5715dac7821fe1c49a1c1ecd8f6b12192f3f1d1c
[sfrench/cifs-2.6.git] / fs / buffer.c
1 /*
2  *  linux/fs/buffer.c
3  *
4  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
5  */
6
7 /*
8  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9  *
10  * Removed a lot of unnecessary code and simplified things now that
11  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12  *
13  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
14  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
15  *
16  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
17  *
18  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
19  */
20
21 #include <linux/kernel.h>
22 #include <linux/sched/signal.h>
23 #include <linux/syscalls.h>
24 #include <linux/fs.h>
25 #include <linux/iomap.h>
26 #include <linux/mm.h>
27 #include <linux/percpu.h>
28 #include <linux/slab.h>
29 #include <linux/capability.h>
30 #include <linux/blkdev.h>
31 #include <linux/file.h>
32 #include <linux/quotaops.h>
33 #include <linux/highmem.h>
34 #include <linux/export.h>
35 #include <linux/backing-dev.h>
36 #include <linux/writeback.h>
37 #include <linux/hash.h>
38 #include <linux/suspend.h>
39 #include <linux/buffer_head.h>
40 #include <linux/task_io_accounting_ops.h>
41 #include <linux/bio.h>
42 #include <linux/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  * Install a buffer_head into this cpu's LRU.  If not already in the LRU, it is
1285  * inserted at the front, and the buffer_head at the back if any is evicted.
1286  * Or, if already in the LRU it is moved to the front.
1287  */
1288 static void bh_lru_install(struct buffer_head *bh)
1289 {
1290         struct buffer_head *evictee = bh;
1291         struct bh_lru *b;
1292         int i;
1293
1294         check_irqs_on();
1295         bh_lru_lock();
1296
1297         b = this_cpu_ptr(&bh_lrus);
1298         for (i = 0; i < BH_LRU_SIZE; i++) {
1299                 swap(evictee, b->bhs[i]);
1300                 if (evictee == bh) {
1301                         bh_lru_unlock();
1302                         return;
1303                 }
1304         }
1305
1306         get_bh(bh);
1307         bh_lru_unlock();
1308         brelse(evictee);
1309 }
1310
1311 /*
1312  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1313  */
1314 static struct buffer_head *
1315 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1316 {
1317         struct buffer_head *ret = NULL;
1318         unsigned int i;
1319
1320         check_irqs_on();
1321         bh_lru_lock();
1322         for (i = 0; i < BH_LRU_SIZE; i++) {
1323                 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1324
1325                 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1326                     bh->b_size == size) {
1327                         if (i) {
1328                                 while (i) {
1329                                         __this_cpu_write(bh_lrus.bhs[i],
1330                                                 __this_cpu_read(bh_lrus.bhs[i - 1]));
1331                                         i--;
1332                                 }
1333                                 __this_cpu_write(bh_lrus.bhs[0], bh);
1334                         }
1335                         get_bh(bh);
1336                         ret = bh;
1337                         break;
1338                 }
1339         }
1340         bh_lru_unlock();
1341         return ret;
1342 }
1343
1344 /*
1345  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1346  * it in the LRU and mark it as accessed.  If it is not present then return
1347  * NULL
1348  */
1349 struct buffer_head *
1350 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1351 {
1352         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1353
1354         if (bh == NULL) {
1355                 /* __find_get_block_slow will mark the page accessed */
1356                 bh = __find_get_block_slow(bdev, block);
1357                 if (bh)
1358                         bh_lru_install(bh);
1359         } else
1360                 touch_buffer(bh);
1361
1362         return bh;
1363 }
1364 EXPORT_SYMBOL(__find_get_block);
1365
1366 /*
1367  * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1368  * which corresponds to the passed block_device, block and size. The
1369  * returned buffer has its reference count incremented.
1370  *
1371  * __getblk_gfp() will lock up the machine if grow_dev_page's
1372  * try_to_free_buffers() attempt is failing.  FIXME, perhaps?
1373  */
1374 struct buffer_head *
1375 __getblk_gfp(struct block_device *bdev, sector_t block,
1376              unsigned size, gfp_t gfp)
1377 {
1378         struct buffer_head *bh = __find_get_block(bdev, block, size);
1379
1380         might_sleep();
1381         if (bh == NULL)
1382                 bh = __getblk_slow(bdev, block, size, gfp);
1383         return bh;
1384 }
1385 EXPORT_SYMBOL(__getblk_gfp);
1386
1387 /*
1388  * Do async read-ahead on a buffer..
1389  */
1390 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1391 {
1392         struct buffer_head *bh = __getblk(bdev, block, size);
1393         if (likely(bh)) {
1394                 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1395                 brelse(bh);
1396         }
1397 }
1398 EXPORT_SYMBOL(__breadahead);
1399
1400 /**
1401  *  __bread_gfp() - reads a specified block and returns the bh
1402  *  @bdev: the block_device to read from
1403  *  @block: number of block
1404  *  @size: size (in bytes) to read
1405  *  @gfp: page allocation flag
1406  *
1407  *  Reads a specified block, and returns buffer head that contains it.
1408  *  The page cache can be allocated from non-movable area
1409  *  not to prevent page migration if you set gfp to zero.
1410  *  It returns NULL if the block was unreadable.
1411  */
1412 struct buffer_head *
1413 __bread_gfp(struct block_device *bdev, sector_t block,
1414                    unsigned size, gfp_t gfp)
1415 {
1416         struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1417
1418         if (likely(bh) && !buffer_uptodate(bh))
1419                 bh = __bread_slow(bh);
1420         return bh;
1421 }
1422 EXPORT_SYMBOL(__bread_gfp);
1423
1424 /*
1425  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1426  * This doesn't race because it runs in each cpu either in irq
1427  * or with preempt disabled.
1428  */
1429 static void invalidate_bh_lru(void *arg)
1430 {
1431         struct bh_lru *b = &get_cpu_var(bh_lrus);
1432         int i;
1433
1434         for (i = 0; i < BH_LRU_SIZE; i++) {
1435                 brelse(b->bhs[i]);
1436                 b->bhs[i] = NULL;
1437         }
1438         put_cpu_var(bh_lrus);
1439 }
1440
1441 static bool has_bh_in_lru(int cpu, void *dummy)
1442 {
1443         struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1444         int i;
1445         
1446         for (i = 0; i < BH_LRU_SIZE; i++) {
1447                 if (b->bhs[i])
1448                         return 1;
1449         }
1450
1451         return 0;
1452 }
1453
1454 void invalidate_bh_lrus(void)
1455 {
1456         on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1457 }
1458 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1459
1460 void set_bh_page(struct buffer_head *bh,
1461                 struct page *page, unsigned long offset)
1462 {
1463         bh->b_page = page;
1464         BUG_ON(offset >= PAGE_SIZE);
1465         if (PageHighMem(page))
1466                 /*
1467                  * This catches illegal uses and preserves the offset:
1468                  */
1469                 bh->b_data = (char *)(0 + offset);
1470         else
1471                 bh->b_data = page_address(page) + offset;
1472 }
1473 EXPORT_SYMBOL(set_bh_page);
1474
1475 /*
1476  * Called when truncating a buffer on a page completely.
1477  */
1478
1479 /* Bits that are cleared during an invalidate */
1480 #define BUFFER_FLAGS_DISCARD \
1481         (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1482          1 << BH_Delay | 1 << BH_Unwritten)
1483
1484 static void discard_buffer(struct buffer_head * bh)
1485 {
1486         unsigned long b_state, b_state_old;
1487
1488         lock_buffer(bh);
1489         clear_buffer_dirty(bh);
1490         bh->b_bdev = NULL;
1491         b_state = bh->b_state;
1492         for (;;) {
1493                 b_state_old = cmpxchg(&bh->b_state, b_state,
1494                                       (b_state & ~BUFFER_FLAGS_DISCARD));
1495                 if (b_state_old == b_state)
1496                         break;
1497                 b_state = b_state_old;
1498         }
1499         unlock_buffer(bh);
1500 }
1501
1502 /**
1503  * block_invalidatepage - invalidate part or all of a buffer-backed page
1504  *
1505  * @page: the page which is affected
1506  * @offset: start of the range to invalidate
1507  * @length: length of the range to invalidate
1508  *
1509  * block_invalidatepage() is called when all or part of the page has become
1510  * invalidated by a truncate operation.
1511  *
1512  * block_invalidatepage() does not have to release all buffers, but it must
1513  * ensure that no dirty buffer is left outside @offset and that no I/O
1514  * is underway against any of the blocks which are outside the truncation
1515  * point.  Because the caller is about to free (and possibly reuse) those
1516  * blocks on-disk.
1517  */
1518 void block_invalidatepage(struct page *page, unsigned int offset,
1519                           unsigned int length)
1520 {
1521         struct buffer_head *head, *bh, *next;
1522         unsigned int curr_off = 0;
1523         unsigned int stop = length + offset;
1524
1525         BUG_ON(!PageLocked(page));
1526         if (!page_has_buffers(page))
1527                 goto out;
1528
1529         /*
1530          * Check for overflow
1531          */
1532         BUG_ON(stop > PAGE_SIZE || stop < length);
1533
1534         head = page_buffers(page);
1535         bh = head;
1536         do {
1537                 unsigned int next_off = curr_off + bh->b_size;
1538                 next = bh->b_this_page;
1539
1540                 /*
1541                  * Are we still fully in range ?
1542                  */
1543                 if (next_off > stop)
1544                         goto out;
1545
1546                 /*
1547                  * is this block fully invalidated?
1548                  */
1549                 if (offset <= curr_off)
1550                         discard_buffer(bh);
1551                 curr_off = next_off;
1552                 bh = next;
1553         } while (bh != head);
1554
1555         /*
1556          * We release buffers only if the entire page is being invalidated.
1557          * The get_block cached value has been unconditionally invalidated,
1558          * so real IO is not possible anymore.
