Merge branch 'pruss-fix' into fixes
[platform/kernel/linux-starfive.git] / fs / buffer.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  linux/fs/buffer.c
4  *
5  *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
6  */
7
8 /*
9  * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
10  *
11  * Removed a lot of unnecessary code and simplified things now that
12  * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
13  *
14  * Speed up hash, lru, and free list operations.  Use gfp() for allocating
15  * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
16  *
17  * Added 32k buffer block sizes - these are required older ARM systems. - RMK
18  *
19  * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
20  */
21
22 #include <linux/kernel.h>
23 #include <linux/sched/signal.h>
24 #include <linux/syscalls.h>
25 #include <linux/fs.h>
26 #include <linux/iomap.h>
27 #include <linux/mm.h>
28 #include <linux/percpu.h>
29 #include <linux/slab.h>
30 #include <linux/capability.h>
31 #include <linux/blkdev.h>
32 #include <linux/file.h>
33 #include <linux/quotaops.h>
34 #include <linux/highmem.h>
35 #include <linux/export.h>
36 #include <linux/backing-dev.h>
37 #include <linux/writeback.h>
38 #include <linux/hash.h>
39 #include <linux/suspend.h>
40 #include <linux/buffer_head.h>
41 #include <linux/task_io_accounting_ops.h>
42 #include <linux/bio.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 <linux/sched/mm.h>
49 #include <trace/events/block.h>
50 #include <linux/fscrypt.h>
51
52 #include "internal.h"
53
54 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
55 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
56                          enum rw_hint hint, struct writeback_control *wbc);
57
58 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
59
60 inline void touch_buffer(struct buffer_head *bh)
61 {
62         trace_block_touch_buffer(bh);
63         mark_page_accessed(bh->b_page);
64 }
65 EXPORT_SYMBOL(touch_buffer);
66
67 void __lock_buffer(struct buffer_head *bh)
68 {
69         wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
70 }
71 EXPORT_SYMBOL(__lock_buffer);
72
73 void unlock_buffer(struct buffer_head *bh)
74 {
75         clear_bit_unlock(BH_Lock, &bh->b_state);
76         smp_mb__after_atomic();
77         wake_up_bit(&bh->b_state, BH_Lock);
78 }
79 EXPORT_SYMBOL(unlock_buffer);
80
81 /*
82  * Returns if the page has dirty or writeback buffers. If all the buffers
83  * are unlocked and clean then the PageDirty information is stale. If
84  * any of the pages are locked, it is assumed they are locked for IO.
85  */
86 void buffer_check_dirty_writeback(struct page *page,
87                                      bool *dirty, bool *writeback)
88 {
89         struct buffer_head *head, *bh;
90         *dirty = false;
91         *writeback = false;
92
93         BUG_ON(!PageLocked(page));
94
95         if (!page_has_buffers(page))
96                 return;
97
98         if (PageWriteback(page))
99                 *writeback = true;
100
101         head = page_buffers(page);
102         bh = head;
103         do {
104                 if (buffer_locked(bh))
105                         *writeback = true;
106
107                 if (buffer_dirty(bh))
108                         *dirty = true;
109
110                 bh = bh->b_this_page;
111         } while (bh != head);
112 }
113 EXPORT_SYMBOL(buffer_check_dirty_writeback);
114
115 /*
116  * Block until a buffer comes unlocked.  This doesn't stop it
117  * from becoming locked again - you have to lock it yourself
118  * if you want to preserve its state.
119  */
120 void __wait_on_buffer(struct buffer_head * bh)
121 {
122         wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
123 }
124 EXPORT_SYMBOL(__wait_on_buffer);
125
126 static void buffer_io_error(struct buffer_head *bh, char *msg)
127 {
128         if (!test_bit(BH_Quiet, &bh->b_state))
129                 printk_ratelimited(KERN_ERR
130                         "Buffer I/O error on dev %pg, logical block %llu%s\n",
131                         bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
132 }
133
134 /*
135  * End-of-IO handler helper function which does not touch the bh after
136  * unlocking it.
137  * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
138  * a race there is benign: unlock_buffer() only use the bh's address for
139  * hashing after unlocking the buffer, so it doesn't actually touch the bh
140  * itself.
141  */
142 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
143 {
144         if (uptodate) {
145                 set_buffer_uptodate(bh);
146         } else {
147                 /* This happens, due to failed read-ahead attempts. */
148                 clear_buffer_uptodate(bh);
149         }
150         unlock_buffer(bh);
151 }
152
153 /*
154  * Default synchronous end-of-IO handler..  Just mark it up-to-date and
155  * unlock the buffer. This is what ll_rw_block uses too.
156  */
157 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
158 {
159         __end_buffer_read_notouch(bh, uptodate);
160         put_bh(bh);
161 }
162 EXPORT_SYMBOL(end_buffer_read_sync);
163
164 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
165 {
166         if (uptodate) {
167                 set_buffer_uptodate(bh);
168         } else {
169                 buffer_io_error(bh, ", lost sync page write");
170                 mark_buffer_write_io_error(bh);
171                 clear_buffer_uptodate(bh);
172         }
173         unlock_buffer(bh);
174         put_bh(bh);
175 }
176 EXPORT_SYMBOL(end_buffer_write_sync);
177
178 /*
179  * Various filesystems appear to want __find_get_block to be non-blocking.
180  * But it's the page lock which protects the buffers.  To get around this,
181  * we get exclusion from try_to_free_buffers with the blockdev mapping's
182  * private_lock.
183  *
184  * Hack idea: for the blockdev mapping, private_lock contention
185  * may be quite high.  This code could TryLock the page, and if that
186  * succeeds, there is no need to take private_lock.
187  */
188 static struct buffer_head *
189 __find_get_block_slow(struct block_device *bdev, sector_t block)
190 {
191         struct inode *bd_inode = bdev->bd_inode;
192         struct address_space *bd_mapping = bd_inode->i_mapping;
193         struct buffer_head *ret = NULL;
194         pgoff_t index;
195         struct buffer_head *bh;
196         struct buffer_head *head;
197         struct page *page;
198         int all_mapped = 1;
199         static DEFINE_RATELIMIT_STATE(last_warned, HZ, 1);
200
201         index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
202         page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
203         if (!page)
204                 goto out;
205
206         spin_lock(&bd_mapping->private_lock);
207         if (!page_has_buffers(page))
208                 goto out_unlock;
209         head = page_buffers(page);
210         bh = head;
211         do {
212                 if (!buffer_mapped(bh))
213                         all_mapped = 0;
214                 else if (bh->b_blocknr == block) {
215                         ret = bh;
216                         get_bh(bh);
217                         goto out_unlock;
218                 }
219                 bh = bh->b_this_page;
220         } while (bh != head);
221
222         /* we might be here because some of the buffers on this page are
223          * not mapped.  This is due to various races between
224          * file io on the block device and getblk.  It gets dealt with
225          * elsewhere, don't buffer_error if we had some unmapped buffers
226          */
227         ratelimit_set_flags(&last_warned, RATELIMIT_MSG_ON_RELEASE);
228         if (all_mapped && __ratelimit(&last_warned)) {
229                 printk("__find_get_block_slow() failed. block=%llu, "
230                        "b_blocknr=%llu, b_state=0x%08lx, b_size=%zu, "
231                        "device %pg blocksize: %d\n",
232                        (unsigned long long)block,
233                        (unsigned long long)bh->b_blocknr,
234                        bh->b_state, bh->b_size, bdev,
235                        1 << bd_inode->i_blkbits);
236         }
237 out_unlock:
238         spin_unlock(&bd_mapping->private_lock);
239         put_page(page);
240 out:
241         return ret;
242 }
243
244 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
245 {
246         unsigned long flags;
247         struct buffer_head *first;
248         struct buffer_head *tmp;
249         struct page *page;
250         int page_uptodate = 1;
251
252         BUG_ON(!buffer_async_read(bh));
253
254         page = bh->b_page;
255         if (uptodate) {
256                 set_buffer_uptodate(bh);
257         } else {
258                 clear_buffer_uptodate(bh);
259                 buffer_io_error(bh, ", async page read");
260                 SetPageError(page);
261         }
262
263         /*
264          * Be _very_ careful from here on. Bad things can happen if
265          * two buffer heads end IO at almost the same time and both
266          * decide that the page is now completely done.
267          */
268         first = page_buffers(page);
269         spin_lock_irqsave(&first->b_uptodate_lock, flags);
270         clear_buffer_async_read(bh);
271         unlock_buffer(bh);
272         tmp = bh;
273         do {
274                 if (!buffer_uptodate(tmp))
275                         page_uptodate = 0;
276                 if (buffer_async_read(tmp)) {
277                         BUG_ON(!buffer_locked(tmp));
278                         goto still_busy;
279                 }
280                 tmp = tmp->b_this_page;
281         } while (tmp != bh);
282         spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
283
284         /*
285          * If none of the buffers had errors and they are all
286          * uptodate then we can set the page uptodate.
287          */
288         if (page_uptodate && !PageError(page))
289                 SetPageUptodate(page);
290         unlock_page(page);
291         return;
292
293 still_busy:
294         spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
295         return;
296 }
297
298 struct decrypt_bh_ctx {
299         struct work_struct work;
300         struct buffer_head *bh;
301 };
302
303 static void decrypt_bh(struct work_struct *work)
304 {
305         struct decrypt_bh_ctx *ctx =
306                 container_of(work, struct decrypt_bh_ctx, work);
307         struct buffer_head *bh = ctx->bh;
308         int err;
309
310         err = fscrypt_decrypt_pagecache_blocks(bh->b_page, bh->b_size,
311                                                bh_offset(bh));
312         end_buffer_async_read(bh, err == 0);
313         kfree(ctx);
314 }
315
316 /*
317  * I/O completion handler for block_read_full_page() - pages
318  * which come unlocked at the end of I/O.
