty(struct page *page, struct address_space *mapping,
575                              int warn)
576 {
577         unsigned long flags;
578
579         xa_lock_irqsave(&mapping->i_pages, flags);
580         if (page->mapping) {    /* Race with truncate? */
581                 WARN_ON_ONCE(warn && !PageUptodate(page));
582                 account_page_dirtied(page, mapping);
583                 __xa_set_mark(&mapping->i_pages, page_index(page),
584                                 PAGECACHE_TAG_DIRTY);
585         }
586         xa_unlock_irqrestore(&mapping->i_pages, flags);
587 }
588 EXPORT_SYMBOL_GPL(__set_page_dirty);
589
590 /*
591  * Add a page to the dirty page list.
592  *
593  * It is a sad fact of life that this function is called from several places
594  * deeply under spinlocking.  It may not sleep.
595  *
596  * If the page has buffers, the uptodate buffers are set dirty, to preserve
597  * dirty-state coherency between the page and the buffers.  It the page does
598  * not have buffers then when they are later attached they will all be set
599  * dirty.
600  *
601  * The buffers are dirtied before the page is dirtied.  There's a small race
602  * window in which a writepage caller may see the page cleanness but not the
603  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
604  * before the buffers, a concurrent writepage caller could clear the page dirty
605  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
606  * page on the dirty page list.
607  *
608  * We use private_lock to lock against try_to_free_buffers while using the
609  * page's buffer list.  Also use this to protect against clean buffers being
610  * added to the page after it was set dirty.
611  *
612  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
613  * address_space though.
614  */
615 int __set_page_dirty_buffers(struct page *page)
616 {
617         int newly_dirty;
618         struct address_space *mapping = page_mapping(page);
619
620         if (unlikely(!mapping))
621                 return !TestSetPageDirty(page);
622
623         spin_lock(&mapping->private_lock);
624         if (page_has_buffers(page)) {
625                 struct buffer_head *head = page_buffers(page);
626                 struct buffer_head *bh = head;
627
628                 do {
629                         set_buffer_dirty(bh);
630                         bh = bh->b_this_page;
631                 } while (bh != head);
632         }
633         /*
634          * Lock out page->mem_cgroup migration to keep PageDirty
635          * synchronized with per-memcg dirty page counters.
636          */
637         lock_page_memcg(page);
638         newly_dirty = !TestSetPageDirty(page);
639         spin_unlock(&mapping->private_lock);
640
641         if (newly_dirty)
642                 __set_page_dirty(page, mapping, 1);
643
644         unlock_page_memcg(page);
645
646         if (newly_dirty)
647                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
648
649         return newly_dirty;
650 }
651 EXPORT_SYMBOL(__set_page_dirty_buffers);
652
653 /*
654  * Write out and wait upon a list of buffers.
655  *
656  * We have conflicting pressures: we want to make sure that all
657  * initially dirty buffers get waited on, but that any subsequently
658  * dirtied buffers don't.  After all, we don't want fsync to last
659  * forever if somebody is actively writing to the file.
660  *
661  * Do this in two main stages: first we copy dirty buffers to a
662  * temporary inode list, queueing the writes as we go.  Then we clean
663  * up, waiting for those writes to complete.
664  * 
665  * During this second stage, any subsequent updates to the file may end
666  * up refiling the buffer on the original inode's dirty list again, so
667  * there is a chance we will end up with a buffer queued for write but
668  * not yet completed on that list.  So, as a final cleanup we go through
669  * the osync code to catch these locked, dirty buffers without requeuing
670  * any newly dirty buffers for write.
671  */
672 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
673 {
674         struct buffer_head *bh;
675         struct list_head tmp;
676         struct address_space *mapping;
677         int err = 0, err2;
678         struct blk_plug plug;
679
680         INIT_LIST_HEAD(&tmp);
681         blk_start_plug(&plug);
682
683         spin_lock(lock);
684         while (!list_empty(list)) {
685                 bh = BH_ENTRY(list->next);
686                 mapping = bh->b_assoc_map;
687                 __remove_assoc_queue(bh);
688                 /* Avoid race with mark_buffer_dirty_inode() which does
689                  * a lockless check and we rely on seeing the dirty bit */
690                 smp_mb();
691                 if (buffer_dirty(bh) || buffer_locked(bh)) {
692                         list_add(&bh->b_assoc_buffers, &tmp);
693                         bh->b_assoc_map = mapping;
694                         if (buffer_dirty(bh)) {
695                                 get_bh(bh);
696                                 spin_unlock(lock);
697                                 /*
698                                  * Ensure any pending I/O completes so that
699                                  * write_dirty_buffer() actually writes the
700                                  * current contents - it is a noop if I/O is
701                                  * still in flight on potentially older
702                                  * contents.
703                                  */
704                                 write_dirty_buffer(bh, REQ_SYNC);
705
706                                 /*
707                                  * Kick off IO for the previous mapping. Note
708                                  * that we will not run the very last mapping,
709                                  * wait_on_buffer() will do that for us
710                                  * through sync_buffer().
711                                  */
712                                 brelse(bh);
713                                 spin_lock(lock);
714                         }
715                 }
716         }
717
718         spin_unlock(lock);
719         blk_finish_plug(&plug);
720         spin_lock(lock);
721
722         while (!list_empty(&tmp)) {
723                 bh = BH_ENTRY(tmp.prev);
724                 get_bh(bh);
725                 mapping = bh->b_assoc_map;
726                 __remove_assoc_queue(bh);
727                 /* Avoid race with mark_buffer_dirty_inode() which does
728                  * a lockless check and we rely on seeing the dirty bit */
729                 smp_mb();
730                 if (buffer_dirty(bh)) {
731                         list_add(&bh->b_assoc_buffers,
732                                  &mapping->private_list);
733                         bh->b_assoc_map = mapping;
734                 }
735                 spin_unlock(lock);
736                 wait_on_buffer(bh);
737                 if (!buffer_uptodate(bh))
738                         err = -EIO;
739                 brelse(bh);
740                 spin_lock(lock);
741         }
742         
743         spin_unlock(lock);
744         err2 = osync_buffers_list(lock, list);
745         if (err)
746                 return err;
747         else
748                 return err2;
749 }
750
751 /*
752  * Invalidate any and all dirty buffers on a given inode.  We are
753  * probably unmounting the fs, but that doesn't mean we have already
754  * done a sync().  Just drop the buffers from the inode list.
755  *
756  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
757  * assumes that all the buffers are against the blockdev.  Not true
758  * for reiserfs.
759  */
760 void invalidate_inode_buffers(struct inode *inode)
761 {
762         if (inode_has_buffers(inode)) {
763                 struct address_space *mapping = &inode->i_data;
764                 struct list_head *list = &mapping->private_list;
765                 struct address_space *buffer_mapping = mapping->private_data;
766
767                 spin_lock(&buffer_mapping->private_lock);
768                 while (!list_empty(list))
769                         __remove_assoc_queue(BH_ENTRY(list->next));
770                 spin_unlock(&buffer_mapping->private_lock);
771         }
772 }
773 EXPORT_SYMBOL(invalidate_inode_buffers);
774
775 /*
776  * Remove any clean buffers from the inode's buffer list.  This is called
777  * when we're trying to free the inode itself.  Those buffers can pin it.
778  *
779  * Returns true if all buffers were removed.
780  */
781 int remove_inode_buffers(struct inode *inode)
782 {
783         int ret = 1;
784
785         if (inode_has_buffers(inode)) {
786                 struct address_space *mapping = &inode->i_data;
787                 struct list_head *list = &mapping->private_list;
788                 struct address_space *buffer_mapping = mapping->private_data;
789
790                 spin_lock(&buffer_mapping->private_lock);
791                 while (!list_empty(list)) {
792                         struct buffer_head *bh = BH_ENTRY(list->next);
793                         if (buffer_dirty(bh)) {
794                                 ret = 0;
795                                 break;
796                         }
797                         __remove_assoc_queue(bh);
798                 }
799                 spin_unlock(&buffer_mapping->private_lock);
800         }
801         return ret;
802 }
803
804 /*
805  * Create the appropriate buffers when given a page for data area and
806  * the size of each buffer.. Use the bh->b_this_page linked list to
807  * follow the buffers created.  Return NULL if unable to create more
808  * buffers.
809  *
810  * The retry flag is used to differentiate async IO (paging, swapping)
811  * which may not fail from ordinary buffer allocations.
812  */
813 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
814                 bool retry)
815 {
816         struct buffer_head *bh, *head;
817         gfp_t gfp = GFP_NOFS | __GFP_ACCOUNT;
818         long offset;
819         struct mem_cgroup *memcg;
820
821         if (retry)
822                 gfp |= __GFP_NOFAIL;
823
824         memcg = get_mem_cgroup_from_page(page);
825         memalloc_use_memcg(memcg);
826
827         head = NULL;
828         offset = PAGE_SIZE;
829         while ((offset -= size) >= 0) {
830                 bh = alloc_buffer_head(gfp);
831                 if (!bh)
832                         goto no_grow;
833
834                 bh->b_this_page = head;
835                 bh->b_blocknr = -1;
836                 head = bh;
837
838                 bh->b_size = size;
839
840                 /* Link the buffer to its page */
841                 set_bh_page(bh, page, offset);
842         }
843 out:
844         memalloc_unuse_memcg();
845         mem_cgroup_put(memcg);
846         return head;
847 /*
848  * In case anything failed, we just free everything we got.
849  */
850 no_grow:
851         if (head) {
852                 do {
853                         bh = head;
854                         head = head->b_this_page;
855                         free_buffer_head(bh);
856                 } while (head);
857         }
858
859         goto out;
860 }
861 EXPORT_SYMBOL_GPL(alloc_page_buffers);
862
863 static inline void
864 link_dev_buffers(struct page *page, struct buffer_head *head)
865 {
866         struct buffer_head *bh, *tail;
867
868         bh = head;
869         do {
870                 tail = bh;
871                 bh = bh->b_this_page;
872         } while (bh);
873         tail->b_this_page = head;
874         attach_page_buffers(page, head);
875 }
876
877 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
878 {
879         sector_t retval = ~((sector_t)0);
880         loff_t sz = i_size_read(bdev->bd_inode);
881
882         if (sz) {
883                 unsigned int sizebits = blksize_bits(size);
884                 retval = (sz >> sizebits);
885         }
886         return retval;
887 }
888
889 /*
890  * Initialise the state of a blockdev page's buffers.
891  */ 
892 static sector_t
893 init_page_buffers(struct page *page, struct block_device *bdev,
894                         sector_t block, int size)
895 {
896         struct buffer_head *head = page_buffers(page);
897         struct buffer_head *bh = head;
898         int uptodate = PageUptodate(page);
899         sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
900
901         do {
902                 if (!buffer_mapped(bh)) {
903                         bh->b_end_io = NULL;
904                         bh->b_private = NULL;
905                         bh->b_bdev = bdev;
906                         bh->b_blocknr = block;
907                         if (uptodate)
908                                 set_buffer_uptodate(bh);
909                         if (block < end_block)
910                                 set_buffer_mapped(bh);
911                 }
912                 block++;
913                 bh = bh->b_this_page;
914         } while (bh != head);
915
916         /*
917          * Caller needs to validate requested block against end of device.
918          */
919         return end_block;
920 }
921
922 /*
923  * Create the page-cache page that contains the requested block.
924  *
925  * This is used purely for blockdev mappings.