1559          */
1560         if (offset == 0)
1561                 try_to_release_page(page, 0);
1562 out:
1563         return;
1564 }
1565 EXPORT_SYMBOL(block_invalidatepage);
1566
1567
1568 /*
1569  * We attach and possibly dirty the buffers atomically wrt
1570  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1571  * is already excluded via the page lock.
1572  */
1573 void create_empty_buffers(struct page *page,
1574                         unsigned long blocksize, unsigned long b_state)
1575 {
1576         struct buffer_head *bh, *head, *tail;
1577
1578         head = alloc_page_buffers(page, blocksize, 1);
1579         bh = head;
1580         do {
1581                 bh->b_state |= b_state;
1582                 tail = bh;
1583                 bh = bh->b_this_page;
1584         } while (bh);
1585         tail->b_this_page = head;
1586
1587         spin_lock(&page->mapping->private_lock);
1588         if (PageUptodate(page) || PageDirty(page)) {
1589                 bh = head;
1590                 do {
1591                         if (PageDirty(page))
1592                                 set_buffer_dirty(bh);
1593                         if (PageUptodate(page))
1594                                 set_buffer_uptodate(bh);
1595                         bh = bh->b_this_page;
1596                 } while (bh != head);
1597         }
1598         attach_page_buffers(page, head);
1599         spin_unlock(&page->mapping->private_lock);
1600 }
1601 EXPORT_SYMBOL(create_empty_buffers);
1602
1603 /**
1604  * clean_bdev_aliases: clean a range of buffers in block device
1605  * @bdev: Block device to clean buffers in
1606  * @block: Start of a range of blocks to clean
1607  * @len: Number of blocks to clean
1608  *
1609  * We are taking a range of blocks for data and we don't want writeback of any
1610  * buffer-cache aliases starting from return from this function and until the
1611  * moment when something will explicitly mark the buffer dirty (hopefully that
1612  * will not happen until we will free that block ;-) We don't even need to mark
1613  * it not-uptodate - nobody can expect anything from a newly allocated buffer
1614  * anyway. We used to use unmap_buffer() for such invalidation, but that was
1615  * wrong. We definitely don't want to mark the alias unmapped, for example - it
1616  * would confuse anyone who might pick it with bread() afterwards...
1617  *
1618  * Also..  Note that bforget() doesn't lock the buffer.  So there can be
1619  * writeout I/O going on against recently-freed buffers.  We don't wait on that
1620  * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1621  * need to.  That happens here.
1622  */
1623 void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1624 {
1625         struct inode *bd_inode = bdev->bd_inode;
1626         struct address_space *bd_mapping = bd_inode->i_mapping;
1627         struct pagevec pvec;
1628         pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1629         pgoff_t end;
1630         int i;
1631         struct buffer_head *bh;
1632         struct buffer_head *head;
1633
1634         end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1635         pagevec_init(&pvec, 0);
1636         while (index <= end && pagevec_lookup(&pvec, bd_mapping, index,
1637                         min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
1638                 for (i = 0; i < pagevec_count(&pvec); i++) {
1639                         struct page *page = pvec.pages[i];
1640
1641                         index = page->index;
1642                         if (index > end)
1643                                 break;
1644                         if (!page_has_buffers(page))
1645                                 continue;
1646                         /*
1647                          * We use page lock instead of bd_mapping->private_lock
1648                          * to pin buffers here since we can afford to sleep and
1649                          * it scales better than a global spinlock lock.
1650                          */
1651                         lock_page(page);
1652                         /* Recheck when the page is locked which pins bhs */
1653                         if (!page_has_buffers(page))
1654                                 goto unlock_page;
1655                         head = page_buffers(page);
1656                         bh = head;
1657                         do {
1658                                 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1659                                         goto next;
1660                                 if (bh->b_blocknr >= block + len)
1661                                         break;
1662                                 clear_buffer_dirty(bh);
1663                                 wait_on_buffer(bh);
1664                                 clear_buffer_req(bh);
1665 next:
1666                                 bh = bh->b_this_page;
1667                         } while (bh != head);
1668 unlock_page:
1669                         unlock_page(page);
1670                 }
1671                 pagevec_release(&pvec);
1672                 cond_resched();
1673                 index++;
1674         }
1675 }
1676 EXPORT_SYMBOL(clean_bdev_aliases);
1677
1678 /*
1679  * Size is a power-of-two in the range 512..PAGE_SIZE,
1680  * and the case we care about most is PAGE_SIZE.
1681  *
1682  * So this *could* possibly be written with those
1683  * constraints in mind (relevant mostly if some
1684  * architecture has a slow bit-scan instruction)
1685  */
1686 static inline int block_size_bits(unsigned int blocksize)
1687 {
1688         return ilog2(blocksize);
1689 }
1690
1691 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1692 {
1693         BUG_ON(!PageLocked(page));
1694
1695         if (!page_has_buffers(page))
1696                 create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1697         return page_buffers(page);
1698 }
1699
1700 /*
1701  * NOTE! All mapped/uptodate combinations are valid:
1702  *
1703  *      Mapped  Uptodate        Meaning
1704  *
1705  *      No      No              "unknown" - must do get_block()
1706  *      No      Yes             "hole" - zero-filled
1707  *      Yes     No              "allocated" - allocated on disk, not read in
1708  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1709  *
1710  * "Dirty" is valid only with the last case (mapped+uptodate).
1711  */
1712
1713 /*
1714  * While block_write_full_page is writing back the dirty buffers under
1715  * the page lock, whoever dirtied the buffers may decide to clean them
1716  * again at any time.  We handle that by only looking at the buffer
1717  * state inside lock_buffer().
1718  *
1719  * If block_write_full_page() is called for regular writeback
1720  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1721  * locked buffer.   This only can happen if someone has written the buffer
1722  * directly, with submit_bh().  At the address_space level PageWriteback
1723  * prevents this contention from occurring.
1724  *
1725  * If block_write_full_page() is called with wbc->sync_mode ==
1726  * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1727  * causes the writes to be flagged as synchronous writes.
1728  */
1729 int __block_write_full_page(struct inode *inode, struct page *page,
1730                         get_block_t *get_block, struct writeback_control *wbc,
1731                         bh_end_io_t *handler)
1732 {
1733         int err;
1734         sector_t block;
1735         sector_t last_block;
1736         struct buffer_head *bh, *head;
1737         unsigned int blocksize, bbits;
1738         int nr_underway = 0;
1739         int write_flags = wbc_to_write_flags(wbc);
1740
1741         head = create_page_buffers(page, inode,
1742                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1743
1744         /*
1745          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1746          * here, and the (potentially unmapped) buffers may become dirty at
1747          * any time.  If a buffer becomes dirty here after we've inspected it
1748          * then we just miss that fact, and the page stays dirty.
1749          *
1750          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1751          * handle that here by just cleaning them.
1752          */
1753
1754         bh = head;
1755         blocksize = bh->b_size;
1756         bbits = block_size_bits(blocksize);
1757
1758         block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1759         last_block = (i_size_read(inode) - 1) >> bbits;
1760
1761         /*
1762          * Get all the dirty buffers mapped to disk addresses and
1763          * handle any aliases from the underlying blockdev's mapping.
1764          */
1765         do {
1766                 if (block > last_block) {
1767                         /*
1768                          * mapped buffers outside i_size will occur, because
1769                          * this page can be outside i_size when there is a
1770                          * truncate in progress.
1771                          */
1772                         /*
1773                          * The buffer was zeroed by block_write_full_page()
1774                          */
1775                         clear_buffer_dirty(bh);
1776                         set_buffer_uptodate(bh);
1777                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1778                            buffer_dirty(bh)) {
1779                         WARN_ON(bh->b_size != blocksize);
1780                         err = get_block(inode, block, bh, 1);
1781                         if (err)
1782                                 goto recover;
1783                         clear_buffer_delay(bh);
1784                         if (buffer_new(bh)) {
1785                                 /* blockdev mappings never come here */
1786                                 clear_buffer_new(bh);
1787                                 clean_bdev_bh_alias(bh);
1788                         }
1789                 }
1790                 bh = bh->b_this_page;
1791                 block++;
1792         } while (bh != head);
1793
1794         do {
1795                 if (!buffer_mapped(bh))
1796                         continue;
1797                 /*
1798                  * If it's a fully non-blocking write attempt and we cannot
1799                  * lock the buffer then redirty the page.  Note that this can
1800                  * potentially cause a busy-wait loop from writeback threads
1801                  * and kswapd activity, but those code paths have their own
1802                  * higher-level throttling.
1803                  */
1804                 if (wbc->sync_mode != WB_SYNC_NONE) {
1805                         lock_buffer(bh);
1806                 } else if (!trylock_buffer(bh)) {
1807                         redirty_page_for_writepage(wbc, page);
1808                         continue;
1809                 }
1810                 if (test_clear_buffer_dirty(bh)) {
1811                         mark_buffer_async_write_endio(bh, handler);
1812                 } else {
1813                         unlock_buffer(bh);
1814                 }
1815         } while ((bh = bh->b_this_page) != head);
1816
1817         /*
1818          * The page and its buffers are protected by PageWriteback(), so we can
1819          * drop the bh refcounts early.
1820          */
1821         BUG_ON(PageWriteback(page));
1822         set_page_writeback(page);
1823
1824         do {
1825                 struct buffer_head *next = bh->b_this_page;
1826                 if (buffer_async_write(bh)) {
1827                         submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1828                                         inode->i_write_hint, wbc);
1829                         nr_underway++;
1830                 }
1831                 bh = next;
1832         } while (bh != head);
1833         unlock_page(page);
1834
1835         err = 0;
1836 done:
1837         if (nr_underway == 0) {
1838                 /*
1839                  * The page was marked dirty, but the buffers were
1840                  * clean.  Someone wrote them back by hand with
1841                  * ll_rw_block/submit_bh.  A rare case.
1842                  */
1843                 end_page_writeback(page);
1844
1845                 /*
1846                  * The page and buffer_heads can be released at any time from
1847                  * here on.