319  */
320 static void end_buffer_async_read_io(struct buffer_head *bh, int uptodate)
321 {
322         /* Decrypt if needed */
323         if (uptodate &&
324             fscrypt_inode_uses_fs_layer_crypto(bh->b_page->mapping->host)) {
325                 struct decrypt_bh_ctx *ctx = kmalloc(sizeof(*ctx), GFP_ATOMIC);
326
327                 if (ctx) {
328                         INIT_WORK(&ctx->work, decrypt_bh);
329                         ctx->bh = bh;
330                         fscrypt_enqueue_decrypt_work(&ctx->work);
331                         return;
332                 }
333                 uptodate = 0;
334         }
335         end_buffer_async_read(bh, uptodate);
336 }
337
338 /*
339  * Completion handler for block_write_full_page() - pages which are unlocked
340  * during I/O, and which have PageWriteback cleared upon I/O completion.
341  */
342 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
343 {
344         unsigned long flags;
345         struct buffer_head *first;
346         struct buffer_head *tmp;
347         struct page *page;
348
349         BUG_ON(!buffer_async_write(bh));
350
351         page = bh->b_page;
352         if (uptodate) {
353                 set_buffer_uptodate(bh);
354         } else {
355                 buffer_io_error(bh, ", lost async page write");
356                 mark_buffer_write_io_error(bh);
357                 clear_buffer_uptodate(bh);
358                 SetPageError(page);
359         }
360
361         first = page_buffers(page);
362         spin_lock_irqsave(&first->b_uptodate_lock, flags);
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         spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
375         end_page_writeback(page);
376         return;
377
378 still_busy:
379         spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
380         return;
381 }
382 EXPORT_SYMBOL(end_buffer_async_write);
383
384 /*
385  * If a page's buffers are under async readin (end_buffer_async_read
386  * completion) then there is a possibility that another thread of
387  * control could lock one of the buffers after it has completed
388  * but while some of the other buffers have not completed.  This
389  * locked buffer would confuse end_buffer_async_read() into not unlocking
390  * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
391  * that this buffer is not under async I/O.
392  *
393  * The page comes unlocked when it has no locked buffer_async buffers
394  * left.
395  *
396  * PageLocked prevents anyone starting new async I/O reads any of
397  * the buffers.
398  *
399  * PageWriteback is used to prevent simultaneous writeout of the same
400  * page.
401  *
402  * PageLocked prevents anyone from starting writeback of a page which is
403  * under read I/O (PageWriteback is only ever set against a locked page).
404  */
405 static void mark_buffer_async_read(struct buffer_head *bh)
406 {
407         bh->b_end_io = end_buffer_async_read_io;
408         set_buffer_async_read(bh);
409 }
410
411 static void mark_buffer_async_write_endio(struct buffer_head *bh,
412                                           bh_end_io_t *handler)
413 {
414         bh->b_end_io = handler;
415         set_buffer_async_write(bh);
416 }
417
418 void mark_buffer_async_write(struct buffer_head *bh)
419 {
420         mark_buffer_async_write_endio(bh, end_buffer_async_write);
421 }
422 EXPORT_SYMBOL(mark_buffer_async_write);
423
424
425 /*
426  * fs/buffer.c contains helper functions for buffer-backed address space's
427  * fsync functions.  A common requirement for buffer-based filesystems is
428  * that certain data from the backing blockdev needs to be written out for
429  * a successful fsync().  For example, ext2 indirect blocks need to be
430  * written back and waited upon before fsync() returns.
431  *
432  * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
433  * inode_has_buffers() and invalidate_inode_buffers() are provided for the
434  * management of a list of dependent buffers at ->i_mapping->private_list.
435  *
436  * Locking is a little subtle: try_to_free_buffers() will remove buffers
437  * from their controlling inode's queue when they are being freed.  But
438  * try_to_free_buffers() will be operating against the *blockdev* mapping
439  * at the time, not against the S_ISREG file which depends on those buffers.
440  * So the locking for private_list is via the private_lock in the address_space
441  * which backs the buffers.  Which is different from the address_space 
442  * against which the buffers are listed.  So for a particular address_space,
443  * mapping->private_lock does *not* protect mapping->private_list!  In fact,
444  * mapping->private_list will always be protected by the backing blockdev's
445  * ->private_lock.
446  *
447  * Which introduces a requirement: all buffers on an address_space's
448  * ->private_list must be from the same address_space: the blockdev's.
449  *
450  * address_spaces which do not place buffers at ->private_list via these
451  * utility functions are free to use private_lock and private_list for
452  * whatever they want.  The only requirement is that list_empty(private_list)
453  * be true at clear_inode() time.
454  *
455  * FIXME: clear_inode should not call invalidate_inode_buffers().  The
456  * filesystems should do that.  invalidate_inode_buffers() should just go
457  * BUG_ON(!list_empty).
458  *
459  * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
460  * take an address_space, not an inode.  And it should be called
461  * mark_buffer_dirty_fsync() to clearly define why those buffers are being
462  * queued up.
463  *
464  * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
465  * list if it is already on a list.  Because if the buffer is on a list,
466  * it *must* already be on the right one.  If not, the filesystem is being
467  * silly.  This will save a ton of locking.  But first we have to ensure
468  * that buffers are taken *off* the old inode's list when they are freed
469  * (presumably in truncate).  That requires careful auditing of all
470  * filesystems (do it inside bforget()).  It could also be done by bringing
471  * b_inode back.
472  */
473
474 /*
475  * The buffer's backing address_space's private_lock must be held
476  */
477 static void __remove_assoc_queue(struct buffer_head *bh)
478 {
479         list_del_init(&bh->b_assoc_buffers);
480         WARN_ON(!bh->b_assoc_map);
481         bh->b_assoc_map = NULL;
482 }
483
484 int inode_has_buffers(struct inode *inode)
485 {
486         return !list_empty(&inode->i_data.private_list);
487 }
488
489 /*
490  * osync is designed to support O_SYNC io.  It waits synchronously for
491  * all already-submitted IO to complete, but does not queue any new
492  * writes to the disk.
493  *
494  * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
495  * you dirty the buffers, and then use osync_inode_buffers to wait for
496  * completion.  Any other dirty buffers which are not yet queued for
497  * write will not be flushed to disk by the osync.
498  */
499 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
500 {
501         struct buffer_head *bh;
502         struct list_head *p;
503         int err = 0;
504
505         spin_lock(lock);
506 repeat:
507         list_for_each_prev(p, list) {
508                 bh = BH_ENTRY(p);
509                 if (buffer_locked(bh)) {
510                         get_bh(bh);
511                         spin_unlock(lock);
512                         wait_on_buffer(bh);
513                         if (!buffer_uptodate(bh))
514                                 err = -EIO;
515                         brelse(bh);
516                         spin_lock(lock);
517                         goto repeat;
518                 }
519         }
520         spin_unlock(lock);
521         return err;
522 }
523
524 void emergency_thaw_bdev(struct super_block *sb)
525 {
526         while (sb->s_bdev && !thaw_bdev(sb->s_bdev))
527                 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
528 }
529
530 /**
531  * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
532  * @mapping: the mapping which wants those buffers written
533  *
534  * Starts I/O against the buffers at mapping->private_list, and waits upon
535  * that I/O.
536  *
537  * Basically, this is a convenience function for fsync().
538  * @mapping is a file or directory which needs those buffers to be written for
539  * a successful fsync().
540  */
541 int sync_mapping_buffers(struct address_space *mapping)
542 {
543         struct address_space *buffer_mapping = mapping->private_data;
544
545         if (buffer_mapping == NULL || list_empty(&mapping->private_list))
546                 return 0;
547
548         return fsync_buffers_list(&buffer_mapping->private_lock,
549                                         &mapping->private_list);
550 }
551 EXPORT_SYMBOL(sync_mapping_buffers);
552
553 /*
554  * Called when we've recently written block `bblock', and it is known that
555  * `bblock' was for a buffer_boundary() buffer.  This means that the block at
556  * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
557  * dirty, schedule it for IO.  So that indirects merge nicely with their data.
558  */
559 void write_boundary_block(struct block_device *bdev,
560                         sector_t bblock, unsigned blocksize)
561 {
562         struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
563         if (bh) {
564                 if (buffer_dirty(bh))
565                         ll_rw_block(REQ_OP_WRITE, 0, 1, &bh);
566                 put_bh(bh);
567         }
568 }
569
570 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
571 {
572         struct address_space *mapping = inode->i_mapping;
573         struct address_space *buffer_mapping = bh->b_page->mapping;
574
575         mark_buffer_dirty(bh);
576         if (!mapping->private_data) {
577                 mapping->private_data = buffer_mapping;
578         } else {
579                 BUG_ON(mapping->private_data != buffer_mapping);
580         }
581         if (!bh->b_assoc_map) {
582                 spin_lock(&buffer_mapping->private_lock);
583                 list_move_tail(&bh->b_assoc_buffers,
584                                 &mapping->private_list);
585                 bh->b_assoc_map = mapping;
586                 spin_unlock(&buffer_mapping->private_lock);
587         }
588 }
589 EXPORT_SYMBOL(mark_buffer_dirty_inode);
590
591 /*
592  * Add a page to the dirty page list.
593  *
594  * It is a sad fact of life that this function is called from several places
595  * deeply under spinlocking.  It may not sleep.
596  *
597  * If the page has buffers, the uptodate buffers are set dirty, to preserve
598  * dirty-state coherency between the page and the buffers.  It the page does
599  * not have buffers then when they are later attached they will all be set
600  * dirty.
601  *
602  * The buffers are dirtied before the page is dirtied.  There's a small race
603  * window in which a writepage caller may see the page cleanness but not the
604  * buffer dirtiness.  That's fine.  If this code were to set the page dirty
605  * before the buffers, a concurrent writepage caller could clear the page dirty
606  * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
607  * page on the dirty page list.
608  *
609  * We use private_lock to lock against try_to_free_buffers while using the
610  * page's buffer list.  Also use this to protect against clean buffers being
611  * added to the page after it was set dirty.