926  */
927 static int
928 grow_dev_page(struct block_device *bdev, sector_t block,
929               pgoff_t index, int size, int sizebits, gfp_t gfp)
930 {
931         struct inode *inode = bdev->bd_inode;
932         struct page *page;
933         struct buffer_head *bh;
934         sector_t end_block;
935         int ret = 0;            /* Will call free_more_memory() */
936         gfp_t gfp_mask;
937
938         gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
939
940         /*
941          * XXX: __getblk_slow() can not really deal with failure and
942          * will endlessly loop on improvised global reclaim.  Prefer
943          * looping in the allocator rather than here, at least that
944          * code knows what it's doing.
945          */
946         gfp_mask |= __GFP_NOFAIL;
947
948         page = find_or_create_page(inode->i_mapping, index, gfp_mask);
949
950         BUG_ON(!PageLocked(page));
951
952         if (page_has_buffers(page)) {
953                 bh = page_buffers(page);
954                 if (bh->b_size == size) {
955                         end_block = init_page_buffers(page, bdev,
956                                                 (sector_t)index << sizebits,
957                                                 size);
958                         goto done;
959                 }
960                 if (!try_to_free_buffers(page))
961                         goto failed;
962         }
963
964         /*
965          * Allocate some buffers for this page
966          */
967         bh = alloc_page_buffers(page, size, true);
968
969         /*
970          * Link the page to the buffers and initialise them.  Take the
971          * lock to be atomic wrt __find_get_block(), which does not
972          * run under the page lock.
973          */
974         spin_lock(&inode->i_mapping->private_lock);
975         link_dev_buffers(page, bh);
976         end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
977                         size);
978         spin_unlock(&inode->i_mapping->private_lock);
979 done:
980         ret = (block < end_block) ? 1 : -ENXIO;
981 failed:
982         unlock_page(page);
983         put_page(page);
984         return ret;
985 }
986
987 /*
988  * Create buffers for the specified block device block's page.  If
989  * that page was dirty, the buffers are set dirty also.
990  */
991 static int
992 grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
993 {
994         pgoff_t index;
995         int sizebits;
996
997         sizebits = -1;
998         do {
999                 sizebits++;
1000         } while ((size << sizebits) < PAGE_SIZE);
1001
1002         index = block >> sizebits;
1003
1004         /*
1005          * Check for a block which wants to lie outside our maximum possible
1006          * pagecache index.  (this comparison is done using sector_t types).
1007          */
1008         if (unlikely(index != block >> sizebits)) {
1009                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1010                         "device %pg\n",
1011                         __func__, (unsigned long long)block,
1012                         bdev);
1013                 return -EIO;
1014         }
1015
1016         /* Create a page with the proper size buffers.. */
1017         return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1018 }
1019
1020 static struct buffer_head *
1021 __getblk_slow(struct block_device *bdev, sector_t block,
1022              unsigned size, gfp_t gfp)
1023 {
1024         /* Size must be multiple of hard sectorsize */
1025         if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1026                         (size < 512 || size > PAGE_SIZE))) {
1027                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1028                                         size);
1029                 printk(KERN_ERR "logical block size: %d\n",
1030                                         bdev_logical_block_size(bdev));
1031
1032                 dump_stack();
1033                 return NULL;
1034         }
1035
1036         for (;;) {
1037                 struct buffer_head *bh;
1038                 int ret;
1039
1040                 bh = __find_get_block(bdev, block, size);
1041                 if (bh)
1042                         return bh;
1043
1044                 ret = grow_buffers(bdev, block, size, gfp);
1045                 if (ret < 0)
1046                         return NULL;
1047         }
1048 }
1049
1050 /*
1051  * The relationship between dirty buffers and dirty pages:
1052  *
1053  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1054  * the page is tagged dirty in the page cache.
1055  *
1056  * At all times, the dirtiness of the buffers represents the dirtiness of
1057  * subsections of the page.  If the page has buffers, the page dirty bit is
1058  * merely a hint about the true dirty state.
1059  *
1060  * When a page is set dirty in its entirety, all its buffers are marked dirty
1061  * (if the page has buffers).
1062  *
1063  * When a buffer is marked dirty, its page is dirtied, but the page's other
1064  * buffers are not.
1065  *
1066  * Also.  When blockdev buffers are explicitly read with bread(), they
1067  * individually become uptodate.  But their backing page remains not
1068  * uptodate - even if all of its buffers are uptodate.  A subsequent
1069  * block_read_full_page() against that page will discover all the uptodate
1070  * buffers, will set the page uptodate and will perform no I/O.
1071  */
1072
1073 /**
1074  * mark_buffer_dirty - mark a buffer_head as needing writeout
1075  * @bh: the buffer_head to mark dirty
1076  *
1077  * mark_buffer_dirty() will set the dirty bit against the buffer, then set
1078  * its backing page dirty, then tag the page as dirty in the page cache
1079  * and then attach the address_space's inode to its superblock's dirty
1080  * inode list.
1081  *
1082  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1083  * i_pages lock and mapping->host->i_lock.
1084  */
1085 void mark_buffer_dirty(struct buffer_head *bh)
1086 {
1087         WARN_ON_ONCE(!buffer_uptodate(bh));
1088
1089         trace_block_dirty_buffer(bh);
1090
1091         /*
1092          * Very *carefully* optimize the it-is-already-dirty case.
1093          *
1094          * Don't let the final "is it dirty" escape to before we
1095          * perhaps modified the buffer.
1096          */
1097         if (buffer_dirty(bh)) {
1098                 smp_mb();
1099                 if (buffer_dirty(bh))
1100                         return;
1101         }
1102
1103         if (!test_set_buffer_dirty(bh)) {
1104                 struct page *page = bh->b_page;
1105                 struct address_space *mapping = NULL;
1106
1107                 lock_page_memcg(page);
1108                 if (!TestSetPageDirty(page)) {
1109                         mapping = page_mapping(page);
1110                         if (mapping)
1111                                 __set_page_dirty(page, mapping, 0);
1112                 }
1113                 unlock_page_memcg(page);
1114                 if (mapping)
1115                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1116         }
1117 }
1118 EXPORT_SYMBOL(mark_buffer_dirty);
1119
1120 void mark_buffer_write_io_error(struct buffer_head *bh)
1121 {
1122         set_buffer_write_io_error(bh);
1123         /* FIXME: do we need to set this in both places? */
1124         if (bh->b_page && bh->b_page->mapping)
1125                 mapping_set_error(bh->b_page->mapping, -EIO);
1126         if (bh->b_assoc_map)
1127                 mapping_set_error(bh->b_assoc_map, -EIO);
1128 }
1129 EXPORT_SYMBOL(mark_buffer_write_io_error);
1130
1131 /*
1132  * Decrement a buffer_head's reference count.  If all buffers against a page
1133  * have zero reference count, are clean and unlocked, and if the page is clean
1134  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1135  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1136  * a page but it ends up not being freed, and buffers may later be reattached).
1137  */
1138 void __brelse(struct buffer_head * buf)
1139 {
1140         if (atomic_read(&buf->b_count)) {
1141                 put_bh(buf);
1142                 return;
1143         }
1144         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1145 }
1146 EXPORT_SYMBOL(__brelse);
1147
1148 /*
1149  * bforget() is like brelse(), except it discards any
1150  * potentially dirty data.
1151  */
1152 void __bforget(struct buffer_head *bh)
1153 {
1154         clear_buffer_dirty(bh);
1155         if (bh->b_assoc_map) {
1156                 struct address_space *buffer_mapping = bh->b_page->mapping;
1157
1158                 spin_lock(&buffer_mapping->private_lock);
1159                 list_del_init(&bh->b_assoc_buffers);
1160                 bh->b_assoc_map = NULL;
1161                 spin_unlock(&buffer_mapping->private_lock);
1162         }
1163         __brelse(bh);
1164 }
1165 EXPORT_SYMBOL(__bforget);
1166
1167 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1168 {
1169         lock_buffer(bh);
1170         if (buffer_uptodate(bh)) {
1171                 unlock_buffer(bh);
1172                 return bh;
1173         } else {
1174                 get_bh(bh);
1175                 bh->b_end_io = end_buffer_read_sync;
1176                 submit_bh(REQ_OP_READ, 0, bh);
1177                 wait_on_buffer(bh);
1178                 if (buffer_uptodate(bh))
1179                         return bh;
1180         }
1181         brelse(bh);
1182         return NULL;
1183 }
1184
1185 /*
1186  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1187  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1188  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1189  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1190  * CPU's LRUs at the same time.
1191  *
1192  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1193  * sb_find_get_block().
1194  *
1195  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1196  * a local interrupt disable for that.
1197  */
1198
1199 #define BH_LRU_SIZE     16
1200
1201 struct bh_lru {
1202         struct buffer_head *bhs[BH_LRU_SIZE];
1203 };
1204
1205 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1206
1207 #ifdef CONFIG_SMP
1208 #define bh_lru_lock()   local_irq_disable()
1209 #define bh_lru_unlock() local_irq_enable()
1210 #else
1211 #define bh_lru_lock()   preempt_disable()
1212 #define bh_lru_unlock() preempt_enable()
1213 #endif
1214
1215 static inline void check_irqs_on(void)
1216 {
1217 #ifdef irqs_disabled
1218         BUG_ON(irqs_disabled());
1219 #endif
1220 }
1221
1222 /*
1223  * Install a buffer_head into this cpu's LRU.  If not already in the LRU, it is
1224  * inserted at the front, and the buffer_head at the back if any is evicted.
1225  * Or, if already in the LRU it is moved to the front.
1226  */
1227 static void bh_lru_install(struct buffer_head *bh)
1228 {
1229         struct buffer_head *evictee = bh;
1230         struct bh_lru *b;
1231         int i;
1232
1233         check_irqs_on();
1234         bh_lru_lock();
1235
1236         b = this_cpu_ptr(&bh_lrus);
1237         for (i = 0; i < BH_LRU_SIZE; i++) {
1238                 swap(evictee, b->bhs[i]);
1239                 if (evictee == bh) {
1240                         bh_lru_unlock();
1241                         return;
1242                 }
1243         }
1244
1245         get_bh(bh);
1246         bh_lru_unlock();
1247         brelse(evictee);
1248 }
1249
1250 /*
1251  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1252  */
1253 static struct buffer_head *
1254 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1255 {
1256         struct buffer_head *ret = NULL;
1257         unsigned int i;
1258
1259         check_irqs_on();
1260         bh_lru_lock();
1261         for (i = 0; i < BH_LRU_SIZE; i++) {
1262                 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1263
1264                 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1265                     bh->b_size == size) {
1266                         if (i) {
1267                                 while (i) {
1268                                         __this_cpu_write(bh_lrus.bhs[i],
1269                                                 __this_cpu_read(bh_lrus.bhs[i - 1]));
1270                                         i--;
1271                                 }
1272                                 __this_cpu_write(bh_lrus.bhs[0], bh);
1273                         }
1274                         get_bh(bh);
1275                         ret = bh;
1276                         break;
1277                 }
1278         }
1279         bh_lru_unlock();
1280         return ret;
1281 }
1282
1283 /*
1284  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1285  * it in the LRU and mark it as accessed.  If it is not present then return
1286  * NULL
1287  */
1288 struct buffer_head *
1289 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1290 {
1291         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1292
1293         if (bh == NULL) {
1294                 /* __find_get_block_slow will mark the page accessed */
1295                 bh = __find_get_block_slow(bdev, block);
1296                 if (bh)
1297                         bh_lru_install(bh);
1298         } else
1299                 touch_buffer(bh);
1300
1301         return bh;
1302 }
1303 EXPORT_SYMBOL(__find_get_block);
1304
1305 /*
1306  * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1307  * which corresponds to the passed block_device, block and size. The
1308  * returned buffer has its reference count incremented.
1309  *
1310  * __getblk_gfp() will lock up the machine if grow_dev_page's
1311  * try_to_free_buffers() attempt is failing.  FIXME, perhaps?