1848                  */
1849         }
1850         return err;
1851
1852 recover:
1853         /*
1854          * ENOSPC, or some other error.  We may already have added some
1855          * blocks to the file, so we need to write these out to avoid
1856          * exposing stale data.
1857          * The page is currently locked and not marked for writeback
1858          */
1859         bh = head;
1860         /* Recovery: lock and submit the mapped buffers */
1861         do {
1862                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1863                     !buffer_delay(bh)) {
1864                         lock_buffer(bh);
1865                         mark_buffer_async_write_endio(bh, handler);
1866                 } else {
1867                         /*
1868                          * The buffer may have been set dirty during
1869                          * attachment to a dirty page.
1870                          */
1871                         clear_buffer_dirty(bh);
1872                 }
1873         } while ((bh = bh->b_this_page) != head);
1874         SetPageError(page);
1875         BUG_ON(PageWriteback(page));
1876         mapping_set_error(page->mapping, err);
1877         set_page_writeback(page);
1878         do {
1879                 struct buffer_head *next = bh->b_this_page;
1880                 if (buffer_async_write(bh)) {
1881                         clear_buffer_dirty(bh);
1882                         submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1883                                         inode->i_write_hint, wbc);
1884                         nr_underway++;
1885                 }
1886                 bh = next;
1887         } while (bh != head);
1888         unlock_page(page);
1889         goto done;
1890 }
1891 EXPORT_SYMBOL(__block_write_full_page);
1892
1893 /*
1894  * If a page has any new buffers, zero them out here, and mark them uptodate
1895  * and dirty so they'll be written out (in order to prevent uninitialised
1896  * block data from leaking). And clear the new bit.
1897  */
1898 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1899 {
1900         unsigned int block_start, block_end;
1901         struct buffer_head *head, *bh;
1902
1903         BUG_ON(!PageLocked(page));
1904         if (!page_has_buffers(page))
1905                 return;
1906
1907         bh = head = page_buffers(page);
1908         block_start = 0;
1909         do {
1910                 block_end = block_start + bh->b_size;
1911
1912                 if (buffer_new(bh)) {
1913                         if (block_end > from && block_start < to) {
1914                                 if (!PageUptodate(page)) {
1915                                         unsigned start, size;
1916
1917                                         start = max(from, block_start);
1918                                         size = min(to, block_end) - start;
1919
1920                                         zero_user(page, start, size);
1921                                         set_buffer_uptodate(bh);
1922                                 }
1923
1924                                 clear_buffer_new(bh);
1925                                 mark_buffer_dirty(bh);
1926                         }
1927                 }
1928
1929                 block_start = block_end;
1930                 bh = bh->b_this_page;
1931         } while (bh != head);
1932 }
1933 EXPORT_SYMBOL(page_zero_new_buffers);
1934
1935 static void
1936 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1937                 struct iomap *iomap)
1938 {
1939         loff_t offset = block << inode->i_blkbits;
1940
1941         bh->b_bdev = iomap->bdev;
1942
1943         /*
1944          * Block points to offset in file we need to map, iomap contains
1945          * the offset at which the map starts. If the map ends before the
1946          * current block, then do not map the buffer and let the caller
1947          * handle it.
1948          */
1949         BUG_ON(offset >= iomap->offset + iomap->length);
1950
1951         switch (iomap->type) {
1952         case IOMAP_HOLE:
1953                 /*
1954                  * If the buffer is not up to date or beyond the current EOF,
1955                  * we need to mark it as new to ensure sub-block zeroing is
1956                  * executed if necessary.
1957                  */
1958                 if (!buffer_uptodate(bh) ||
1959                     (offset >= i_size_read(inode)))
1960                         set_buffer_new(bh);
1961                 break;
1962         case IOMAP_DELALLOC:
1963                 if (!buffer_uptodate(bh) ||
1964                     (offset >= i_size_read(inode)))
1965                         set_buffer_new(bh);
1966                 set_buffer_uptodate(bh);
1967                 set_buffer_mapped(bh);
1968                 set_buffer_delay(bh);
1969                 break;
1970         case IOMAP_UNWRITTEN:
1971                 /*
1972                  * For unwritten regions, we always need to ensure that
1973                  * sub-block writes cause the regions in the block we are not
1974                  * writing to are zeroed. Set the buffer as new to ensure this.
1975                  */
1976                 set_buffer_new(bh);
1977                 set_buffer_unwritten(bh);
1978                 /* FALLTHRU */
1979         case IOMAP_MAPPED:
1980                 if (offset >= i_size_read(inode))
1981                         set_buffer_new(bh);
1982                 bh->b_blocknr = (iomap->blkno >> (inode->i_blkbits - 9)) +
1983                                 ((offset - iomap->offset) >> inode->i_blkbits);
1984                 set_buffer_mapped(bh);
1985                 break;
1986         }
1987 }
1988
1989 int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
1990                 get_block_t *get_block, struct iomap *iomap)
1991 {
1992         unsigned from = pos & (PAGE_SIZE - 1);
1993         unsigned to = from + len;
1994         struct inode *inode = page->mapping->host;
1995         unsigned block_start, block_end;
1996         sector_t block;
1997         int err = 0;
1998         unsigned blocksize, bbits;
1999         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
2000
2001         BUG_ON(!PageLocked(page));
2002         BUG_ON(from > PAGE_SIZE);
2003         BUG_ON(to > PAGE_SIZE);
2004         BUG_ON(from > to);
2005
2006         head = create_page_buffers(page, inode, 0);
2007         blocksize = head->b_size;
2008         bbits = block_size_bits(blocksize);
2009
2010         block = (sector_t)page->index << (PAGE_SHIFT - bbits);
2011
2012         for(bh = head, block_start = 0; bh != head || !block_start;
2013             block++, block_start=block_end, bh = bh->b_this_page) {
2014                 block_end = block_start + blocksize;
2015                 if (block_end <= from || block_start >= to) {
2016                         if (PageUptodate(page)) {
2017                                 if (!buffer_uptodate(bh))
2018                                         set_buffer_uptodate(bh);
2019                         }
2020                         continue;
2021                 }
2022                 if (buffer_new(bh))
2023                         clear_buffer_new(bh);
2024                 if (!buffer_mapped(bh)) {
2025                         WARN_ON(bh->b_size != blocksize);
2026                         if (get_block) {
2027                                 err = get_block(inode, block, bh, 1);
2028                                 if (err)
2029                                         break;
2030                         } else {
2031                                 iomap_to_bh(inode, block, bh, iomap);
2032                         }
2033
2034                         if (buffer_new(bh)) {
2035                                 clean_bdev_bh_alias(bh);
2036                                 if (PageUptodate(page)) {
2037                                         clear_buffer_new(bh);
2038                                         set_buffer_uptodate(bh);
2039                                         mark_buffer_dirty(bh);
2040                                         continue;
2041                                 }
2042                                 if (block_end > to || block_start < from)
2043                                         zero_user_segments(page,
2044                                                 to, block_end,
2045                                                 block_start, from);
2046                                 continue;
2047                         }
2048                 }
2049                 if (PageUptodate(page)) {
2050                         if (!buffer_uptodate(bh))
2051                                 set_buffer_uptodate(bh);
2052                         continue; 
2053                 }
2054                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2055                     !buffer_unwritten(bh) &&
2056                      (block_start < from || block_end > to)) {
2057                         ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2058                         *wait_bh++=bh;
2059                 }
2060         }
2061         /*
2062          * If we issued read requests - let them complete.
2063          */
2064         while(wait_bh > wait) {
2065                 wait_on_buffer(*--wait_bh);
2066                 if (!buffer_uptodate(*wait_bh))
2067                         err = -EIO;
2068         }
2069         if (unlikely(err))
2070                 page_zero_new_buffers(page, from, to);
2071         return err;
2072 }
2073
2074 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2075                 get_block_t *get_block)
2076 {
2077         return __block_write_begin_int(page, pos, len, get_block, NULL);
2078 }
2079 EXPORT_SYMBOL(__block_write_begin);
2080
2081 static int __block_commit_write(struct inode *inode, struct page *page,
2082                 unsigned from, unsigned to)
2083 {
2084         unsigned block_start, block_end;
2085         int partial = 0;
2086         unsigned blocksize;
2087         struct buffer_head *bh, *head;
2088
2089         bh = head = page_buffers(page);
2090         blocksize = bh->b_size;
2091
2092         block_start = 0;
2093         do {
2094                 block_end = block_start + blocksize;
2095                 if (block_end <= from || block_start >= to) {
2096                         if (!buffer_uptodate(bh))
2097                                 partial = 1;
2098                 } else {
2099                         set_buffer_uptodate(bh);
2100                         mark_buffer_dirty(bh);
2101                 }
2102                 clear_buffer_new(bh);
2103
2104                 block_start = block_end;
2105                 bh = bh->b_this_page;
2106         } while (bh != head);
2107
2108         /*
2109          * If this is a partial write which happened to make all buffers
2110          * uptodate then we can optimize away a bogus readpage() for
2111          * the next read(). Here we 'discover' whether the page went
2112          * uptodate as a result of this (potentially partial) write.
2113          */
2114         if (!partial)
2115                 SetPageUptodate(page);
2116         return 0;
2117 }
2118
2119 /*
2120  * block_write_begin takes care of the basic task of block allocation and
2121  * bringing partial write blocks uptodate first.
2122  *
2123  * The filesystem needs to handle block truncation upon failure.