612  *
613  * FIXME: may need to call ->reservepage here as well.  That's rather up to the
614  * address_space though.
615  */
616 int __set_page_dirty_buffers(struct page *page)
617 {
618         int newly_dirty;
619         struct address_space *mapping = page_mapping(page);
620
621         if (unlikely(!mapping))
622                 return !TestSetPageDirty(page);
623
624         spin_lock(&mapping->private_lock);
625         if (page_has_buffers(page)) {
626                 struct buffer_head *head = page_buffers(page);
627                 struct buffer_head *bh = head;
628
629                 do {
630                         set_buffer_dirty(bh);
631                         bh = bh->b_this_page;
632                 } while (bh != head);
633         }
634         /*
635          * Lock out page's memcg migration to keep PageDirty
636          * synchronized with per-memcg dirty page counters.
637          */
638         lock_page_memcg(page);
639         newly_dirty = !TestSetPageDirty(page);
640         spin_unlock(&mapping->private_lock);
641
642         if (newly_dirty)
643                 __set_page_dirty(page, mapping, 1);
644
645         unlock_page_memcg(page);
646
647         if (newly_dirty)
648                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
649
650         return newly_dirty;
651 }
652 EXPORT_SYMBOL(__set_page_dirty_buffers);
653
654 /*
655  * Write out and wait upon a list of buffers.
656  *
657  * We have conflicting pressures: we want to make sure that all
658  * initially dirty buffers get waited on, but that any subsequently
659  * dirtied buffers don't.  After all, we don't want fsync to last
660  * forever if somebody is actively writing to the file.
661  *
662  * Do this in two main stages: first we copy dirty buffers to a
663  * temporary inode list, queueing the writes as we go.  Then we clean
664  * up, waiting for those writes to complete.
665  * 
666  * During this second stage, any subsequent updates to the file may end
667  * up refiling the buffer on the original inode's dirty list again, so
668  * there is a chance we will end up with a buffer queued for write but
669  * not yet completed on that list.  So, as a final cleanup we go through
670  * the osync code to catch these locked, dirty buffers without requeuing
671  * any newly dirty buffers for write.
672  */
673 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
674 {
675         struct buffer_head *bh;
676         struct list_head tmp;
677         struct address_space *mapping;
678         int err = 0, err2;
679         struct blk_plug plug;
680
681         INIT_LIST_HEAD(&tmp);
682         blk_start_plug(&plug);
683
684         spin_lock(lock);
685         while (!list_empty(list)) {
686                 bh = BH_ENTRY(list->next);
687                 mapping = bh->b_assoc_map;
688                 __remove_assoc_queue(bh);
689                 /* Avoid race with mark_buffer_dirty_inode() which does
690                  * a lockless check and we rely on seeing the dirty bit */
691                 smp_mb();
692                 if (buffer_dirty(bh) || buffer_locked(bh)) {
693                         list_add(&bh->b_assoc_buffers, &tmp);
694                         bh->b_assoc_map = mapping;
695                         if (buffer_dirty(bh)) {
696                                 get_bh(bh);
697                                 spin_unlock(lock);
698                                 /*
699                                  * Ensure any pending I/O completes so that
700                                  * write_dirty_buffer() actually writes the
701                                  * current contents - it is a noop if I/O is
702                                  * still in flight on potentially older
703                                  * contents.
704                                  */
705                                 write_dirty_buffer(bh, REQ_SYNC);
706
707                                 /*
708                                  * Kick off IO for the previous mapping. Note
709                                  * that we will not run the very last mapping,
710                                  * wait_on_buffer() will do that for us
711                                  * through sync_buffer().
712                                  */
713                                 brelse(bh);
714                                 spin_lock(lock);
715                         }
716                 }
717         }
718
719         spin_unlock(lock);
720         blk_finish_plug(&plug);
721         spin_lock(lock);
722
723         while (!list_empty(&tmp)) {
724                 bh = BH_ENTRY(tmp.prev);
725                 get_bh(bh);
726                 mapping = bh->b_assoc_map;
727                 __remove_assoc_queue(bh);
728                 /* Avoid race with mark_buffer_dirty_inode() which does
729                  * a lockless check and we rely on seeing the dirty bit */
730                 smp_mb();
731                 if (buffer_dirty(bh)) {
732                         list_add(&bh->b_assoc_buffers,
733                                  &mapping->private_list);
734                         bh->b_assoc_map = mapping;
735                 }
736                 spin_unlock(lock);
737                 wait_on_buffer(bh);
738                 if (!buffer_uptodate(bh))
739                         err = -EIO;
740                 brelse(bh);
741                 spin_lock(lock);
742         }
743         
744         spin_unlock(lock);
745         err2 = osync_buffers_list(lock, list);
746         if (err)
747                 return err;
748         else
749                 return err2;
750 }
751
752 /*
753  * Invalidate any and all dirty buffers on a given inode.  We are
754  * probably unmounting the fs, but that doesn't mean we have already
755  * done a sync().  Just drop the buffers from the inode list.
756  *
757  * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
758  * assumes that all the buffers are against the blockdev.  Not true
759  * for reiserfs.
760  */
761 void invalidate_inode_buffers(struct inode *inode)
762 {
763         if (inode_has_buffers(inode)) {
764                 struct address_space *mapping = &inode->i_data;
765                 struct list_head *list = &mapping->private_list;
766                 struct address_space *buffer_mapping = mapping->private_data;
767
768                 spin_lock(&buffer_mapping->private_lock);
769                 while (!list_empty(list))
770                         __remove_assoc_queue(BH_ENTRY(list->next));
771                 spin_unlock(&buffer_mapping->private_lock);
772         }
773 }
774 EXPORT_SYMBOL(invalidate_inode_buffers);
775
776 /*
777  * Remove any clean buffers from the inode's buffer list.  This is called
778  * when we're trying to free the inode itself.  Those buffers can pin it.
779  *
780  * Returns true if all buffers were removed.
781  */
782 int remove_inode_buffers(struct inode *inode)
783 {
784         int ret = 1;
785
786         if (inode_has_buffers(inode)) {
787                 struct address_space *mapping = &inode->i_data;
788                 struct list_head *list = &mapping->private_list;
789                 struct address_space *buffer_mapping = mapping->private_data;
790
791                 spin_lock(&buffer_mapping->private_lock);
792                 while (!list_empty(list)) {
793                         struct buffer_head *bh = BH_ENTRY(list->next);
794                         if (buffer_dirty(bh)) {
795                                 ret = 0;
796                                 break;
797                         }
798                         __remove_assoc_queue(bh);
799                 }
800                 spin_unlock(&buffer_mapping->private_lock);
801         }
802         return ret;
803 }
804
805 /*
806  * Create the appropriate buffers when given a page for data area and
807  * the size of each buffer.. Use the bh->b_this_page linked list to
808  * follow the buffers created.  Return NULL if unable to create more
809  * buffers.
810  *
811  * The retry flag is used to differentiate async IO (paging, swapping)
812  * which may not fail from ordinary buffer allocations.
813  */
814 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
815                 bool retry)
816 {
817         struct buffer_head *bh, *head;
818         gfp_t gfp = GFP_NOFS | __GFP_ACCOUNT;
819         long offset;
820         struct mem_cgroup *memcg, *old_memcg;
821
822         if (retry)
823                 gfp |= __GFP_NOFAIL;
824
825         /* The page lock pins the memcg */
826         memcg = page_memcg(page);
827         old_memcg = set_active_memcg(memcg);
828
829         head = NULL;
830         offset = PAGE_SIZE;
831         while ((offset -= size) >= 0) {
832                 bh = alloc_buffer_head(gfp);
833                 if (!bh)
834                         goto no_grow;
835
836                 bh->b_this_page = head;
837                 bh->b_blocknr = -1;
838                 head = bh;
839
840                 bh->b_size = size;
841
842                 /* Link the buffer to its page */
843                 set_bh_page(bh, page, offset);
844         }
845 out:
846         set_active_memcg(old_memcg);
847         return head;
848 /*
849  * In case anything failed, we just free everything we got.
850  */
851 no_grow:
852         if (head) {
853                 do {
854                         bh = head;
855                         head = head->b_this_page;
856                         free_buffer_head(bh);
857                 } while (head);
858         }
859
860         goto out;
861 }
862 EXPORT_SYMBOL_GPL(alloc_page_buffers);
863
864 static inline void
865 link_dev_buffers(struct page *page, struct buffer_head *head)
866 {
867         struct buffer_head *bh, *tail;
868
869         bh = head;
870         do {
871                 tail = bh;
872                 bh = bh->b_this_page;
873         } while (bh);
874         tail->b_this_page = head;
875         attach_page_private(page, head);
876 }
877
878 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
879 {
880         sector_t retval = ~((sector_t)0);
881         loff_t sz = i_size_read(bdev->bd_inode);
882
883         if (sz) {
884                 unsigned int sizebits = blksize_bits(size);
885                 retval = (sz >> sizebits);
886         }
887         return retval;
888 }
889
890 /*
891  * Initialise the state of a blockdev page's buffers.
892  */ 
893 static sector_t
894 init_page_buffers(struct page *page, struct block_device *bdev,
895                         sector_t block, int size)
896 {
897         struct buffer_head *head = page_buffers(page);
898         struct buffer_head *bh = head;
899         int uptodate = PageUptodate(page);
900         sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
901
902         do {
903                 if (!buffer_mapped(bh)) {
904                         bh->b_end_io = NULL;
905                         bh->b_private = NULL;
906                         bh->b_bdev = bdev;
907                         bh->b_blocknr = block;
908                         if (uptodate)
909                                 set_buffer_uptodate(bh);
910                         if (block < end_block)
911                                 set_buffer_mapped(bh);
912                 }
913                 block++;
914                 bh = bh->b_this_page;
915         } while (bh != head);
916
917         /*
918          * Caller needs to validate requested block against end of device.
