Merge tag 'xfs-4.13-merge-5' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux
[muen/linux.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/notifier.h>
43 #include <linux/cpu.h>
44 #include <linux/bitops.h>
45 #include <linux/mpage.h>
46 #include <linux/bit_spinlock.h>
47 #include <linux/pagevec.h>
48 #include <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 void init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
57 {
58         bh->b_end_io = handler;
59         bh->b_private = private;
60 }
61 EXPORT_SYMBOL(init_buffer);
62
63 inline void touch_buffer(struct buffer_head *bh)
64 {
65         trace_block_touch_buffer(bh);
66         mark_page_accessed(bh->b_page);
67 }
68 EXPORT_SYMBOL(touch_buffer);
69
70 void __lock_buffer(struct buffer_head *bh)
71 {
72         wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
73 }
74 EXPORT_SYMBOL(__lock_buffer);
75
76 void unlock_buffer(struct buffer_head *bh)
77 {
78         clear_bit_unlock(BH_Lock, &bh->b_state);
79         smp_mb__after_atomic();
80         wake_up_bit(&bh->b_state, BH_Lock);
81 }
82 EXPORT_SYMBOL(unlock_buffer);
83
84 /*
85  * Returns if the page has dirty or writeback buffers. If all the buffers
86  * are unlocked and clean then the PageDirty information is stale. If
87  * any of the pages are locked, it is assumed they are locked for IO.
88  */
89 void buffer_check_dirty_writeback(struct page *page,
90                                      bool *dirty, bool *writeback)
91 {
92         struct buffer_head *head, *bh;
93         *dirty = false;
94         *writeback = false;
95
96         BUG_ON(!PageLocked(page));
97
98         if (!page_has_buffers(page))
99                 return;
100
101         if (PageWriteback(page))
102                 *writeback = true;
103
104         head = page_buffers(page);
105         bh = head;
106         do {
107                 if (buffer_locked(bh))
108                         *writeback = true;
109
110                 if (buffer_dirty(bh))
111                         *dirty = true;
112
113                 bh = bh->b_this_page;
114         } while (bh != head);
115 }
116 EXPORT_SYMBOL(buffer_check_dirty_writeback);
117
118 /*
119  * Block until a buffer comes unlocked.  This doesn't stop it
120  * from becoming locked again - you have to lock it yourself
121  * if you want to preserve its state.
122  */
123 void __wait_on_buffer(struct buffer_head * bh)
124 {
125         wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
126 }
127 EXPORT_SYMBOL(__wait_on_buffer);
128
129 static void
130 __clear_page_buffers(struct page *page)
131 {
132         ClearPagePrivate(page);
133         set_page_private(page, 0);
134         put_page(page);
135 }
136
137 static void buffer_io_error(struct buffer_head *bh, char *msg)
138 {
139         if (!test_bit(BH_Quiet, &bh->b_state))
140                 printk_ratelimited(KERN_ERR
141                         "Buffer I/O error on dev %pg, logical block %llu%s\n",
142                         bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
143 }
144
145 /*
146  * End-of-IO handler helper function which does not touch the bh after
147  * unlocking it.
148  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
149  * a race there is benign: unlock_buffer() only use the bh's address for
150  * hashing after unlocking the buffer, so it doesn't actually touch the bh
151  * itself.
152  */
153 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
154 {
155         if (uptodate) {
156                 set_buffer_uptodate(bh);
157         } else {
158                 /* This happens, due to failed read-ahead attempts. */
159                 clear_buffer_uptodate(bh);
160         }
161         unlock_buffer(bh);
162 }
163
164 /*
165  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
166  * unlock the buffer. This is what ll_rw_block uses too.
167  */
168 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
169 {
170         __end_buffer_read_notouch(bh, uptodate);
171         put_bh(bh);
172 }
173 EXPORT_SYMBOL(end_buffer_read_sync);
174
175 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
176 {
177         if (uptodate) {
178                 set_buffer_uptodate(bh);
179         } else {
180                 buffer_io_error(bh, ", lost sync page write");
181                 mark_buffer_write_io_error(bh);
182                 clear_buffer_uptodate(bh);
183         }
184         unlock_buffer(bh);
185         put_bh(bh);
186 }
187 EXPORT_SYMBOL(end_buffer_write_sync);
188
189 /*
190  * Various filesystems appear to want __find_get_block to be non-blocking.
191  * But it's the page lock which protects the buffers.  To get around this,
192  * we get exclusion from try_to_free_buffers with the blockdev mapping's
193  * private_lock.
194  *
195  * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
196  * may be quite high.  This code could TryLock the page, and if that
197  * succeeds, there is no need to take private_lock. (But if
198  * private_lock is contended then so is mapping->tree_lock).
199  */
200 static struct buffer_head *
201 __find_get_block_slow(struct block_device *bdev, sector_t block)
202 {
203         struct inode *bd_inode = bdev->bd_inode;
204         struct address_space *bd_mapping = bd_inode->i_mapping;
205         struct buffer_head *ret = NULL;
206         pgoff_t index;
207         struct buffer_head *bh;
208         struct buffer_head *head;
209         struct page *page;
210         int all_mapped = 1;
211
212         index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
213         page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
214         if (!page)
215                 goto out;
216
217         spin_lock(&bd_mapping->private_lock);
218         if (!page_has_buffers(page))
219                 goto out_unlock;
220         head = page_buffers(page);
221         bh = head;
222         do {
223                 if (!buffer_mapped(bh))
224                         all_mapped = 0;
225                 else if (bh->b_blocknr == block) {
226                         ret = bh;
227                         get_bh(bh);
228                         goto out_unlock;
229                 }
230                 bh = bh->b_this_page;
231         } while (bh != head);
232
233         /* we might be here because some of the buffers on this page are
234          * not mapped.  This is due to various races between
235          * file io on the block device and getblk.  It gets dealt with
236          * elsewhere, don't buffer_error if we had some unmapped buffers
237          */
238         if (all_mapped) {
239                 printk("__find_get_block_slow() failed. "
240                         "block=%llu, b_blocknr=%llu\n",
241                         (unsigned long long)block,
242                         (unsigned long long)bh->b_blocknr);
243                 printk("b_state=0x%08lx, b_size=%zu\n",
244                         bh->b_state, bh->b_size);
245                 printk("device %pg blocksize: %d\n", bdev,
246                         1 << bd_inode->i_blkbits);
247         }
248 out_unlock:
249         spin_unlock(&bd_mapping->private_lock);
250         put_page(page);
251 out:
252         return ret;
253 }
254
255 /*
256  * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
257  */
258 static void free_more_memory(void)
259 {
260         struct zoneref *z;
261         int nid;
262
263         wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
264         yield();
265
266         for_each_online_node(nid) {
267
268                 z = first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
269                                                 gfp_zone(GFP_NOFS), NULL);
270                 if (z->zone)
271                         try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
272                                                 GFP_NOFS, NULL);
273         }
274 }
275
276 /*
277  * I/O completion handler for block_read_full_page() - pages
278  * which come unlocked at the end of I/O.
279  */
280 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
281 {
282         unsigned long flags;
283         struct buffer_head *first;
284         struct buffer_head *tmp;
285         struct page *page;
286         int page_uptodate = 1;
287
288         BUG_ON(!buffer_async_read(bh));
289
290         page = bh->b_page;
291         if (uptodate) {
292                 set_buffer_uptodate(bh);
293         } else {
294                 clear_buffer_uptodate(bh);
295                 buffer_io_error(bh, ", async page read");
296                 SetPageError(page);
297         }
298
299         /*
300          * Be _very_ careful from here on. Bad things can happen if
301          * two buffer heads end IO at almost the same time and both
302          * decide that the page is now completely done.
303          */
304         first = page_buffers(page);
305         local_irq_save(flags);
306         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
307         clear_buffer_async_read(bh);
308         unlock_buffer(bh);
309         tmp = bh;
310         do {
311                 if (!buffer_uptodate(tmp))
312                         page_uptodate = 0;
313                 if (buffer_async_read(tmp)) {
314                         BUG_ON(!buffer_locked(tmp));
315                         goto still_busy;
316                 }
317                 tmp = tmp->b_this_page;
318         } while (tmp != bh);
319         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
320         local_irq_restore(flags);
321
322         /*
323          * If none of the buffers had errors and they are all
324          * uptodate then we can set the page uptodate.
325          */
326         if (page_uptodate && !PageError(page))
327                 SetPageUptodate(page);
328         unlock_page(page);
329         return;
330
331 still_busy:
332         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
333         local_irq_restore(flags);
334         return;
335 }
336
337 /*
338  * Completion handler for block_write_full_page() - pages which are unlocked
339  * during I/O, and which have PageWriteback cleared upon I/O completion.
340  */
341 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
342 {
343         unsigned long flags;
344         struct buffer_head *first;
345         struct buffer_head *tmp;
346         struct page *page;
347
348         BUG_ON(!buffer_async_write(bh));
349
350         page = bh->b_page;
351         if (uptodate) {
352                 set_buffer_uptodate(bh);
353         } else {
354                 buffer_io_error(bh, ", lost async page write");
355                 mark_buffer_write_io_error(bh);
356                 clear_buffer_uptodate(bh);
357                 SetPageError(page);
358         }
359
360         first = page_buffers(page);
361         local_irq_save(flags);
362         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
363
364         clear_buffer_async_write(bh);
365         unlock_buffer(bh);
366         tmp = bh->b_this_page;
367         while (tmp != bh) {
368                 if (buffer_async_write(tmp)) {
369                         BUG_ON(!buffer_locked(tmp));
370                         goto still_busy;
371                 }
372                 tmp = tmp->b_this_page;
373         }
374         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
375         local_irq_restore(flags);
376         end_page_writeback(page);
377         return;
378
379 still_busy:
380         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
381         local_irq_restore(flags);
382         return;
383 }
384 EXPORT_SYMBOL(end_buffer_async_write);
385
386 /*
387  * If a page's buffers are under async readin (end_buffer_async_read
388  * completion) then there is a possibility that another thread of
389  * control could lock one of the buffers after it has completed
390  * but while some of the other buffers have not completed.  This
391  * locked buffer would confuse end_buffer_async_read() into not unlocking
392  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
393  * that this buffer is not under async I/O.
394  *
395  * The page comes unlocked when it has no locked buffer_async buffers
396  * left.
397  *
398  * PageLocked prevents anyone starting new async I/O reads any of
399  * the buffers.
400  *
401  * PageWriteback is used to prevent simultaneous writeout of the same
402  * page.
403  *
404  * PageLocked prevents anyone from starting writeback of a page which is
405  * under read I/O (PageWriteback is only ever set against a locked page).
406  */
407 static void mark_buffer_async_read(struct buffer_head *bh)
408 {
409         bh->b_end_io = end_buffer_async_read;
410         set_buffer_async_read(bh);
411 }
412
413 static void mark_buffer_async_write_endio(struct buffer_head *bh,
414                                           bh_end_io_t *handler)
415 {
416         bh->b_end_io = handler;
417         set_buffer_async_write(bh);
418 }
419
420 void mark_buffer_async_write(struct buffer_head *bh)
421 {
422         mark_buffer_async_write_endio(bh, end_buffer_async_write);
423 }
424 EXPORT_SYMBOL(mark_buffer_async_write);
425
426
427 /*
428  * fs/buffer.c contains helper functions for buffer-backed address space's
429  * fsync functions.  A common requirement for buffer-based filesystems is
430  * that certain data from the backing blockdev needs to be written out for
431  * a successful fsync().  For example, ext2 indirect blocks need to be
432  * written back and waited upon before fsync() returns.
433  *
434  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
435  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
436  * management of a list of dependent buffers at ->i_mapping->private_list.
437  *
438  * Locking is a little subtle: try_to_free_buffers() will remove buffers
439  * from their controlling inode's queue when they are being freed.  But
440  * try_to_free_buffers() will be operating against the *blockdev* mapping
441  * at the time, not against the S_ISREG file which depends on those buffers.
442  * So the locking for private_list is via the private_lock in the address_space
443  * which backs the buffers.  Which is different from the address_space 
444  * against which the buffers are listed.  So for a particular address_space,
445  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
446  * mapping->private_list will always be protected by the backing blockdev's
447  * ->private_lock.
448  *
449  * Which introduces a requirement: all buffers on an address_space's
450  * ->private_list must be from the same address_space: the blockdev's.
451  *
452  * address_spaces which do not place buffers at ->private_list via these
453  * utility functions are free to use private_lock and private_list for
454  * whatever they want.  The only requirement is that list_empty(private_list)
455  * be true at clear_inode() time.
456  *
457  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
458  * filesystems should do that.  invalidate_inode_buffers() should just go
459  * BUG_ON(!list_empty).
460  *
461  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
462  * take an address_space, not an inode.  And it should be called
463  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
464  * queued up.
465  *
466  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
467  * list if it is already on a list.  Because if the buffer is on a list,
468  * it *must* already be on the right one.  If not, the filesystem is being
469  * silly.  This will save a ton of locking.  But first we have to ensure
470  * that buffers are taken *off* the old inode's list when they are freed
471  * (presumably in truncate).  That requires careful auditing of all
472  * filesystems (do it inside bforget()).  It could also be done by bringing
473  * b_inode back.
474  */
475
476 /*
477  * The buffer's backing address_space's private_lock must be held
478  */
479 static void __remove_assoc_queue(struct buffer_head *bh)
480 {
481         list_del_init(&bh->b_assoc_buffers);
482         WARN_ON(!bh->b_assoc_map);
483         bh->b_assoc_map = NULL;
484 }
485
486 int inode_has_buffers(struct inode *inode)
487 {
488         return !list_empty(&inode->i_data.private_list);
489 }
490
491 /*
492  * osync is designed to support O_SYNC io.  It waits synchronously for
493  * all already-submitted IO to complete, but does not queue any new
494  * writes to the disk.
495  *
496  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
497  * you dirty the buffers, and then use osync_inode_buffers to wait for
498  * completion.  Any other dirty buffers which are not yet queued for
499  * write will not be flushed to disk by the osync.