1312  */
1313 struct buffer_head *
1314 __getblk_gfp(struct block_device *bdev, sector_t block,
1315              unsigned size, gfp_t gfp)
1316 {
1317         struct buffer_head *bh = __find_get_block(bdev, block, size);
1318
1319         might_sleep();
1320         if (bh == NULL)
1321                 bh = __getblk_slow(bdev, block, size, gfp);
1322         return bh;
1323 }
1324 EXPORT_SYMBOL(__getblk_gfp);
1325
1326 /*
1327  * Do async read-ahead on a buffer..
1328  */
1329 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1330 {
1331         struct buffer_head *bh = __getblk(bdev, block, size);
1332         if (likely(bh)) {
1333                 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1334                 brelse(bh);
1335         }
1336 }
1337 EXPORT_SYMBOL(__breadahead);
1338
1339 /**
1340  *  __bread_gfp() - reads a specified block and returns the bh
1341  *  @bdev: the block_device to read from
1342  *  @block: number of block
1343  *  @size: size (in bytes) to read
1344  *  @gfp: page allocation flag
1345  *
1346  *  Reads a specified block, and returns buffer head that contains it.
1347  *  The page cache can be allocated from non-movable area
1348  *  not to prevent page migration if you set gfp to zero.
1349  *  It returns NULL if the block was unreadable.
1350  */
1351 struct buffer_head *
1352 __bread_gfp(struct block_device *bdev, sector_t block,
1353                    unsigned size, gfp_t gfp)
1354 {
1355         struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1356
1357         if (likely(bh) && !buffer_uptodate(bh))
1358                 bh = __bread_slow(bh);
1359         return bh;
1360 }
1361 EXPORT_SYMBOL(__bread_gfp);
1362
1363 /*
1364  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1365  * This doesn't race because it runs in each cpu either in irq
1366  * or with preempt disabled.
1367  */
1368 static void invalidate_bh_lru(void *arg)
1369 {
1370         struct bh_lru *b = &get_cpu_var(bh_lrus);
1371         int i;
1372
1373         for (i = 0; i < BH_LRU_SIZE; i++) {
1374                 brelse(b->bhs[i]);
1375                 b->bhs[i] = NULL;
1376         }
1377         put_cpu_var(bh_lrus);
1378 }
1379
1380 static bool has_bh_in_lru(int cpu, void *dummy)
1381 {
1382         struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1383         int i;
1384         
1385         for (i = 0; i < BH_LRU_SIZE; i++) {
1386                 if (b->bhs[i])
1387                         return 1;
1388         }
1389
1390         return 0;
1391 }
1392
1393 void invalidate_bh_lrus(void)
1394 {
1395         on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1396 }
1397 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1398
1399 void set_bh_page(struct buffer_head *bh,
1400                 struct page *page, unsigned long offset)
1401 {
1402         bh->b_page = page;
1403         BUG_ON(offset >= PAGE_SIZE);
1404         if (PageHighMem(page))
1405                 /*
1406                  * This catches illegal uses and preserves the offset:
1407                  */
1408                 bh->b_data = (char *)(0 + offset);
1409         else
1410                 bh->b_data = page_address(page) + offset;
1411 }
1412 EXPORT_SYMBOL(set_bh_page);
1413
1414 /*
1415  * Called when truncating a buffer on a page completely.
1416  */
1417
1418 /* Bits that are cleared during an invalidate */
1419 #define BUFFER_FLAGS_DISCARD \
1420         (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1421          1 << BH_Delay | 1 << BH_Unwritten)
1422
1423 static void discard_buffer(struct buffer_head * bh)
1424 {
1425         unsigned long b_state, b_state_old;
1426
1427         lock_buffer(bh);
1428         clear_buffer_dirty(bh);
1429         bh->b_bdev = NULL;
1430         b_state = bh->b_state;
1431         for (;;) {
1432                 b_state_old = cmpxchg(&bh->b_state, b_state,
1433                                       (b_state & ~BUFFER_FLAGS_DISCARD));
1434                 if (b_state_old == b_state)
1435                         break;
1436                 b_state = b_state_old;
1437         }
1438         unlock_buffer(bh);
1439 }
1440
1441 /**
1442  * block_invalidatepage - invalidate part or all of a buffer-backed page
1443  *
1444  * @page: the page which is affected
1445  * @offset: start of the range to invalidate
1446  * @length: length of the range to invalidate
1447  *
1448  * block_invalidatepage() is called when all or part of the page has become
1449  * invalidated by a truncate operation.
1450  *
1451  * block_invalidatepage() does not have to release all buffers, but it must
1452  * ensure that no dirty buffer is left outside @offset and that no I/O
1453  * is underway against any of the blocks which are outside the truncation
1454  * point.  Because the caller is about to free (and possibly reuse) those
1455  * blocks on-disk.
1456  */
1457 void block_invalidatepage(struct page *page, unsigned int offset,
1458                           unsigned int length)
1459 {
1460         struct buffer_head *head, *bh, *next;
1461         unsigned int curr_off = 0;
1462         unsigned int stop = length + offset;
1463
1464         BUG_ON(!PageLocked(page));
1465         if (!page_has_buffers(page))
1466                 goto out;
1467
1468         /*
1469          * Check for overflow
1470          */
1471         BUG_ON(stop > PAGE_SIZE || stop < length);
1472
1473         head = page_buffers(page);
1474         bh = head;
1475         do {
1476                 unsigned int next_off = curr_off + bh->b_size;
1477                 next = bh->b_this_page;
1478
1479                 /*
1480                  * Are we still fully in range ?
1481                  */
1482                 if (next_off > stop)
1483                         goto out;
1484
1485                 /*
1486                  * is this block fully invalidated?
1487                  */
1488                 if (offset <= curr_off)
1489                         discard_buffer(bh);
1490                 curr_off = next_off;
1491                 bh = next;
1492         } while (bh != head);
1493
1494         /*
1495          * We release buffers only if the entire page is being invalidated.
1496          * The get_block cached value has been unconditionally invalidated,
1497          * so real IO is not possible anymore.
1498          */
1499         if (length == PAGE_SIZE)
1500                 try_to_release_page(page, 0);
1501 out:
1502         return;
1503 }
1504 EXPORT_SYMBOL(block_invalidatepage);
1505
1506
1507 /*
1508  * We attach and possibly dirty the buffers atomically wrt
1509  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1510  * is already excluded via the page lock.
1511  */
1512 void create_empty_buffers(struct page *page,
1513                         unsigned long blocksize, unsigned long b_state)
1514 {
1515         struct buffer_head *bh, *head, *tail;
1516
1517         head = alloc_page_buffers(page, blocksize, true);
1518         bh = head;
1519         do {
1520                 bh->b_state |= b_state;
1521                 tail = bh;
1522                 bh = bh->b_this_page;
1523         } while (bh);
1524         tail->b_this_page = head;
1525
1526         spin_lock(&page->mapping->private_lock);
1527         if (PageUptodate(page) || PageDirty(page)) {
1528                 bh = head;
1529                 do {
1530                         if (PageDirty(page))
1531                                 set_buffer_dirty(bh);
1532                         if (PageUptodate(page))
1533                                 set_buffer_uptodate(bh);
1534                         bh = bh->b_this_page;
1535                 } while (bh != head);
1536         }
1537         attach_page_buffers(page, head);
1538         spin_unlock(&page->mapping->private_lock);
1539 }
1540 EXPORT_SYMBOL(create_empty_buffers);
1541
1542 /**
1543  * clean_bdev_aliases: clean a range of buffers in block device
1544  * @bdev: Block device to clean buffers in
1545  * @block: Start of a range of blocks to clean
1546  * @len: Number of blocks to clean
1547  *
1548  * We are taking a range of blocks for data and we don't want writeback of any
1549  * buffer-cache aliases starting from return from this function and until the
1550  * moment when something will explicitly mark the buffer dirty (hopefully that
1551  * will not happen until we will free that block ;-) We don't even need to mark
1552  * it not-uptodate - nobody can expect anything from a newly allocated buffer
1553  * anyway. We used to use unmap_buffer() for such invalidation, but that was
1554  * wrong. We definitely don't want to mark the alias unmapped, for example - it
1555  * would confuse anyone who might pick it with bread() afterwards...
1556  *
1557  * Also..  Note that bforget() doesn't lock the buffer.  So there can be
1558  * writeout I/O going on against recently-freed buffers.  We don't wait on that
1559  * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1560  * need to.  That happens here.
1561  */
1562 void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1563 {
1564         struct inode *bd_inode = bdev->bd_inode;
1565         struct address_space *bd_mapping = bd_inode->i_mapping;
1566         struct pagevec pvec;
1567         pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1568         pgoff_t end;
1569         int i, count;
1570         struct buffer_head *bh;
1571         struct buffer_head *head;
1572
1573         end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1574         pagevec_init(&pvec);
1575         while (pagevec_lookup_range(&pvec, bd_mapping, &index, end)) {
1576                 count = pagevec_count(&pvec);
1577                 for (i = 0; i < count; i++) {
1578                         struct page *page = pvec.pages[i];
1579
1580                         if (!page_has_buffers(page))
1581                                 continue;
1582                         /*
1583                          * We use page lock instead of bd_mapping->private_lock
1584                          * to pin buffers here since we can afford to sleep and
1585                          * it scales better than a global spinlock lock.
1586                          */
1587                         lock_page(page);
1588                         /* Recheck when the page is locked which pins bhs */
1589                         if (!page_has_buffers(page))
1590                                 goto unlock_page;
1591                         head = page_buffers(page);
1592                         bh = head;
1593                         do {
1594                                 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1595                                         goto next;
1596                                 if (bh->b_blocknr >= block + len)
1597                                         break;
1598                                 clear_buffer_dirty(bh);
1599                                 wait_on_buffer(bh);
1600                                 clear_buffer_req(bh);
1601 next:
1602                                 bh = bh->b_this_page;
1603                         } while (bh != head);
1604 unlock_page:
1605                         unlock_page(page);
1606                 }
1607                 pagevec_release(&pvec);
1608                 cond_resched();
1609                 /* End of range already reached? */
1610                 if (index > end || !index)
1611                         break;
1612         }
1613 }
1614 EXPORT_SYMBOL(clean_bdev_aliases);
1615
1616 /*
1617  * Size is a power-of-two in the range 512..PAGE_SIZE,
1618  * and the case we care about most is PAGE_SIZE.
1619  *
1620  * So this *could* possibly be written with those
1621  * constraints in mind (relevant mostly if some
1622  * architecture has a slow bit-scan instruction)
1623  */
1624 static inline int block_size_bits(unsigned int blocksize)
1625 {
1626         return ilog2(blocksize);
1627 }
1628
1629 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1630 {
1631         BUG_ON(!PageLocked(page));
1632
1633         if (!page_has_buffers(page))
1634                 create_empty_buffers(page, 1 << READ_ONCE(inode->i_blkbits),
1635                                      b_state);
1636         return page_buffers(page);
1637 }
1638
1639 /*
1640  * NOTE! All mapped/uptodate combinations are valid:
1641  *
1642  *      Mapped  Uptodate        Meaning
1643  *
1644  *      No      No              "unknown" - must do get_block()
1645  *      No      Yes             "hole" - zero-filled
1646  *      Yes     No              "allocated" - allocated on disk, not read in
1647  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1648  *
1649  * "Dirty" is valid only with the last case (mapped+uptodate).
1650  */
1651
1652 /*
1653  * While block_write_full_page is writing back the dirty buffers under
1654  * the page lock, whoever dirtied the buffers may decide to clean them
1655  * again at any time.  We handle that by only looking at the buffer
1656  * state inside lock_buffer().