919          */
920         return end_block;
921 }
922
923 /*
924  * Create the page-cache page that contains the requested block.
925  *
926  * This is used purely for blockdev mappings.
927  */
928 static int
929 grow_dev_page(struct block_device *bdev, sector_t block,
930               pgoff_t index, int size, int sizebits, gfp_t gfp)
931 {
932         struct inode *inode = bdev->bd_inode;
933         struct page *page;
934         struct buffer_head *bh;
935         sector_t end_block;
936         int ret = 0;
937         gfp_t gfp_mask;
938
939         gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
940
941         /*
942          * XXX: __getblk_slow() can not really deal with failure and
943          * will endlessly loop on improvised global reclaim.  Prefer
944          * looping in the allocator rather than here, at least that
945          * code knows what it's doing.
946          */
947         gfp_mask |= __GFP_NOFAIL;
948
949         page = find_or_create_page(inode->i_mapping, index, gfp_mask);
950
951         BUG_ON(!PageLocked(page));
952
953         if (page_has_buffers(page)) {
954                 bh = page_buffers(page);
955                 if (bh->b_size == size) {
956                         end_block = init_page_buffers(page, bdev,
957                                                 (sector_t)index << sizebits,
958                                                 size);
959                         goto done;
960                 }
961                 if (!try_to_free_buffers(page))
962                         goto failed;
963         }
964
965         /*
966          * Allocate some buffers for this page
967          */
968         bh = alloc_page_buffers(page, size, true);
969
970         /*
971          * Link the page to the buffers and initialise them.  Take the
972          * lock to be atomic wrt __find_get_block(), which does not
973          * run under the page lock.
974          */
975         spin_lock(&inode->i_mapping->private_lock);
976         link_dev_buffers(page, bh);
977         end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
978                         size);
979         spin_unlock(&inode->i_mapping->private_lock);
980 done:
981         ret = (block < end_block) ? 1 : -ENXIO;
982 failed:
983         unlock_page(page);
984         put_page(page);
985         return ret;
986 }
987
988 /*
989  * Create buffers for the specified block device block's page.  If
990  * that page was dirty, the buffers are set dirty also.
991  */
992 static int
993 grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
994 {
995         pgoff_t index;
996         int sizebits;
997
998         sizebits = PAGE_SHIFT - __ffs(size);
999         index = block >> sizebits;
1000
1001         /*
1002          * Check for a block which wants to lie outside our maximum possible
1003          * pagecache index.  (this comparison is done using sector_t types).
1004          */
1005         if (unlikely(index != block >> sizebits)) {
1006                 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1007                         "device %pg\n",
1008                         __func__, (unsigned long long)block,
1009                         bdev);
1010                 return -EIO;
1011         }
1012
1013         /* Create a page with the proper size buffers.. */
1014         return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1015 }
1016
1017 static struct buffer_head *
1018 __getblk_slow(struct block_device *bdev, sector_t block,
1019              unsigned size, gfp_t gfp)
1020 {
1021         /* Size must be multiple of hard sectorsize */
1022         if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1023                         (size < 512 || size > PAGE_SIZE))) {
1024                 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1025                                         size);
1026                 printk(KERN_ERR "logical block size: %d\n",
1027                                         bdev_logical_block_size(bdev));
1028
1029                 dump_stack();
1030                 return NULL;
1031         }
1032
1033         for (;;) {
1034                 struct buffer_head *bh;
1035                 int ret;
1036
1037                 bh = __find_get_block(bdev, block, size);
1038                 if (bh)
1039                         return bh;
1040
1041                 ret = grow_buffers(bdev, block, size, gfp);
1042                 if (ret < 0)
1043                         return NULL;
1044         }
1045 }
1046
1047 /*
1048  * The relationship between dirty buffers and dirty pages:
1049  *
1050  * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1051  * the page is tagged dirty in the page cache.
1052  *
1053  * At all times, the dirtiness of the buffers represents the dirtiness of
1054  * subsections of the page.  If the page has buffers, the page dirty bit is
1055  * merely a hint about the true dirty state.
1056  *
1057  * When a page is set dirty in its entirety, all its buffers are marked dirty
1058  * (if the page has buffers).
1059  *
1060  * When a buffer is marked dirty, its page is dirtied, but the page's other
1061  * buffers are not.
1062  *
1063  * Also.  When blockdev buffers are explicitly read with bread(), they
1064  * individually become uptodate.  But their backing page remains not
1065  * uptodate - even if all of its buffers are uptodate.  A subsequent
1066  * block_read_full_page() against that page will discover all the uptodate
1067  * buffers, will set the page uptodate and will perform no I/O.
1068  */
1069
1070 /**
1071  * mark_buffer_dirty - mark a buffer_head as needing writeout
1072  * @bh: the buffer_head to mark dirty
1073  *
1074  * mark_buffer_dirty() will set the dirty bit against the buffer, then set
1075  * its backing page dirty, then tag the page as dirty in the page cache
1076  * and then attach the address_space's inode to its superblock's dirty
1077  * inode list.
1078  *
1079  * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1080  * i_pages lock and mapping->host->i_lock.
1081  */
1082 void mark_buffer_dirty(struct buffer_head *bh)
1083 {
1084         WARN_ON_ONCE(!buffer_uptodate(bh));
1085
1086         trace_block_dirty_buffer(bh);
1087
1088         /*
1089          * Very *carefully* optimize the it-is-already-dirty case.
1090          *
1091          * Don't let the final "is it dirty" escape to before we
1092          * perhaps modified the buffer.
1093          */
1094         if (buffer_dirty(bh)) {
1095                 smp_mb();
1096                 if (buffer_dirty(bh))
1097                         return;
1098         }
1099
1100         if (!test_set_buffer_dirty(bh)) {
1101                 struct page *page = bh->b_page;
1102                 struct address_space *mapping = NULL;
1103
1104                 lock_page_memcg(page);
1105                 if (!TestSetPageDirty(page)) {
1106                         mapping = page_mapping(page);
1107                         if (mapping)
1108                                 __set_page_dirty(page, mapping, 0);
1109                 }
1110                 unlock_page_memcg(page);
1111                 if (mapping)
1112                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1113         }
1114 }
1115 EXPORT_SYMBOL(mark_buffer_dirty);
1116
1117 void mark_buffer_write_io_error(struct buffer_head *bh)
1118 {
1119         struct super_block *sb;
1120
1121         set_buffer_write_io_error(bh);
1122         /* FIXME: do we need to set this in both places? */
1123         if (bh->b_page && bh->b_page->mapping)
1124                 mapping_set_error(bh->b_page->mapping, -EIO);
1125         if (bh->b_assoc_map)
1126                 mapping_set_error(bh->b_assoc_map, -EIO);
1127         rcu_read_lock();
1128         sb = READ_ONCE(bh->b_bdev->bd_super);
1129         if (sb)
1130                 errseq_set(&sb->s_wb_err, -EIO);
1131         rcu_read_unlock();
1132 }
1133 EXPORT_SYMBOL(mark_buffer_write_io_error);
1134
1135 /*
1136  * Decrement a buffer_head's reference count.  If all buffers against a page
1137  * have zero reference count, are clean and unlocked, and if the page is clean
1138  * and unlocked then try_to_free_buffers() may strip the buffers from the page
1139  * in preparation for freeing it (sometimes, rarely, buffers are removed from
1140  * a page but it ends up not being freed, and buffers may later be reattached).
1141  */
1142 void __brelse(struct buffer_head * buf)
1143 {
1144         if (atomic_read(&buf->b_count)) {
1145                 put_bh(buf);
1146                 return;
1147         }
1148         WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1149 }
1150 EXPORT_SYMBOL(__brelse);
1151
1152 /*
1153  * bforget() is like brelse(), except it discards any
1154  * potentially dirty data.
1155  */
1156 void __bforget(struct buffer_head *bh)
1157 {
1158         clear_buffer_dirty(bh);
1159         if (bh->b_assoc_map) {
1160                 struct address_space *buffer_mapping = bh->b_page->mapping;
1161
1162                 spin_lock(&buffer_mapping->private_lock);
1163                 list_del_init(&bh->b_assoc_buffers);
1164                 bh->b_assoc_map = NULL;
1165                 spin_unlock(&buffer_mapping->private_lock);
1166         }
1167         __brelse(bh);
1168 }
1169 EXPORT_SYMBOL(__bforget);
1170
1171 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1172 {
1173         lock_buffer(bh);
1174         if (buffer_uptodate(bh)) {
1175                 unlock_buffer(bh);
1176                 return bh;
1177         } else {
1178                 get_bh(bh);
1179                 bh->b_end_io = end_buffer_read_sync;
1180                 submit_bh(REQ_OP_READ, 0, bh);
1181                 wait_on_buffer(bh);
1182                 if (buffer_uptodate(bh))
1183                         return bh;
1184         }
1185         brelse(bh);
1186         return NULL;
1187 }
1188
1189 /*
1190  * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1191  * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1192  * refcount elevated by one when they're in an LRU.  A buffer can only appear
1193  * once in a particular CPU's LRU.  A single buffer can be present in multiple
1194  * CPU's LRUs at the same time.
1195  *
1196  * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1197  * sb_find_get_block().
1198  *
1199  * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1200  * a local interrupt disable for that.
1201  */
1202
1203 #define BH_LRU_SIZE     16
1204
1205 struct bh_lru {
1206         struct buffer_head *bhs[BH_LRU_SIZE];
1207 };
1208
1209 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1210
1211 #ifdef CONFIG_SMP
1212 #define bh_lru_lock()   local_irq_disable()
1213 #define bh_lru_unlock() local_irq_enable()
1214 #else
1215 #define bh_lru_lock()   preempt_disable()
1216 #define bh_lru_unlock() preempt_enable()
1217 #endif
1218
1219 static inline void check_irqs_on(void)
1220 {
1221 #ifdef irqs_disabled
1222         BUG_ON(irqs_disabled());
1223 #endif
1224 }
1225
1226 /*
1227  * Install a buffer_head into this cpu's LRU.  If not already in the LRU, it is
1228  * inserted at the front, and the buffer_head at the back if any is evicted.