2124  */
2125 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2126                 unsigned flags, struct page **pagep, get_block_t *get_block)
2127 {
2128         pgoff_t index = pos >> PAGE_SHIFT;
2129         struct page *page;
2130         int status;
2131
2132         page = grab_cache_page_write_begin(mapping, index, flags);
2133         if (!page)
2134                 return -ENOMEM;
2135
2136         status = __block_write_begin(page, pos, len, get_block);
2137         if (unlikely(status)) {
2138                 unlock_page(page);
2139                 put_page(page);
2140                 page = NULL;
2141         }
2142
2143         *pagep = page;
2144         return status;
2145 }
2146 EXPORT_SYMBOL(block_write_begin);
2147
2148 int block_write_end(struct file *file, struct address_space *mapping,
2149                         loff_t pos, unsigned len, unsigned copied,
2150                         struct page *page, void *fsdata)
2151 {
2152         struct inode *inode = mapping->host;
2153         unsigned start;
2154
2155         start = pos & (PAGE_SIZE - 1);
2156
2157         if (unlikely(copied < len)) {
2158                 /*
2159                  * The buffers that were written will now be uptodate, so we
2160                  * don't have to worry about a readpage reading them and
2161                  * overwriting a partial write. However if we have encountered
2162                  * a short write and only partially written into a buffer, it
2163                  * will not be marked uptodate, so a readpage might come in and
2164                  * destroy our partial write.
2165                  *
2166                  * Do the simplest thing, and just treat any short write to a
2167                  * non uptodate page as a zero-length write, and force the
2168                  * caller to redo the whole thing.
2169                  */
2170                 if (!PageUptodate(page))
2171                         copied = 0;
2172
2173                 page_zero_new_buffers(page, start+copied, start+len);
2174         }
2175         flush_dcache_page(page);
2176
2177         /* This could be a short (even 0-length) commit */
2178         __block_commit_write(inode, page, start, start+copied);
2179
2180         return copied;
2181 }
2182 EXPORT_SYMBOL(block_write_end);
2183
2184 int generic_write_end(struct file *file, struct address_space *mapping,
2185                         loff_t pos, unsigned len, unsigned copied,
2186                         struct page *page, void *fsdata)
2187 {
2188         struct inode *inode = mapping->host;
2189         loff_t old_size = inode->i_size;
2190         int i_size_changed = 0;
2191
2192         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2193
2194         /*
2195          * No need to use i_size_read() here, the i_size
2196          * cannot change under us because we hold i_mutex.
2197          *
2198          * But it's important to update i_size while still holding page lock:
2199          * page writeout could otherwise come in and zero beyond i_size.
2200          */
2201         if (pos+copied > inode->i_size) {
2202                 i_size_write(inode, pos+copied);
2203                 i_size_changed = 1;
2204         }
2205
2206         unlock_page(page);
2207         put_page(page);
2208
2209         if (old_size < pos)
2210                 pagecache_isize_extended(inode, old_size, pos);
2211         /*
2212          * Don't mark the inode dirty under page lock. First, it unnecessarily
2213          * makes the holding time of page lock longer. Second, it forces lock
2214          * ordering of page lock and transaction start for journaling
2215          * filesystems.
2216          */
2217         if (i_size_changed)
2218                 mark_inode_dirty(inode);
2219
2220         return copied;
2221 }
2222 EXPORT_SYMBOL(generic_write_end);
2223
2224 /*
2225  * block_is_partially_uptodate checks whether buffers within a page are
2226  * uptodate or not.
2227  *
2228  * Returns true if all buffers which correspond to a file portion
2229  * we want to read are uptodate.
2230  */
2231 int block_is_partially_uptodate(struct page *page, unsigned long from,
2232                                         unsigned long count)
2233 {
2234         unsigned block_start, block_end, blocksize;
2235         unsigned to;
2236         struct buffer_head *bh, *head;
2237         int ret = 1;
2238
2239         if (!page_has_buffers(page))
2240                 return 0;
2241
2242         head = page_buffers(page);
2243         blocksize = head->b_size;
2244         to = min_t(unsigned, PAGE_SIZE - from, count);
2245         to = from + to;
2246         if (from < blocksize && to > PAGE_SIZE - blocksize)
2247                 return 0;
2248
2249         bh = head;
2250         block_start = 0;
2251         do {
2252                 block_end = block_start + blocksize;
2253                 if (block_end > from && block_start < to) {
2254                         if (!buffer_uptodate(bh)) {
2255                                 ret = 0;
2256                                 break;
2257                         }
2258                         if (block_end >= to)
2259                                 break;
2260                 }
2261                 block_start = block_end;
2262                 bh = bh->b_this_page;
2263         } while (bh != head);
2264
2265         return ret;
2266 }
2267 EXPORT_SYMBOL(block_is_partially_uptodate);
2268
2269 /*
2270  * Generic "read page" function for block devices that have the normal
2271  * get_block functionality. This is most of the block device filesystems.
2272  * Reads the page asynchronously --- the unlock_buffer() and
2273  * set/clear_buffer_uptodate() functions propagate buffer state into the
2274  * page struct once IO has completed.
2275  */
2276 int block_read_full_page(struct page *page, get_block_t *get_block)
2277 {
2278         struct inode *inode = page->mapping->host;
2279         sector_t iblock, lblock;
2280         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2281         unsigned int blocksize, bbits;
2282         int nr, i;
2283         int fully_mapped = 1;
2284
2285         head = create_page_buffers(page, inode, 0);
2286         blocksize = head->b_size;
2287         bbits = block_size_bits(blocksize);
2288
2289         iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2290         lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2291         bh = head;
2292         nr = 0;
2293         i = 0;
2294
2295         do {
2296                 if (buffer_uptodate(bh))
2297                         continue;
2298
2299                 if (!buffer_mapped(bh)) {
2300                         int err = 0;
2301
2302                         fully_mapped = 0;
2303                         if (iblock < lblock) {
2304                                 WARN_ON(bh->b_size != blocksize);
2305                                 err = get_block(inode, iblock, bh, 0);
2306                                 if (err)
2307                                         SetPageError(page);
2308                         }
2309                         if (!buffer_mapped(bh)) {
2310                                 zero_user(page, i * blocksize, blocksize);
2311                                 if (!err)
2312                                         set_buffer_uptodate(bh);
2313                                 continue;
2314                         }
2315                         /*
2316                          * get_block() might have updated the buffer
2317                          * synchronously
2318                          */
2319                         if (buffer_uptodate(bh))
2320                                 continue;
2321                 }
2322                 arr[nr++] = bh;
2323         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2324
2325         if (fully_mapped)
2326                 SetPageMappedToDisk(page);
2327
2328         if (!nr) {
2329                 /*
2330                  * All buffers are uptodate - we can set the page uptodate
2331                  * as well. But not if get_block() returned an error.
2332                  */
2333                 if (!PageError(page))
2334                         SetPageUptodate(page);
2335                 unlock_page(page);
2336                 return 0;
2337         }
2338
2339         /* Stage two: lock the buffers */
2340         for (i = 0; i < nr; i++) {
2341                 bh = arr[i];
2342                 lock_buffer(bh);
2343                 mark_buffer_async_read(bh);
2344         }
2345
2346         /*
2347          * Stage 3: start the IO.  Check for uptodateness
2348          * inside the buffer lock in case another process reading
2349          * the underlying blockdev brought it uptodate (the sct fix).
2350          */
2351         for (i = 0; i < nr; i++) {
2352                 bh = arr[i];
2353                 if (buffer_uptodate(bh))
2354                         end_buffer_async_read(bh, 1);
2355                 else
2356                         submit_bh(REQ_OP_READ, 0, bh);
2357         }
2358         return 0;
2359 }
2360 EXPORT_SYMBOL(block_read_full_page);
2361
2362 /* utility function for filesystems that need to do work on expanding
2363  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2364  * deal with the hole.  
2365  */
2366 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2367 {
2368         struct address_space *mapping = inode->i_mapping;
2369         struct page *page;
2370         void *fsdata;
2371         int err;
2372
2373         err = inode_newsize_ok(inode, size);
2374         if (err)
2375                 goto out;
2376
2377         err = pagecache_write_begin(NULL, mapping, size, 0,
2378                                     AOP_FLAG_CONT_EXPAND, &page, &fsdata);
2379         if (err)
2380                 goto out;
2381
2382         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2383         BUG_ON(err > 0);
2384
2385 out:
2386         return err;
2387 }
2388 EXPORT_SYMBOL(generic_cont_expand_simple);
2389
2390 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2391                             loff_t pos, loff_t *bytes)
2392 {
2393         struct inode *inode = mapping->host;
2394         unsigned int blocksize = i_blocksize(inode);
2395         struct page *page;
2396         void *fsdata;
2397         pgoff_t index, curidx;
2398         loff_t curpos;
2399         unsigned zerofrom, offset, len;
2400         int err = 0;
2401
2402         index = pos >> PAGE_SHIFT;
2403         offset = pos & ~PAGE_MASK;
2404
2405         while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2406                 zerofrom = curpos & ~PAGE_MASK;
2407                 if (zerofrom & (blocksize-1)) {
2408                         *bytes |= (blocksize-1);
2409                         (*bytes)++;
2410                 }
2411                 len = PAGE_SIZE - zerofrom;
2412
2413                 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2414                                             &page, &fsdata);
2415                 if (err)
2416                         goto out;
2417                 zero_user(page, zerofrom, len);
2418                 err = pagecache_write_end(file, mapping, curpos, len, len,
2419                                                 page, fsdata);
2420                 if (err < 0)
2421                         goto out;
2422                 BUG_ON(err != len);
2423                 err = 0;
2424
2425                 balance_dirty_pages_ratelimited(mapping);
2426
2427                 if (unlikely(fatal_signal_pending(current))) {
2428                         err = -EINTR;
2429                         goto out;
2430                 }
2431         }
2432
2433         /* page covers the boundary, find the boundary offset */
2434         if (index == curidx) {
2435                 zerofrom = curpos & ~PAGE_MASK;
2436                 /* if we will expand the thing last block will be filled */
2437                 if (offset <= zerofrom) {
2438                         goto out;
2439                 }
2440                 if (zerofrom & (blocksize-1)) {
2441                         *bytes |= (blocksize-1);
2442                         (*bytes)++;
2443                 }
2444                 len = offset - zerofrom;
2445
2446                 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2447                                             &page, &fsdata);
2448                 if (err)
2449                         goto out;
2450                 zero_user(page, zerofrom, len);
2451                 err = pagecache_write_end(file, mapping, curpos, len, len,
2452                                                 page, fsdata);
2453                 if (err < 0)
2454                         goto out;
2455                 BUG_ON(err != len);
2456                 err = 0;
2457         }
2458 out:
2459         return err;
2460 }
2461
2462 /*
2463  * For moronic filesystems that do not allow holes in file.