1657  *
1658  * If block_write_full_page() is called for regular writeback
1659  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1660  * locked buffer.   This only can happen if someone has written the buffer
1661  * directly, with submit_bh().  At the address_space level PageWriteback
1662  * prevents this contention from occurring.
1663  *
1664  * If block_write_full_page() is called with wbc->sync_mode ==
1665  * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1666  * causes the writes to be flagged as synchronous writes.
1667  */
1668 int __block_write_full_page(struct inode *inode, struct page *page,
1669                         get_block_t *get_block, struct writeback_control *wbc,
1670                         bh_end_io_t *handler)
1671 {
1672         int err;
1673         sector_t block;
1674         sector_t last_block;
1675         struct buffer_head *bh, *head;
1676         unsigned int blocksize, bbits;
1677         int nr_underway = 0;
1678         int write_flags = wbc_to_write_flags(wbc);
1679
1680         head = create_page_buffers(page, inode,
1681                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1682
1683         /*
1684          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1685          * here, and the (potentially unmapped) buffers may become dirty at
1686          * any time.  If a buffer becomes dirty here after we've inspected it
1687          * then we just miss that fact, and the page stays dirty.
1688          *
1689          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1690          * handle that here by just cleaning them.
1691          */
1692
1693         bh = head;
1694         blocksize = bh->b_size;
1695         bbits = block_size_bits(blocksize);
1696
1697         block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1698         last_block = (i_size_read(inode) - 1) >> bbits;
1699
1700         /*
1701          * Get all the dirty buffers mapped to disk addresses and
1702          * handle any aliases from the underlying blockdev's mapping.
1703          */
1704         do {
1705                 if (block > last_block) {
1706                         /*
1707                          * mapped buffers outside i_size will occur, because
1708                          * this page can be outside i_size when there is a
1709                          * truncate in progress.
1710                          */
1711                         /*
1712                          * The buffer was zeroed by block_write_full_page()
1713                          */
1714                         clear_buffer_dirty(bh);
1715                         set_buffer_uptodate(bh);
1716                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1717                            buffer_dirty(bh)) {
1718                         WARN_ON(bh->b_size != blocksize);
1719                         err = get_block(inode, block, bh, 1);
1720                         if (err)
1721                                 goto recover;
1722                         clear_buffer_delay(bh);
1723                         if (buffer_new(bh)) {
1724                                 /* blockdev mappings never come here */
1725                                 clear_buffer_new(bh);
1726                                 clean_bdev_bh_alias(bh);
1727                         }
1728                 }
1729                 bh = bh->b_this_page;
1730                 block++;
1731         } while (bh != head);
1732
1733         do {
1734                 if (!buffer_mapped(bh))
1735                         continue;
1736                 /*
1737                  * If it's a fully non-blocking write attempt and we cannot
1738                  * lock the buffer then redirty the page.  Note that this can
1739                  * potentially cause a busy-wait loop from writeback threads
1740                  * and kswapd activity, but those code paths have their own
1741                  * higher-level throttling.
1742                  */
1743                 if (wbc->sync_mode != WB_SYNC_NONE) {
1744                         lock_buffer(bh);
1745                 } else if (!trylock_buffer(bh)) {
1746                         redirty_page_for_writepage(wbc, page);
1747                         continue;
1748                 }
1749                 if (test_clear_buffer_dirty(bh)) {
1750                         mark_buffer_async_write_endio(bh, handler);
1751                 } else {
1752                         unlock_buffer(bh);
1753                 }
1754         } while ((bh = bh->b_this_page) != head);
1755
1756         /*
1757          * The page and its buffers are protected by PageWriteback(), so we can
1758          * drop the bh refcounts early.
1759          */
1760         BUG_ON(PageWriteback(page));
1761         set_page_writeback(page);
1762
1763         do {
1764                 struct buffer_head *next = bh->b_this_page;
1765                 if (buffer_async_write(bh)) {
1766                         submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1767                                         inode->i_write_hint, wbc);
1768                         nr_underway++;
1769                 }
1770                 bh = next;
1771         } while (bh != head);
1772         unlock_page(page);
1773
1774         err = 0;
1775 done:
1776         if (nr_underway == 0) {
1777                 /*
1778                  * The page was marked dirty, but the buffers were
1779                  * clean.  Someone wrote them back by hand with
1780                  * ll_rw_block/submit_bh.  A rare case.
1781                  */
1782                 end_page_writeback(page);
1783
1784                 /*
1785                  * The page and buffer_heads can be released at any time from
1786                  * here on.
1787                  */
1788         }
1789         return err;
1790
1791 recover:
1792         /*
1793          * ENOSPC, or some other error.  We may already have added some
1794          * blocks to the file, so we need to write these out to avoid
1795          * exposing stale data.
1796          * The page is currently locked and not marked for writeback
1797          */
1798         bh = head;
1799         /* Recovery: lock and submit the mapped buffers */
1800         do {
1801                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1802                     !buffer_delay(bh)) {
1803                         lock_buffer(bh);
1804                         mark_buffer_async_write_endio(bh, handler);
1805                 } else {
1806                         /*
1807                          * The buffer may have been set dirty during
1808                          * attachment to a dirty page.
1809                          */
1810                         clear_buffer_dirty(bh);
1811                 }
1812         } while ((bh = bh->b_this_page) != head);
1813         SetPageError(page);
1814         BUG_ON(PageWriteback(page));
1815         mapping_set_error(page->mapping, err);
1816         set_page_writeback(page);
1817         do {
1818                 struct buffer_head *next = bh->b_this_page;
1819                 if (buffer_async_write(bh)) {
1820                         clear_buffer_dirty(bh);
1821                         submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1822                                         inode->i_write_hint, wbc);
1823                         nr_underway++;
1824                 }
1825                 bh = next;
1826         } while (bh != head);
1827         unlock_page(page);
1828         goto done;
1829 }
1830 EXPORT_SYMBOL(__block_write_full_page);
1831
1832 /*
1833  * If a page has any new buffers, zero them out here, and mark them uptodate
1834  * and dirty so they'll be written out (in order to prevent uninitialised
1835  * block data from leaking). And clear the new bit.
1836  */
1837 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1838 {
1839         unsigned int block_start, block_end;
1840         struct buffer_head *head, *bh;
1841
1842         BUG_ON(!PageLocked(page));
1843         if (!page_has_buffers(page))
1844                 return;
1845
1846         bh = head = page_buffers(page);
1847         block_start = 0;
1848         do {
1849                 block_end = block_start + bh->b_size;
1850
1851                 if (buffer_new(bh)) {
1852                         if (block_end > from && block_start < to) {
1853                                 if (!PageUptodate(page)) {
1854                                         unsigned start, size;
1855
1856                                         start = max(from, block_start);
1857                                         size = min(to, block_end) - start;
1858
1859                                         zero_user(page, start, size);
1860                                         set_buffer_uptodate(bh);
1861                                 }
1862
1863                                 clear_buffer_new(bh);
1864                                 mark_buffer_dirty(bh);
1865                         }
1866                 }
1867
1868                 block_start = block_end;
1869                 bh = bh->b_this_page;
1870         } while (bh != head);
1871 }
1872 EXPORT_SYMBOL(page_zero_new_buffers);
1873
1874 static void
1875 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1876                 struct iomap *iomap)
1877 {
1878         loff_t offset = block << inode->i_blkbits;
1879
1880         bh->b_bdev = iomap->bdev;
1881
1882         /*
1883          * Block points to offset in file we need to map, iomap contains
1884          * the offset at which the map starts. If the map ends before the
1885          * current block, then do not map the buffer and let the caller
1886          * handle it.
1887          */
1888         BUG_ON(offset >= iomap->offset + iomap->length);
1889
1890         switch (iomap->type) {
1891         case IOMAP_HOLE:
1892                 /*
1893                  * If the buffer is not up to date or beyond the current EOF,
1894                  * we need to mark it as new to ensure sub-block zeroing is
1895                  * executed if necessary.
1896                  */
1897                 if (!buffer_uptodate(bh) ||
1898                     (offset >= i_size_read(inode)))
1899                         set_buffer_new(bh);
1900                 break;
1901         case IOMAP_DELALLOC:
1902                 if (!buffer_uptodate(bh) ||
1903                     (offset >= i_size_read(inode)))
1904                         set_buffer_new(bh);
1905                 set_buffer_uptodate(bh);
1906                 set_buffer_mapped(bh);
1907                 set_buffer_delay(bh);
1908                 break;
1909         case IOMAP_UNWRITTEN:
1910                 /*
1911                  * For unwritten regions, we always need to ensure that regions
1912                  * in the block we are not writing to are zeroed. Mark the
1913                  * buffer as new to ensure this.
1914                  */
1915                 set_buffer_new(bh);
1916                 set_buffer_unwritten(bh);
1917                 /* FALLTHRU */
1918         case IOMAP_MAPPED:
1919                 if ((iomap->flags & IOMAP_F_NEW) ||
1920                     offset >= i_size_read(inode))
1921                         set_buffer_new(bh);
1922                 bh->b_blocknr = (iomap->addr + offset - iomap->offset) >>
1923                                 inode->i_blkbits;
1924                 set_buffer_mapped(bh);
1925                 break;
1926         }
1927 }
1928
1929 int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
1930                 get_block_t *get_block, struct iomap *iomap)
1931 {
1932         unsigned from = pos & (PAGE_SIZE - 1);
1933         unsigned to = from + len;
1934         struct inode *inode = page->mapping->host;
1935         unsigned block_start, block_end;
1936         sector_t block;
1937         int err = 0;
1938         unsigned blocksize, bbits;
1939         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1940
1941         BUG_ON(!PageLocked(page));
1942         BUG_ON(from > PAGE_SIZE);
1943         BUG_ON(to > PAGE_SIZE);
1944         BUG_ON(from > to);
1945
1946         head = create_page_buffers(page, inode, 0);
1947         blocksize = head->b_size;
1948         bbits = block_size_bits(blocksize);
1949
1950         block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1951
1952         for(bh = head, block_start = 0; bh != head || !block_start;
1953             block++, block_start=block_end, bh = bh->b_this_page) {
1954                 block_end = block_start + blocksize;
1955                 if (block_end <= from || block_start >= to) {
1956                         if (PageUptodate(page)) {
1957                                 if (!buffer_uptodate(bh))
1958                                         set_buffer_uptodate(bh);
1959                         }
1960                         continue;
1961                 }
1962                 if (buffer_new(bh))
1963                         clear_buffer_new(bh);
1964                 if (!buffer_mapped(bh)) {
1965                         WARN_ON(bh->b_size != blocksize);
1966                         if (get_block) {
1967                                 err = get_block(inode, block, bh, 1);
1968                                 if (err)
1969                                         break;
1970                         } else {
1971                                 iomap_to_bh(inode, block, bh, iomap);
1972                         }
1973
1974                         if (buffer_new(bh)) {
1975                                 clean_bdev_bh_alias(bh);
1976                                 if (PageUptodate(page)) {
1977                                         clear_buffer_new(bh);
1978                                         set_buffer_uptodate(bh);
1979                                         mark_buffer_dirty(bh);
1980                                         continue;
1981                                 }
1982                                 if (block_end > to || block_start < from)
1983                                         zero_user_segments(page,
1984                                                 to, block_end,
1985                                                 block_start, from);
1986                                 continue;
1987                         }
1988                 }
1989                 if (PageUptodate(page)) {
1990                         if (!buffer_uptodate(bh))
1991                                 set_buffer_uptodate(bh);
1992                         continue; 
1993                 }
1994                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1995                     !buffer_unwritten(bh) &&
1996                      (block_start < from || block_end > to)) {
1997                         ll_rw_block(REQ_OP_READ, 0, 1, &bh);
1998                         *wait_bh++=bh;
1999                 }
2000         }
2001         /*
2002          * If we issued read requests - let them complete.