1229  * Or, if already in the LRU it is moved to the front.
1230  */
1231 static void bh_lru_install(struct buffer_head *bh)
1232 {
1233         struct buffer_head *evictee = bh;
1234         struct bh_lru *b;
1235         int i;
1236
1237         check_irqs_on();
1238         /*
1239          * the refcount of buffer_head in bh_lru prevents dropping the
1240          * attached page(i.e., try_to_free_buffers) so it could cause
1241          * failing page migration.
1242          * Skip putting upcoming bh into bh_lru until migration is done.
1243          */
1244         if (lru_cache_disabled())
1245                 return;
1246
1247         bh_lru_lock();
1248
1249         b = this_cpu_ptr(&bh_lrus);
1250         for (i = 0; i < BH_LRU_SIZE; i++) {
1251                 swap(evictee, b->bhs[i]);
1252                 if (evictee == bh) {
1253                         bh_lru_unlock();
1254                         return;
1255                 }
1256         }
1257
1258         get_bh(bh);
1259         bh_lru_unlock();
1260         brelse(evictee);
1261 }
1262
1263 /*
1264  * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1265  */
1266 static struct buffer_head *
1267 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1268 {
1269         struct buffer_head *ret = NULL;
1270         unsigned int i;
1271
1272         check_irqs_on();
1273         bh_lru_lock();
1274         for (i = 0; i < BH_LRU_SIZE; i++) {
1275                 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1276
1277                 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1278                     bh->b_size == size) {
1279                         if (i) {
1280                                 while (i) {
1281                                         __this_cpu_write(bh_lrus.bhs[i],
1282                                                 __this_cpu_read(bh_lrus.bhs[i - 1]));
1283                                         i--;
1284                                 }
1285                                 __this_cpu_write(bh_lrus.bhs[0], bh);
1286                         }
1287                         get_bh(bh);
1288                         ret = bh;
1289                         break;
1290                 }
1291         }
1292         bh_lru_unlock();
1293         return ret;
1294 }
1295
1296 /*
1297  * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1298  * it in the LRU and mark it as accessed.  If it is not present then return
1299  * NULL
1300  */
1301 struct buffer_head *
1302 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1303 {
1304         struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1305
1306         if (bh == NULL) {
1307                 /* __find_get_block_slow will mark the page accessed */
1308                 bh = __find_get_block_slow(bdev, block);
1309                 if (bh)
1310                         bh_lru_install(bh);
1311         } else
1312                 touch_buffer(bh);
1313
1314         return bh;
1315 }
1316 EXPORT_SYMBOL(__find_get_block);
1317
1318 /*
1319  * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1320  * which corresponds to the passed block_device, block and size. The
1321  * returned buffer has its reference count incremented.
1322  *
1323  * __getblk_gfp() will lock up the machine if grow_dev_page's
1324  * try_to_free_buffers() attempt is failing.  FIXME, perhaps?
1325  */
1326 struct buffer_head *
1327 __getblk_gfp(struct block_device *bdev, sector_t block,
1328              unsigned size, gfp_t gfp)
1329 {
1330         struct buffer_head *bh = __find_get_block(bdev, block, size);
1331
1332         might_sleep();
1333         if (bh == NULL)
1334                 bh = __getblk_slow(bdev, block, size, gfp);
1335         return bh;
1336 }
1337 EXPORT_SYMBOL(__getblk_gfp);
1338
1339 /*
1340  * Do async read-ahead on a buffer..
1341  */
1342 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1343 {
1344         struct buffer_head *bh = __getblk(bdev, block, size);
1345         if (likely(bh)) {
1346                 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1347                 brelse(bh);
1348         }
1349 }
1350 EXPORT_SYMBOL(__breadahead);
1351
1352 void __breadahead_gfp(struct block_device *bdev, sector_t block, unsigned size,
1353                       gfp_t gfp)
1354 {
1355         struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1356         if (likely(bh)) {
1357                 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1358                 brelse(bh);
1359         }
1360 }
1361 EXPORT_SYMBOL(__breadahead_gfp);
1362
1363 /**
1364  *  __bread_gfp() - reads a specified block and returns the bh
1365  *  @bdev: the block_device to read from
1366  *  @block: number of block
1367  *  @size: size (in bytes) to read
1368  *  @gfp: page allocation flag
1369  *
1370  *  Reads a specified block, and returns buffer head that contains it.
1371  *  The page cache can be allocated from non-movable area
1372  *  not to prevent page migration if you set gfp to zero.
1373  *  It returns NULL if the block was unreadable.
1374  */
1375 struct buffer_head *
1376 __bread_gfp(struct block_device *bdev, sector_t block,
1377                    unsigned size, gfp_t gfp)
1378 {
1379         struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1380
1381         if (likely(bh) && !buffer_uptodate(bh))
1382                 bh = __bread_slow(bh);
1383         return bh;
1384 }
1385 EXPORT_SYMBOL(__bread_gfp);
1386
1387 static void __invalidate_bh_lrus(struct bh_lru *b)
1388 {
1389         int i;
1390
1391         for (i = 0; i < BH_LRU_SIZE; i++) {
1392                 brelse(b->bhs[i]);
1393                 b->bhs[i] = NULL;
1394         }
1395 }
1396 /*
1397  * invalidate_bh_lrus() is called rarely - but not only at unmount.
1398  * This doesn't race because it runs in each cpu either in irq
1399  * or with preempt disabled.
1400  */
1401 static void invalidate_bh_lru(void *arg)
1402 {
1403         struct bh_lru *b = &get_cpu_var(bh_lrus);
1404
1405         __invalidate_bh_lrus(b);
1406         put_cpu_var(bh_lrus);
1407 }
1408
1409 bool has_bh_in_lru(int cpu, void *dummy)
1410 {
1411         struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1412         int i;
1413         
1414         for (i = 0; i < BH_LRU_SIZE; i++) {
1415                 if (b->bhs[i])
1416                         return true;
1417         }
1418
1419         return false;
1420 }
1421
1422 void invalidate_bh_lrus(void)
1423 {
1424         on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1);
1425 }
1426 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1427
1428 void invalidate_bh_lrus_cpu(int cpu)
1429 {
1430         struct bh_lru *b;
1431
1432         bh_lru_lock();
1433         b = per_cpu_ptr(&bh_lrus, cpu);
1434         __invalidate_bh_lrus(b);
1435         bh_lru_unlock();
1436 }
1437
1438 void set_bh_page(struct buffer_head *bh,
1439                 struct page *page, unsigned long offset)
1440 {
1441         bh->b_page = page;
1442         BUG_ON(offset >= PAGE_SIZE);
1443         if (PageHighMem(page))
1444                 /*
1445                  * This catches illegal uses and preserves the offset:
1446                  */
1447                 bh->b_data = (char *)(0 + offset);
1448         else
1449                 bh->b_data = page_address(page) + offset;
1450 }
1451 EXPORT_SYMBOL(set_bh_page);
1452
1453 /*
1454  * Called when truncating a buffer on a page completely.
1455  */
1456
1457 /* Bits that are cleared during an invalidate */
1458 #define BUFFER_FLAGS_DISCARD \
1459         (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1460          1 << BH_Delay | 1 << BH_Unwritten)
1461
1462 static void discard_buffer(struct buffer_head * bh)
1463 {
1464         unsigned long b_state, b_state_old;
1465
1466         lock_buffer(bh);
1467         clear_buffer_dirty(bh);
1468         bh->b_bdev = NULL;
1469         b_state = bh->b_state;
1470         for (;;) {
1471                 b_state_old = cmpxchg(&bh->b_state, b_state,
1472                                       (b_state & ~BUFFER_FLAGS_DISCARD));
1473                 if (b_state_old == b_state)
1474                         break;
1475                 b_state = b_state_old;
1476         }
1477         unlock_buffer(bh);
1478 }
1479
1480 /**
1481  * block_invalidatepage - invalidate part or all of a buffer-backed page
1482  *
1483  * @page: the page which is affected
1484  * @offset: start of the range to invalidate
1485  * @length: length of the range to invalidate
1486  *
1487  * block_invalidatepage() is called when all or part of the page has become
1488  * invalidated by a truncate operation.
1489  *
1490  * block_invalidatepage() does not have to release all buffers, but it must
1491  * ensure that no dirty buffer is left outside @offset and that no I/O
1492  * is underway against any of the blocks which are outside the truncation
1493  * point.  Because the caller is about to free (and possibly reuse) those
1494  * blocks on-disk.
1495  */
1496 void block_invalidatepage(struct page *page, unsigned int offset,
1497                           unsigned int length)
1498 {
1499         struct buffer_head *head, *bh, *next;
1500         unsigned int curr_off = 0;
1501         unsigned int stop = length + offset;
1502
1503         BUG_ON(!PageLocked(page));
1504         if (!page_has_buffers(page))
1505                 goto out;
1506
1507         /*
1508          * Check for overflow
1509          */
1510         BUG_ON(stop > PAGE_SIZE || stop < length);
1511
1512         head = page_buffers(page);
1513         bh = head;
1514         do {
1515                 unsigned int next_off = curr_off + bh->b_size;
1516                 next = bh->b_this_page;
1517
1518                 /*
1519                  * Are we still fully in range ?
1520                  */
1521                 if (next_off > stop)
1522                         goto out;
1523
1524                 /*
1525                  * is this block fully invalidated?
1526                  */
1527                 if (offset <= curr_off)
1528                         discard_buffer(bh);
1529                 curr_off = next_off;
1530                 bh = next;
1531         } while (bh != head);
1532
1533         /*
1534          * We release buffers only if the entire page is being invalidated.
1535          * The get_block cached value has been unconditionally invalidated,
1536          * so real IO is not possible anymore.