2464  * We may have to extend the file.
2465  */
2466 int cont_write_begin(struct file *file, struct address_space *mapping,
2467                         loff_t pos, unsigned len, unsigned flags,
2468                         struct page **pagep, void **fsdata,
2469                         get_block_t *get_block, loff_t *bytes)
2470 {
2471         struct inode *inode = mapping->host;
2472         unsigned int blocksize = i_blocksize(inode);
2473         unsigned int zerofrom;
2474         int err;
2475
2476         err = cont_expand_zero(file, mapping, pos, bytes);
2477         if (err)
2478                 return err;
2479
2480         zerofrom = *bytes & ~PAGE_MASK;
2481         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2482                 *bytes |= (blocksize-1);
2483                 (*bytes)++;
2484         }
2485
2486         return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2487 }
2488 EXPORT_SYMBOL(cont_write_begin);
2489
2490 int block_commit_write(struct page *page, unsigned from, unsigned to)
2491 {
2492         struct inode *inode = page->mapping->host;
2493         __block_commit_write(inode,page,from,to);
2494         return 0;
2495 }
2496 EXPORT_SYMBOL(block_commit_write);
2497
2498 /*
2499  * block_page_mkwrite() is not allowed to change the file size as it gets
2500  * called from a page fault handler when a page is first dirtied. Hence we must
2501  * be careful to check for EOF conditions here. We set the page up correctly
2502  * for a written page which means we get ENOSPC checking when writing into
2503  * holes and correct delalloc and unwritten extent mapping on filesystems that
2504  * support these features.
2505  *
2506  * We are not allowed to take the i_mutex here so we have to play games to
2507  * protect against truncate races as the page could now be beyond EOF.  Because
2508  * truncate writes the inode size before removing pages, once we have the
2509  * page lock we can determine safely if the page is beyond EOF. If it is not
2510  * beyond EOF, then the page is guaranteed safe against truncation until we
2511  * unlock the page.
2512  *
2513  * Direct callers of this function should protect against filesystem freezing
2514  * using sb_start_pagefault() - sb_end_pagefault() functions.
2515  */
2516 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2517                          get_block_t get_block)
2518 {
2519         struct page *page = vmf->page;
2520         struct inode *inode = file_inode(vma->vm_file);
2521         unsigned long end;
2522         loff_t size;
2523         int ret;
2524
2525         lock_page(page);
2526         size = i_size_read(inode);
2527         if ((page->mapping != inode->i_mapping) ||
2528             (page_offset(page) > size)) {
2529                 /* We overload EFAULT to mean page got truncated */
2530                 ret = -EFAULT;
2531                 goto out_unlock;
2532         }
2533
2534         /* page is wholly or partially inside EOF */
2535         if (((page->index + 1) << PAGE_SHIFT) > size)
2536                 end = size & ~PAGE_MASK;
2537         else
2538                 end = PAGE_SIZE;
2539
2540         ret = __block_write_begin(page, 0, end, get_block);
2541         if (!ret)
2542                 ret = block_commit_write(page, 0, end);
2543
2544         if (unlikely(ret < 0))
2545                 goto out_unlock;
2546         set_page_dirty(page);
2547         wait_for_stable_page(page);
2548         return 0;
2549 out_unlock:
2550         unlock_page(page);
2551         return ret;
2552 }
2553 EXPORT_SYMBOL(block_page_mkwrite);
2554
2555 /*
2556  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2557  * immediately, while under the page lock.  So it needs a special end_io
2558  * handler which does not touch the bh after unlocking it.
2559  */
2560 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2561 {
2562         __end_buffer_read_notouch(bh, uptodate);
2563 }
2564
2565 /*
2566  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2567  * the page (converting it to circular linked list and taking care of page
2568  * dirty races).
2569  */
2570 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2571 {
2572         struct buffer_head *bh;
2573
2574         BUG_ON(!PageLocked(page));
2575
2576         spin_lock(&page->mapping->private_lock);
2577         bh = head;
2578         do {
2579                 if (PageDirty(page))
2580                         set_buffer_dirty(bh);
2581                 if (!bh->b_this_page)
2582                         bh->b_this_page = head;
2583                 bh = bh->b_this_page;
2584         } while (bh != head);
2585         attach_page_buffers(page, head);
2586         spin_unlock(&page->mapping->private_lock);
2587 }
2588
2589 /*
2590  * On entry, the page is fully not uptodate.
2591  * On exit the page is fully uptodate in the areas outside (from,to)
2592  * The filesystem needs to handle block truncation upon failure.
2593  */
2594 int nobh_write_begin(struct address_space *mapping,
2595                         loff_t pos, unsigned len, unsigned flags,
2596                         struct page **pagep, void **fsdata,
2597                         get_block_t *get_block)
2598 {
2599         struct inode *inode = mapping->host;
2600         const unsigned blkbits = inode->i_blkbits;
2601         const unsigned blocksize = 1 << blkbits;
2602         struct buffer_head *head, *bh;
2603         struct page *page;
2604         pgoff_t index;
2605         unsigned from, to;
2606         unsigned block_in_page;
2607         unsigned block_start, block_end;
2608         sector_t block_in_file;
2609         int nr_reads = 0;
2610         int ret = 0;
2611         int is_mapped_to_disk = 1;
2612
2613         index = pos >> PAGE_SHIFT;
2614         from = pos & (PAGE_SIZE - 1);
2615         to = from + len;
2616
2617         page = grab_cache_page_write_begin(mapping, index, flags);
2618         if (!page)
2619                 return -ENOMEM;
2620         *pagep = page;
2621         *fsdata = NULL;
2622
2623         if (page_has_buffers(page)) {
2624                 ret = __block_write_begin(page, pos, len, get_block);
2625                 if (unlikely(ret))
2626                         goto out_release;
2627                 return ret;
2628         }
2629
2630         if (PageMappedToDisk(page))
2631                 return 0;
2632
2633         /*
2634          * Allocate buffers so that we can keep track of state, and potentially
2635          * attach them to the page if an error occurs. In the common case of
2636          * no error, they will just be freed again without ever being attached
2637          * to the page (which is all OK, because we're under the page lock).
2638          *
2639          * Be careful: the buffer linked list is a NULL terminated one, rather
2640          * than the circular one we're used to.
2641          */
2642         head = alloc_page_buffers(page, blocksize, 0);
2643         if (!head) {
2644                 ret = -ENOMEM;
2645                 goto out_release;
2646         }
2647
2648         block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2649
2650         /*
2651          * We loop across all blocks in the page, whether or not they are
2652          * part of the affected region.  This is so we can discover if the
2653          * page is fully mapped-to-disk.
2654          */
2655         for (block_start = 0, block_in_page = 0, bh = head;
2656                   block_start < PAGE_SIZE;
2657                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2658                 int create;
2659
2660                 block_end = block_start + blocksize;
2661                 bh->b_state = 0;
2662                 create = 1;
2663                 if (block_start >= to)
2664                         create = 0;
2665                 ret = get_block(inode, block_in_file + block_in_page,
2666                                         bh, create);
2667                 if (ret)
2668                         goto failed;
2669                 if (!buffer_mapped(bh))
2670                         is_mapped_to_disk = 0;
2671                 if (buffer_new(bh))
2672                         clean_bdev_bh_alias(bh);
2673                 if (PageUptodate(page)) {
2674                         set_buffer_uptodate(bh);
2675                         continue;
2676                 }
2677                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2678                         zero_user_segments(page, block_start, from,
2679                                                         to, block_end);
2680                         continue;
2681                 }
2682                 if (buffer_uptodate(bh))
2683                         continue;       /* reiserfs does this */
2684                 if (block_start < from || block_end > to) {
2685                         lock_buffer(bh);
2686                         bh->b_end_io = end_buffer_read_nobh;
2687                         submit_bh(REQ_OP_READ, 0, bh);
2688                         nr_reads++;
2689                 }
2690         }
2691
2692         if (nr_reads) {
2693                 /*
2694                  * The page is locked, so these buffers are protected from
2695                  * any VM or truncate activity.  Hence we don't need to care
2696                  * for the buffer_head refcounts.
2697                  */
2698                 for (bh = head; bh; bh = bh->b_this_page) {
2699                         wait_on_buffer(bh);
2700                         if (!buffer_uptodate(bh))
2701                                 ret = -EIO;
2702                 }
2703                 if (ret)
2704                         goto failed;
2705         }
2706
2707         if (is_mapped_to_disk)
2708                 SetPageMappedToDisk(page);
2709
2710         *fsdata = head; /* to be released by nobh_write_end */
2711
2712         return 0;
2713
2714 failed:
2715         BUG_ON(!ret);
2716         /*
2717          * Error recovery is a bit difficult. We need to zero out blocks that
2718          * were newly allocated, and dirty them to ensure they get written out.
2719          * Buffers need to be attached to the page at this point, otherwise
2720          * the handling of potential IO errors during writeout would be hard
2721          * (could try doing synchronous writeout, but what if that fails too?)