2003          */
2004         while(wait_bh > wait) {
2005                 wait_on_buffer(*--wait_bh);
2006                 if (!buffer_uptodate(*wait_bh))
2007                         err = -EIO;
2008         }
2009         if (unlikely(err))
2010                 page_zero_new_buffers(page, from, to);
2011         return err;
2012 }
2013
2014 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2015                 get_block_t *get_block)
2016 {
2017         return __block_write_begin_int(page, pos, len, get_block, NULL);
2018 }
2019 EXPORT_SYMBOL(__block_write_begin);
2020
2021 static int __block_commit_write(struct inode *inode, struct page *page,
2022                 unsigned from, unsigned to)
2023 {
2024         unsigned block_start, block_end;
2025         int partial = 0;
2026         unsigned blocksize;
2027         struct buffer_head *bh, *head;
2028
2029         bh = head = page_buffers(page);
2030         blocksize = bh->b_size;
2031
2032         block_start = 0;
2033         do {
2034                 block_end = block_start + blocksize;
2035                 if (block_end <= from || block_start >= to) {
2036                         if (!buffer_uptodate(bh))
2037                                 partial = 1;
2038                 } else {
2039                         set_buffer_uptodate(bh);
2040                         mark_buffer_dirty(bh);
2041                 }
2042                 clear_buffer_new(bh);
2043
2044                 block_start = block_end;
2045                 bh = bh->b_this_page;
2046         } while (bh != head);
2047
2048         /*
2049          * If this is a partial write which happened to make all buffers
2050          * uptodate then we can optimize away a bogus readpage() for
2051          * the next read(). Here we 'discover' whether the page went
2052          * uptodate as a result of this (potentially partial) write.
2053          */
2054         if (!partial)
2055                 SetPageUptodate(page);
2056         return 0;
2057 }
2058
2059 /*
2060  * block_write_begin takes care of the basic task of block allocation and
2061  * bringing partial write blocks uptodate first.
2062  *
2063  * The filesystem needs to handle block truncation upon failure.
2064  */
2065 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2066                 unsigned flags, struct page **pagep, get_block_t *get_block)
2067 {
2068         pgoff_t index = pos >> PAGE_SHIFT;
2069         struct page *page;
2070         int status;
2071
2072         page = grab_cache_page_write_begin(mapping, index, flags);
2073         if (!page)
2074                 return -ENOMEM;
2075
2076         status = __block_write_begin(page, pos, len, get_block);
2077         if (unlikely(status)) {
2078                 unlock_page(page);
2079                 put_page(page);
2080                 page = NULL;
2081         }
2082
2083         *pagep = page;
2084         return status;
2085 }
2086 EXPORT_SYMBOL(block_write_begin);
2087
2088 void __generic_write_end(struct inode *inode, loff_t pos, unsigned copied,
2089                 struct page *page)
2090 {
2091         loff_t old_size = inode->i_size;
2092         bool i_size_changed = false;
2093
2094         /*
2095          * No need to use i_size_read() here, the i_size cannot change under us
2096          * because we hold i_rwsem.
2097          *
2098          * But it's important to update i_size while still holding page lock:
2099          * page writeout could otherwise come in and zero beyond i_size.
2100          */
2101         if (pos + copied > inode->i_size) {
2102                 i_size_write(inode, pos + copied);
2103                 i_size_changed = true;
2104         }
2105
2106         unlock_page(page);
2107
2108         if (old_size < pos)
2109                 pagecache_isize_extended(inode, old_size, pos);
2110         /*
2111          * Don't mark the inode dirty under page lock. First, it unnecessarily
2112          * makes the holding time of page lock longer. Second, it forces lock
2113          * ordering of page lock and transaction start for journaling
2114          * filesystems.
2115          */
2116         if (i_size_changed)
2117                 mark_inode_dirty(inode);
2118 }
2119
2120 int block_write_end(struct file *file, struct address_space *mapping,
2121                         loff_t pos, unsigned len, unsigned copied,
2122                         struct page *page, void *fsdata)
2123 {
2124         struct inode *inode = mapping->host;
2125         unsigned start;
2126
2127         start = pos & (PAGE_SIZE - 1);
2128
2129         if (unlikely(copied < len)) {
2130                 /*
2131                  * The buffers that were written will now be uptodate, so we
2132                  * don't have to worry about a readpage reading them and
2133                  * overwriting a partial write. However if we have encountered
2134                  * a short write and only partially written into a buffer, it
2135                  * will not be marked uptodate, so a readpage might come in and
2136                  * destroy our partial write.
2137                  *
2138                  * Do the simplest thing, and just treat any short write to a
2139                  * non uptodate page as a zero-length write, and force the
2140                  * caller to redo the whole thing.
2141                  */
2142                 if (!PageUptodate(page))
2143                         copied = 0;
2144
2145                 page_zero_new_buffers(page, start+copied, start+len);
2146         }
2147         flush_dcache_page(page);
2148
2149         /* This could be a short (even 0-length) commit */
2150         __block_commit_write(inode, page, start, start+copied);
2151
2152         return copied;
2153 }
2154 EXPORT_SYMBOL(block_write_end);
2155
2156 int generic_write_end(struct file *file, struct address_space *mapping,
2157                         loff_t pos, unsigned len, unsigned copied,
2158                         struct page *page, void *fsdata)
2159 {
2160         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2161         __generic_write_end(mapping->host, pos, copied, page);
2162         put_page(page);
2163         return copied;
2164 }
2165 EXPORT_SYMBOL(generic_write_end);
2166
2167 /*
2168  * block_is_partially_uptodate checks whether buffers within a page are
2169  * uptodate or not.
2170  *
2171  * Returns true if all buffers which correspond to a file portion
2172  * we want to read are uptodate.
2173  */
2174 int block_is_partially_uptodate(struct page *page, unsigned long from,
2175                                         unsigned long count)
2176 {
2177         unsigned block_start, block_end, blocksize;
2178         unsigned to;
2179         struct buffer_head *bh, *head;
2180         int ret = 1;
2181
2182         if (!page_has_buffers(page))
2183                 return 0;
2184
2185         head = page_buffers(page);
2186         blocksize = head->b_size;
2187         to = min_t(unsigned, PAGE_SIZE - from, count);
2188         to = from + to;
2189         if (from < blocksize && to > PAGE_SIZE - blocksize)
2190                 return 0;
2191
2192         bh = head;
2193         block_start = 0;
2194         do {
2195                 block_end = block_start + blocksize;
2196                 if (block_end > from && block_start < to) {
2197                         if (!buffer_uptodate(bh)) {
2198                                 ret = 0;
2199                                 break;
2200                         }
2201                         if (block_end >= to)
2202                                 break;
2203                 }
2204                 block_start = block_end;
2205                 bh = bh->b_this_page;
2206         } while (bh != head);
2207
2208         return ret;
2209 }
2210 EXPORT_SYMBOL(block_is_partially_uptodate);
2211
2212 /*
2213  * Generic "read page" function for block devices that have the normal
2214  * get_block functionality. This is most of the block device filesystems.
2215  * Reads the page asynchronously --- the unlock_buffer() and
2216  * set/clear_buffer_uptodate() functions propagate buffer state into the
2217  * page struct once IO has completed.
2218  */
2219 int block_read_full_page(struct page *page, get_block_t *get_block)
2220 {
2221         struct inode *inode = page->mapping->host;
2222         sector_t iblock, lblock;
2223         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2224         unsigned int blocksize, bbits;
2225         int nr, i;
2226         int fully_mapped = 1;
2227
2228         head = create_page_buffers(page, inode, 0);
2229         blocksize = head->b_size;
2230         bbits = block_size_bits(blocksize);
2231
2232         iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2233         lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2234         bh = head;
2235         nr = 0;
2236         i = 0;
2237
2238         do {
2239                 if (buffer_uptodate(bh))
2240                         continue;
2241
2242                 if (!buffer_mapped(bh)) {
2243                         int err = 0;
2244
2245                         fully_mapped = 0;
2246                         if (iblock < lblock) {
2247                                 WARN_ON(bh->b_size != blocksize);
2248                                 err = get_block(inode, iblock, bh, 0);
2249                                 if (err)
2250                                         SetPageError(page);
2251                         }
2252                         if (!buffer_mapped(bh)) {
2253                                 zero_user(page, i * blocksize, blocksize);
2254                                 if (!err)
2255                                         set_buffer_uptodate(bh);
2256                                 continue;
2257                         }
2258                         /*
2259                          * get_block() might have updated the buffer
2260                          * synchronously
2261                          */
2262                         if (buffer_uptodate(bh))
2263                                 continue;
2264                 }
2265                 arr[nr++] = bh;
2266         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2267
2268         if (fully_mapped)
2269                 SetPageMappedToDisk(page);
2270
2271         if (!nr) {
2272                 /*
2273                  * All buffers are uptodate - we can set the page uptodate
2274                  * as well. But not if get_block() returned an error.
2275                  */
2276                 if (!PageError(page))
2277                         SetPageUptodate(page);
2278                 unlock_page(page);
2279                 return 0;
2280         }
2281
2282         /* Stage two: lock the buffers */
2283         for (i = 0; i < nr; i++) {
2284                 bh = arr[i];
2285                 lock_buffer(bh);
2286                 mark_buffer_async_read(bh);
2287         }
2288
2289         /*
2290          * Stage 3: start the IO.  Check for uptodateness
2291          * inside the buffer lock in case another process reading
2292          * the underlying blockdev brought it uptodate (the sct fix).
2293          */
2294         for (i = 0; i < nr; i++) {
2295                 bh = arr[i];
2296                 if (buffer_uptodate(bh))
2297                         end_buffer_async_read(bh, 1);
2298                 else
2299                         submit_bh(REQ_OP_READ, 0, bh);
2300         }
2301         return 0;
2302 }
2303 EXPORT_SYMBOL(block_read_full_page);
2304
2305 /* utility function for filesystems that need to do work on expanding
2306  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2307  * deal with the hole.  