1537          */
1538         if (length == PAGE_SIZE)
1539                 try_to_release_page(page, 0);
1540 out:
1541         return;
1542 }
1543 EXPORT_SYMBOL(block_invalidatepage);
1544
1545
1546 /*
1547  * We attach and possibly dirty the buffers atomically wrt
1548  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1549  * is already excluded via the page lock.
1550  */
1551 void create_empty_buffers(struct page *page,
1552                         unsigned long blocksize, unsigned long b_state)
1553 {
1554         struct buffer_head *bh, *head, *tail;
1555
1556         head = alloc_page_buffers(page, blocksize, true);
1557         bh = head;
1558         do {
1559                 bh->b_state |= b_state;
1560                 tail = bh;
1561                 bh = bh->b_this_page;
1562         } while (bh);
1563         tail->b_this_page = head;
1564
1565         spin_lock(&page->mapping->private_lock);
1566         if (PageUptodate(page) || PageDirty(page)) {
1567                 bh = head;
1568                 do {
1569                         if (PageDirty(page))
1570                                 set_buffer_dirty(bh);
1571                         if (PageUptodate(page))
1572                                 set_buffer_uptodate(bh);
1573                         bh = bh->b_this_page;
1574                 } while (bh != head);
1575         }
1576         attach_page_private(page, head);
1577         spin_unlock(&page->mapping->private_lock);
1578 }
1579 EXPORT_SYMBOL(create_empty_buffers);
1580
1581 /**
1582  * clean_bdev_aliases: clean a range of buffers in block device
1583  * @bdev: Block device to clean buffers in
1584  * @block: Start of a range of blocks to clean
1585  * @len: Number of blocks to clean
1586  *
1587  * We are taking a range of blocks for data and we don't want writeback of any
1588  * buffer-cache aliases starting from return from this function and until the
1589  * moment when something will explicitly mark the buffer dirty (hopefully that
1590  * will not happen until we will free that block ;-) We don't even need to mark
1591  * it not-uptodate - nobody can expect anything from a newly allocated buffer
1592  * anyway. We used to use unmap_buffer() for such invalidation, but that was
1593  * wrong. We definitely don't want to mark the alias unmapped, for example - it
1594  * would confuse anyone who might pick it with bread() afterwards...
1595  *
1596  * Also..  Note that bforget() doesn't lock the buffer.  So there can be
1597  * writeout I/O going on against recently-freed buffers.  We don't wait on that
1598  * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1599  * need to.  That happens here.
1600  */
1601 void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1602 {
1603         struct inode *bd_inode = bdev->bd_inode;
1604         struct address_space *bd_mapping = bd_inode->i_mapping;
1605         struct pagevec pvec;
1606         pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1607         pgoff_t end;
1608         int i, count;
1609         struct buffer_head *bh;
1610         struct buffer_head *head;
1611
1612         end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1613         pagevec_init(&pvec);
1614         while (pagevec_lookup_range(&pvec, bd_mapping, &index, end)) {
1615                 count = pagevec_count(&pvec);
1616                 for (i = 0; i < count; i++) {
1617                         struct page *page = pvec.pages[i];
1618
1619                         if (!page_has_buffers(page))
1620                                 continue;
1621                         /*
1622                          * We use page lock instead of bd_mapping->private_lock
1623                          * to pin buffers here since we can afford to sleep and
1624                          * it scales better than a global spinlock lock.
1625                          */
1626                         lock_page(page);
1627                         /* Recheck when the page is locked which pins bhs */
1628                         if (!page_has_buffers(page))
1629                                 goto unlock_page;
1630                         head = page_buffers(page);
1631                         bh = head;
1632                         do {
1633                                 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1634                                         goto next;
1635                                 if (bh->b_blocknr >= block + len)
1636                                         break;
1637                                 clear_buffer_dirty(bh);
1638                                 wait_on_buffer(bh);
1639                                 clear_buffer_req(bh);
1640 next:
1641                                 bh = bh->b_this_page;
1642                         } while (bh != head);
1643 unlock_page:
1644                         unlock_page(page);
1645                 }
1646                 pagevec_release(&pvec);
1647                 cond_resched();
1648                 /* End of range already reached? */
1649                 if (index > end || !index)
1650                         break;
1651         }
1652 }
1653 EXPORT_SYMBOL(clean_bdev_aliases);
1654
1655 /*
1656  * Size is a power-of-two in the range 512..PAGE_SIZE,
1657  * and the case we care about most is PAGE_SIZE.
1658  *
1659  * So this *could* possibly be written with those
1660  * constraints in mind (relevant mostly if some
1661  * architecture has a slow bit-scan instruction)
1662  */
1663 static inline int block_size_bits(unsigned int blocksize)
1664 {
1665         return ilog2(blocksize);
1666 }
1667
1668 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1669 {
1670         BUG_ON(!PageLocked(page));
1671
1672         if (!page_has_buffers(page))
1673                 create_empty_buffers(page, 1 << READ_ONCE(inode->i_blkbits),
1674                                      b_state);
1675         return page_buffers(page);
1676 }
1677
1678 /*
1679  * NOTE! All mapped/uptodate combinations are valid:
1680  *
1681  *      Mapped  Uptodate        Meaning
1682  *
1683  *      No      No              "unknown" - must do get_block()
1684  *      No      Yes             "hole" - zero-filled
1685  *      Yes     No              "allocated" - allocated on disk, not read in
1686  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1687  *
1688  * "Dirty" is valid only with the last case (mapped+uptodate).
1689  */
1690
1691 /*
1692  * While block_write_full_page is writing back the dirty buffers under
1693  * the page lock, whoever dirtied the buffers may decide to clean them
1694  * again at any time.  We handle that by only looking at the buffer
1695  * state inside lock_buffer().
1696  *
1697  * If block_write_full_page() is called for regular writeback
1698  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1699  * locked buffer.   This only can happen if someone has written the buffer
1700  * directly, with submit_bh().  At the address_space level PageWriteback
1701  * prevents this contention from occurring.
1702  *
1703  * If block_write_full_page() is called with wbc->sync_mode ==
1704  * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1705  * causes the writes to be flagged as synchronous writes.
1706  */
1707 int __block_write_full_page(struct inode *inode, struct page *page,
1708                         get_block_t *get_block, struct writeback_control *wbc,
1709                         bh_end_io_t *handler)
1710 {
1711         int err;
1712         sector_t block;
1713         sector_t last_block;
1714         struct buffer_head *bh, *head;
1715         unsigned int blocksize, bbits;
1716         int nr_underway = 0;
1717         int write_flags = wbc_to_write_flags(wbc);
1718
1719         head = create_page_buffers(page, inode,
1720                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1721
1722         /*
1723          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1724          * here, and the (potentially unmapped) buffers may become dirty at
1725          * any time.  If a buffer becomes dirty here after we've inspected it
1726          * then we just miss that fact, and the page stays dirty.
1727          *
1728          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1729          * handle that here by just cleaning them.
1730          */
1731
1732         bh = head;
1733         blocksize = bh->b_size;
1734         bbits = block_size_bits(blocksize);
1735
1736         block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1737         last_block = (i_size_read(inode) - 1) >> bbits;
1738
1739         /*
1740          * Get all the dirty buffers mapped to disk addresses and
1741          * handle any aliases from the underlying blockdev's mapping.
1742          */
1743         do {
1744                 if (block > last_block) {
1745                         /*
1746                          * mapped buffers outside i_size will occur, because
1747                          * this page can be outside i_size when there is a
1748                          * truncate in progress.
1749                          */
1750                         /*
1751                          * The buffer was zeroed by block_write_full_page()
1752                          */
1753                         clear_buffer_dirty(bh);
1754                         set_buffer_uptodate(bh);
1755                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1756                            buffer_dirty(bh)) {
1757                         WARN_ON(bh->b_size != blocksize);
1758                         err = get_block(inode, block, bh, 1);
1759                         if (err)
1760                                 goto recover;
1761                         clear_buffer_delay(bh);
1762                         if (buffer_new(bh)) {
1763                                 /* blockdev mappings never come here */
1764                                 clear_buffer_new(bh);
1765                                 clean_bdev_bh_alias(bh);
1766                         }
1767                 }
1768                 bh = bh->b_this_page;
1769                 block++;
1770         } while (bh != head);
1771
1772         do {
1773                 if (!buffer_mapped(bh))
1774                         continue;
1775                 /*
1776                  * If it's a fully non-blocking write attempt and we cannot
1777                  * lock the buffer then redirty the page.  Note that this can
1778                  * potentially cause a busy-wait loop from writeback threads
1779                  * and kswapd activity, but those code paths have their own
1780                  * higher-level throttling.
1781                  */
1782                 if (wbc->sync_mode != WB_SYNC_NONE) {
1783                         lock_buffer(bh);
1784                 } else if (!trylock_buffer(bh)) {
1785                         redirty_page_for_writepage(wbc, page);
1786                         continue;
1787                 }
1788                 if (test_clear_buffer_dirty(bh)) {
1789                         mark_buffer_async_write_endio(bh, handler);
1790                 } else {
1791                         unlock_buffer(bh);
1792                 }
1793         } while ((bh = bh->b_this_page) != head);
1794
1795         /*
1796          * The page and its buffers are protected by PageWriteback(), so we can
1797          * drop the bh refcounts early.
1798          */
1799         BUG_ON(PageWriteback(page));
1800         set_page_writeback(page);
1801
1802         do {
1803                 struct buffer_head *next = bh->b_this_page;
1804                 if (buffer_async_write(bh)) {
1805                         submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1806                                         inode->i_write_hint, wbc);
1807                         nr_underway++;
1808                 }
1809                 bh = next;
1810         } while (bh != head);
1811         unlock_page(page);
1812
1813         err = 0;
1814 done:
1815         if (nr_underway == 0) {
1816                 /*
1817                  * The page was marked dirty, but the buffers were
1818                  * clean.  Someone wrote them back by hand with
1819                  * ll_rw_block/submit_bh.  A rare case.