2722          */
2723         attach_nobh_buffers(page, head);
2724         page_zero_new_buffers(page, from, to);
2725
2726 out_release:
2727         unlock_page(page);
2728         put_page(page);
2729         *pagep = NULL;
2730
2731         return ret;
2732 }
2733 EXPORT_SYMBOL(nobh_write_begin);
2734
2735 int nobh_write_end(struct file *file, struct address_space *mapping,
2736                         loff_t pos, unsigned len, unsigned copied,
2737                         struct page *page, void *fsdata)
2738 {
2739         struct inode *inode = page->mapping->host;
2740         struct buffer_head *head = fsdata;
2741         struct buffer_head *bh;
2742         BUG_ON(fsdata != NULL && page_has_buffers(page));
2743
2744         if (unlikely(copied < len) && head)
2745                 attach_nobh_buffers(page, head);
2746         if (page_has_buffers(page))
2747                 return generic_write_end(file, mapping, pos, len,
2748                                         copied, page, fsdata);
2749
2750         SetPageUptodate(page);
2751         set_page_dirty(page);
2752         if (pos+copied > inode->i_size) {
2753                 i_size_write(inode, pos+copied);
2754                 mark_inode_dirty(inode);
2755         }
2756
2757         unlock_page(page);
2758         put_page(page);
2759
2760         while (head) {
2761                 bh = head;
2762                 head = head->b_this_page;
2763                 free_buffer_head(bh);
2764         }
2765
2766         return copied;
2767 }
2768 EXPORT_SYMBOL(nobh_write_end);
2769
2770 /*
2771  * nobh_writepage() - based on block_full_write_page() except
2772  * that it tries to operate without attaching bufferheads to
2773  * the page.
2774  */
2775 int nobh_writepage(struct page *page, get_block_t *get_block,
2776                         struct writeback_control *wbc)
2777 {
2778         struct inode * const inode = page->mapping->host;
2779         loff_t i_size = i_size_read(inode);
2780         const pgoff_t end_index = i_size >> PAGE_SHIFT;
2781         unsigned offset;
2782         int ret;
2783
2784         /* Is the page fully inside i_size? */
2785         if (page->index < end_index)
2786                 goto out;
2787
2788         /* Is the page fully outside i_size? (truncate in progress) */
2789         offset = i_size & (PAGE_SIZE-1);
2790         if (page->index >= end_index+1 || !offset) {
2791                 /*
2792                  * The page may have dirty, unmapped buffers.  For example,
2793                  * they may have been added in ext3_writepage().  Make them
2794                  * freeable here, so the page does not leak.
2795                  */
2796 #if 0
2797                 /* Not really sure about this  - do we need this ? */
2798                 if (page->mapping->a_ops->invalidatepage)
2799                         page->mapping->a_ops->invalidatepage(page, offset);
2800 #endif
2801                 unlock_page(page);
2802                 return 0; /* don't care */
2803         }
2804
2805         /*
2806          * The page straddles i_size.  It must be zeroed out on each and every
2807          * writepage invocation because it may be mmapped.  "A file is mapped
2808          * in multiples of the page size.  For a file that is not a multiple of
2809          * the  page size, the remaining memory is zeroed when mapped, and
2810          * writes to that region are not written out to the file."
2811          */
2812         zero_user_segment(page, offset, PAGE_SIZE);
2813 out:
2814         ret = mpage_writepage(page, get_block, wbc);
2815         if (ret == -EAGAIN)
2816                 ret = __block_write_full_page(inode, page, get_block, wbc,
2817                                               end_buffer_async_write);
2818         return ret;
2819 }
2820 EXPORT_SYMBOL(nobh_writepage);
2821
2822 int nobh_truncate_page(struct address_space *mapping,
2823                         loff_t from, get_block_t *get_block)
2824 {
2825         pgoff_t index = from >> PAGE_SHIFT;
2826         unsigned offset = from & (PAGE_SIZE-1);
2827         unsigned blocksize;
2828         sector_t iblock;
2829         unsigned length, pos;
2830         struct inode *inode = mapping->host;
2831         struct page *page;
2832         struct buffer_head map_bh;
2833         int err;
2834
2835         blocksize = i_blocksize(inode);
2836         length = offset & (blocksize - 1);
2837
2838         /* Block boundary? Nothing to do */
2839         if (!length)
2840                 return 0;
2841
2842         length = blocksize - length;
2843         iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2844
2845         page = grab_cache_page(mapping, index);
2846         err = -ENOMEM;
2847         if (!page)
2848                 goto out;
2849
2850         if (page_has_buffers(page)) {
2851 has_buffers:
2852                 unlock_page(page);
2853                 put_page(page);
2854                 return block_truncate_page(mapping, from, get_block);
2855         }
2856
2857         /* Find the buffer that contains "offset" */
2858         pos = blocksize;
2859         while (offset >= pos) {
2860                 iblock++;
2861                 pos += blocksize;
2862         }
2863
2864         map_bh.b_size = blocksize;
2865         map_bh.b_state = 0;
2866         err = get_block(inode, iblock, &map_bh, 0);
2867         if (err)
2868                 goto unlock;
2869         /* unmapped? It's a hole - nothing to do */
2870         if (!buffer_mapped(&map_bh))
2871                 goto unlock;
2872
2873         /* Ok, it's mapped. Make sure it's up-to-date */
2874         if (!PageUptodate(page)) {
2875                 err = mapping->a_ops->readpage(NULL, page);
2876                 if (err) {
2877                         put_page(page);
2878                         goto out;
2879                 }
2880                 lock_page(page);
2881                 if (!PageUptodate(page)) {
2882                         err = -EIO;
2883                         goto unlock;
2884                 }
2885                 if (page_has_buffers(page))
2886                         goto has_buffers;
2887         }
2888         zero_user(page, offset, length);
2889         set_page_dirty(page);
2890         err = 0;
2891
2892 unlock:
2893         unlock_page(page);
2894         put_page(page);
2895 out:
2896         return err;
2897 }
2898 EXPORT_SYMBOL(nobh_truncate_page);
2899
2900 int block_truncate_page(struct address_space *mapping,
2901                         loff_t from, get_block_t *get_block)
2902 {
2903         pgoff_t index = from >> PAGE_SHIFT;
2904         unsigned offset = from & (PAGE_SIZE-1);
2905         unsigned blocksize;
2906         sector_t iblock;
2907         unsigned length, pos;
2908         struct inode *inode = mapping->host;
2909         struct page *page;
2910         struct buffer_head *bh;
2911         int err;
2912
2913         blocksize = i_blocksize(inode);
2914         length = offset & (blocksize - 1);
2915
2916         /* Block boundary? Nothing to do */
2917         if (!length)
2918                 return 0;
2919
2920         length = blocksize - length;
2921         iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2922         
2923         page = grab_cache_page(mapping, index);
2924         err = -ENOMEM;
2925         if (!page)
2926                 goto out;
2927
2928         if (!page_has_buffers(page))
2929                 create_empty_buffers(page, blocksize, 0);
2930
2931         /* Find the buffer that contains "offset" */
2932         bh = page_buffers(page);
2933         pos = blocksize;
2934         while (offset >= pos) {
2935                 bh = bh->b_this_page;
2936                 iblock++;
2937                 pos += blocksize;
2938         }
2939
2940         err = 0;
2941         if (!buffer_mapped(bh)) {
2942                 WARN_ON(bh->b_size != blocksize);
2943                 err = get_block(inode, iblock, bh, 0);
2944                 if (err)
2945                         goto unlock;
2946                 /* unmapped? It's a hole - nothing to do */
2947                 if (!buffer_mapped(bh))
2948                         goto unlock;
2949         }
2950
2951         /* Ok, it's mapped. Make sure it's up-to-date */
2952         if (PageUptodate(page))
2953                 set_buffer_uptodate(bh);
2954
2955         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2956                 err = -EIO;
2957                 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2958                 wait_on_buffer(bh);
2959                 /* Uhhuh. Read error. Complain and punt. */
2960                 if (!buffer_uptodate(bh))
2961                         goto unlock;
2962         }
2963
2964         zero_user(page, offset, length);
2965         mark_buffer_dirty(bh);
2966         err = 0;
2967
2968 unlock:
2969         unlock_page(page);
2970         put_page(page);
2971 out:
2972         return err;
2973 }
2974 EXPORT_SYMBOL(block_truncate_page);
2975
2976 /*
2977  * The generic ->writepage function for buffer-backed address_spaces
2978  */
2979 int block_write_full_page(struct page *page, get_block_t *get_block,
2980                         struct writeback_control *wbc)
2981 {
2982         struct inode * const inode = page->mapping->host;
2983         loff_t i_size = i_size_read(inode);
2984         const pgoff_t end_index = i_size >> PAGE_SHIFT;
2985         unsigned offset;
2986
2987         /* Is the page fully inside i_size? */
2988         if (page->index < end_index)
2989                 return __block_write_full_page(inode, page, get_block, wbc,
2990                                                end_buffer_async_write);
2991
2992         /* Is the page fully outside i_size? (truncate in progress) */
2993         offset = i_size & (PAGE_SIZE-1);
2994         if (page->index >= end_index+1 || !offset) {
2995                 /*
2996                  * The page may have dirty, unmapped buffers.  For example,
2997                  * they may have been added in ext3_writepage().  Make them
2998                  * freeable here, so the page does not leak.
2999                  */
3000                 do_invalidatepage(page, 0, PAGE_SIZE);
3001                 unlock_page(page);
3002                 return 0; /* don't care */
3003         }
3004
3005         /*
3006          * The page straddles i_size.  It must be zeroed out on each and every
3007          * writepage invocation because it may be mmapped.  "A file is mapped
3008          * in multiples of the page size.  For a file that is not a multiple of
3009          * the  page size, the remaining memory is zeroed when mapped, and
3010          * writes to that region are not written out to the file."
3011          */
3012         zero_user_segment(page, offset, PAGE_SIZE);
3013         return __block_write_full_page(inode, page, get_block, wbc,
3014                                                         end_buffer_async_write);
3015 }
3016 EXPORT_SYMBOL(block_write_full_page);
3017
3018 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
3019                             get_block_t *get_block)
3020 {
3021         struct inode *inode = mapping->host;
3022         struct buffer_head tmp = {
3023                 .b_size = i_blocksize(inode),
3024         };
3025
3026         get_block(inode, block, &tmp, 0);
3027         return tmp.b_blocknr;
3028 }
3029 EXPORT_SYMBOL(generic_block_bmap);
3030
3031 static void end_bio_bh_io_sync(struct bio *bio)
3032 {
3033         struct buffer_head *bh = bio->bi_private;
3034
3035         if (unlikely(bio_flagged(bio, BIO_QUIET)))
3036                 set_bit(BH_Quiet, &bh->b_state);
3037
3038         bh->b_end_io(bh, !bio->bi_status);
3039         bio_put(bio);
3040 }
3041
3042 /*
3043  * This allows us to do IO even on the odd last sectors
3044  * of a device, even if the block size is some multiple
3045  * of the physical sector size.