500  */
501 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
502 {
503         struct buffer_head *bh;
504         struct list_head *p;
505         int err = 0;
506
507         spin_lock(lock);
508 repeat:
509         list_for_each_prev(p, list) {
510                 bh = BH_ENTRY(p);
511                 if (buffer_locked(bh)) {
512                         get_bh(bh);
513                         spin_unlock(lock);
514                         wait_on_buffer(bh);
515                         if (!buffer_uptodate(bh))
516                                 err = -EIO;
517                         brelse(bh);
518                         spin_lock(lock);
519                         goto repeat;
520                 }
521         }
522         spin_unlock(lock);
523         return err;
524 }
525
526 static void do_thaw_one(struct super_block *sb, void *unused)
527 {
528         while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
529                 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
530 }
531
532 static void do_thaw_all(struct work_struct *work)
533 {
534         iterate_supers(do_thaw_one, NULL);
535         kfree(work);
536         printk(KERN_WARNING "Emergency Thaw complete\n");
537 }
538
539 /**
540  * emergency_thaw_all -- forcibly thaw every frozen filesystem
541  *
542  * Used for emergency unfreeze of all filesystems via SysRq
543  */
544 void emergency_thaw_all(void)
545 {
546         struct work_struct *work;
547
548         work = kmalloc(sizeof(*work), GFP_ATOMIC);
549         if (work) {
550                 INIT_WORK(work, do_thaw_all);
551                 schedule_work(work);
552         }
553 }
554
555 /**
556  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
557  * @mapping: the mapping which wants those buffers written
558  *
559  * Starts I/O against the buffers at mapping->private_list, and waits upon
560  * that I/O.
561  *
562  * Basically, this is a convenience function for fsync().
563  * @mapping is a file or directory which needs those buffers to be written for
564  * a successful fsync().
565  */
566 int sync_mapping_buffers(struct address_space *mapping)
567 {
568         struct address_space *buffer_mapping = mapping->private_data;
569
570         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
571                 return 0;
572
573         return fsync_buffers_list(&buffer_mapping->private_lock,
574                                         &mapping->private_list);
575 }
576 EXPORT_SYMBOL(sync_mapping_buffers);
577
578 /*
579  * Called when we've recently written block `bblock', and it is known that
580  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
581  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
582  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
583  */
584 void write_boundary_block(struct block_device *bdev,
585                         sector_t bblock, unsigned blocksize)
586 {
587         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
588         if (bh) {
589                 if (buffer_dirty(bh))
590                         ll_rw_block(REQ_OP_WRITE, 0, 1, &bh);
591                 put_bh(bh);
592         }
593 }
594
595 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
596 {
597         struct address_space *mapping = inode->i_mapping;
598         struct address_space *buffer_mapping = bh->b_page->mapping;
599
600         mark_buffer_dirty(bh);
601         if (!mapping->private_data) {
602                 mapping->private_data = buffer_mapping;
603         } else {
604                 BUG_ON(mapping->private_data != buffer_mapping);
605         }
606         if (!bh->b_assoc_map) {
607                 spin_lock(&buffer_mapping->private_lock);
608                 list_move_tail(&bh->b_assoc_buffers,
609                                 &mapping->private_list);
610                 bh->b_assoc_map = mapping;
611                 spin_unlock(&buffer_mapping->private_lock);
612         }
613 }
614 EXPORT_SYMBOL(mark_buffer_dirty_inode);
615
616 /*
617  * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
618  * dirty.
619  *
620  * If warn is true, then emit a warning if the page is not uptodate and has
621  * not been truncated.
622  *
623  * The caller must hold lock_page_memcg().
624  */
625 static void __set_page_dirty(struct page *page, struct address_space *mapping,
626                              int warn)
627 {
628         unsigned long flags;
629
630         spin_lock_irqsave(&mapping->tree_lock, flags);
631         if (page->mapping) {    /* Race with truncate? */
632                 WARN_ON_ONCE(warn && !PageUptodate(page));
633                 account_page_dirtied(page, mapping);
634                 radix_tree_tag_set(&mapping->page_tree,
635                                 page_index(page), PAGECACHE_TAG_DIRTY);
636         }
637         spin_unlock_irqrestore(&mapping->tree_lock, flags);
638 }
639
640 /*
641  * Add a page to the dirty page list.
642  *
643  * It is a sad fact of life that this function is called from several places
644  * deeply under spinlocking.  It may not sleep.
645  *
646  * If the page has buffers, the uptodate buffers are set dirty, to preserve
647  * dirty-state coherency between the page and the buffers.  It the page does
648  * not have buffers then when they are later attached they will all be set
649  * dirty.
650  *
651  * The buffers are dirtied before the page is dirtied.  There's a small race
652  * window in which a writepage caller may see the page cleanness but not the
653  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
654  * before the buffers, a concurrent writepage caller could clear the page dirty
655  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
656  * page on the dirty page list.
657  *
658  * We use private_lock to lock against try_to_free_buffers while using the
659  * page's buffer list.  Also use this to protect against clean buffers being
660  * added to the page after it was set dirty.
661  *
662  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
663  * address_space though.
664  */
665 int __set_page_dirty_buffers(struct page *page)
666 {
667         int newly_dirty;
668         struct address_space *mapping = page_mapping(page);
669
670         if (unlikely(!mapping))
671                 return !TestSetPageDirty(page);
672
673         spin_lock(&mapping->private_lock);
674         if (page_has_buffers(page)) {
675                 struct buffer_head *head = page_buffers(page);
676                 struct buffer_head *bh = head;
677
678                 do {
679                         set_buffer_dirty(bh);
680                         bh = bh->b_this_page;
681                 } while (bh != head);
682         }
683         /*
684          * Lock out page->mem_cgroup migration to keep PageDirty
685          * synchronized with per-memcg dirty page counters.
686          */
687         lock_page_memcg(page);
688         newly_dirty = !TestSetPageDirty(page);
689         spin_unlock(&mapping->private_lock);
690
691         if (newly_dirty)
692                 __set_page_dirty(page, mapping, 1);
693
694         unlock_page_memcg(page);
695
696         if (newly_dirty)
697                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
698
699         return newly_dirty;
700 }
701 EXPORT_SYMBOL(__set_page_dirty_buffers);
702
703 /*
704  * Write out and wait upon a list of buffers.
705  *
706  * We have conflicting pressures: we want to make sure that all
707  * initially dirty buffers get waited on, but that any subsequently
708  * dirtied buffers don't.  After all, we don't want fsync to last
709  * forever if somebody is actively writing to the file.
710  *
711  * Do this in two main stages: first we copy dirty buffers to a
712  * temporary inode list, queueing the writes as we go.  Then we clean
713  * up, waiting for those writes to complete.
714  * 
715  * During this second stage, any subsequent updates to the file may end
716  * up refiling the buffer on the original inode's dirty list again, so
717  * there is a chance we will end up with a buffer queued for write but
718  * not yet completed on that list.  So, as a final cleanup we go through
719  * the osync code to catch these locked, dirty buffers without requeuing
720  * any newly dirty buffers for write.
721  */
722 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
723 {
724         struct buffer_head *bh;
725         struct list_head tmp;
726         struct address_space *mapping;
727         int err = 0, err2;
728         struct blk_plug plug;
729
730         INIT_LIST_HEAD(&tmp);
731         blk_start_plug(&plug);
732
733         spin_lock(lock);
734         while (!list_empty(list)) {
735                 bh = BH_ENTRY(list->next);
736                 mapping = bh->b_assoc_map;
737                 __remove_assoc_queue(bh);
738                 /* Avoid race with mark_buffer_dirty_inode() which does
739                  * a lockless check and we rely on seeing the dirty bit */
740                 smp_mb();
741                 if (buffer_dirty(bh) || buffer_locked(bh)) {
742                         list_add(&bh->b_assoc_buffers, &tmp);
743                         bh->b_assoc_map = mapping;
744                         if (buffer_dirty(bh)) {
745                                 get_bh(bh);
746                                 spin_unlock(lock);
747                                 /*
748                                  * Ensure any pending I/O completes so that
749                                  * write_dirty_buffer() actually writes the
750                                  * current contents - it is a noop if I/O is
751                                  * still in flight on potentially older
752                                  * contents.
753                                  */
754                                 write_dirty_buffer(bh, REQ_SYNC);
755
756                                 /*
757                                  * Kick off IO for the previous mapping. Note
758                                  * that we will not run the very last mapping,
759                                  * wait_on_buffer() will do that for us
760                                  * through sync_buffer().
761                                  */
762                                 brelse(bh);
763                                 spin_lock(lock);
764                         }
765                 }
766         }
767
768         spin_unlock(lock);
769         blk_finish_plug(&plug);
770         spin_lock(lock);
771
772         while (!list_empty(&tmp)) {
773                 bh = BH_ENTRY(tmp.prev);
774                 get_bh(bh);
775                 mapping = bh->b_assoc_map;
776                 __remove_assoc_queue(bh);
777                 /* Avoid race with mark_buffer_dirty_inode() which does
778                  * a lockless check and we rely on seeing the dirty bit */
779                 smp_mb();
780                 if (buffer_dirty(bh)) {
781                         list_add(&bh->b_assoc_buffers,
782                                  &mapping->private_list);
783                         bh->b_assoc_map = mapping;
784                 }
785                 spin_unlock(lock);
786                 wait_on_buffer(bh);
787                 if (!buffer_uptodate(bh))
788                         err = -EIO;
789                 brelse(bh);
790                 spin_lock(lock);
791         }
792         
793         spin_unlock(lock);
794         err2 = osync_buffers_list(lock, list);
795         if (err)
796                 return err;
797         else
798                 return err2;
799 }
800
801 /*
802  * Invalidate any and all dirty buffers on a given inode.  We are
803  * probably unmounting the fs, but that doesn't mean we have already
804  * done a sync().  Just drop the buffers from the inode list.
805  *
806  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
807  * assumes that all the buffers are against the blockdev.  Not true
808  * for reiserfs.
809  */
810 void invalidate_inode_buffers(struct inode *inode)
811 {
812         if (inode_has_buffers(inode)) {
813                 struct address_space *mapping = &inode->i_data;
814                 struct list_head *list = &mapping->private_list;
815                 struct address_space *buffer_mapping = mapping->private_data;
816
817                 spin_lock(&buffer_mapping->private_lock);
818                 while (!list_empty(list))
819                         __remove_assoc_queue(BH_ENTRY(list->next));
820                 spin_unlock(&buffer_mapping->private_lock);
821         }
822 }
823 EXPORT_SYMBOL(invalidate_inode_buffers);
824
825 /*
826  * Remove any clean buffers from the inode's buffer list.  This is called
827  * when we're trying to free the inode itself.  Those buffers can pin it.
828  *
829  * Returns true if all buffers were removed.
830  */
831 int remove_inode_buffers(struct inode *inode)
832 {
833         int ret = 1;
834
835         if (inode_has_buffers(inode)) {
836                 struct address_space *mapping = &inode->i_data;
837                 struct list_head *list = &mapping->private_list;
838                 struct address_space *buffer_mapping = mapping->private_data;
839
840                 spin_lock(&buffer_mapping->private_lock);
841                 while (!list_empty(list)) {
842                         struct buffer_head *bh = BH_ENTRY(list->next);
843                         if (buffer_dirty(bh)) {
844                                 ret = 0;
845                                 break;
846                         }
847                         __remove_assoc_queue(bh);
848                 }
849                 spin_unlock(&buffer_mapping->private_lock);
850         }
851         return ret;
852 }
853
854 /*
855  * Create the appropriate buffers when given a page for data area and
856  * the size of each buffer.. Use the bh->b_this_page linked list to
857  * follow the buffers created.  Return NULL if unable to create more
858  * buffers.
859  *
860  * The retry flag is used to differentiate async IO (paging, swapping)
861  * which may not fail from ordinary buffer allocations.
862  */
863 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
864                 int retry)
865 {
866         struct buffer_head *bh, *head;
867         long offset;
868
869 try_again:
870         head = NULL;
871         offset = PAGE_SIZE;
872         while ((offset -= size) >= 0) {
873                 bh = alloc_buffer_head(GFP_NOFS);
874                 if (!bh)
875                         goto no_grow;
876
877                 bh->b_this_page = head;
878                 bh->b_blocknr = -1;
879                 head = bh;
880
881                 bh->b_size = size;
882
883                 /* Link the buffer to its page */
884                 set_bh_page(bh, page, offset);
885         }
886         return head;
887 /*
888  * In case anything failed, we just free everything we got.
889  */
890 no_grow:
891         if (head) {
892                 do {
893                         bh = head;
894                         head = head->b_this_page;
895                         free_buffer_head(bh);
896                 } while (head);
897         }
898
899         /*
900          * Return failure for non-async IO requests.  Async IO requests
901          * are not allowed to fail, so we have to wait until buffer heads
902          * become available.  But we don't want tasks sleeping with 
903          * partially complete buffers, so all were released above.
904          */
905         if (!retry)
906                 return NULL;
907
908         /* We're _really_ low on memory. Now we just
909          * wait for old buffer heads to become free due to
910          * finishing IO.  Since this is an async request and
911          * the reserve list is empty, we're sure there are 
912          * async buffer heads in use.
913          */
914         free_more_memory();
915         goto try_again;
916 }
917 EXPORT_SYMBOL_GPL(alloc_page_buffers);
918
919 static inline void
920 link_dev_buffers(struct page *page, struct buffer_head *head)
921 {
922         struct buffer_head *bh, *tail;
923
924         bh = head;
925         do {
926                 tail = bh;
927                 bh = bh->b_this_page;
928         } while (bh);
929         tail->b_this_page = head;
930         attach_page_buffers(page, head);
931 }
932
933 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
934 {
935         sector_t retval = ~((sector_t)0);
936         loff_t sz = i_size_read(bdev->bd_inode);
937
938         if (sz) {
939                 unsigned int sizebits = blksize_bits(size);
940                 retval = (sz >> sizebits);
941         }
942         return retval;
943 }
944
945 /*
946  * Initialise the state of a blockdev page's buffers.
947  */ 
948 static sector_t
949 init_page_buffers(struct page *page, struct block_device *bdev,
950                         sector_t block, int size)
951 {
952         struct buffer_head *head = page_buffers(page);
953         struct buffer_head *bh = head;
954         int uptodate = PageUptodate(page);
955         sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
956
957         do {
958                 if (!buffer_mapped(bh)) {
959                         init_buffer(bh, NULL, NULL);
960                         bh->b_bdev = bdev;
961                         bh->b_blocknr = block;
962                         if (uptodate)
963                                 set_buffer_uptodate(bh);
964                         if (block < end_block)
965                                 set_buffer_mapped(bh);
966                 }
967                 block++;
968                 bh = bh->b_this_page;
969         } while (bh != head);
970
971         /*
972          * Caller needs to validate requested block against end of device.