1820                  */
1821                 end_page_writeback(page);
1822
1823                 /*
1824                  * The page and buffer_heads can be released at any time from
1825                  * here on.
1826                  */
1827         }
1828         return err;
1829
1830 recover:
1831         /*
1832          * ENOSPC, or some other error.  We may already have added some
1833          * blocks to the file, so we need to write these out to avoid
1834          * exposing stale data.
1835          * The page is currently locked and not marked for writeback
1836          */
1837         bh = head;
1838         /* Recovery: lock and submit the mapped buffers */
1839         do {
1840                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1841                     !buffer_delay(bh)) {
1842                         lock_buffer(bh);
1843                         mark_buffer_async_write_endio(bh, handler);
1844                 } else {
1845                         /*
1846                          * The buffer may have been set dirty during
1847                          * attachment to a dirty page.
1848                          */
1849                         clear_buffer_dirty(bh);
1850                 }
1851         } while ((bh = bh->b_this_page) != head);
1852         SetPageError(page);
1853         BUG_ON(PageWriteback(page));
1854         mapping_set_error(page->mapping, err);
1855         set_page_writeback(page);
1856         do {
1857                 struct buffer_head *next = bh->b_this_page;
1858                 if (buffer_async_write(bh)) {
1859                         clear_buffer_dirty(bh);
1860                         submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1861                                         inode->i_write_hint, wbc);
1862                         nr_underway++;
1863                 }
1864                 bh = next;
1865         } while (bh != head);
1866         unlock_page(page);
1867         goto done;
1868 }
1869 EXPORT_SYMBOL(__block_write_full_page);
1870
1871 /*
1872  * If a page has any new buffers, zero them out here, and mark them uptodate
1873  * and dirty so they'll be written out (in order to prevent uninitialised
1874  * block data from leaking). And clear the new bit.
1875  */
1876 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1877 {
1878         unsigned int block_start, block_end;
1879         struct buffer_head *head, *bh;
1880
1881         BUG_ON(!PageLocked(page));
1882         if (!page_has_buffers(page))
1883                 return;
1884
1885         bh = head = page_buffers(page);
1886         block_start = 0;
1887         do {
1888                 block_end = block_start + bh->b_size;
1889
1890                 if (buffer_new(bh)) {
1891                         if (block_end > from && block_start < to) {
1892                                 if (!PageUptodate(page)) {
1893                                         unsigned start, size;
1894
1895                                         start = max(from, block_start);
1896                                         size = min(to, block_end) - start;
1897
1898                                         zero_user(page, start, size);
1899                                         set_buffer_uptodate(bh);
1900                                 }
1901
1902                                 clear_buffer_new(bh);
1903                                 mark_buffer_dirty(bh);
1904                         }
1905                 }
1906
1907                 block_start = block_end;
1908                 bh = bh->b_this_page;
1909         } while (bh != head);
1910 }
1911 EXPORT_SYMBOL(page_zero_new_buffers);
1912
1913 static void
1914 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1915                 const struct iomap *iomap)
1916 {
1917         loff_t offset = block << inode->i_blkbits;
1918
1919         bh->b_bdev = iomap->bdev;
1920
1921         /*
1922          * Block points to offset in file we need to map, iomap contains
1923          * the offset at which the map starts. If the map ends before the
1924          * current block, then do not map the buffer and let the caller
1925          * handle it.
1926          */
1927         BUG_ON(offset >= iomap->offset + iomap->length);
1928
1929         switch (iomap->type) {
1930         case IOMAP_HOLE:
1931                 /*
1932                  * If the buffer is not up to date or beyond the current EOF,
1933                  * we need to mark it as new to ensure sub-block zeroing is
1934                  * executed if necessary.
1935                  */
1936                 if (!buffer_uptodate(bh) ||
1937                     (offset >= i_size_read(inode)))
1938                         set_buffer_new(bh);
1939                 break;
1940         case IOMAP_DELALLOC:
1941                 if (!buffer_uptodate(bh) ||
1942                     (offset >= i_size_read(inode)))
1943                         set_buffer_new(bh);
1944                 set_buffer_uptodate(bh);
1945                 set_buffer_mapped(bh);
1946                 set_buffer_delay(bh);
1947                 break;
1948         case IOMAP_UNWRITTEN:
1949                 /*
1950                  * For unwritten regions, we always need to ensure that regions
1951                  * in the block we are not writing to are zeroed. Mark the
1952                  * buffer as new to ensure this.
1953                  */
1954                 set_buffer_new(bh);
1955                 set_buffer_unwritten(bh);
1956                 fallthrough;
1957         case IOMAP_MAPPED:
1958                 if ((iomap->flags & IOMAP_F_NEW) ||
1959                     offset >= i_size_read(inode))
1960                         set_buffer_new(bh);
1961                 bh->b_blocknr = (iomap->addr + offset - iomap->offset) >>
1962                                 inode->i_blkbits;
1963                 set_buffer_mapped(bh);
1964                 break;
1965         }
1966 }
1967
1968 int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
1969                 get_block_t *get_block, const struct iomap *iomap)
1970 {
1971         unsigned from = pos & (PAGE_SIZE - 1);
1972         unsigned to = from + len;
1973         struct inode *inode = page->mapping->host;
1974         unsigned block_start, block_end;
1975         sector_t block;
1976         int err = 0;
1977         unsigned blocksize, bbits;
1978         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1979
1980         BUG_ON(!PageLocked(page));
1981         BUG_ON(from > PAGE_SIZE);
1982         BUG_ON(to > PAGE_SIZE);
1983         BUG_ON(from > to);
1984
1985         head = create_page_buffers(page, inode, 0);
1986         blocksize = head->b_size;
1987         bbits = block_size_bits(blocksize);
1988
1989         block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1990
1991         for(bh = head, block_start = 0; bh != head || !block_start;
1992             block++, block_start=block_end, bh = bh->b_this_page) {
1993                 block_end = block_start + blocksize;
1994                 if (block_end <= from || block_start >= to) {
1995                         if (PageUptodate(page)) {
1996                                 if (!buffer_uptodate(bh))
1997                                         set_buffer_uptodate(bh);
1998                         }
1999                         continue;
2000                 }
2001                 if (buffer_new(bh))
2002                         clear_buffer_new(bh);
2003                 if (!buffer_mapped(bh)) {
2004                         WARN_ON(bh->b_size != blocksize);
2005                         if (get_block) {
2006                                 err = get_block(inode, block, bh, 1);
2007                                 if (err)
2008                                         break;
2009                         } else {
2010                                 iomap_to_bh(inode, block, bh, iomap);
2011                         }
2012
2013                         if (buffer_new(bh)) {
2014                                 clean_bdev_bh_alias(bh);
2015                                 if (PageUptodate(page)) {
2016                                         clear_buffer_new(bh);
2017                                         set_buffer_uptodate(bh);
2018                                         mark_buffer_dirty(bh);
2019                                         continue;
2020                                 }
2021                                 if (block_end > to || block_start < from)
2022                                         zero_user_segments(page,
2023                                                 to, block_end,
2024                                                 block_start, from);
2025                                 continue;
2026                         }
2027                 }
2028                 if (PageUptodate(page)) {
2029                         if (!buffer_uptodate(bh))
2030                                 set_buffer_uptodate(bh);
2031                         continue; 
2032                 }
2033                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2034                     !buffer_unwritten(bh) &&
2035                      (block_start < from || block_end > to)) {
2036                         ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2037                         *wait_bh++=bh;
2038                 }
2039         }
2040         /*
2041          * If we issued read requests - let them complete.
2042          */
2043         while(wait_bh > wait) {
2044                 wait_on_buffer(*--wait_bh);
2045                 if (!buffer_uptodate(*wait_bh))
2046                         err = -EIO;
2047         }
2048         if (unlikely(err))
2049                 page_zero_new_buffers(page, from, to);
2050         return err;
2051 }
2052
2053 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2054                 get_block_t *get_block)
2055 {
2056         return __block_write_begin_int(page, pos, len, get_block, NULL);
2057 }
2058 EXPORT_SYMBOL(__block_write_begin);
2059
2060 static int __block_commit_write(struct inode *inode, struct page *page,
2061                 unsigned from, unsigned to)
2062 {
2063         unsigned block_start, block_end;
2064         int partial = 0;
2065         unsigned blocksize;
2066         struct buffer_head *bh, *head;
2067
2068         bh = head = page_buffers(page);
2069         blocksize = bh->b_size;
2070
2071         block_start = 0;
2072         do {
2073                 block_end = block_start + blocksize;
2074                 if (block_end <= from || block_start >= to) {
2075                         if (!buffer_uptodate(bh))
2076                                 partial = 1;
2077                 } else {
2078                         set_buffer_uptodate(bh);
2079                         mark_buffer_dirty(bh);
2080                 }
2081                 if (buffer_new(bh))
2082                         clear_buffer_new(bh);
2083
2084                 block_start = block_end;
2085                 bh = bh->b_this_page;
2086         } while (bh != head);
2087
2088         /*
2089          * If this is a partial write which happened to make all buffers
2090          * uptodate then we can optimize away a bogus readpage() for
2091          * the next read(). Here we 'discover' whether the page went
2092          * uptodate as a result of this (potentially partial) write.
2093          */
2094         if (!partial)
2095                 SetPageUptodate(page);
2096         return 0;
2097 }
2098
2099 /*
2100  * block_write_begin takes care of the basic task of block allocation and
2101  * bringing partial write blocks uptodate first.
2102  *
2103  * The filesystem needs to handle block truncation upon failure.