2308  */
2309 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2310 {
2311         struct address_space *mapping = inode->i_mapping;
2312         struct page *page;
2313         void *fsdata;
2314         int err;
2315
2316         err = inode_newsize_ok(inode, size);
2317         if (err)
2318                 goto out;
2319
2320         err = pagecache_write_begin(NULL, mapping, size, 0,
2321                                     AOP_FLAG_CONT_EXPAND, &page, &fsdata);
2322         if (err)
2323                 goto out;
2324
2325         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2326         BUG_ON(err > 0);
2327
2328 out:
2329         return err;
2330 }
2331 EXPORT_SYMBOL(generic_cont_expand_simple);
2332
2333 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2334                             loff_t pos, loff_t *bytes)
2335 {
2336         struct inode *inode = mapping->host;
2337         unsigned int blocksize = i_blocksize(inode);
2338         struct page *page;
2339         void *fsdata;
2340         pgoff_t index, curidx;
2341         loff_t curpos;
2342         unsigned zerofrom, offset, len;
2343         int err = 0;
2344
2345         index = pos >> PAGE_SHIFT;
2346         offset = pos & ~PAGE_MASK;
2347
2348         while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2349                 zerofrom = curpos & ~PAGE_MASK;
2350                 if (zerofrom & (blocksize-1)) {
2351                         *bytes |= (blocksize-1);
2352                         (*bytes)++;
2353                 }
2354                 len = PAGE_SIZE - zerofrom;
2355
2356                 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2357                                             &page, &fsdata);
2358                 if (err)
2359                         goto out;
2360                 zero_user(page, zerofrom, len);
2361                 err = pagecache_write_end(file, mapping, curpos, len, len,
2362                                                 page, fsdata);
2363                 if (err < 0)
2364                         goto out;
2365                 BUG_ON(err != len);
2366                 err = 0;
2367
2368                 balance_dirty_pages_ratelimited(mapping);
2369
2370                 if (fatal_signal_pending(current)) {
2371                         err = -EINTR;
2372                         goto out;
2373                 }
2374         }
2375
2376         /* page covers the boundary, find the boundary offset */
2377         if (index == curidx) {
2378                 zerofrom = curpos & ~PAGE_MASK;
2379                 /* if we will expand the thing last block will be filled */
2380                 if (offset <= zerofrom) {
2381                         goto out;
2382                 }
2383                 if (zerofrom & (blocksize-1)) {
2384                         *bytes |= (blocksize-1);
2385                         (*bytes)++;
2386                 }
2387                 len = offset - zerofrom;
2388
2389                 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2390                                             &page, &fsdata);
2391                 if (err)
2392                         goto out;
2393                 zero_user(page, zerofrom, len);
2394                 err = pagecache_write_end(file, mapping, curpos, len, len,
2395                                                 page, fsdata);
2396                 if (err < 0)
2397                         goto out;
2398                 BUG_ON(err != len);
2399                 err = 0;
2400         }
2401 out:
2402         return err;
2403 }
2404
2405 /*
2406  * For moronic filesystems that do not allow holes in file.
2407  * We may have to extend the file.
2408  */
2409 int cont_write_begin(struct file *file, struct address_space *mapping,
2410                         loff_t pos, unsigned len, unsigned flags,
2411                         struct page **pagep, void **fsdata,
2412                         get_block_t *get_block, loff_t *bytes)
2413 {
2414         struct inode *inode = mapping->host;
2415         unsigned int blocksize = i_blocksize(inode);
2416         unsigned int zerofrom;
2417         int err;
2418
2419         err = cont_expand_zero(file, mapping, pos, bytes);
2420         if (err)
2421                 return err;
2422
2423         zerofrom = *bytes & ~PAGE_MASK;
2424         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2425                 *bytes |= (blocksize-1);
2426                 (*bytes)++;
2427         }
2428
2429         return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2430 }
2431 EXPORT_SYMBOL(cont_write_begin);
2432
2433 int block_commit_write(struct page *page, unsigned from, unsigned to)
2434 {
2435         struct inode *inode = page->mapping->host;
2436         __block_commit_write(inode,page,from,to);
2437         return 0;
2438 }
2439 EXPORT_SYMBOL(block_commit_write);
2440
2441 /*
2442  * block_page_mkwrite() is not allowed to change the file size as it gets
2443  * called from a page fault handler when a page is first dirtied. Hence we must
2444  * be careful to check for EOF conditions here. We set the page up correctly
2445  * for a written page which means we get ENOSPC checking when writing into
2446  * holes and correct delalloc and unwritten extent mapping on filesystems that
2447  * support these features.
2448  *
2449  * We are not allowed to take the i_mutex here so we have to play games to
2450  * protect against truncate races as the page could now be beyond EOF.  Because
2451  * truncate writes the inode size before removing pages, once we have the
2452  * page lock we can determine safely if the page is beyond EOF. If it is not
2453  * beyond EOF, then the page is guaranteed safe against truncation until we
2454  * unlock the page.
2455  *
2456  * Direct callers of this function should protect against filesystem freezing
2457  * using sb_start_pagefault() - sb_end_pagefault() functions.
2458  */
2459 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2460                          get_block_t get_block)
2461 {
2462         struct page *page = vmf->page;
2463         struct inode *inode = file_inode(vma->vm_file);
2464         unsigned long end;
2465         loff_t size;
2466         int ret;
2467
2468         lock_page(page);
2469         size = i_size_read(inode);
2470         if ((page->mapping != inode->i_mapping) ||
2471             (page_offset(page) > size)) {
2472                 /* We overload EFAULT to mean page got truncated */
2473                 ret = -EFAULT;
2474                 goto out_unlock;
2475         }
2476
2477         /* page is wholly or partially inside EOF */
2478         if (((page->index + 1) << PAGE_SHIFT) > size)
2479                 end = size & ~PAGE_MASK;
2480         else
2481                 end = PAGE_SIZE;
2482
2483         ret = __block_write_begin(page, 0, end, get_block);
2484         if (!ret)
2485                 ret = block_commit_write(page, 0, end);
2486
2487         if (unlikely(ret < 0))
2488                 goto out_unlock;
2489         set_page_dirty(page);
2490         wait_for_stable_page(page);
2491         return 0;
2492 out_unlock:
2493         unlock_page(page);
2494         return ret;
2495 }
2496 EXPORT_SYMBOL(block_page_mkwrite);
2497
2498 /*
2499  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2500  * immediately, while under the page lock.  So it needs a special end_io
2501  * handler which does not touch the bh after unlocking it.
2502  */
2503 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2504 {
2505         __end_buffer_read_notouch(bh, uptodate);
2506 }
2507
2508 /*
2509  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2510  * the page (converting it to circular linked list and taking care of page
2511  * dirty races).
2512  */
2513 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2514 {
2515         struct buffer_head *bh;
2516
2517         BUG_ON(!PageLocked(page));
2518
2519         spin_lock(&page->mapping->private_lock);
2520         bh = head;
2521         do {
2522                 if (PageDirty(page))
2523                         set_buffer_dirty(bh);
2524                 if (!bh->b_this_page)
2525                         bh->b_this_page = head;
2526                 bh = bh->b_this_page;
2527         } while (bh != head);
2528         attach_page_buffers(page, head);
2529         spin_unlock(&page->mapping->private_lock);
2530 }
2531
2532 /*
2533  * On entry, the page is fully not uptodate.
2534  * On exit the page is fully uptodate in the areas outside (from,to)
2535  * The filesystem needs to handle block truncation upon failure.
2536  */
2537 int nobh_write_begin(struct address_space *mapping,
2538                         loff_t pos, unsigned len, unsigned flags,
2539                         struct page **pagep, void **fsdata,
2540                         get_block_t *get_block)
2541 {
2542         struct inode *inode = mapping->host;
2543         const unsigned blkbits = inode->i_blkbits;
2544         const unsigned blocksize = 1 << blkbits;
2545         struct buffer_head *head, *bh;
2546         struct page *page;
2547         pgoff_t index;
2548         unsigned from, to;
2549         unsigned block_in_page;
2550         unsigned block_start, block_end;
2551         sector_t block_in_file;
2552         int nr_reads = 0;
2553         int ret = 0;
2554         int is_mapped_to_disk = 1;
2555
2556         index = pos >> PAGE_SHIFT;
2557         from = pos & (PAGE_SIZE - 1);
2558         to = from + len;
2559
2560         page = grab_cache_page_write_begin(mapping, index, flags);
2561         if (!page)
2562                 return -ENOMEM;
2563         *pagep = page;
2564         *fsdata = NULL;
2565
2566         if (page_has_buffers(page)) {
2567                 ret = __block_write_begin(page, pos, len, get_block);
2568                 if (unlikely(ret))
2569                         goto out_release;
2570                 return ret;
2571         }
2572
2573         if (PageMappedToDisk(page))
2574                 return 0;
2575
2576         /*
2577          * Allocate buffers so that we can keep track of state, and potentially
2578          * attach them to the page if an error occurs. In the common case of
2579          * no error, they will just be freed again without ever being attached
2580          * to the page (which is all OK, because we're under the page lock).
2581          *
2582          * Be careful: the buffer linked list is a NULL terminated one, rather
2583          * than the circular one we're used to.
2584          */
2585         head = alloc_page_buffers(page, blocksize, false);
2586         if (!head) {
2587                 ret = -ENOMEM;
2588                 goto out_release;
2589         }
2590
2591         block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2592
2593         /*
2594          * We loop across all blocks in the page, whether or not they are
2595          * part of the affected region.  This is so we can discover if the
2596          * page is fully mapped-to-disk.
2597          */
2598         for (block_start = 0, block_in_page = 0, bh = head;
2599                   block_start < PAGE_SIZE;
2600                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2601                 int create;
2602
2603                 block_end = block_start + blocksize;
2604                 bh->b_state = 0;
2605                 create = 1;
2606                 if (block_start >= to)
2607                         create = 0;
2608                 ret = get_block(inode, block_in_file + block_in_page,
2609                                         bh, create);
2610                 if (ret)
2611                         goto failed;
2612                 if (!buffer_mapped(bh))
2613                         is_mapped_to_disk = 0;
2614                 if (buffer_new(bh))
2615                         clean_bdev_bh_alias(bh);
2616                 if (PageUptodate(page)) {
2617                         set_buffer_uptodate(bh);
2618                         continue;
2619                 }
2620                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2621                         zero_user_segments(page, block_start, from,
2622                                                         to, block_end);
2623                         continue;
2624                 }
2625                 if (buffer_uptodate(bh))
2626                         continue;       /* reiserfs does this */
2627                 if (block_start < from || block_end > to) {
2628                         lock_buffer(bh);
2629                         bh->b_end_io = end_buffer_read_nobh;
2630                         submit_bh(REQ_OP_READ, 0, bh);
2631                         nr_reads++;
2632                 }
2633         }
2634
2635         if (nr_reads) {
2636                 /*
2637                  * The page is locked, so these buffers are protected from
2638                  * any VM or truncate activity.  Hence we don't need to care
2639                  * for the buffer_head refcounts.
2640                  */
2641                 for (bh = head; bh; bh = bh->b_this_page) {
2642                         wait_on_buffer(bh);
2643                         if (!buffer_uptodate(bh))
2644                                 ret = -EIO;
2645                 }
2646                 if (ret)
2647                         goto failed;
2648         }
2649
2650         if (is_mapped_to_disk)
2651                 SetPageMappedToDisk(page);
2652
2653         *fsdata = head; /* to be released by nobh_write_end */
2654
2655         return 0;
2656
2657 failed:
2658         BUG_ON(!ret);
2659         /*
2660          * Error recovery is a bit difficult. We need to zero out blocks that
2661          * were newly allocated, and dirty them to ensure they get written out.
2662          * Buffers need to be attached to the page at this point, otherwise
2663          * the handling of potential IO errors during writeout would be hard
2664          * (could try doing synchronous writeout, but what if that fails too?)
2665          */
2666         attach_nobh_buffers(page, head);
2667         page_zero_new_buffers(page, from, to);
2668
2669 out_release:
2670         unlock_page(page);
2671         put_page(page);
2672         *pagep = NULL;
2673
2674         return ret;
2675 }
2676 EXPORT_SYMBOL(nobh_write_begin);
2677
2678 int nobh_write_end(struct file *file, struct address_space *mapping,
2679                         loff_t pos, unsigned len, unsigned copied,
2680                         struct page *page, void *fsdata)
2681 {
2682         struct inode *inode = page->mapping->host;
2683         struct buffer_head *head = fsdata;
2684         struct buffer_head *bh;
2685         BUG_ON(fsdata != NULL && page_has_buffers(page));
2686
2687         if (unlikely(copied < len) && head)
2688                 attach_nobh_buffers(page, head);
2689         if (page_has_buffers(page))
2690                 return generic_write_end(file, mapping, pos, len,
2691                                         copied, page, fsdata);
2692
2693         SetPageUptodate(page);
2694         set_page_dirty(page);
2695         if (pos+copied > inode->i_size) {
2696                 i_size_write(inode, pos+copied);
2697                 mark_inode_dirty(inode);
2698         }
2699
2700         unlock_page(page);
2701         put_page(page);
2702
2703         while (head) {
2704                 bh = head;
2705                 head = head->b_this_page;
2706                 free_buffer_head(bh);
2707         }
2708
2709         return copied;
2710 }
2711 EXPORT_SYMBOL(nobh_write_end);
2712
2713 /*
2714  * nobh_writepage() - based on block_full_write_page() except
2715  * that it tries to operate without attaching bufferheads to
2716  * the page.