973          */
974         return end_block;
975 }
976
977 /*
978  * Create the page-cache page that contains the requested block.
979  *
980  * This is used purely for blockdev mappings.
981  */
982 static int
983 grow_dev_page(struct block_device *bdev, sector_t block,
984               pgoff_t index, int size, int sizebits, gfp_t gfp)
985 {
986         struct inode *inode = bdev->bd_inode;
987         struct page *page;
988         struct buffer_head *bh;
989         sector_t end_block;
990         int ret = 0;            /* Will call free_more_memory() */
991         gfp_t gfp_mask;
992
993         gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
994
995         /*
996          * XXX: __getblk_slow() can not really deal with failure and
997          * will endlessly loop on improvised global reclaim.  Prefer
998          * looping in the allocator rather than here, at least that
999          * code knows what it's doing.
1000          */
1001         gfp_mask |= __GFP_NOFAIL;
1002
1003         page = find_or_create_page(inode->i_mapping, index, gfp_mask);
1004         if (!page)
1005                 return ret;
1006
1007         BUG_ON(!PageLocked(page));
1008
1009         if (page_has_buffers(page)) {
1010                 bh = page_buffers(page);
1011                 if (bh->b_size == size) {
1012                         end_block = init_page_buffers(page, bdev,
1013                                                 (sector_t)index << sizebits,
1014                                                 size);
1015                         goto done;
1016                 }
1017                 if (!try_to_free_buffers(page))
1018                         goto failed;
1019         }
1020
1021         /*
1022          * Allocate some buffers for this page
1023          */
1024         bh = alloc_page_buffers(page, size, 0);
1025         if (!bh)
1026                 goto failed;
1027
1028         /*
1029          * Link the page to the buffers and initialise them.  Take the
1030          * lock to be atomic wrt __find_get_block(), which does not
1031          * run under the page lock.
1032          */
1033         spin_lock(&inode->i_mapping->private_lock);
1034         link_dev_buffers(page, bh);
1035         end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1036                         size);
1037         spin_unlock(&inode->i_mapping->private_lock);
1038 done:
1039         ret = (block < end_block) ? 1 : -ENXIO;
1040 failed:
1041         unlock_page(page);
1042         put_page(page);
1043         return ret;
1044 }
1045
1046 /*
1047  * Create buffers for the specified block device block's page.  If
1048  * that page was dirty, the buffers are set dirty also.
1049  */
1050 static int
1051 grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
1052 {
1053         pgoff_t index;
1054         int sizebits;
1055
1056         sizebits = -1;
1057         do {
1058                 sizebits++;
1059         } while ((size << sizebits) < PAGE_SIZE);
1060
1061         index = block >> sizebits;
1062
1063         /*
1064          * Check for a block which wants to lie outside our maximum possible
1065          * pagecache index.  (this comparison is done using sector_t types).
1066          */
1067         if (unlikely(index != block >> sizebits)) {
1068                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1069                         "device %pg\n",
1070                         __func__, (unsigned long long)block,
1071                         bdev);
1072                 return -EIO;
1073         }
1074
1075         /* Create a page with the proper size buffers.. */
1076         return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1077 }
1078
1079 static struct buffer_head *
1080 __getblk_slow(struct block_device *bdev, sector_t block,
1081              unsigned size, gfp_t gfp)
1082 {
1083         /* Size must be multiple of hard sectorsize */
1084         if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1085                         (size < 512 || size > PAGE_SIZE))) {
1086                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1087                                         size);
1088                 printk(KERN_ERR "logical block size: %d\n",
1089                                         bdev_logical_block_size(bdev));
1090
1091                 dump_stack();
1092                 return NULL;
1093         }
1094
1095         for (;;) {
1096                 struct buffer_head *bh;
1097                 int ret;
1098
1099                 bh = __find_get_block(bdev, block, size);
1100                 if (bh)
1101                         return bh;
1102
1103                 ret = grow_buffers(bdev, block, size, gfp);
1104                 if (ret < 0)
1105                         return NULL;
1106                 if (ret == 0)
1107                         free_more_memory();
1108         }
1109 }
1110
1111 /*
1112  * The relationship between dirty buffers and dirty pages:
1113  *
1114  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1115  * the page is tagged dirty in its radix tree.
1116  *
1117  * At all times, the dirtiness of the buffers represents the dirtiness of
1118  * subsections of the page.  If the page has buffers, the page dirty bit is
1119  * merely a hint about the true dirty state.
1120  *
1121  * When a page is set dirty in its entirety, all its buffers are marked dirty
1122  * (if the page has buffers).
1123  *
1124  * When a buffer is marked dirty, its page is dirtied, but the page's other
1125  * buffers are not.
1126  *
1127  * Also.  When blockdev buffers are explicitly read with bread(), they
1128  * individually become uptodate.  But their backing page remains not
1129  * uptodate - even if all of its buffers are uptodate.  A subsequent
1130  * block_read_full_page() against that page will discover all the uptodate
1131  * buffers, will set the page uptodate and will perform no I/O.
1132  */
1133
1134 /**
1135  * mark_buffer_dirty - mark a buffer_head as needing writeout
1136  * @bh: the buffer_head to mark dirty
1137  *
1138  * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1139  * backing page dirty, then tag the page as dirty in its address_space's radix
1140  * tree and then attach the address_space's inode to its superblock's dirty
1141  * inode list.
1142  *
1143  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1144  * mapping->tree_lock and mapping->host->i_lock.
1145  */
1146 void mark_buffer_dirty(struct buffer_head *bh)
1147 {
1148         WARN_ON_ONCE(!buffer_uptodate(bh));
1149
1150         trace_block_dirty_buffer(bh);
1151
1152         /*
1153          * Very *carefully* optimize the it-is-already-dirty case.
1154          *
1155          * Don't let the final "is it dirty" escape to before we
1156          * perhaps modified the buffer.
1157          */
1158         if (buffer_dirty(bh)) {
1159                 smp_mb();
1160                 if (buffer_dirty(bh))
1161                         return;
1162         }
1163
1164         if (!test_set_buffer_dirty(bh)) {
1165                 struct page *page = bh->b_page;
1166                 struct address_space *mapping = NULL;
1167
1168                 lock_page_memcg(page);
1169                 if (!TestSetPageDirty(page)) {
1170                         mapping = page_mapping(page);
1171                         if (mapping)
1172                                 __set_page_dirty(page, mapping, 0);
1173                 }
1174                 unlock_page_memcg(page);
1175                 if (mapping)
1176                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1177         }
1178 }
1179 EXPORT_SYMBOL(mark_buffer_dirty);
1180
1181 void mark_buffer_write_io_error(struct buffer_head *bh)
1182 {
1183         set_buffer_write_io_error(bh);
1184         /* FIXME: do we need to set this in both places? */
1185         if (bh->b_page && bh->b_page->mapping)
1186                 mapping_set_error(bh->b_page->mapping, -EIO);
1187         if (bh->b_assoc_map)
1188                 mapping_set_error(bh->b_assoc_map, -EIO);
1189 }
1190 EXPORT_SYMBOL(mark_buffer_write_io_error);
1191
1192 /*
1193  * Decrement a buffer_head's reference count.  If all buffers against a page
1194  * have zero reference count, are clean and unlocked, and if the page is clean
1195  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1196  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1197  * a page but it ends up not being freed, and buffers may later be reattached).
1198  */
1199 void __brelse(struct buffer_head * buf)
1200 {
1201         if (atomic_read(&buf->b_count)) {
1202                 put_bh(buf);
1203                 return;
1204         }
1205         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1206 }
1207 EXPORT_SYMBOL(__brelse);
1208
1209 /*
1210  * bforget() is like brelse(), except it discards any
1211  * potentially dirty data.
1212  */
1213 void __bforget(struct buffer_head *bh)
1214 {
1215         clear_buffer_dirty(bh);
1216         if (bh->b_assoc_map) {
1217                 struct address_space *buffer_mapping = bh->b_page->mapping;
1218
1219                 spin_lock(&buffer_mapping->private_lock);
1220                 list_del_init(&bh->b_assoc_buffers);
1221                 bh->b_assoc_map = NULL;
1222                 spin_unlock(&buffer_mapping->private_lock);
1223         }
1224         __brelse(bh);
1225 }
1226 EXPORT_SYMBOL(__bforget);
1227
1228 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1229 {
1230         lock_buffer(bh);
1231         if (buffer_uptodate(bh)) {
1232                 unlock_buffer(bh);
1233                 return bh;
1234         } else {
1235                 get_bh(bh);
1236                 bh->b_end_io = end_buffer_read_sync;
1237                 submit_bh(REQ_OP_READ, 0, bh);
1238                 wait_on_buffer(bh);
1239                 if (buffer_uptodate(bh))
1240                         return bh;
1241         }
1242         brelse(bh);
1243         return NULL;
1244 }
1245
1246 /*
1247  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1248  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1249  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1250  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1251  * CPU's LRUs at the same time.
1252  *
1253  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1254  * sb_find_get_block().
1255  *
1256  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1257  * a local interrupt disable for that.
1258  */
1259
1260 #define BH_LRU_SIZE     16
1261
1262 struct bh_lru {
1263         struct buffer_head *bhs[BH_LRU_SIZE];
1264 };
1265
1266 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1267
1268 #ifdef CONFIG_SMP
1269 #define bh_lru_lock()   local_irq_disable()
1270 #define bh_lru_unlock() local_irq_enable()
1271 #else
1272 #define bh_lru_lock()   preempt_disable()
1273 #define bh_lru_unlock() preempt_enable()
1274 #endif
1275
1276 static inline void check_irqs_on(void)
1277 {
1278 #ifdef irqs_disabled
1279         BUG_ON(irqs_disabled());
1280 #endif
1281 }
1282
1283 /*
1284  * The LRU management algorithm is dopey-but-simple.  Sorry.
1285  */
1286 static void bh_lru_install(struct buffer_head *bh)
1287 {
1288         struct buffer_head *evictee = NULL;
1289
1290         check_irqs_on();
1291         bh_lru_lock();
1292         if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1293                 struct buffer_head *bhs[BH_LRU_SIZE];
1294                 int in;
1295                 int out = 0;
1296
1297                 get_bh(bh);
1298                 bhs[out++] = bh;
1299                 for (in = 0; in < BH_LRU_SIZE; in++) {
1300                         struct buffer_head *bh2 =
1301                                 __this_cpu_read(bh_lrus.bhs[in]);
1302
1303                         if (bh2 == bh) {
1304                                 __brelse(bh2);
1305                         } else {
1306                                 if (out >= BH_LRU_SIZE) {
1307                                         BUG_ON(evictee != NULL);
1308                                         evictee = bh2;
1309                                 } else {
1310                                         bhs[out++] = bh2;
1311                                 }
1312                         }
1313                 }
1314                 while (out < BH_LRU_SIZE)
1315                         bhs[out++] = NULL;
1316                 memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1317         }
1318         bh_lru_unlock();
1319
1320         if (evictee)
1321                 __brelse(evictee);
1322 }
1323
1324 /*
1325  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1326  */
1327 static struct buffer_head *
1328 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1329 {
1330         struct buffer_head *ret = NULL;
1331         unsigned int i;
1332
1333         check_irqs_on();
1334         bh_lru_lock();
1335         for (i = 0; i < BH_LRU_SIZE; i++) {
1336                 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1337
1338                 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1339                     bh->b_size == size) {
1340                         if (i) {
1341                                 while (i) {
1342                                         __this_cpu_write(bh_lrus.bhs[i],
1343                                                 __this_cpu_read(bh_lrus.bhs[i - 1]));
1344                                         i--;
1345                                 }
1346                                 __this_cpu_write(bh_lrus.bhs[0], bh);
1347                         }
1348                         get_bh(bh);
1349                         ret = bh;
1350                         break;
1351                 }
1352         }
1353         bh_lru_unlock();
1354         return ret;
1355 }
1356
1357 /*
1358  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1359  * it in the LRU and mark it as accessed.  If it is not present then return
1360  * NULL
1361  */
1362 struct buffer_head *
1363 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1364 {
1365         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1366
1367         if (bh == NULL) {
1368                 /* __find_get_block_slow will mark the page accessed */
1369                 bh = __find_get_block_slow(bdev, block);
1370                 if (bh)
1371                         bh_lru_install(bh);
1372         } else
1373                 touch_buffer(bh);
1374
1375         return bh;
1376 }
1377 EXPORT_SYMBOL(__find_get_block);
1378
1379 /*
1380  * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1381  * which corresponds to the passed block_device, block and size. The
1382  * returned buffer has its reference count incremented.
1383  *
1384  * __getblk_gfp() will lock up the machine if grow_dev_page's
1385  * try_to_free_buffers() attempt is failing.  FIXME, perhaps?
1386  */
1387 struct buffer_head *
1388 __getblk_gfp(struct block_device *bdev, sector_t block,
1389              unsigned size, gfp_t gfp)
1390 {
1391         struct buffer_head *bh = __find_get_block(bdev, block, size);
1392
1393         might_sleep();
1394         if (bh == NULL)
1395                 bh = __getblk_slow(bdev, block, size, gfp);
1396         return bh;
1397 }
1398 EXPORT_SYMBOL(__getblk_gfp);
1399
1400 /*
1401  * Do async read-ahead on a buffer..
1402  */
1403 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1404 {
1405         struct buffer_head *bh = __getblk(bdev, block, size);
1406         if (likely(bh)) {
1407                 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1408                 brelse(bh);
1409         }
1410 }
1411 EXPORT_SYMBOL(__breadahead);
1412
1413 /**
1414  *  __bread_gfp() - reads a specified block and returns the bh
1415  *  @bdev: the block_device to read from
1416  *  @block: number of block
1417  *  @size: size (in bytes) to read
1418  *  @gfp: page allocation flag
1419  *
1420  *  Reads a specified block, and returns buffer head that contains it.
1421  *  The page cache can be allocated from non-movable area
1422  *  not to prevent page migration if you set gfp to zero.
1423  *  It returns NULL if the block was unreadable.