3046  *
3047  * We'll just truncate the bio to the size of the device,
3048  * and clear the end of the buffer head manually.
3049  *
3050  * Truly out-of-range accesses will turn into actual IO
3051  * errors, this only handles the "we need to be able to
3052  * do IO at the final sector" case.
3053  */
3054 void guard_bio_eod(int op, struct bio *bio)
3055 {
3056         sector_t maxsector;
3057         struct bio_vec *bvec = &bio->bi_io_vec[bio->bi_vcnt - 1];
3058         unsigned truncated_bytes;
3059
3060         maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
3061         if (!maxsector)
3062                 return;
3063
3064         /*
3065          * If the *whole* IO is past the end of the device,
3066          * let it through, and the IO layer will turn it into
3067          * an EIO.
3068          */
3069         if (unlikely(bio->bi_iter.bi_sector >= maxsector))
3070                 return;
3071
3072         maxsector -= bio->bi_iter.bi_sector;
3073         if (likely((bio->bi_iter.bi_size >> 9) <= maxsector))
3074                 return;
3075
3076         /* Uhhuh. We've got a bio that straddles the device size! */
3077         truncated_bytes = bio->bi_iter.bi_size - (maxsector << 9);
3078
3079         /* Truncate the bio.. */
3080         bio->bi_iter.bi_size -= truncated_bytes;
3081         bvec->bv_len -= truncated_bytes;
3082
3083         /* ..and clear the end of the buffer for reads */
3084         if (op == REQ_OP_READ) {
3085                 zero_user(bvec->bv_page, bvec->bv_offset + bvec->bv_len,
3086                                 truncated_bytes);
3087         }
3088 }
3089
3090 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3091                          enum rw_hint write_hint, struct writeback_control *wbc)
3092 {
3093         struct bio *bio;
3094
3095         BUG_ON(!buffer_locked(bh));
3096         BUG_ON(!buffer_mapped(bh));
3097         BUG_ON(!bh->b_end_io);
3098         BUG_ON(buffer_delay(bh));
3099         BUG_ON(buffer_unwritten(bh));
3100
3101         /*
3102          * Only clear out a write error when rewriting
3103          */
3104         if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3105                 clear_buffer_write_io_error(bh);
3106
3107         /*
3108          * from here on down, it's all bio -- do the initial mapping,
3109          * submit_bio -> generic_make_request may further map this bio around
3110          */
3111         bio = bio_alloc(GFP_NOIO, 1);
3112
3113         if (wbc) {
3114                 wbc_init_bio(wbc, bio);
3115                 wbc_account_io(wbc, bh->b_page, bh->b_size);
3116         }
3117
3118         bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3119         bio->bi_bdev = bh->b_bdev;
3120         bio->bi_write_hint = write_hint;
3121
3122         bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3123         BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3124
3125         bio->bi_end_io = end_bio_bh_io_sync;
3126         bio->bi_private = bh;
3127
3128         /* Take care of bh's that straddle the end of the device */
3129         guard_bio_eod(op, bio);
3130
3131         if (buffer_meta(bh))
3132                 op_flags |= REQ_META;
3133         if (buffer_prio(bh))
3134                 op_flags |= REQ_PRIO;
3135         bio_set_op_attrs(bio, op, op_flags);
3136
3137         submit_bio(bio);
3138         return 0;
3139 }
3140
3141 int submit_bh(int op, int op_flags, struct buffer_head *bh)
3142 {
3143         return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3144 }
3145 EXPORT_SYMBOL(submit_bh);
3146
3147 /**
3148  * ll_rw_block: low-level access to block devices (DEPRECATED)
3149  * @op: whether to %READ or %WRITE
3150  * @op_flags: req_flag_bits
3151  * @nr: number of &struct buffer_heads in the array
3152  * @bhs: array of pointers to &struct buffer_head
3153  *
3154  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3155  * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3156  * @op_flags contains flags modifying the detailed I/O behavior, most notably
3157  * %REQ_RAHEAD.
3158  *
3159  * This function drops any buffer that it cannot get a lock on (with the
3160  * BH_Lock state bit), any buffer that appears to be clean when doing a write
3161  * request, and any buffer that appears to be up-to-date when doing read
3162  * request.  Further it marks as clean buffers that are processed for
3163  * writing (the buffer cache won't assume that they are actually clean
3164  * until the buffer gets unlocked).
3165  *
3166  * ll_rw_block sets b_end_io to simple completion handler that marks
3167  * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3168  * any waiters. 
3169  *
3170  * All of the buffers must be for the same device, and must also be a
3171  * multiple of the current approved size for the device.
3172  */
3173 void ll_rw_block(int op, int op_flags,  int nr, struct buffer_head *bhs[])
3174 {
3175         int i;
3176
3177         for (i = 0; i < nr; i++) {
3178                 struct buffer_head *bh = bhs[i];
3179
3180                 if (!trylock_buffer(bh))
3181                         continue;
3182                 if (op == WRITE) {
3183                         if (test_clear_buffer_dirty(bh)) {
3184                                 bh->b_end_io = end_buffer_write_sync;
3185                                 get_bh(bh);
3186                                 submit_bh(op, op_flags, bh);
3187                                 continue;
3188                         }
3189                 } else {
3190                         if (!buffer_uptodate(bh)) {
3191                                 bh->b_end_io = end_buffer_read_sync;
3192                                 get_bh(bh);
3193                                 submit_bh(op, op_flags, bh);
3194                                 continue;
3195                         }
3196                 }
3197                 unlock_buffer(bh);
3198         }
3199 }
3200 EXPORT_SYMBOL(ll_rw_block);
3201
3202 void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3203 {
3204         lock_buffer(bh);
3205         if (!test_clear_buffer_dirty(bh)) {
3206                 unlock_buffer(bh);
3207                 return;
3208         }
3209         bh->b_end_io = end_buffer_write_sync;
3210         get_bh(bh);
3211         submit_bh(REQ_OP_WRITE, op_flags, bh);
3212 }
3213 EXPORT_SYMBOL(write_dirty_buffer);
3214
3215 /*
3216  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3217  * and then start new I/O and then wait upon it.  The caller must have a ref on
3218  * the buffer_head.
3219  */
3220 int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3221 {
3222         int ret = 0;
3223
3224         WARN_ON(atomic_read(&bh->b_count) < 1);
3225         lock_buffer(bh);
3226         if (test_clear_buffer_dirty(bh)) {
3227                 get_bh(bh);
3228                 bh->b_end_io = end_buffer_write_sync;
3229                 ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3230                 wait_on_buffer(bh);
3231                 if (!ret && !buffer_uptodate(bh))
3232                         ret = -EIO;
3233         } else {
3234                 unlock_buffer(bh);
3235         }
3236         return ret;
3237 }
3238 EXPORT_SYMBOL(__sync_dirty_buffer);
3239
3240 int sync_dirty_buffer(struct buffer_head *bh)
3241 {
3242         return __sync_dirty_buffer(bh, REQ_SYNC);
3243 }
3244 EXPORT_SYMBOL(sync_dirty_buffer);
3245
3246 /*
3247  * try_to_free_buffers() checks if all the buffers on this particular page
3248  * are unused, and releases them if so.
3249  *
3250  * Exclusion against try_to_free_buffers may be obtained by either
3251  * locking the page or by holding its mapping's private_lock.
3252  *
3253  * If the page is dirty but all the buffers are clean then we need to
3254  * be sure to mark the page clean as well.  This is because the page
3255  * may be against a block device, and a later reattachment of buffers
3256  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3257  * filesystem data on the same device.
3258  *
3259  * The same applies to regular filesystem pages: if all the buffers are
3260  * clean then we set the page clean and proceed.  To do that, we require
3261  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3262  * private_lock.
3263  *
3264  * try_to_free_buffers() is non-blocking.
3265  */
3266 static inline int buffer_busy(struct buffer_head *bh)
3267 {
3268         return atomic_read(&bh->b_count) |
3269                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3270 }
3271
3272 static int
3273 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3274 {
3275         struct buffer_head *head = page_buffers(page);
3276         struct buffer_head *bh;
3277
3278         bh = head;
3279         do {
3280                 if (buffer_busy(bh))
3281                         goto failed;
3282                 bh = bh->b_this_page;
3283         } while (bh != head);
3284
3285         do {
3286                 struct buffer_head *next = bh->b_this_page;
3287
3288                 if (bh->b_assoc_map)
3289                         __remove_assoc_queue(bh);
3290                 bh = next;
3291         } while (bh != head);
3292         *buffers_to_free = head;
3293         __clear_page_buffers(page);
3294         return 1;
3295 failed:
3296         return 0;
3297 }
3298
3299 int try_to_free_buffers(struct page *page)
3300 {
3301         struct address_space * const mapping = page->mapping;
3302         struct buffer_head *buffers_to_free = NULL;
3303         int ret = 0;
3304
3305         BUG_ON(!PageLocked(page));
3306         if (PageWriteback(page))
3307                 return 0;
3308
3309         if (mapping == NULL) {          /* can this still happen? */
3310                 ret = drop_buffers(page, &buffers_to_free);
3311                 goto out;
3312         }
3313
3314         spin_lock(&mapping->private_lock);
3315         ret = drop_buffers(page, &buffers_to_free);
3316
3317         /*
3318          * If the filesystem writes its buffers by hand (eg ext3)
3319          * then we can have clean buffers against a dirty page.  We
3320          * clean the page here; otherwise the VM will never notice
3321          * that the filesystem did any IO at all.
3322          *
3323          * Also, during truncate, discard_buffer will have marked all
3324          * the page's buffers clean.  We discover that here and clean
3325          * the page also.