2717  */
2718 int nobh_writepage(struct page *page, get_block_t *get_block,
2719                         struct writeback_control *wbc)
2720 {
2721         struct inode * const inode = page->mapping->host;
2722         loff_t i_size = i_size_read(inode);
2723         const pgoff_t end_index = i_size >> PAGE_SHIFT;
2724         unsigned offset;
2725         int ret;
2726
2727         /* Is the page fully inside i_size? */
2728         if (page->index < end_index)
2729                 goto out;
2730
2731         /* Is the page fully outside i_size? (truncate in progress) */
2732         offset = i_size & (PAGE_SIZE-1);
2733         if (page->index >= end_index+1 || !offset) {
2734                 /*
2735                  * The page may have dirty, unmapped buffers.  For example,
2736                  * they may have been added in ext3_writepage().  Make them
2737                  * freeable here, so the page does not leak.
2738                  */
2739 #if 0
2740                 /* Not really sure about this  - do we need this ? */
2741                 if (page->mapping->a_ops->invalidatepage)
2742                         page->mapping->a_ops->invalidatepage(page, offset);
2743 #endif
2744                 unlock_page(page);
2745                 return 0; /* don't care */
2746         }
2747
2748         /*
2749          * The page straddles i_size.  It must be zeroed out on each and every
2750          * writepage invocation because it may be mmapped.  "A file is mapped
2751          * in multiples of the page size.  For a file that is not a multiple of
2752          * the  page size, the remaining memory is zeroed when mapped, and
2753          * writes to that region are not written out to the file."
2754          */
2755         zero_user_segment(page, offset, PAGE_SIZE);
2756 out:
2757         ret = mpage_writepage(page, get_block, wbc);
2758         if (ret == -EAGAIN)
2759                 ret = __block_write_full_page(inode, page, get_block, wbc,
2760                                               end_buffer_async_write);
2761         return ret;
2762 }
2763 EXPORT_SYMBOL(nobh_writepage);
2764
2765 int nobh_truncate_page(struct address_space *mapping,
2766                         loff_t from, get_block_t *get_block)
2767 {
2768         pgoff_t index = from >> PAGE_SHIFT;
2769         unsigned offset = from & (PAGE_SIZE-1);
2770         unsigned blocksize;
2771         sector_t iblock;
2772         unsigned length, pos;
2773         struct inode *inode = mapping->host;
2774         struct page *page;
2775         struct buffer_head map_bh;
2776         int err;
2777
2778         blocksize = i_blocksize(inode);
2779         length = offset & (blocksize - 1);
2780
2781         /* Block boundary? Nothing to do */
2782         if (!length)
2783                 return 0;
2784
2785         length = blocksize - length;
2786         iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2787
2788         page = grab_cache_page(mapping, index);
2789         err = -ENOMEM;
2790         if (!page)
2791                 goto out;
2792
2793         if (page_has_buffers(page)) {
2794 has_buffers:
2795                 unlock_page(page);
2796                 put_page(page);
2797                 return block_truncate_page(mapping, from, get_block);
2798         }
2799
2800         /* Find the buffer that contains "offset" */
2801         pos = blocksize;
2802         while (offset >= pos) {
2803                 iblock++;
2804                 pos += blocksize;
2805         }
2806
2807         map_bh.b_size = blocksize;
2808         map_bh.b_state = 0;
2809         err = get_block(inode, iblock, &map_bh, 0);
2810         if (err)
2811                 goto unlock;
2812         /* unmapped? It's a hole - nothing to do */
2813         if (!buffer_mapped(&map_bh))
2814                 goto unlock;
2815
2816         /* Ok, it's mapped. Make sure it's up-to-date */
2817         if (!PageUptodate(page)) {
2818                 err = mapping->a_ops->readpage(NULL, page);
2819                 if (err) {
2820                         put_page(page);
2821                         goto out;
2822                 }
2823                 lock_page(page);
2824                 if (!PageUptodate(page)) {
2825                         err = -EIO;
2826                         goto unlock;
2827                 }
2828                 if (page_has_buffers(page))
2829                         goto has_buffers;
2830         }
2831         zero_user(page, offset, length);
2832         set_page_dirty(page);
2833         err = 0;
2834
2835 unlock:
2836         unlock_page(page);
2837         put_page(page);
2838 out:
2839         return err;
2840 }
2841 EXPORT_SYMBOL(nobh_truncate_page);
2842
2843 int block_truncate_page(struct address_space *mapping,
2844                         loff_t from, get_block_t *get_block)
2845 {
2846         pgoff_t index = from >> PAGE_SHIFT;
2847         unsigned offset = from & (PAGE_SIZE-1);
2848         unsigned blocksize;
2849         sector_t iblock;
2850         unsigned length, pos;
2851         struct inode *inode = mapping->host;
2852         struct page *page;
2853         struct buffer_head *bh;
2854         int err;
2855
2856         blocksize = i_blocksize(inode);
2857         length = offset & (blocksize - 1);
2858
2859         /* Block boundary? Nothing to do */
2860         if (!length)
2861                 return 0;
2862
2863         length = blocksize - length;
2864         iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2865         
2866         page = grab_cache_page(mapping, index);
2867         err = -ENOMEM;
2868         if (!page)
2869                 goto out;
2870
2871         if (!page_has_buffers(page))
2872                 create_empty_buffers(page, blocksize, 0);
2873
2874         /* Find the buffer that contains "offset" */
2875         bh = page_buffers(page);
2876         pos = blocksize;
2877         while (offset >= pos) {
2878                 bh = bh->b_this_page;
2879                 iblock++;
2880                 pos += blocksize;
2881         }
2882
2883         err = 0;
2884         if (!buffer_mapped(bh)) {
2885                 WARN_ON(bh->b_size != blocksize);
2886                 err = get_block(inode, iblock, bh, 0);
2887                 if (err)
2888                         goto unlock;
2889                 /* unmapped? It's a hole - nothing to do */
2890                 if (!buffer_mapped(bh))
2891                         goto unlock;
2892         }
2893
2894         /* Ok, it's mapped. Make sure it's up-to-date */
2895         if (PageUptodate(page))
2896                 set_buffer_uptodate(bh);
2897
2898         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2899                 err = -EIO;
2900                 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2901                 wait_on_buffer(bh);
2902                 /* Uhhuh. Read error. Complain and punt. */
2903                 if (!buffer_uptodate(bh))
2904                         goto unlock;
2905         }
2906
2907         zero_user(page, offset, length);
2908         mark_buffer_dirty(bh);
2909         err = 0;
2910
2911 unlock:
2912         unlock_page(page);
2913         put_page(page);
2914 out:
2915         return err;
2916 }
2917 EXPORT_SYMBOL(block_truncate_page);
2918
2919 /*
2920  * The generic ->writepage function for buffer-backed address_spaces
2921  */
2922 int block_write_full_page(struct page *page, get_block_t *get_block,
2923                         struct writeback_control *wbc)
2924 {
2925         struct inode * const inode = page->mapping->host;
2926         loff_t i_size = i_size_read(inode);
2927         const pgoff_t end_index = i_size >> PAGE_SHIFT;
2928         unsigned offset;
2929
2930         /* Is the page fully inside i_size? */
2931         if (page->index < end_index)
2932                 return __block_write_full_page(inode, page, get_block, wbc,
2933                                                end_buffer_async_write);
2934
2935         /* Is the page fully outside i_size? (truncate in progress) */
2936         offset = i_size & (PAGE_SIZE-1);
2937         if (page->index >= end_index+1 || !offset) {
2938                 /*
2939                  * The page may have dirty, unmapped buffers.  For example,
2940                  * they may have been added in ext3_writepage().  Make them
2941                  * freeable here, so the page does not leak.
2942                  */
2943                 do_invalidatepage(page, 0, PAGE_SIZE);
2944                 unlock_page(page);
2945                 return 0; /* don't care */
2946         }
2947
2948         /*
2949          * The page straddles i_size.  It must be zeroed out on each and every
2950          * writepage invocation because it may be mmapped.  "A file is mapped
2951          * in multiples of the page size.  For a file that is not a multiple of
2952          * the  page size, the remaining memory is zeroed when mapped, and
2953          * writes to that region are not written out to the file."
2954          */
2955         zero_user_segment(page, offset, PAGE_SIZE);
2956         return __block_write_full_page(inode, page, get_block, wbc,
2957                                                         end_buffer_async_write);
2958 }
2959 EXPORT_SYMBOL(block_write_full_page);
2960
2961 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2962                             get_block_t *get_block)
2963 {
2964         struct inode *inode = mapping->host;
2965         struct buffer_head tmp = {
2966                 .b_size = i_blocksize(inode),
2967         };
2968
2969         get_block(inode, block, &tmp, 0);
2970         return tmp.b_blocknr;
2971 }
2972 EXPORT_SYMBOL(generic_block_bmap);
2973
2974 static void end_bio_bh_io_sync(struct bio *bio)
2975 {
2976         struct buffer_head *bh = bio->bi_private;
2977
2978         if (unlikely(bio_flagged(bio, BIO_QUIET)))
2979                 set_bit(BH_Quiet, &bh->b_state);
2980
2981         bh->b_end_io(bh, !bio->bi_status);
2982         bio_put(bio);
2983 }
2984
2985 /*
2986  * This allows us to do IO even on the odd last sectors
2987  * of a device, even if the block size is some multiple
2988  * of the physical sector size.
2989  *
2990  * We'll just truncate the bio to the size of the device,
2991  * and clear the end of the buffer head manually.
2992  *
2993  * Truly out-of-range accesses will turn into actual IO
2994  * errors, this only handles the "we need to be able to
2995  * do IO at the final sector" case.
2996  */
2997 void guard_bio_eod(int op, struct bio *bio)
2998 {
2999         sector_t maxsector;
3000         struct bio_vec *bvec = bio_last_bvec_all(bio);
3001         unsigned truncated_bytes;
3002         struct hd_struct *part;
3003
3004         rcu_read_lock();
3005         part = __disk_get_part(bio->bi_disk, bio->bi_partno);
3006         if (part)
3007                 maxsector = part_nr_sects_read(part);
3008         else
3009                 maxsector = get_capacity(bio->bi_disk);
3010         rcu_read_unlock();
3011
3012         if (!maxsector)
3013                 return;
3014
3015         /*
3016          * If the *whole* IO is past the end of the device,
3017          * let it through, and the IO layer will turn it into
3018          * an EIO.