1424  */
1425 struct buffer_head *
1426 __bread_gfp(struct block_device *bdev, sector_t block,
1427                    unsigned size, gfp_t gfp)
1428 {
1429         struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1430
1431         if (likely(bh) && !buffer_uptodate(bh))
1432                 bh = __bread_slow(bh);
1433         return bh;
1434 }
1435 EXPORT_SYMBOL(__bread_gfp);
1436
1437 /*
1438  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1439  * This doesn't race because it runs in each cpu either in irq
1440  * or with preempt disabled.
1441  */
1442 static void invalidate_bh_lru(void *arg)
1443 {
1444         struct bh_lru *b = &get_cpu_var(bh_lrus);
1445         int i;
1446
1447         for (i = 0; i < BH_LRU_SIZE; i++) {
1448                 brelse(b->bhs[i]);
1449                 b->bhs[i] = NULL;
1450         }
1451         put_cpu_var(bh_lrus);
1452 }
1453
1454 static bool has_bh_in_lru(int cpu, void *dummy)
1455 {
1456         struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1457         int i;
1458         
1459         for (i = 0; i < BH_LRU_SIZE; i++) {
1460                 if (b->bhs[i])
1461                         return 1;
1462         }
1463
1464         return 0;
1465 }
1466
1467 void invalidate_bh_lrus(void)
1468 {
1469         on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1470 }
1471 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1472
1473 void set_bh_page(struct buffer_head *bh,
1474                 struct page *page, unsigned long offset)
1475 {
1476         bh->b_page = page;
1477         BUG_ON(offset >= PAGE_SIZE);
1478         if (PageHighMem(page))
1479                 /*
1480                  * This catches illegal uses and preserves the offset:
1481                  */
1482                 bh->b_data = (char *)(0 + offset);
1483         else
1484                 bh->b_data = page_address(page) + offset;
1485 }
1486 EXPORT_SYMBOL(set_bh_page);
1487
1488 /*
1489  * Called when truncating a buffer on a page completely.
1490  */
1491
1492 /* Bits that are cleared during an invalidate */
1493 #define BUFFER_FLAGS_DISCARD \
1494         (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1495          1 << BH_Delay | 1 << BH_Unwritten)
1496
1497 static void discard_buffer(struct buffer_head * bh)
1498 {
1499         unsigned long b_state, b_state_old;
1500
1501         lock_buffer(bh);
1502         clear_buffer_dirty(bh);
1503         bh->b_bdev = NULL;
1504         b_state = bh->b_state;
1505         for (;;) {
1506                 b_state_old = cmpxchg(&bh->b_state, b_state,
1507                                       (b_state & ~BUFFER_FLAGS_DISCARD));
1508                 if (b_state_old == b_state)
1509                         break;
1510                 b_state = b_state_old;
1511         }
1512         unlock_buffer(bh);
1513 }
1514
1515 /**
1516  * block_invalidatepage - invalidate part or all of a buffer-backed page
1517  *
1518  * @page: the page which is affected
1519  * @offset: start of the range to invalidate
1520  * @length: length of the range to invalidate
1521  *
1522  * block_invalidatepage() is called when all or part of the page has become
1523  * invalidated by a truncate operation.
1524  *
1525  * block_invalidatepage() does not have to release all buffers, but it must
1526  * ensure that no dirty buffer is left outside @offset and that no I/O
1527  * is underway against any of the blocks which are outside the truncation
1528  * point.  Because the caller is about to free (and possibly reuse) those
1529  * blocks on-disk.
1530  */
1531 void block_invalidatepage(struct page *page, unsigned int offset,
1532                           unsigned int length)
1533 {
1534         struct buffer_head *head, *bh, *next;
1535         unsigned int curr_off = 0;
1536         unsigned int stop = length + offset;
1537
1538         BUG_ON(!PageLocked(page));
1539         if (!page_has_buffers(page))
1540                 goto out;
1541
1542         /*
1543          * Check for overflow
1544          */
1545         BUG_ON(stop > PAGE_SIZE || stop < length);
1546
1547         head = page_buffers(page);
1548         bh = head;
1549         do {
1550                 unsigned int next_off = curr_off + bh->b_size;
1551                 next = bh->b_this_page;
1552
1553                 /*
1554                  * Are we still fully in range ?
1555                  */
1556                 if (next_off > stop)
1557                         goto out;
1558
1559                 /*
1560                  * is this block fully invalidated?
1561                  */
1562                 if (offset <= curr_off)
1563                         discard_buffer(bh);
1564                 curr_off = next_off;
1565                 bh = next;
1566         } while (bh != head);
1567
1568         /*
1569          * We release buffers only if the entire page is being invalidated.
1570          * The get_block cached value has been unconditionally invalidated,
1571          * so real IO is not possible anymore.
1572          */
1573         if (offset == 0)
1574                 try_to_release_page(page, 0);
1575 out:
1576         return;
1577 }
1578 EXPORT_SYMBOL(block_invalidatepage);
1579
1580
1581 /*
1582  * We attach and possibly dirty the buffers atomically wrt
1583  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1584  * is already excluded via the page lock.
1585  */
1586 void create_empty_buffers(struct page *page,
1587                         unsigned long blocksize, unsigned long b_state)
1588 {
1589         struct buffer_head *bh, *head, *tail;
1590
1591         head = alloc_page_buffers(page, blocksize, 1);
1592         bh = head;
1593         do {
1594                 bh->b_state |= b_state;
1595                 tail = bh;
1596                 bh = bh->b_this_page;
1597         } while (bh);
1598         tail->b_this_page = head;
1599
1600         spin_lock(&page->mapping->private_lock);
1601         if (PageUptodate(page) || PageDirty(page)) {
1602                 bh = head;
1603                 do {
1604                         if (PageDirty(page))
1605                                 set_buffer_dirty(bh);
1606                         if (PageUptodate(page))
1607                                 set_buffer_uptodate(bh);
1608                         bh = bh->b_this_page;
1609                 } while (bh != head);
1610         }
1611         attach_page_buffers(page, head);
1612         spin_unlock(&page->mapping->private_lock);
1613 }
1614 EXPORT_SYMBOL(create_empty_buffers);
1615
1616 /**
1617  * clean_bdev_aliases: clean a range of buffers in block device
1618  * @bdev: Block device to clean buffers in
1619  * @block: Start of a range of blocks to clean
1620  * @len: Number of blocks to clean
1621  *
1622  * We are taking a range of blocks for data and we don't want writeback of any
1623  * buffer-cache aliases starting from return from this function and until the
1624  * moment when something will explicitly mark the buffer dirty (hopefully that
1625  * will not happen until we will free that block ;-) We don't even need to mark
1626  * it not-uptodate - nobody can expect anything from a newly allocated buffer
1627  * anyway. We used to use unmap_buffer() for such invalidation, but that was
1628  * wrong. We definitely don't want to mark the alias unmapped, for example - it
1629  * would confuse anyone who might pick it with bread() afterwards...
1630  *
1631  * Also..  Note that bforget() doesn't lock the buffer.  So there can be
1632  * writeout I/O going on against recently-freed buffers.  We don't wait on that
1633  * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1634  * need to.  That happens here.
1635  */
1636 void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1637 {
1638         struct inode *bd_inode = bdev->bd_inode;
1639         struct address_space *bd_mapping = bd_inode->i_mapping;
1640         struct pagevec pvec;
1641         pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1642         pgoff_t end;
1643         int i;
1644         struct buffer_head *bh;
1645         struct buffer_head *head;
1646
1647         end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1648         pagevec_init(&pvec, 0);
1649         while (index <= end && pagevec_lookup(&pvec, bd_mapping, index,
1650                         min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
1651                 for (i = 0; i < pagevec_count(&pvec); i++) {
1652                         struct page *page = pvec.pages[i];
1653
1654                         index = page->index;
1655                         if (index > end)
1656                                 break;
1657                         if (!page_has_buffers(page))
1658                                 continue;
1659                         /*
1660                          * We use page lock instead of bd_mapping->private_lock
1661                          * to pin buffers here since we can afford to sleep and
1662                          * it scales better than a global spinlock lock.
1663                          */
1664                         lock_page(page);
1665                         /* Recheck when the page is locked which pins bhs */
1666                         if (!page_has_buffers(page))
1667                                 goto unlock_page;
1668                         head = page_buffers(page);
1669                         bh = head;
1670                         do {
1671                                 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1672                                         goto next;
1673                                 if (bh->b_blocknr >= block + len)
1674                                         break;
1675                                 clear_buffer_dirty(bh);
1676                                 wait_on_buffer(bh);
1677                                 clear_buffer_req(bh);
1678 next:
1679                                 bh = bh->b_this_page;
1680                         } while (bh != head);
1681 unlock_page:
1682                         unlock_page(page);
1683                 }
1684                 pagevec_release(&pvec);
1685                 cond_resched();
1686                 index++;
1687         }
1688 }
1689 EXPORT_SYMBOL(clean_bdev_aliases);
1690
1691 /*
1692  * Size is a power-of-two in the range 512..PAGE_SIZE,
1693  * and the case we care about most is PAGE_SIZE.
1694  *
1695  * So this *could* possibly be written with those
1696  * constraints in mind (relevant mostly if some
1697  * architecture has a slow bit-scan instruction)
1698  */
1699 static inline int block_size_bits(unsigned int blocksize)
1700 {
1701         return ilog2(blocksize);
1702 }
1703
1704 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1705 {
1706         BUG_ON(!PageLocked(page));
1707
1708         if (!page_has_buffers(page))
1709                 create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1710         return page_buffers(page);
1711 }
1712
1713 /*
1714  * NOTE! All mapped/uptodate combinations are valid:
1715  *
1716  *      Mapped  Uptodate        Meaning
1717  *
1718  *      No      No              "unknown" - must do get_block()
1719  *      No      Yes             "hole" - zero-filled
1720  *      Yes     No              "allocated" - allocated on disk, not read in
1721  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1722  *
1723  * "Dirty" is valid only with the last case (mapped+uptodate).
1724  */
1725
1726 /*
1727  * While block_write_full_page is writing back the dirty buffers under
1728  * the page lock, whoever dirtied the buffers may decide to clean them
1729  * again at any time.  We handle that by only looking at the buffer
1730  * state inside lock_buffer().
1731  *
1732  * If block_write_full_page() is called for regular writeback
1733  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1734  * locked buffer.   This only can happen if someone has written the buffer
1735  * directly, with submit_bh().  At the address_space level PageWriteback
1736  * prevents this contention from occurring.
1737  *
1738  * If block_write_full_page() is called with wbc->sync_mode ==
1739  * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1740  * causes the writes to be flagged as synchronous writes.
1741  */
1742 int __block_write_full_page(struct inode *inode, struct page *page,
1743                         get_block_t *get_block, struct writeback_control *wbc,
1744                         bh_end_io_t *handler)
1745 {
1746         int err;
1747         sector_t block;
1748         sector_t last_block;
1749         struct buffer_head *bh, *head;
1750         unsigned int blocksize, bbits;
1751         int nr_underway = 0;
1752         int write_flags = wbc_to_write_flags(wbc);
1753
1754         head = create_page_buffers(page, inode,
1755                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1756
1757         /*
1758          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1759          * here, and the (potentially unmapped) buffers may become dirty at
1760          * any time.  If a buffer becomes dirty here after we've inspected it
1761          * then we just miss that fact, and the page stays dirty.
1762          *
1763          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1764          * handle that here by just cleaning them.
1765          */
1766
1767         bh = head;
1768         blocksize = bh->b_size;
1769         bbits = block_size_bits(blocksize);
1770
1771         block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1772         last_block = (i_size_read(inode) - 1) >> bbits;
1773
1774         /*
1775          * Get all the dirty buffers mapped to disk addresses and
1776          * handle any aliases from the underlying blockdev's mapping.
1777          */
1778         do {
1779                 if (block > last_block) {
1780                         /*
1781                          * mapped buffers outside i_size will occur, because
1782                          * this page can be outside i_size when there is a
1783                          * truncate in progress.
1784                          */
1785                         /*
1786                          * The buffer was zeroed by block_write_full_page()
1787                          */
1788                         clear_buffer_dirty(bh);
1789                         set_buffer_uptodate(bh);
1790                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1791                            buffer_dirty(bh)) {
1792                         WARN_ON(bh->b_size != blocksize);
1793                         err = get_block(inode, block, bh, 1);
1794                         if (err)
1795                                 goto recover;
1796                         clear_buffer_delay(bh);
1797                         if (buffer_new(bh)) {
1798                                 /* blockdev mappings never come here */
1799                                 clear_buffer_new(bh);
1800                                 clean_bdev_bh_alias(bh);
1801                         }
1802                 }
1803                 bh = bh->b_this_page;
1804                 block++;
1805         } while (bh != head);
1806
1807         do {
1808                 if (!buffer_mapped(bh))
1809                         continue;
1810                 /*
1811                  * If it's a fully non-blocking write attempt and we cannot
1812                  * lock the buffer then redirty the page.  Note that this can
1813                  * potentially cause a busy-wait loop from writeback threads
1814                  * and kswapd activity, but those code paths have their own
1815                  * higher-level throttling.
1816                  */
1817                 if (wbc->sync_mode != WB_SYNC_NONE) {
1818                         lock_buffer(bh);
1819                 } else if (!trylock_buffer(bh)) {
1820                         redirty_page_for_writepage(wbc, page);
1821                         continue;
1822                 }
1823                 if (test_clear_buffer_dirty(bh)) {
1824                         mark_buffer_async_write_endio(bh, handler);
1825                 } else {
1826                         unlock_buffer(bh);
1827                 }
1828         } while ((bh = bh->b_this_page) != head);
1829
1830         /*
1831          * The page and its buffers are protected by PageWriteback(), so we can
1832          * drop the bh refcounts early.
1833          */
1834         BUG_ON(PageWriteback(page));
1835         set_page_writeback(page);
1836
1837         do {
1838                 struct buffer_head *next = bh->b_this_page;
1839                 if (buffer_async_write(bh)) {
1840                         submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1841                                         inode->i_write_hint, wbc);
1842                         nr_underway++;
1843                 }
1844                 bh = next;
1845         } while (bh != head);
1846         unlock_page(page);
1847
1848         err = 0;
1849 done:
1850         if (nr_underway == 0) {
1851                 /*
1852                  * The page was marked dirty, but the buffers were
1853                  * clean.  Someone wrote them back by hand with
1854                  * ll_rw_block/submit_bh.  A rare case.
1855                  */
1856                 end_page_writeback(page);
1857
1858                 /*
1859                  * The page and buffer_heads can be released at any time from
1860                  * here on.