3326          *
3327          * private_lock must be held over this entire operation in order
3328          * to synchronise against __set_page_dirty_buffers and prevent the
3329          * dirty bit from being lost.
3330          */
3331         if (ret)
3332                 cancel_dirty_page(page);
3333         spin_unlock(&mapping->private_lock);
3334 out:
3335         if (buffers_to_free) {
3336                 struct buffer_head *bh = buffers_to_free;
3337
3338                 do {
3339                         struct buffer_head *next = bh->b_this_page;
3340                         free_buffer_head(bh);
3341                         bh = next;
3342                 } while (bh != buffers_to_free);
3343         }
3344         return ret;
3345 }
3346 EXPORT_SYMBOL(try_to_free_buffers);
3347
3348 /*
3349  * There are no bdflush tunables left.  But distributions are
3350  * still running obsolete flush daemons, so we terminate them here.
3351  *
3352  * Use of bdflush() is deprecated and will be removed in a future kernel.
3353  * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3354  */
3355 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3356 {
3357         static int msg_count;
3358
3359         if (!capable(CAP_SYS_ADMIN))
3360                 return -EPERM;
3361
3362         if (msg_count < 5) {
3363                 msg_count++;
3364                 printk(KERN_INFO
3365                         "warning: process `%s' used the obsolete bdflush"
3366                         " system call\n", current->comm);
3367                 printk(KERN_INFO "Fix your initscripts?\n");
3368         }
3369
3370         if (func == 1)
3371                 do_exit(0);
3372         return 0;
3373 }
3374
3375 /*
3376  * Buffer-head allocation
3377  */
3378 static struct kmem_cache *bh_cachep __read_mostly;
3379
3380 /*
3381  * Once the number of bh's in the machine exceeds this level, we start
3382  * stripping them in writeback.
3383  */
3384 static unsigned long max_buffer_heads;
3385
3386 int buffer_heads_over_limit;
3387
3388 struct bh_accounting {
3389         int nr;                 /* Number of live bh's */
3390         int ratelimit;          /* Limit cacheline bouncing */
3391 };
3392
3393 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3394
3395 static void recalc_bh_state(void)
3396 {
3397         int i;
3398         int tot = 0;
3399
3400         if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3401                 return;
3402         __this_cpu_write(bh_accounting.ratelimit, 0);
3403         for_each_online_cpu(i)
3404                 tot += per_cpu(bh_accounting, i).nr;
3405         buffer_heads_over_limit = (tot > max_buffer_heads);
3406 }
3407
3408 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3409 {
3410         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3411         if (ret) {
3412                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3413                 preempt_disable();
3414                 __this_cpu_inc(bh_accounting.nr);
3415                 recalc_bh_state();
3416                 preempt_enable();
3417         }
3418         return ret;
3419 }
3420 EXPORT_SYMBOL(alloc_buffer_head);
3421
3422 void free_buffer_head(struct buffer_head *bh)
3423 {
3424         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3425         kmem_cache_free(bh_cachep, bh);
3426         preempt_disable();
3427         __this_cpu_dec(bh_accounting.nr);
3428         recalc_bh_state();
3429         preempt_enable();
3430 }
3431 EXPORT_SYMBOL(free_buffer_head);
3432
3433 static int buffer_exit_cpu_dead(unsigned int cpu)
3434 {
3435         int i;
3436         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3437
3438         for (i = 0; i < BH_LRU_SIZE; i++) {
3439                 brelse(b->bhs[i]);
3440                 b->bhs[i] = NULL;
3441         }
3442         this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3443         per_cpu(bh_accounting, cpu).nr = 0;
3444         return 0;
3445 }
3446
3447 /**
3448  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3449  * @bh: struct buffer_head
3450  *
3451  * Return true if the buffer is up-to-date and false,
3452  * with the buffer locked, if not.
3453  */
3454 int bh_uptodate_or_lock(struct buffer_head *bh)
3455 {
3456         if (!buffer_uptodate(bh)) {
3457                 lock_buffer(bh);
3458                 if (!buffer_uptodate(bh))
3459                         return 0;
3460                 unlock_buffer(bh);
3461         }
3462         return 1;
3463 }
3464 EXPORT_SYMBOL(bh_uptodate_or_lock);
3465
3466 /**
3467  * bh_submit_read - Submit a locked buffer for reading
3468  * @bh: struct buffer_head
3469  *
3470  * Returns zero on success and -EIO on error.
3471  */
3472 int bh_submit_read(struct buffer_head *bh)
3473 {
3474         BUG_ON(!buffer_locked(bh));
3475
3476         if (buffer_uptodate(bh)) {
3477                 unlock_buffer(bh);
3478                 return 0;
3479         }
3480
3481         get_bh(bh);
3482         bh->b_end_io = end_buffer_read_sync;
3483         submit_bh(REQ_OP_READ, 0, bh);
3484         wait_on_buffer(bh);
3485         if (buffer_uptodate(bh))
3486                 return 0;
3487         return -EIO;
3488 }
3489 EXPORT_SYMBOL(bh_submit_read);
3490
3491 /*
3492  * Seek for SEEK_DATA / SEEK_HOLE within @page, starting at @lastoff.
3493  *
3494  * Returns the offset within the file on success, and -ENOENT otherwise.
3495  */
3496 static loff_t
3497 page_seek_hole_data(struct page *page, loff_t lastoff, int whence)
3498 {
3499         loff_t offset = page_offset(page);
3500         struct buffer_head *bh, *head;
3501         bool seek_data = whence == SEEK_DATA;
3502
3503         if (lastoff < offset)
3504                 lastoff = offset;
3505
3506         bh = head = page_buffers(page);
3507         do {
3508                 offset += bh->b_size;
3509                 if (lastoff >= offset)
3510                         continue;
3511
3512                 /*
3513                  * Unwritten extents that have data in the page cache covering
3514                  * them can be identified by the BH_Unwritten state flag.
3515                  * Pages with multiple buffers might have a mix of holes, data
3516                  * and unwritten extents - any buffer with valid data in it
3517                  * should have BH_Uptodate flag set on it.
3518                  */
3519
3520                 if ((buffer_unwritten(bh) || buffer_uptodate(bh)) == seek_data)
3521                         return lastoff;
3522
3523                 lastoff = offset;
3524         } while ((bh = bh->b_this_page) != head);
3525         return -ENOENT;
3526 }
3527
3528 /*
3529  * Seek for SEEK_DATA / SEEK_HOLE in the page cache.
3530  *
3531  * Within unwritten extents, the page cache determines which parts are holes
3532  * and which are data: unwritten and uptodate buffer heads count as data;
3533  * everything else counts as a hole.
3534  *
3535  * Returns the resulting offset on successs, and -ENOENT otherwise.
3536  */
3537 loff_t
3538 page_cache_seek_hole_data(struct inode *inode, loff_t offset, loff_t length,
3539                           int whence)
3540 {
3541         pgoff_t index = offset >> PAGE_SHIFT;
3542         pgoff_t end = DIV_ROUND_UP(offset + length, PAGE_SIZE);
3543         loff_t lastoff = offset;
3544         struct pagevec pvec;
3545
3546         if (length <= 0)
3547                 return -ENOENT;
3548
3549         pagevec_init(&pvec, 0);
3550
3551         do {
3552                 unsigned want, nr_pages, i;
3553
3554                 want = min_t(unsigned, end - index, PAGEVEC_SIZE);
3555                 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, index, want);
3556                 if (nr_pages == 0)
3557                         break;
3558
3559                 for (i = 0; i < nr_pages; i++) {
3560                         struct page *page = pvec.pages[i];
3561
3562                         /*
3563                          * At this point, the page may be truncated or
3564                          * invalidated (changing page->mapping to NULL), or
3565                          * even swizzled back from swapper_space to tmpfs file
3566                          * mapping.  However, page->index will not change
3567                          * because we have a reference on the page.
3568                          *
3569                          * If current page offset is beyond where we've ended,
3570                          * we've found a hole.
3571                          */
3572                         if (whence == SEEK_HOLE &&
3573                             lastoff < page_offset(page))
3574                                 goto check_range;
3575
3576                         /* Searching done if the page index is out of range. */
3577                         if (page->index >= end)
3578                                 goto not_found;
3579
3580                         lock_page(page);
3581                         if (likely(page->mapping == inode->i_mapping) &&
3582                             page_has_buffers(page)) {
3583                                 lastoff = page_seek_hole_data(page, lastoff, whence);
3584                                 if (lastoff >= 0) {
3585                                         unlock_page(page);
3586                                         goto check_range;
3587                                 }
3588                         }
3589                         unlock_page(page);
3590                         lastoff = page_offset(page) + PAGE_SIZE;
3591                 }
3592
3593                 /* Searching done if fewer pages returned than wanted. */
3594                 if (nr_pages < want)
3595                         break;
3596
3597                 index = pvec.pages[i - 1]->index + 1;
3598                 pagevec_release(&pvec);
3599         } while (index < end);
3600
3601         /* When no page at lastoff and we are not done, we found a hole. */
3602         if (whence != SEEK_HOLE)
3603                 goto not_found;
3604
3605 check_range:
3606         if (lastoff < offset + length)
3607                 goto out;
3608 not_found:
3609         lastoff = -ENOENT;
3610 out:
3611         pagevec_release(&pvec);
3612         return lastoff;
3613 }
3614
3615 void __init buffer_init(void)
3616 {
3617         unsigned long nrpages;
3618         int ret;
3619
3620         bh_cachep = kmem_cache_create("buffer_head",
3621                         sizeof(struct buffer_head), 0,
3622                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3623                                 SLAB_MEM_SPREAD),
3624                                 NULL);
3625
3626         /*
3627          * Limit the bh occupancy to 10% of ZONE_NORMAL
3628          */
3629         nrpages = (nr_free_buffer_pages() * 10) / 100;
3630         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3631         ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3632                                         NULL, buffer_exit_cpu_dead);
3633         WARN_ON(ret < 0);
3634 }