3019          */
3020         if (unlikely(bio->bi_iter.bi_sector >= maxsector))
3021                 return;
3022
3023         maxsector -= bio->bi_iter.bi_sector;
3024         if (likely((bio->bi_iter.bi_size >> 9) <= maxsector))
3025                 return;
3026
3027         /* Uhhuh. We've got a bio that straddles the device size! */
3028         truncated_bytes = bio->bi_iter.bi_size - (maxsector << 9);
3029
3030         /*
3031          * The bio contains more than one segment which spans EOD, just return
3032          * and let IO layer turn it into an EIO
3033          */
3034         if (truncated_bytes > bvec->bv_len)
3035                 return;
3036
3037         /* Truncate the bio.. */
3038         bio->bi_iter.bi_size -= truncated_bytes;
3039         bvec->bv_len -= truncated_bytes;
3040
3041         /* ..and clear the end of the buffer for reads */
3042         if (op == REQ_OP_READ) {
3043                 struct bio_vec bv;
3044
3045                 mp_bvec_last_segment(bvec, &bv);
3046                 zero_user(bv.bv_page, bv.bv_offset + bv.bv_len,
3047                                 truncated_bytes);
3048         }
3049 }
3050
3051 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3052                          enum rw_hint write_hint, struct writeback_control *wbc)
3053 {
3054         struct bio *bio;
3055
3056         BUG_ON(!buffer_locked(bh));
3057         BUG_ON(!buffer_mapped(bh));
3058         BUG_ON(!bh->b_end_io);
3059         BUG_ON(buffer_delay(bh));
3060         BUG_ON(buffer_unwritten(bh));
3061
3062         /*
3063          * Only clear out a write error when rewriting
3064          */
3065         if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3066                 clear_buffer_write_io_error(bh);
3067
3068         /*
3069          * from here on down, it's all bio -- do the initial mapping,
3070          * submit_bio -> generic_make_request may further map this bio around
3071          */
3072         bio = bio_alloc(GFP_NOIO, 1);
3073
3074         bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3075         bio_set_dev(bio, bh->b_bdev);
3076         bio->bi_write_hint = write_hint;
3077
3078         bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3079         BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3080
3081         bio->bi_end_io = end_bio_bh_io_sync;
3082         bio->bi_private = bh;
3083
3084         /* Take care of bh's that straddle the end of the device */
3085         guard_bio_eod(op, bio);
3086
3087         if (buffer_meta(bh))
3088                 op_flags |= REQ_META;
3089         if (buffer_prio(bh))
3090                 op_flags |= REQ_PRIO;
3091         bio_set_op_attrs(bio, op, op_flags);
3092
3093         if (wbc) {
3094                 wbc_init_bio(wbc, bio);
3095                 wbc_account_io(wbc, bh->b_page, bh->b_size);
3096         }
3097
3098         submit_bio(bio);
3099         return 0;
3100 }
3101
3102 int submit_bh(int op, int op_flags, struct buffer_head *bh)
3103 {
3104         return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3105 }
3106 EXPORT_SYMBOL(submit_bh);
3107
3108 /**
3109  * ll_rw_block: low-level access to block devices (DEPRECATED)
3110  * @op: whether to %READ or %WRITE
3111  * @op_flags: req_flag_bits
3112  * @nr: number of &struct buffer_heads in the array
3113  * @bhs: array of pointers to &struct buffer_head
3114  *
3115  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3116  * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3117  * @op_flags contains flags modifying the detailed I/O behavior, most notably
3118  * %REQ_RAHEAD.
3119  *
3120  * This function drops any buffer that it cannot get a lock on (with the
3121  * BH_Lock state bit), any buffer that appears to be clean when doing a write
3122  * request, and any buffer that appears to be up-to-date when doing read
3123  * request.  Further it marks as clean buffers that are processed for
3124  * writing (the buffer cache won't assume that they are actually clean
3125  * until the buffer gets unlocked).
3126  *
3127  * ll_rw_block sets b_end_io to simple completion handler that marks
3128  * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3129  * any waiters. 
3130  *
3131  * All of the buffers must be for the same device, and must also be a
3132  * multiple of the current approved size for the device.
3133  */
3134 void ll_rw_block(int op, int op_flags,  int nr, struct buffer_head *bhs[])
3135 {
3136         int i;
3137
3138         for (i = 0; i < nr; i++) {
3139                 struct buffer_head *bh = bhs[i];
3140
3141                 if (!trylock_buffer(bh))
3142                         continue;
3143                 if (op == WRITE) {
3144                         if (test_clear_buffer_dirty(bh)) {
3145                                 bh->b_end_io = end_buffer_write_sync;
3146                                 get_bh(bh);
3147                                 submit_bh(op, op_flags, bh);
3148                                 continue;
3149                         }
3150                 } else {
3151                         if (!buffer_uptodate(bh)) {
3152                                 bh->b_end_io = end_buffer_read_sync;
3153                                 get_bh(bh);
3154                                 submit_bh(op, op_flags, bh);
3155                                 continue;
3156                         }
3157                 }
3158                 unlock_buffer(bh);
3159         }
3160 }
3161 EXPORT_SYMBOL(ll_rw_block);
3162
3163 void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3164 {
3165         lock_buffer(bh);
3166         if (!test_clear_buffer_dirty(bh)) {
3167                 unlock_buffer(bh);
3168                 return;
3169         }
3170         bh->b_end_io = end_buffer_write_sync;
3171         get_bh(bh);
3172         submit_bh(REQ_OP_WRITE, op_flags, bh);
3173 }
3174 EXPORT_SYMBOL(write_dirty_buffer);
3175
3176 /*
3177  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3178  * and then start new I/O and then wait upon it.  The caller must have a ref on
3179  * the buffer_head.
3180  */
3181 int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3182 {
3183         int ret = 0;
3184
3185         WARN_ON(atomic_read(&bh->b_count) < 1);
3186         lock_buffer(bh);
3187         if (test_clear_buffer_dirty(bh)) {
3188                 get_bh(bh);
3189                 bh->b_end_io = end_buffer_write_sync;
3190                 ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3191                 wait_on_buffer(bh);
3192                 if (!ret && !buffer_uptodate(bh))
3193                         ret = -EIO;
3194         } else {
3195                 unlock_buffer(bh);
3196         }
3197         return ret;
3198 }
3199 EXPORT_SYMBOL(__sync_dirty_buffer);
3200
3201 int sync_dirty_buffer(struct buffer_head *bh)
3202 {
3203         return __sync_dirty_buffer(bh, REQ_SYNC);
3204 }
3205 EXPORT_SYMBOL(sync_dirty_buffer);
3206
3207 /*
3208  * try_to_free_buffers() checks if all the buffers on this particular page
3209  * are unused, and releases them if so.
3210  *
3211  * Exclusion against try_to_free_buffers may be obtained by either
3212  * locking the page or by holding its mapping's private_lock.
3213  *
3214  * If the page is dirty but all the buffers are clean then we need to
3215  * be sure to mark the page clean as well.  This is because the page
3216  * may be against a block device, and a later reattachment of buffers
3217  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3218  * filesystem data on the same device.
3219  *
3220  * The same applies to regular filesystem pages: if all the buffers are
3221  * clean then we set the page clean and proceed.  To do that, we require
3222  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3223  * private_lock.
3224  *
3225  * try_to_free_buffers() is non-blocking.
3226  */
3227 static inline int buffer_busy(struct buffer_head *bh)
3228 {
3229         return atomic_read(&bh->b_count) |
3230                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3231 }
3232
3233 static int
3234 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3235 {
3236         struct buffer_head *head = page_buffers(page);
3237         struct buffer_head *bh;
3238
3239         bh = head;
3240         do {
3241                 if (buffer_busy(bh))
3242                         goto failed;
3243                 bh = bh->b_this_page;
3244         } while (bh != head);
3245
3246         do {
3247                 struct buffer_head *next = bh->b_this_page;
3248
3249                 if (bh->b_assoc_map)
3250                         __remove_assoc_queue(bh);
3251                 bh = next;
3252         } while (bh != head);
3253         *buffers_to_free = head;
3254         __clear_page_buffers(page);
3255         return 1;
3256 failed:
3257         return 0;
3258 }
3259
3260 int try_to_free_buffers(struct page *page)
3261 {
3262         struct address_space * const mapping = page->mapping;
3263         struct buffer_head *buffers_to_free = NULL;
3264         int ret = 0;
3265
3266         BUG_ON(!PageLocked(page));
3267         if (PageWriteback(page))
3268                 return 0;
3269
3270         if (mapping == NULL) {          /* can this still happen? */
3271                 ret = drop_buffers(page, &buffers_to_free);
3272                 goto out;
3273         }
3274
3275         spin_lock(&mapping->private_lock);
3276         ret = drop_buffers(page, &buffers_to_free);
3277
3278         /*
3279          * If the filesystem writes its buffers by hand (eg ext3)
3280          * then we can have clean buffers against a dirty page.  We
3281          * clean the page here; otherwise the VM will never notice
3282          * that the filesystem did any IO at all.
3283          *
3284          * Also, during truncate, discard_buffer will have marked all
3285          * the page's buffers clean.  We discover that here and clean
3286          * the page also.
3287          *
3288          * private_lock must be held over this entire operation in order
3289          * to synchronise against __set_page_dirty_buffers and prevent the
3290          * dirty bit from being lost.
3291          */
3292         if (ret)
3293                 cancel_dirty_page(page);
3294         spin_unlock(&mapping->private_lock);
3295 out:
3296         if (buffers_to_free) {
3297                 struct buffer_head *bh = buffers_to_free;
3298
3299                 do {
3300                         struct buffer_head *next = bh->b_this_page;
3301                         free_buffer_head(bh);
3302                         bh = next;
3303                 } while (bh != buffers_to_free);
3304         }
3305         return ret;
3306 }
3307 EXPORT_SYMBOL(try_to_free_buffers);
3308
3309 /*
3310  * There are no bdflush tunables left.  But distributions are
3311  * still running obsolete flush daemons, so we terminate them here.
3312  *
3313  * Use of bdflush() is deprecated and will be removed in a future kernel.
3314  * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3315  */
3316 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3317 {
3318         static int msg_count;
3319
3320         if (!capable(CAP_SYS_ADMIN))
3321                 return -EPERM;
3322
3323         if (msg_count < 5) {
3324                 msg_count++;
3325                 printk(KERN_INFO
3326                         "warning: process `%s' used the obsolete bdflush"
3327                         " system call\n", current->comm);
3328                 printk(KERN_INFO "Fix your initscripts?\n");
3329         }
3330
3331         if (func == 1)
3332                 do_exit(0);
3333         return 0;
3334 }
3335
3336 /*
3337  * Buffer-head allocation
3338  */
3339 static struct kmem_cache *bh_cachep __read_mostly;
3340
3341 /*
3342  * Once the number of bh's in the machine exceeds this level, we start
3343  * stripping them in writeback.
3344  */
3345 static unsigned long max_buffer_heads;
3346
3347 int buffer_heads_over_limit;
3348
3349 struct bh_accounting {
3350         int nr;                 /* Number of live bh's */
3351         int ratelimit;          /* Limit cacheline bouncing */
3352 };
3353
3354 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3355
3356 static void recalc_bh_state(void)
3357 {
3358         int i;
3359         int tot = 0;
3360
3361         if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3362                 return;
3363         __this_cpu_write(bh_accounting.ratelimit, 0);
3364         for_each_online_cpu(i)
3365                 tot += per_cpu(bh_accounting, i).nr;
3366         buffer_heads_over_limit = (tot > max_buffer_heads);
3367 }
3368
3369 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3370 {
3371         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3372         if (ret) {
3373                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3374                 preempt_disable();
3375                 __this_cpu_inc(bh_accounting.nr);
3376                 recalc_bh_state();
3377                 preempt_enable();
3378         }
3379         return ret;
3380 }
3381 EXPORT_SYMBOL(alloc_buffer_head);
3382
3383 void free_buffer_head(struct buffer_head *bh)
3384 {
3385         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3386         kmem_cache_free(bh_cachep, bh);
3387         preempt_disable();
3388         __this_cpu_dec(bh_accounting.nr);
3389         recalc_bh_state();
3390         preempt_enable();
3391 }
3392 EXPORT_SYMBOL(free_buffer_head);
3393
3394 static int buffer_exit_cpu_dead(unsigned int cpu)
3395 {
3396         int i;
33