1861                  */
1862         }
1863         return err;
1864
1865 recover:
1866         /*
1867          * ENOSPC, or some other error.  We may already have added some
1868          * blocks to the file, so we need to write these out to avoid
1869          * exposing stale data.
1870          * The page is currently locked and not marked for writeback
1871          */
1872         bh = head;
1873         /* Recovery: lock and submit the mapped buffers */
1874         do {
1875                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1876                     !buffer_delay(bh)) {
1877                         lock_buffer(bh);
1878                         mark_buffer_async_write_endio(bh, handler);
1879                 } else {
1880                         /*
1881                          * The buffer may have been set dirty during
1882                          * attachment to a dirty page.
1883                          */
1884                         clear_buffer_dirty(bh);
1885                 }
1886         } while ((bh = bh->b_this_page) != head);
1887         SetPageError(page);
1888         BUG_ON(PageWriteback(page));
1889         mapping_set_error(page->mapping, err);
1890         set_page_writeback(page);
1891         do {
1892                 struct buffer_head *next = bh->b_this_page;
1893                 if (buffer_async_write(bh)) {
1894                         clear_buffer_dirty(bh);
1895                         submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1896                                         inode->i_write_hint, wbc);
1897                         nr_underway++;
1898                 }
1899                 bh = next;
1900         } while (bh != head);
1901         unlock_page(page);
1902         goto done;
1903 }
1904 EXPORT_SYMBOL(__block_write_full_page);
1905
1906 /*
1907  * If a page has any new buffers, zero them out here, and mark them uptodate
1908  * and dirty so they'll be written out (in order to prevent uninitialised
1909  * block data from leaking). And clear the new bit.
1910  */
1911 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1912 {
1913         unsigned int block_start, block_end;
1914         struct buffer_head *head, *bh;
1915
1916         BUG_ON(!PageLocked(page));
1917         if (!page_has_buffers(page))
1918                 return;
1919
1920         bh = head = page_buffers(page);
1921         block_start = 0;
1922         do {
1923                 block_end = block_start + bh->b_size;
1924
1925                 if (buffer_new(bh)) {
1926                         if (block_end > from && block_start < to) {
1927                                 if (!PageUptodate(page)) {
1928                                         unsigned start, size;
1929
1930                                         start = max(from, block_start);
1931                                         size = min(to, block_end) - start;
1932
1933                                         zero_user(page, start, size);
1934                                         set_buffer_uptodate(bh);
1935                                 }
1936
1937                                 clear_buffer_new(bh);
1938                                 mark_buffer_dirty(bh);
1939                         }
1940                 }
1941
1942                 block_start = block_end;
1943                 bh = bh->b_this_page;
1944         } while (bh != head);
1945 }
1946 EXPORT_SYMBOL(page_zero_new_buffers);
1947
1948 static void
1949 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1950                 struct iomap *iomap)
1951 {
1952         loff_t offset = block << inode->i_blkbits;
1953
1954         bh->b_bdev = iomap->bdev;
1955
1956         /*
1957          * Block points to offset in file we need to map, iomap contains
1958          * the offset at which the map starts. If the map ends before the
1959          * current block, then do not map the buffer and let the caller
1960          * handle it.
1961          */
1962         BUG_ON(offset >= iomap->offset + iomap->length);
1963
1964         switch (iomap->type) {
1965         case IOMAP_HOLE:
1966                 /*
1967                  * If the buffer is not up to date or beyond the current EOF,
1968                  * we need to mark it as new to ensure sub-block zeroing is
1969                  * executed if necessary.
1970                  */
1971                 if (!buffer_uptodate(bh) ||
1972                     (offset >= i_size_read(inode)))
1973                         set_buffer_new(bh);
1974                 break;
1975         case IOMAP_DELALLOC:
1976                 if (!buffer_uptodate(bh) ||
1977                     (offset >= i_size_read(inode)))
1978                         set_buffer_new(bh);
1979                 set_buffer_uptodate(bh);
1980                 set_buffer_mapped(bh);
1981                 set_buffer_delay(bh);
1982                 break;
1983         case IOMAP_UNWRITTEN:
1984                 /*
1985                  * For unwritten regions, we always need to ensure that
1986                  * sub-block writes cause the regions in the block we are not
1987                  * writing to are zeroed. Set the buffer as new to ensure this.
1988                  */
1989                 set_buffer_new(bh);
1990                 set_buffer_unwritten(bh);
1991                 /* FALLTHRU */
1992         case IOMAP_MAPPED:
1993                 if (offset >= i_size_read(inode))
1994                         set_buffer_new(bh);
1995                 bh->b_blocknr = (iomap->blkno >> (inode->i_blkbits - 9)) +
1996                                 ((offset - iomap->offset) >> inode->i_blkbits);
1997                 set_buffer_mapped(bh);
1998                 break;
1999         }
2000 }
2001
2002 int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
2003                 get_block_t *get_block, struct iomap *iomap)
2004 {
2005         unsigned from = pos & (PAGE_SIZE - 1);
2006         unsigned to = from + len;
2007         struct inode *inode = page->mapping->host;
2008         unsigned block_start, block_end;
2009         sector_t block;
2010         int err = 0;
2011         unsigned blocksize, bbits;
2012         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
2013
2014         BUG_ON(!PageLocked(page));
2015         BUG_ON(from > PAGE_SIZE);
2016         BUG_ON(to > PAGE_SIZE);
2017         BUG_ON(from > to);
2018
2019         head = create_page_buffers(page, inode, 0);
2020         blocksize = head->b_size;
2021         bbits = block_size_bits(blocksize);
2022
2023         block = (sector_t)page->index << (PAGE_SHIFT - bbits);
2024
2025         for(bh = head, block_start = 0; bh != head || !block_start;
2026             block++, block_start=block_end, bh = bh->b_this_page) {
2027                 block_end = block_start + blocksize;
2028                 if (block_end <= from || block_start >= to) {
2029                         if (PageUptodate(page)) {
2030                                 if (!buffer_uptodate(bh))
2031                                         set_buffer_uptodate(bh);
2032                         }
2033                         continue;
2034                 }
2035                 if (buffer_new(bh))
2036                         clear_buffer_new(bh);
2037                 if (!buffer_mapped(bh)) {
2038                         WARN_ON(bh->b_size != blocksize);
2039                         if (get_block) {
2040                                 err = get_block(inode, block, bh, 1);
2041                                 if (err)
2042                                         break;
2043                         } else {
2044                                 iomap_to_bh(inode, block, bh, iomap);
2045                         }
2046
2047                         if (buffer_new(bh)) {
2048                                 clean_bdev_bh_alias(bh);
2049                                 if (PageUptodate(page)) {
2050                                         clear_buffer_new(bh);
2051                                         set_buffer_uptodate(bh);
2052                                         mark_buffer_dirty(bh);
2053                                         continue;
2054                                 }
2055                                 if (block_end > to || block_start < from)
2056                                         zero_user_segments(page,
2057                                                 to, block_end,
2058                                                 block_start, from);
2059                                 continue;
2060                         }
2061                 }
2062                 if (PageUptodate(page)) {
2063                         if (!buffer_uptodate(bh))
2064                                 set_buffer_uptodate(bh);
2065                         continue; 
2066                 }
2067                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2068                     !buffer_unwritten(bh) &&
2069                      (block_start < from || block_end > to)) {
2070                         ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2071                         *wait_bh++=bh;
2072                 }
2073         }
2074         /*
2075          * If we issued read requests - let them complete.
2076          */
2077         while(wait_bh > wait) {
2078                 wait_on_buffer(*--wait_bh);
2079                 if (!buffer_uptodate(*wait_bh))
2080                         err = -EIO;
2081         }
2082         if (unlikely(err))
2083                 page_zero_new_buffers(page, from, to);
2084         return err;
2085 }
2086
2087 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2088                 get_block_t *get_block)
2089 {
2090         return __block_write_begin_int(page, pos, len, get_block, NULL);
2091 }
2092 EXPORT_SYMBOL(__block_write_begin);
2093
2094 static int __block_commit_write(struct inode *inode, struct page *page,
2095                 unsigned from, unsigned to)
2096 {
2097         unsigned block_start, block_end;
2098         int partial = 0;
2099         unsigned blocksize;
2100         struct buffer_head *bh, *head;
2101
2102         bh = head = page_buffers(page);
2103         blocksize = bh->b_size;
2104
2105         block_start = 0;
2106         do {
2107                 block_end = block_start + blocksize;
2108                 if (block_end <= from || block_start >= to) {
2109                         if (!buffer_uptodate(bh))
2110                                 partial = 1;
2111                 } else {
2112                         set_buffer_uptodate(bh);
2113                         mark_buffer_dirty(bh);
2114                 }
2115                 clear_buffer_new(bh);
2116
2117                 block_start = block_end;
2118                 bh = bh->b_this_page;
2119         } while (bh != head);
2120
2121         /*
2122          * If this is a partial write which happened to make all buffers
2123          * uptodate then we can optimize away a bogus readpage() for
2124          * the next read(). Here we 'discover' whether the page went
2125          * uptodate as a result of this (potentially partial) write.
2126          */
2127         if (!partial)
2128                 SetPageUptodate(page);
2129         return 0;
2130 }
2131
2132 /*
2133  * block_write_begin takes care of the basic task of block allocation and
2134  * bringing partial write blocks uptodate first.
2135  *
2136  * The filesystem needs to handle block truncation upon failure.
2137  */
2138 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2139                 unsigned flags, struct page **pagep, get_block_t *get_block)
2140 {
2141         pgoff_t index = pos >> PAGE_SHIFT;
2142         struct page *page;
2143         int status;
2144
2145         page = grab_cache_page_write_begin(mapping, index, flags);
2146         if (!page)
2147                 return -ENOMEM;
2148
2149         status = __block_write_begin(page, pos, len, get_block);
2150         if (unlikely(status)) {
2151                 unlock_page(page);
2152                 put_page(page);
2153                 page = NULL;
2154         }
2155
2156         *pagep = page;
2157         return status;
2158 }
2159 EXPORT_SYMBOL(block_write_begin);
2160
2161 int block_write_end(struct file *file, struct address_space *mapping,
2162                         loff_t pos, unsigned len, unsigned copied,
2163                         struct page *page, void *fsdata)
2164 {
2165         struct inode *inode = mapping->host;
2166         unsigned start;
2167
2168         start = pos & (PAGE_SIZE - 1);
2169
2170         if (unlikely(copied < len)) {
2171                 /*
2172                  * The buffers that were written will now be uptodate, so we
2173                  * don't have to worry about a readpage reading them and
2174                  * overwriting a partial write. However if we have encountered
2175                  * a short write and only partially written into a buffer, it
2176                  * will not be marked uptodate, so a readpage might come in and
2177                  * destroy our partial write.
2178                  *
2179                  * Do the simplest thing, and just treat any short write to a
2180                  * non uptodate page as a zero-length write, and force the
2181                  * caller to redo the whole thing.
2182                  */
2183                 if (!PageUptodate(page))
2184                         copied = 0;
2185
2186                 page_zero_new_buffers(page, start+copied, start+len);
2187         }
2188         flush_dcache_page(page);
2189
2190         /* This could be a short (even 0-length) commit */
2191         __block_commit_write(inode, page, start, start+copied);
2192
2193         return copied;
2194 }
2195 EXPORT_SYMBOL(block_write_end);
2196
2197 int generic_write_end(struct file *file, struct address_space *mapping,
2198                         loff_t pos, unsigned len, unsigned copied,
2199                         struct page *page, void *fsdata)
2200 {
2201         struct inode *inode = mapping->host;
2202         loff_t old_size = inode->i_size;
2203         int i_size_changed = 0;
2204
2205         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2206
2207         /*
2208          * No need to use i_size_read() here, the i_size
2209          * cannot change under us because we hold i_mutex.
2210          *
2211          * But it's important to update i_size while still holding page lock:
2212          * page writeout could otherwise come in and zero beyond i_size.
2213          */
2214         if (pos+copied > inode->i_size) {
2215                 i_size_write(inode, pos+copied);
2216                 i_size_changed = 1;
2217         }
2218
2219         unlock_page(page);
2220         put_page(page);
2221
2222         if (old_size < pos)
2223                 pagecache_isize_extended(inode, old_size, pos);
2224         /*
2225          * Don't mark the inode dirty under page lock. First, it unnecessarily
2226          * makes the holding time of page lock longer. Second, it forces lock
2227          * ordering of page lock and transaction start for journaling
2228          * filesystems.
2229          */
2230         if (i_size_changed)
2231                 mark_inode_dirty(inode);
2232
2233         return copied;
2234 }
2235 EXPORT_SYMBOL(generic_write_end);
2236
2237 /*
2238  * block_is_partially_uptodate checks whether buffers within a page are
2239  * uptodate or not.
2240  *
2241  * Returns true if all buffers which correspond to a file portion
2242  * we want to read are uptodate.
2243  */
2244 int block_is_partially_uptodate(struct page *page, unsigned long from,
2245                                         unsigned long count)
2246 {
2247         unsigned block_start, block_end, blocksize;
2248         unsigned to;
2249         struct buffer_head *bh, *head;
2250         int ret = 1;
2251
2252         if (!page_has_buffers(page))
2253                 return 0;
2254
2255         head = page_buffers(page);
2256         blocksize = head->b_size;
2257         to = min_t(unsigned, PAGE_SIZE - from, count);
2258         to = from + to;
2259         if (from < blocksize && to > PAGE_SIZE - blocksize)
2260                 return 0;
2261
2262         bh = head;
2263         block_start = 0;
2264         do {
2265                 block_end = block_start + blocksize;
2266                 if (block_end > from && block_start < to) {
2267                         if (!buffer_uptodate(bh)) {
2268                                 ret = 0;
2269                                 break;
2270                         }
2271                         if (block_end >= to)
2272                                 break;
2273                 }
2274                 block_start = block_end;
2275                 bh = bh->b_this_page;
2276         } while (bh != head);
2277
2278         return ret;
2279 }
2280 EXPORT_SYMBOL(block_is_partially_uptodate);
2281
2282 /*
2283  * Generic "read page" function for block devices that have the normal
2284  * get_block functionality. This is most of the block device filesystems.
2285  * Reads the page asynchronously --- the unlock_buffer() and
2286  * set/clear_buffer_uptodate() functions propagate buffer state into the
2287  * page struct once IO has completed.
2288  */
2289 int block_read_full_page(struct page *page, get_block_t *get_block)
2290 {
2291         struct inode *inode = page->mapping->host;
2292         sector_t iblock, lblock;
2293         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2294         unsigned int blocksize, bbits;
2295         int nr, i;
2296         int fully_mapped = 1;
2297
2298         head = create_page_buffers(page, inode, 0);
2299         blocksize = head->b_size;
2300         bbits = block_size_bits(blocksize);
2301
2302         iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2303         lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2304         bh = head;
2305         nr = 0;
2306         i = 0;
2307
2308         do {
2309                 if (buffer_uptodate(bh))
2310                         continue;
2311
2312                 if (!buffer_mapped(bh)) {
2313                         int err = 0;
2314
2315                         fully_mapped = 0;
2316                         if (iblock < lblock) {
2317                                 WARN_ON(bh->b_size != blocksize);
2318                                 err = get_block(inode, iblock, bh, 0);
2319                                 if (err)
2320                                         SetPageError(page);
2321                         }
2322                         if (!buffer_mapped(bh)) {
2323                                 zero_user(page, i * blocksize, blocksize);
2324                                 if (!err)
2325                                         set_buffer_uptodate(bh);
2326                                 continue;
2327                         }
2328                         /*
2329                          * get_block() might have updated the buffer
2330                          * synchronously
2331                          */
2332                         if (buffer_uptodate(bh))
2333                                 continue;
2334                 }
2335                 arr[nr++] = bh;
2336         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2337
2338         if (fully_mapped)
2339                 SetPageMappedToDisk(page);
2340
2341         if (!nr) {
2342                 /*
2343                  * All buffers are uptodate - we can set the page uptodate
2344                  * as well. But not if get_block() returned an error.
2345                  */
2346                 if (!PageError(page))
2347                         SetPageUptodate(page);
2348                 unlock_page(page);
2349                 return 0;
2350         }
2351
2352         /* Stage two: lock the buffers */
2353         for (i = 0; i < nr; i++) {
2354                 bh = arr[i];
2355                 lock_buffer(bh);
2356                 mark_buffer_async_read(bh);
2357         }
2358
2359         /*
2360          * Stage 3: start the IO.  Check for uptodateness
2361          * inside the buffer lock in case another process reading
2362          * the underlying blockdev brought it uptodate (the sct fix).
2363          */
2364         for (i = 0; i < nr; i++) {
2365                 bh = arr[i];
2366                 if (buffer_uptodate(bh))
2367                         end_buffer_async_read(bh, 1);
2368                 else
2369                         submit_bh(REQ_OP_READ, 0, bh);
2370         }
2371         return 0;
2372 }
2373 EXPORT_SYMBOL(block_read_full_page);
2374
2375 /* utility function for filesystems that need to do work on expanding
2376  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2377  * deal with the hole.  
2378  */
2379 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2380 {
2381         struct address_space *mapping = inode->i_mapping;
2382         struct page *page;
2383         void *fsdata;
2384         int err;
2385
2386         err = inode_newsize_ok(inode, size);
2387         if (err)
2388                 goto out;
2389
2390         err = pagecache_write_begin(NULL, mapping, size, 0,
2391                                     AOP_FLAG_CONT_EXPAND, &page, &fsdata);
2392         if (err)
2393                 goto out;
2394
2395         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2396         BUG_ON(err > 0);
2397
2398 out:
2399         return err;
2400 }
2401 EXPORT_SYMBOL(generic_cont_expand_simple);
2402
2403 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2404                             loff_t pos, loff_t *bytes)
2405 {
2406         struct inode *inode = mapping->host;
2407         unsigned int blocksize = i_blocksize(inode);
2408         struct page *page;
2409         void *fsdata;
2410         pgoff_t index, curidx;
2411         loff_t curpos;
2412         unsigned zerofrom, offset, len;
2413         int err = 0;
2414
2415         index = pos >> PAGE_SHIFT;
2416         offset = pos & ~PAGE_MASK;
2417
2418         while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2419                 zerofrom = curpos & ~PAGE_MASK;
2420                 if (zerofrom & (blocksize-1)) {
2421                         *bytes |= (blocksize-1);
2422                         (*bytes)++;
2423                 }
2424                 len = PAGE_SIZE - zerofrom;
2425
2426                 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2427                                             &page, &fsdata);
2428                 if (err)
2429                         goto out;
2430                 zero_user(page, zerofrom, len);
2431                 err = pagecache_write_end(file, mapping, curpos, len, len,
2432                                                 page, fsdata);
2433                 if (err < 0)
2434                         goto out;
2435                 BUG_ON(err != len);
2436                 err = 0;
2437
2438                 balance_dirty_pages_ratelimited(mapping);
2439
2440                 if (unlikely(fatal_signal_pending(current))) {
2441                         err = -EINTR;
2442                         goto out;
2443                 }
2444         }
2445
2446         /* page covers the boundary, find the boundary offset */
2447         if (index == curidx) {
2448                 zerofrom = curpos & ~PAGE_MASK;
2449                 /* if we will expand the thing last block will be filled */
2450                 if (offset <= zerofrom) {
2451                         goto out;
2452                 }
2453                 if (zerofrom & (blocksize-1)) {
2454                         *bytes |= (blocksize-1);
2455                         (*bytes)++;
2456                 }
2457                 len = offset - zerofrom;
2458
2459                 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2460                                             &page, &fsdata);
2461                 if (err)
2462                         goto out;
2463                 zero_user(page, zerofrom, len);
2464                 err = pagecache_write_end(file, mapping, curpos, len, len,
2465                                                 page, fsdata);
2466                 if (err < 0)
2467                         goto out;
2468                 BUG_ON(err != len);
2469                 err = 0;
2470         }
2471 out:
2472         return err;
2473 }
2474
2475 /*
2476  * For moronic filesystems that do not allow holes in file.
2477  * We may have to extend the file.
2478  */
2479 int cont_write_begin(struct file *file, struct address_space *mapping,
2480                         loff_t pos, unsigned len, unsigned flags,
2481                         struct page **pagep, void **fsdata,
2482                         get_block_t *get_block, loff_t *bytes)
2483 {
2484         struct inode *inode = mapping->host;
2485         unsigned int blocksize = i_blocksize(inode);
2486         unsigned int zerofrom;
2487         int err;
2488
2489         err = cont_expand_zero(file, mapping, pos, bytes);
2490         if (err)
2491                 return err;
2492
2493         zerofrom = *bytes & ~PAGE_MASK;
2494         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2495                 *bytes |= (blocksize-1);
2496                 (*bytes)++;
2497         }
2498
2499         return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2500 }
2501 EXPORT_SYMBOL(cont_write_begin);
2502
2503 int block_commit_write(struct page *page, unsigned from, unsigned to)
2504 {
2505         struct inode *inode = page->mapping->host;
2506         __block_commit_write(inode,page,from,to);
2507         return 0;
2508 }
2509 EXPORT_SYMBOL(block_commit_write);
2510
2511 /*
2512  * block_page_mkwrite() is not allowed to change the file size as it gets
2513  * called from a page fault handler when a page is first dirtied. Hence we must
2514  * be careful to check for EOF conditions here. We set the page up correctly
2515  * for a written page which means we get ENOSPC checking when writing into
2516  * holes and correct delalloc and unwritten extent mapping on filesystems that
2517  * support these features.
2518  *
2519  * We are not allowed to take the i_mutex here so we have to play games to
2520  * protect against truncate races as the page could now be beyond EOF.  Because
2521  * truncate writes the inode size before removing pages, once we have the
2522  * page lock we can determine safely if the page is beyond EOF. If it is not
2523  * beyond EOF, then the page is guaranteed safe against truncation until we
2524  * unlock the page.
2525  *
2526  * Direct callers of this function should protect against filesystem freezing
2527  * using sb_start_pagefault() - sb_end_pagefault() functions.
2528  */
2529 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2530                          get_block_t get_block)
2531 {
2532         struct page *page = vmf->page;
2533         struct inode *inode = file_inode(vma->vm_file);
2534         unsigned long end;
2535         loff_t size;
2536         int ret;
2537
2538         lock_page(page);
2539         size = i_size_read(inode);
2540         if ((page->mapping != inode->i_mapping) ||
2541             (page_offset(page) > size)) {
2542                 /* We overload EFAULT to mean page got truncated */
2543                 ret = -EFAULT;
2544                 goto out_unlock;
2545         }
2546
2547         /* page is wholly or partially inside EOF */
2548         if (((page->index + 1) << PAGE_SHIFT) > size)
2549                 end = size & ~PAGE_MASK;
2550         else
2551                 end = PAGE_SIZE;
2552
2553         ret = __block_write_begin(page, 0, end, get_block);
2554         if (!ret)
2555                 ret = block_commit_write(page, 0, end);
2556
2557         if (unlikely(ret < 0))
2558                 goto out_unlock;
2559         set_page_dirty(page);
2560         wait_for_stable_page(page);
2561         return 0;
2562 out_unlock:
2563         unlock_page(page);
2564         return ret;
2565 }
2566 EXPORT_SYMBOL(block_page_mkwrite);
2567
2568 /*
2569  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2570  * immediately, while under the page lock.  So it needs a special end_io
2571  * handler which does not touch the bh after unlocking it.
2572  */
2573 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2574 {
2575         __end_buffer_read_notouch(bh, uptodate);
2576 }
2577
2578 /*
2579  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2580  * the page (converting it to circular linked list and taking care of page
2581  * dirty races).
2582  */
2583 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2584 {
2585         struct buffer_head *bh;
2586
2587         BUG_ON(!PageLocked(page));
2588
2589         spin_lock(&page->mapping->private_lock);
2590         bh = head;
2591         do {
2592                 if (PageDirty(page))
2593                         set_buffer_dirty(bh);
2594                 if (!bh->b_this_page)
2595                         bh->b_this_page = head;
2596                 bh = bh->b_this_page;
2597         } while (bh != head);
2598         attach_page_buffers(page, head);
2599         spin_unlock(&page->mapping->private_lock);
2600 }
2601
2602 /*
2603  * On entry, the page is fully not uptodate.
2604  * On exit the page is fully uptodate in the areas outside (from,to)
2605  * The filesystem needs to handle block truncation upon failure.
2606  */
2607 int nobh_write_begin(struct address_space *mapping,
2608                         loff_t pos, unsigned len, unsigned flags,
2609                         struct page **pagep, void **fsdata,
2610                         get_block_t *get_block)
2611 {
2612         struct inode *inode = mapping->host;
2613         const unsigned blkbits = inode->i_blkbits;
2614         const unsigned blocksize = 1 << blkbits;
2615         struct buffer_head *head, *bh;
2616         struct page *page;
2617         pgoff_t index;
2618         unsigned from, to;
2619         unsigned block_in_page;
2620         unsigned block_start, block_end;
2621         sector_t block_in_file;
2622         int nr_reads = 0;
2623         int ret = 0;
2624         int is_mapped_to_disk = 1;
2625
2626         index = pos >> PAGE_SHIFT;
2627         from = pos & (PAGE_SIZE - 1);
2628         to = from + len;
2629
2630         page = grab_cache_page_write_begin(mapping, index, flags);
2631         if (!page)
2632                 return -ENOMEM;
2633         *pagep = page;
2634         *fsdata = NULL;
2635
2636         if (page_has_buffers(page)) {
2637                 ret = __block_write_begin(page, pos, len, get_block);
2638                 if (unlikely(ret))
2639                         goto out_release;
2640                 return ret;
2641         }
2642
2643         if (PageMappedToDisk(page))
2644                 return 0;
2645
2646         /*
2647          * Allocate buffers so that we can keep track of state, and potentially
2648          * attach them to the page if an error occurs. In the common case of
2649          * no error, they will just be freed again without ever being attached
2650          * to the page (which is all OK, because we're under the page lock).
2651          *
2652          * Be careful: the buffer linked list is a NULL terminated one, rather
2653          * than the circular one we're used to.
2654          */
2655         head = alloc_page_buffers(page, blocksize, 0);
2656         if (!head) {
2657                 ret = -ENOMEM;
2658                 goto out_release;
2659         }
2660
2661         block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2662
2663         /*
2664          * We loop across all blocks in the page, whether or not they are
2665          * part of the affected region.  This is so we can discover if the
2666          * page is fully mapped-to-disk.
2667          */
2668         for (block_start = 0, block_in_page = 0, bh = head;
2669                   block_start < PAGE_SIZE;
2670                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2671                 int create;
2672
2673                 block_end = block_start + blocksize;
2674                 bh->b_state = 0;
2675                 create = 1;
2676                 if (block_start >= to)
2677                         create = 0;
2678                 ret = get_block(inode, block_in_file + block_in_page,
2679                                         bh, create);
2680                 if (ret)
2681                         goto failed;
2682                 if (!buffer_mapped(bh))
2683                         is_mapped_to_disk = 0;
2684                 if (buffer_new(bh))
2685                         clean_bdev_bh_alias(bh);
2686                 if (PageUptodate(page)) {
2687                         set_buffer_uptodate(bh);
2688                         continue;
2689                 }
2690                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2691                         zero_user_segments(page, block_start, from,
2692                                                         to, block_end);
2693                         continue;
2694                 }
2695                 if (buffer_uptodate(bh))
2696                         continue;       /* reiserfs does this */
2697                 if (block_start < from || block_end > to) {
2698                         lock_buffer(bh);
2699                         bh->b_end_io = end_buffer_read_nobh;
2700                         submit_bh(REQ_OP_READ, 0, bh);
2701                         nr_reads++;
2702                 }
2703         }
2704
2705         if (nr_reads) {
2706                 /*
2707                  * The page is locked, so these buffers are protected from
2708                  * any VM or truncate activity.  Hence we don't need to care
2709                  * for the buffer_head refcounts.
2710                  */
2711                 for (bh = head; bh; bh = bh->b_this_page) {
2712                         wait_on_buffer(bh);
2713                         if (!buffer_uptodate(bh))
2714                                 ret = -EIO;
2715                 }
2716                 if (ret)
2717                         goto failed;
2718         }
2719
2720         if (is_mapped_to_disk)
2721                 SetPageMappedToDisk(page);
2722
2723         *fsdata = head; /* to be released by nobh_write_end */
2724
2725         return 0;
2726
2727 failed:
2728         BUG_ON(!ret);
2729         /*
2730          * Error recovery is a bit difficult. We need to zero out blocks that
2731          * were newly allocated, and dirty them to ensure they get written out.
2732          * Buffers need to be attached to the page at this point, otherwise
2733          * the handling of potential IO errors during writeout would be hard
2734          * (could try doing synchronous writeout, but what if that fails too?)
2735          */
2736         attach_nobh_buffers(page, head);
2737         page_zero_new_buffers(page, from, to);
2738
2739 out_release:
2740         unlock_page(page);
2741         put_page(page);
2742         *pagep = NULL;
2743
2744         return ret;
2745 }
2746 EXPORT_SYMBOL(nobh_write_begin);
2747
2748 int nobh_write_end(struct file *file, struct address_space *mapping,
2749                         loff_t pos, unsigned len, unsigned copied,
2750                         struct page *page, void *fsdata)
2751 {
2752         struct inode *inode = page->mapping->host;
2753         struct buffer_head *head = fsdata;
2754         struct buffer_head *bh;
2755         BUG_ON(fsdata != NULL && page_has_buffers(page));
2756
2757         if (unlikely(copied < len) && head)
2758                 attach_nobh_buffers(page, head);
2759         if (page_has_buffers(page))
2760                 return generic_write_end(file, mapping, pos, len,
2761                                         copied, page, fsdata);
2762
2763         SetPageUptodate(page);
2764         set_page_dirty(page);
2765         if (pos+copied > inode->i_size) {
2766                 i_size_write(inode, pos+copied);
2767                 mark_inode_dirty(inode);
2768         }
2769
2770         unlock_page(page);
2771         put_page(page);
2772
2773         while (head) {
2774                 bh = head;
2775                 head = head->b_this_page;
2776                 free_buffer_head(bh);
2777         }
2778
2779         return copied;
2780 }
2781 EXPORT_SYMBOL(nobh_write_end);
2782
2783 /*
2784  * nobh_writepage() - based on block_full_write_page() except
2785  * that it tries to operate without attaching bufferheads to
2786  * the page.
2787  */
2788 int nobh_writepage(struct page *page, get_block_t *get_block,
2789                         struct writeback_control *wbc)
2790 {
2791         struct inode * const inode = page->mapping->host;
2792         loff_t i_size = i_size_read(inode);
2793         const pgoff_t end_index = i_size >> PAGE_SHIFT;
2794         unsigned offset;
2795         int ret;
2796
2797         /* Is the page fully inside i_size? */
2798         if (page->index < end_index)
2799                 goto out;
2800
2801         /* Is the page fully outside i_size? (truncate in progress) */
2802         offset = i_size & (PAGE_SIZE-1);
2803         if (page->index >= end_index+1 || !offset) {
2804                 /*
2805                  * The page may have dirty, unmapped buffers.  For example,
2806                  * they may have been added in ext3_writepage().  Make them
2807                  * freeable here, so the page does not leak.
2808                  */
2809 #if 0
2810                 /* Not really sure about this  - do we need this ? */
2811                 if (page->mapping->a_ops->invalidatepage)
2812                         page->mapping->a_ops->invalidatepage(page, offset);
2813 #endif
2814                 unlock_page(page);
2815                 return 0; /* don't care */
2816         }
2817
2818         /*
2819          * The page straddles i_size.  It must be zeroed out on each and every
2820          * writepage invocation because it may be mmapped.  "A file is mapped
2821          * in multiples of the page size.  For a file that is not a multiple of
2822          * the  page size, the remaining memory is zeroed when mapped, and
2823          * writes to that region are not written out to the file."
2824          */
2825         zero_user_segment(page, offset, PAGE_SIZE);
2826 out:
2827         ret = mpage_writepage(page, get_block, wbc);
2828         if (ret == -EAGAIN)
2829                 ret = __block_write_full_page(inode, page, get_block, wbc,
2830                                               end_buffer_async_write);
2831         return ret;
2832 }
2833 EXPORT_SYMBOL(nobh_writepage);
2834
2835 int nobh_truncate_page(struct address_space *mapping,
2836                         loff_t from, get_block_t *get_block)
2837 {
2838         pgoff_t index = from >> PAGE_SHIFT;
2839         unsigned offset = from & (PAGE_SIZE-1);
2840         unsigned blocksize;
2841         sector_t iblock;
2842         unsigned length, pos;
2843         struct inode *inode = mapping->host;
2844         struct page *page;
2845         struct buffer_head map_bh;
2846         int err;
2847
2848         blocksize = i_blocksize(inode);
2849         length = offset & (blocksize - 1);
2850
2851         /* Block boundary? Nothing to do */
2852         if (!length)
2853                 return 0;
2854
2855         length = blocksize - length;
2856         iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2857
2858         page = grab_cache_page(mapping, index);
2859         err = -ENOMEM;
2860         if (!page)
2861                 goto out;
2862
2863         if (page_has_buffers(page)) {
2864 has_buffers:
2865                 unlock_page(page);
2866                 put_page(page);
2867                 return block_truncate_page(mapping, from, get_block);
2868         }
2869
2870         /* Find the buffer that contains "offset" */
2871         pos = blocksize;
2872         while (offset >= pos) {
2873                 iblock++;
2874                 pos += blocksize;
2875         }
2876
2877         map_bh.b_size = blocksize;
2878         map_bh.b_state = 0;
2879         err = get_block(inode, iblock, &map_bh, 0);
2880         if (err)
2881                 goto unlock;
2882         /* unmapped? It's a hole - nothing to do */
2883         if (!buffer_mapped(&map_bh))
2884                 goto unlock;
2885
2886         /* Ok, it's mapped. Make sure it's up-to-date */
2887         if (!PageUptodate(page)) {
2888                 err = mapping->a_ops->readpage(NULL, page);
2889                 if (err) {
2890                         put_page(page);
2891                         goto out;
2892                 }
2893                 lock_page(page);
2894                 if (!PageUptodate(page)) {
2895                         err = -EIO;
2896                         goto unlock;
2897                 }
2898                 if (page_has_buffers(page))
2899                         goto has_buffers;
2900         }
2901         zero_user(page, offset, length);
2902         set_page_dirty(page);
2903         err = 0;
2904
2905 unlock:
2906         unlock_page(page);
2907         put_page(page);
2908 out:
2909         return err;
2910 }
2911 EXPORT_SYMBOL(nobh_truncate_page);
2912
2913 int block_truncate_page(struct address_space *mapping,
2914                         loff_t from, get_block_t *get_block)
2915 {
2916         pgoff_t index = from >> PAGE_SHIFT;
2917         unsigned offset = from & (PAGE_SIZE-1);
2918         unsigned blocksize;
2919         sector_t iblock;
2920         unsigned length, pos;
2921         struct inode *inode = mapping->host;
2922         struct page *page;
2923         struct buffer_head *bh;
2924         int err;
2925
2926         blocksize = i_blocksize(inode);
2927         length = offset & (blocksize - 1);
2928
2929         /* Block boundary? Nothing to do */
2930         if (!length)
2931                 return 0;
2932
2933         length = blocksize - length;
2934         iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2935         
2936         page = grab_cache_page(mapping, index);
2937         err = -ENOMEM;
2938         if (!page)
2939                 goto out;
2940
2941         if (!page_has_buffers(page))
2942                 create_empty_buffers(page, blocksize, 0);
2943
2944         /* Find the buffer that contains "offset" */
2945         bh = page_buffers(page);
2946         pos = blocksize;
2947         while (offset >= pos) {
2948                 bh = bh->b_this_page;
2949                 iblock++;
2950                 pos += blocksize;
2951         }
2952
2953         err = 0;
2954         if (!buffer_mapped(bh)) {
2955                 WARN_ON(bh->b_size != blocksize);
2956                 err = get_block(inode, iblock, bh, 0);
2957                 if (err)
2958                         goto unlock;
2959                 /* unmapped? It's a hole - nothing to do */
2960                 if (!buffer_mapped(bh))
2961                         goto unlock;
2962         }
2963
2964         /* Ok, it's mapped. Make sure it's up-to-date */
2965         if (PageUptodate(page))
2966                 set_buffer_uptodate(bh);
2967
2968         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2969                 err = -EIO;
2970                 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2971                 wait_on_buffer(bh);
2972                 /* Uhhuh. Read error. Complain and punt. */
2973                 if (!buffer_uptodate(bh))
2974                         goto unlock;
2975         }
2976
2977         zero_user(page, offset, length);
2978         mark_buffer_dirty(bh);
2979         err = 0;
2980
2981 unlock:
2982         unlock_page(page);
2983         put_page(page);
2984 out:
2985         return err;
2986 }
2987 EXPORT_SYMBOL(block_truncate_page);
2988
2989 /*
2990  * The generic ->writepage function for buffer-backed address_spaces
2991  */
2992 int block_write_full_page(struct page *page, get_block_t *get_block,
2993                         struct writeback_control *wbc)
2994 {
2995         struct inode * const inode = page->mapping->host;
2996         loff_t i_size = i_size_read(inode);
2997         const pgoff_t end_index = i_size >> PAGE_SHIFT;
2998         unsigned offset;
2999
3000         /* Is the page fully inside i_size? */
3001         if (page->index < end_index)
3002                 return __block_write_full_page(inode, page, get_block, wbc,
3003                                                end_buffer_async_write);
3004
3005         /* Is the page fully outside i_size? (truncate in progress) */
3006         offset = i_size & (PAGE_SIZE-1);
3007         if (page->index >= end_index+1 || !offset) {
3008                 /*
3009                  * The page may have dirty, unmapped buffers.  For example,
3010                  * they may have been added in ext3_writepage().  Make them
3011                  * freeable here, so the page does not leak.
3012                  */
3013                 do_invalidatepage(page, 0, PAGE_SIZE);
3014                 unlock_page(page);
3015                 return 0; /* don't care */
3016         }
3017
3018         /*
3019          * The page straddles i_size.  It must be zeroed out on each and every
3020          * writepage invocation because it may be mmapped.  "A file is mapped
3021          * in multiples of the page size.  For a file that is not a multiple of
3022          * the  page size, the remaining memory is zeroed when mapped, and
3023          * writes to that region are not written out to the file."
3024          */
3025         zero_user_segment(page, offset, PAGE_SIZE);
3026         return __block_write_full_page(inode, page, get_block, wbc,
3027                                                         end_buffer_async_write);
3028 }
3029 EXPORT_SYMBOL(block_write_full_page);
3030
3031 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
3032                             get_block_t *get_block)
3033 {
3034         struct inode *inode = mapping->host;
3035         struct buffer_head tmp = {
3036                 .b_size = i_blocksize(inode),
3037         };
3038
3039         get_block(inode, block, &tmp, 0);
3040         return tmp.b_blocknr;
3041 }
3042 EXPORT_SYMBOL(generic_block_bmap);
3043
3044 static void end_bio_bh_io_sync(struct bio *bio)
3045 {
3046         struct buffer_head *bh = bio->bi_private;
3047
3048         if (unlikely(bio_flagged(bio, BIO_QUIET)))
3049                 set_bit(BH_Quiet, &bh->b_state);
3050
3051         bh->b_end_io(bh, !bio->bi_status);
3052         bio_put(bio);
3053 }
3054
3055 /*
3056  * This allows us to do IO even on the odd last sectors
3057  * of a device, even if the block size is some multiple
3058  * of the physical sector size.
3059  *
3060  * We'll just truncate the bio to the size of the device,
3061  * and clear the end of the buffer head manually.
3062  *
3063  * Truly out-of-range accesses will turn into actual IO
3064  * errors, this only handles the "we need to be able to
3065  * do IO at the final sector" case.
3066  */
3067 void guard_bio_eod(int op, struct bio *bio)
3068 {
3069         sector_t maxsector;
3070         struct bio_vec *bvec = &bio->bi_io_vec[bio->bi_vcnt - 1];
3071         unsigned truncated_bytes;
3072
3073         maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
3074         if (!maxsector)
3075                 return;
3076
3077         /*
3078          * If the *whole* IO is past the end of the device,
3079          * let it through, and the IO layer will turn it into
3080          * an EIO.
3081          */
3082         if (unlikely(bio->bi_iter.bi_sector >= maxsector))
3083                 return;
3084
3085         maxsector -= bio->bi_iter.bi_sector;
3086         if (likely((bio->bi_iter.bi_size >> 9) <= maxsector))
3087                 return;
3088
3089         /* Uhhuh. We've got a bio that straddles the device size! */
3090         truncated_bytes = bio->bi_iter.bi_size - (maxsector << 9);
3091
3092         /* Truncate the bio.. */
3093         bio->bi_iter.bi_size -= truncated_bytes;
3094         bvec->bv_len -= truncated_bytes;
3095
3096         /* ..and clear the end of the buffer for reads */
3097         if (op == REQ_OP_READ) {
3098                 zero_user(bvec->bv_page, bvec->bv_offset + bvec->bv_len,
3099                                 truncated_bytes);
3100         }
3101 }
3102
3103 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3104                          enum rw_hint write_hint, struct writeback_control *wbc)
3105 {
3106         struct bio *bio;
3107
3108         BUG_ON(!buffer_locked(bh));
3109         BUG_ON(!buffer_mapped(bh));
3110         BUG_ON(!bh->b_end_io);
3111         BUG_ON(buffer_delay(bh));
3112         BUG_ON(buffer_unwritten(bh));
3113
3114         /*
3115          * Only clear out a write error when rewriting
3116          */
3117         if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3118                 clear_buffer_write_io_error(bh);
3119
3120         /*
3121          * from here on down, it's all bio -- do the initial mapping,
3122          * submit_bio -> generic_make_request may further map this bio around
3123          */
3124         bio = bio_alloc(GFP_NOIO, 1);
3125
3126         if (wbc) {
3127                 wbc_init_bio(wbc, bio);
3128                 wbc_account_io(wbc, bh->b_page, bh->b_size);
3129         }
3130
3131         bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3132         bio->bi_bdev = bh->b_bdev;
3133         bio->bi_write_hint = write_hint;
3134
3135         bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3136         BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3137
3138         bio->bi_end_io = end_bio_bh_io_sync;
3139         bio->bi_private = bh;
3140
3141         /* Take care of bh's that straddle the end of the device */
3142         guard_bio_eod(op, bio);
3143
3144         if (buffer_meta(bh))
3145                 op_flags |= REQ_META;
3146         if (buffer_prio(bh))
3147                 op_flags |= REQ_PRIO;
3148         bio_set_op_attrs(bio, op, op_flags);
3149
3150         submit_bio(bio);
3151         return 0;
3152 }
3153
3154 int submit_bh(int op, int op_flags, struct buffer_head *bh)
3155 {
3156         return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3157 }
3158 EXPORT_SYMBOL(submit_bh);
3159
3160 /**
3161  * ll_rw_block: low-level access to block devices (DEPRECATED)
3162  * @op: whether to %READ or %WRITE
3163  * @op_flags: req_flag_bits
3164  * @nr: number of &struct buffer_heads in the array
3165  * @bhs: array of pointers to &struct buffer_head
3166  *
3167  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3168  * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3169  * @op_flags contains flags modifying the detailed I/O behavior, most notably
3170  * %REQ_RAHEAD.
3171  *
3172  * This function drops any buffer that it cannot get a lock on (with the
3173  * BH_Lock state bit), any buffer that appears to be clean when doing a write
3174  * request, and any buffer that appears to be up-to-date when doing read
3175  * request.  Further it marks as clean buffers that are processed for
3176  * writing (the buffer cache won't assume that they are actually clean
3177  * until the buffer gets unlocked).
3178  *
3179  * ll_rw_block sets b_end_io to simple completion handler that marks
3180  * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3181  * any waiters. 
3182  *
3183  * All of the buffers must be for the same device, and must also be a
3184  * multiple of the current approved size for the device.
3185  */
3186 void ll_rw_block(int op, int op_flags,  int nr, struct buffer_head *bhs[])
3187 {
3188         int i;
3189
3190         for (i = 0; i < nr; i++) {
3191                 struct buffer_head *bh = bhs[i];
3192
3193                 if (!trylock_buffer(bh))
3194                         continue;
3195                 if (op == WRITE) {
3196                         if (test_clear_buffer_dirty(bh)) {
3197                                 bh->b_end_io = end_buffer_write_sync;
3198                                 get_bh(bh);
3199                                 submit_bh(op, op_flags, bh);
3200                                 continue;
3201                         }
3202                 } else {
3203                         if (!buffer_uptodate(bh)) {
3204                                 bh->b_end_io = end_buffer_read_sync;
3205                                 get_bh(bh);
3206                                 submit_bh(op, op_flags, bh);
3207                                 continue;
3208                         }
3209                 }
3210                 unlock_buffer(bh);
3211         }
3212 }
3213 EXPORT_SYMBOL(ll_rw_block);
3214
3215 void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3216 {
3217         lock_buffer(bh);
3218         if (!test_clear_buffer_dirty(bh)) {
3219                 unlock_buffer(bh);
3220                 return;
3221         }
3222         bh->b_end_io = end_buffer_write_sync;
3223         get_bh(bh);
3224         submit_bh(REQ_OP_WRITE, op_flags, bh);
3225 }
3226 EXPORT_SYMBOL(write_dirty_buffer);
3227
3228 /*
3229  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3230  * and then start new I/O and then wait upon it.  The caller must have a ref on
3231  * the buffer_head.
3232  */
3233 int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3234 {
3235         int ret = 0;
3236
3237         WARN_ON(atomic_read(&bh->b_count) < 1);
3238         lock_buffer(bh);
3239         if (test_clear_buffer_dirty(bh)) {
3240                 get_bh(bh);
3241                 bh->b_end_io = end_buffer_write_sync;
3242                 ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3243                 wait_on_buffer(bh);
3244                 if (!ret && !buffer_uptodate(bh))
3245                         ret = -EIO;
3246         } else {
3247                 unlock_buffer(bh);
3248         }
3249         return ret;
3250 }
3251 EXPORT_SYMBOL(__sync_dirty_buffer);
3252
3253 int sync_dirty_buffer(struct buffer_head *bh)
3254 {
3255         return __sync_dirty_buffer(bh, REQ_SYNC);
3256 }
3257 EXPORT_SYMBOL(sync_dirty_buffer);
3258
3259 /*
3260  * try_to_free_buffers() checks if all the buffers on this particular page
3261  * are unused, and releases them if so.
3262  *
3263  * Exclusion against try_to_free_buffers may be obtained by either
3264  * locking the page or by holding its mapping's private_lock.
3265  *
3266  * If the page is dirty but all the buffers are clean then we need to
3267  * be sure to mark the page clean as well.  This is because the page
3268  * may be against a block device, and a later reattachment of buffers
3269  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3270  * filesystem data on the same device.
3271  *
3272  * The same applies to regular filesystem pages: if all the buffers are
3273  * clean then we set the page clean and proceed.  To do that, we require
3274  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3275  * private_lock.
3276  *
3277  * try_to_free_buffers() is non-blocking.
3278  */
3279 static inline int buffer_busy(struct buffer_head *bh)
3280 {
3281         return atomic_read(&bh->b_count) |
3282                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3283 }
3284
3285 static int
3286 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3287 {
3288         struct buffer_head *head = page_buffers(page);
3289         struct buffer_head *bh;
3290
3291         bh = head;
3292         do {
3293                 if (buffer_busy(bh))
3294                         goto failed;
3295                 bh = bh->b_this_page;
3296         } while (bh != head);
3297
3298         do {
3299                 struct buffer_head *next = bh->b_this_page;
3300
3301                 if (bh->b_assoc_map)
3302                         __remove_assoc_queue(bh);
3303                 bh = next;
3304         } while (bh != head);
3305         *buffers_to_free = head;
3306         __clear_page_buffers(page);
3307         return 1;
3308 failed:
3309         return 0;
3310 }
3311
3312 int try_to_free_buffers(struct page *page)
3313 {
3314         struct address_space * const mapping = page->mapping;
3315         struct buffer_head *buffers_to_free = NULL;
3316         int ret = 0;
3317
3318         BUG_ON(!PageLocked(page));
3319         if (PageWriteback(page))
3320                 return 0;
3321
3322         if (mapping == NULL) {          /* can this still happen? */
3323                 ret = drop_buffers(page, &buffers_to_free);
3324                 goto out;
3325         }
3326
3327         spin_lock(&mapping->private_lock);
3328         ret = drop_buffers(page, &buffers_to_free);
3329
3330         /*
3331          * If the filesystem writes its buffers by hand (eg ext3)
3332          * then we can have clean buffers against a dirty page.  We
3333          * clean the page here; otherwise the VM will never notice
3334          * that the filesystem did any IO at all.
3335          *
3336          * Also, during truncate, discard_buffer will have marked all
3337          * the page's buffers clean.  We discover that here and clean
3338          * the page also.
3339          *
3340          * private_lock must be held over this entire operation in order
3341          * to synchronise against __set_page_dirty_buffers and prevent the
3342          * dirty bit from being lost.
3343          */
3344         if (ret)
3345                 cancel_dirty_page(page);
3346         spin_unlock(&mapping->private_lock);
3347 out:
3348         if (buffers_to_free) {
3349                 struct buffer_head *bh = buffers_to_free;
3350
3351                 do {
3352                         struct buffer_head *next = bh->b_this_page;
3353                         free_buffer_head(bh);
3354                         bh = next;
3355                 } while (bh != buffers_to_free);
3356         }
3357         return ret;
3358 }
3359 EXPORT_SYMBOL(try_to_free_buffers);
3360
3361 /*
3362  * There are no bdflush tunables left.  But distributions are
3363  * still running obsolete flush daemons, so we terminate them here.
3364  *
3365  * Use of bdflush() is deprecated and will be removed in a future kernel.
3366  * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3367  */
3368 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3369 {
3370         static int msg_count;
3371
3372         if (!capable(CAP_SYS_ADMIN))
3373                 return -EPERM;
3374
3375         if (msg_count < 5) {
3376                 msg_count++;
3377                 printk(KERN_INFO
3378                         "warning: process `%s' used the obsolete bdflush"
3379                         " system call\n", current->comm);
3380                 printk(KERN_INFO "Fix your initscripts?\n");
3381         }
3382
3383         if (func == 1)
3384                 do_exit(0);
3385         return 0;
3386 }
3387
3388 /*
3389  * Buffer-head allocation
3390  */
3391 static struct kmem_cache *bh_cachep __read_mostly;
3392
3393 /*
3394  * Once the number of bh's in the machine exceeds this level, we start
3395  * stripping them in writeback.
3396  */
3397 static unsigned long max_buffer_heads;
3398
3399 int buffer_heads_over_limit;
3400
3401 struct bh_accounting {
3402         int nr;                 /* Number of live bh's */
3403         int ratelimit;          /* Limit cacheline bouncing */
3404 };
3405
3406 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3407
3408 static void recalc_bh_state(void)
3409 {
3410         int i;
3411         int tot = 0;
3412
3413         if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3414                 return;
3415         __this_cpu_write(bh_accounting.ratelimit, 0);
3416         for_each_online_cpu(i)
3417                 tot += per_cpu(bh_accounting, i).nr;
3418         buffer_heads_over_limit = (tot > max_buffer_heads);
3419 }
3420
3421 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3422 {
3423         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3424         if (ret) {
3425                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3426                 preempt_disable();
3427                 __this_cpu_inc(bh_accounting.nr);
3428                 recalc_bh_state();
3429                 preempt_enable();
3430         }
3431         return ret;
3432 }
3433 EXPORT_SYMBOL(alloc_buffer_head);
3434
3435 void free_buffer_head(struct buffer_head *bh)
3436 {
3437         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3438         kmem_cache_free(bh_cachep, bh);
3439         preempt_disable();
3440         __this_cpu_dec(bh_accounting.nr);
3441         recalc_bh_state();
3442         preempt_enable();
3443 }
3444 EXPORT_SYMBOL(free_buffer_head);
3445
3446 static int buffer_exit_cpu_dead(unsigned int cpu)
3447 {
3448         int i;
3449         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3450
3451         for (i = 0; i < BH_LRU_SIZE; i++) {
3452                 brelse(b->bhs[i]);
3453                 b->bhs[i] = NULL;
3454         }
3455         this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3456         per_cpu(bh_accounting, cpu).nr = 0;
3457         return 0;
3458 }
3459
3460 /**
3461  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3462  * @bh: struct buffer_head
3463  *
3464  * Return true if the buffer is up-to-date and false,
3465  * with the buffer locked, if not.
3466  */
3467 int bh_uptodate_or_lock(struct buffer_head *bh)
3468 {
3469         if (!buffer_uptodate(bh)) {
3470                 lock_buffer(bh);
3471                 if (!buffer_uptodate(bh))
3472                         return 0;
3473                 unlock_buffer(bh);
3474         }
3475         return 1;
3476 }
3477 EXPORT_SYMBOL(bh_uptodate_or_lock);
3478
3479 /**
3480  * bh_submit_read - Submit a locked buffer for reading
3481  * @bh: struct buffer_head
3482  *
3483  * Returns zero on success and -EIO on error.
3484  */
3485 int bh_submit_read(struct buffer_head *bh)
3486 {
3487         BUG_ON(!buffer_locked(bh));
3488
3489         if (buffer_uptodate(bh)) {
3490                 unlock_buffer(bh);
3491                 return 0;
3492         }
3493
3494         get_bh(bh);
3495         bh->b_end_io = end_buffer_read_sync;
3496         submit_bh(REQ_OP_READ, 0, bh);
3497         wait_on_buffer(bh);
3498         if (buffer_uptodate(bh))
3499                 return 0;
3500         return -EIO;
3501 }
3502 EXPORT_SYMBOL(bh_submit_read);
3503
3504 /*
3505  * Seek for SEEK_DATA / SEEK_HOLE within @page, starting at @lastoff.
3506  *
3507  * Returns the offset within the file on success, and -ENOENT otherwise.
3508  */
3509 static loff_t
3510 page_seek_hole_data(struct page *page, loff_t lastoff, int whence)
3511 {
3512         loff_t offset = page_offset(page);
3513         struct buffer_head *bh, *head;
3514         bool seek_data = whence == SEEK_DATA;
3515
3516         if (lastoff < offset)
3517                 lastoff = offset;
3518
3519         bh = head = page_buffers(page);
3520         do {
3521                 offset += bh->b_size;
3522                 if (lastoff >= offset)
3523                         continue;
3524
3525                 /*
3526                  * Unwritten extents that have data in the page cache covering
3527                  * them can be identified by the BH_Unwritten state flag.
3528                  * Pages with multiple buffers might have a mix of holes, data
3529                  * and unwritten extents - any buffer with valid data in it
3530                  * should have BH_Uptodate flag set on it.
3531                  */
3532
3533                 if ((buffer_unwritten(bh) || buffer_uptodate(bh)) == seek_data)
3534                         return lastoff;
3535
3536                 lastoff = offset;
3537         } while ((bh = bh->b_this_page) != head);
3538         return -ENOENT;
3539 }
3540
3541 /*
3542  * Seek for SEEK_DATA / SEEK_HOLE in the page cache.
3543  *
3544  * Within unwritten extents, the page cache determines which parts are holes
3545  * and which are data: unwritten and uptodate buffer heads count as data;
3546  * everything else counts as a hole.
3547  *
3548  * Returns the resulting offset on successs, and -ENOENT otherwise.
3549  */
3550 loff_t
3551 page_cache_seek_hole_data(struct inode *inode, loff_t offset, loff_t length,
3552                           int whence)
3553 {
3554         pgoff_t index = offset >> PAGE_SHIFT;
3555         pgoff_t end = DIV_ROUND_UP(offset + length, PAGE_SIZE);
3556         loff_t lastoff = offset;
3557         struct pagevec pvec;
3558
3559         if (length <= 0)
3560                 return -ENOENT;
3561
3562         pagevec_init(&pvec, 0);
3563
3564         do {
3565                 unsigned want, nr_pages, i;
3566
3567                 want = min_t(unsigned, end - index, PAGEVEC_SIZE);
3568                 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, index, want);
3569                 if (nr_pages == 0)
3570                         break;
3571
3572                 for (i = 0; i < nr_pages; i++) {
3573                         struct page *page = pvec.pages[i];
3574
3575                         /*
3576                          * At this point, the page may be truncated or
3577                          * invalidated (changing page->mapping to NULL), or
3578                          * even swizzled back from swapper_space to tmpfs file
3579                          * mapping.  However, page->index will not change
3580                          * because we have a reference on the page.
3581                          *
3582                          * If current page offset is beyond where we've ended,
3583                          * we've found a hole.