2104  */
2105 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2106                 unsigned flags, struct page **pagep, get_block_t *get_block)
2107 {
2108         pgoff_t index = pos >> PAGE_SHIFT;
2109         struct page *page;
2110         int status;
2111
2112         page = grab_cache_page_write_begin(mapping, index, flags);
2113         if (!page)
2114                 return -ENOMEM;
2115
2116         status = __block_write_begin(page, pos, len, get_block);
2117         if (unlikely(status)) {
2118                 unlock_page(page);
2119                 put_page(page);
2120                 page = NULL;
2121         }
2122
2123         *pagep = page;
2124         return status;
2125 }
2126 EXPORT_SYMBOL(block_write_begin);
2127
2128 int block_write_end(struct file *file, struct address_space *mapping,
2129                         loff_t pos, unsigned len, unsigned copied,
2130                         struct page *page, void *fsdata)
2131 {
2132         struct inode *inode = mapping->host;
2133         unsigned start;
2134
2135         start = pos & (PAGE_SIZE - 1);
2136
2137         if (unlikely(copied < len)) {
2138                 /*
2139                  * The buffers that were written will now be uptodate, so we
2140                  * don't have to worry about a readpage reading them and
2141                  * overwriting a partial write. However if we have encountered
2142                  * a short write and only partially written into a buffer, it
2143                  * will not be marked uptodate, so a readpage might come in and
2144                  * destroy our partial write.
2145                  *
2146                  * Do the simplest thing, and just treat any short write to a
2147                  * non uptodate page as a zero-length write, and force the
2148                  * caller to redo the whole thing.
2149                  */
2150                 if (!PageUptodate(page))
2151                         copied = 0;
2152
2153                 page_zero_new_buffers(page, start+copied, start+len);
2154         }
2155         flush_dcache_page(page);
2156
2157         /* This could be a short (even 0-length) commit */
2158         __block_commit_write(inode, page, start, start+copied);
2159
2160         return copied;
2161 }
2162 EXPORT_SYMBOL(block_write_end);
2163
2164 int generic_write_end(struct file *file, struct address_space *mapping,
2165                         loff_t pos, unsigned len, unsigned copied,
2166                         struct page *page, void *fsdata)
2167 {
2168         struct inode *inode = mapping->host;
2169         loff_t old_size = inode->i_size;
2170         bool i_size_changed = false;
2171
2172         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2173
2174         /*
2175          * No need to use i_size_read() here, the i_size cannot change under us
2176          * because we hold i_rwsem.
2177          *
2178          * But it's important to update i_size while still holding page lock:
2179          * page writeout could otherwise come in and zero beyond i_size.
2180          */
2181         if (pos + copied > inode->i_size) {
2182                 i_size_write(inode, pos + copied);
2183                 i_size_changed = true;
2184         }
2185
2186         unlock_page(page);
2187         put_page(page);
2188
2189         if (old_size < pos)
2190                 pagecache_isize_extended(inode, old_size, pos);
2191         /*
2192          * Don't mark the inode dirty under page lock. First, it unnecessarily
2193          * makes the holding time of page lock longer. Second, it forces lock
2194          * ordering of page lock and transaction start for journaling
2195          * filesystems.
2196          */
2197         if (i_size_changed)
2198                 mark_inode_dirty(inode);
2199         return copied;
2200 }
2201 EXPORT_SYMBOL(generic_write_end);
2202
2203 /*
2204  * block_is_partially_uptodate checks whether buffers within a page are
2205  * uptodate or not.
2206  *
2207  * Returns true if all buffers which correspond to a file portion
2208  * we want to read are uptodate.
2209  */
2210 int block_is_partially_uptodate(struct page *page, unsigned long from,
2211                                         unsigned long count)
2212 {
2213         unsigned block_start, block_end, blocksize;
2214         unsigned to;
2215         struct buffer_head *bh, *head;
2216         int ret = 1;
2217
2218         if (!page_has_buffers(page))
2219                 return 0;
2220
2221         head = page_buffers(page);
2222         blocksize = head->b_size;
2223         to = min_t(unsigned, PAGE_SIZE - from, count);
2224         to = from + to;
2225         if (from < blocksize && to > PAGE_SIZE - blocksize)
2226                 return 0;
2227
2228         bh = head;
2229         block_start = 0;
2230         do {
2231                 block_end = block_start + blocksize;
2232                 if (block_end > from && block_start < to) {
2233                         if (!buffer_uptodate(bh)) {
2234                                 ret = 0;
2235                                 break;
2236                         }
2237                         if (block_end >= to)
2238                                 break;
2239                 }
2240                 block_start = block_end;
2241                 bh = bh->b_this_page;
2242         } while (bh != head);
2243
2244         return ret;
2245 }
2246 EXPORT_SYMBOL(block_is_partially_uptodate);
2247
2248 /*
2249  * Generic "read page" function for block devices that have the normal
2250  * get_block functionality. This is most of the block device filesystems.
2251  * Reads the page asynchronously --- the unlock_buffer() and
2252  * set/clear_buffer_uptodate() functions propagate buffer state into the
2253  * page struct once IO has completed.
2254  */
2255 int block_read_full_page(struct page *page, get_block_t *get_block)
2256 {
2257         struct inode *inode = page->mapping->host;
2258         sector_t iblock, lblock;
2259         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2260         unsigned int blocksize, bbits;
2261         int nr, i;
2262         int fully_mapped = 1;
2263
2264         head = create_page_buffers(page, inode, 0);
2265         blocksize = head->b_size;
2266         bbits = block_size_bits(blocksize);
2267
2268         iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2269         lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2270         bh = head;
2271         nr = 0;
2272         i = 0;
2273
2274         do {
2275                 if (buffer_uptodate(bh))
2276                         continue;
2277
2278                 if (!buffer_mapped(bh)) {
2279                         int err = 0;
2280
2281                         fully_mapped = 0;
2282                         if (iblock < lblock) {
2283                                 WARN_ON(bh->b_size != blocksize);
2284                                 err = get_block(inode, iblock, bh, 0);
2285                                 if (err)
2286                                         SetPageError(page);
2287                         }
2288                         if (!buffer_mapped(bh)) {
2289                                 zero_user(page, i * blocksize, blocksize);
2290                                 if (!err)
2291                                         set_buffer_uptodate(bh);
2292                                 continue;
2293                         }
2294                         /*
2295                          * get_block() might have updated the buffer
2296                          * synchronously
2297                          */
2298                         if (buffer_uptodate(bh))
2299                                 continue;
2300                 }
2301                 arr[nr++] = bh;
2302         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2303
2304         if (fully_mapped)
2305                 SetPageMappedToDisk(page);
2306
2307         if (!nr) {
2308                 /*
2309                  * All buffers are uptodate - we can set the page uptodate
2310                  * as well. But not if get_block() returned an error.
2311                  */
2312                 if (!PageError(page))
2313                         SetPageUptodate(page);
2314                 unlock_page(page);
2315                 return 0;
2316         }
2317
2318         /* Stage two: lock the buffers */
2319         for (i = 0; i < nr; i++) {
2320                 bh = arr[i];
2321                 lock_buffer(bh);
2322                 mark_buffer_async_read(bh);
2323         }
2324
2325         /*
2326          * Stage 3: start the IO.  Check for uptodateness
2327          * inside the buffer lock in case another process reading
2328          * the underlying blockdev brought it uptodate (the sct fix).
2329          */
2330         for (i = 0; i < nr; i++) {
2331                 bh = arr[i];
2332                 if (buffer_uptodate(bh))
2333                         end_buffer_async_read(bh, 1);
2334                 else
2335                         submit_bh(REQ_OP_READ, 0, bh);
2336         }
2337         return 0;
2338 }
2339 EXPORT_SYMBOL(block_read_full_page);
2340
2341 /* utility function for filesystems that need to do work on expanding
2342  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2343  * deal with the hole.  
2344  */
2345 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2346 {
2347         struct address_space *mapping = inode->i_mapping;
2348         struct page *page;
2349         void *fsdata;
2350         int err;
2351
2352         err = inode_newsize_ok(inode, size);
2353         if (err)
2354                 goto out;
2355
2356         err = pagecache_write_begin(NULL, mapping, size, 0,
2357                                     AOP_FLAG_CONT_EXPAND, &page, &fsdata);
2358         if (err)
2359                 goto out;
2360
2361         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2362         BUG_ON(err > 0);
2363
2364 out:
2365         return err;
2366 }
2367 EXPORT_SYMBOL(generic_cont_expand_simple);
2368
2369 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2370                             loff_t pos, loff_t *bytes)
2371 {
2372         struct inode *inode = mapping->host;
2373         unsigned int blocksize = i_blocksize(inode);
2374         struct page *page;
2375         void *fsdata;
2376         pgoff_t index, curidx;
2377         loff_t curpos;
2378         unsigned zerofrom, offset, len;
2379         int err = 0;
2380
2381         index = pos >> PAGE_SHIFT;
2382         offset = pos & ~PAGE_MASK;
2383
2384         while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2385                 zerofrom = curpos & ~PAGE_MASK;
2386                 if (zerofrom & (blocksize-1)) {
2387                         *bytes |= (blocksize-1);
2388                         (*bytes)++;
2389                 }
2390                 len = PAGE_SIZE - zerofrom;
2391
2392                 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2393                                             &page, &fsdata);
2394                 if (err)
2395                         goto out;
2396                 zero_user(page, zerofrom, len);
2397                 err = pagecache_write_end(file, mapping, curpos, len, len,
2398                                                 page, fsdata);
2399                 if (err < 0)
2400                         goto out;
2401                 BUG_ON(err != len);
2402                 err = 0;
2403
2404                 balance_dirty_pages_ratelimited(mapping);
2405
2406                 if (fatal_signal_pending(current)) {
2407                         err = -EINTR;
2408                         goto out;
2409                 }
2410         }
2411
2412         /* page covers the boundary, find the boundary offset */
2413         if (index == curidx) {
2414                 zerofrom = curpos & ~PAGE_MASK;
2415                 /* if we will expand the thing last block will be filled */
2416                 if (offset <= zerofrom) {
2417                         goto out;
2418                 }
2419                 if (zerofrom & (blocksize-1)) {
2420                         *bytes |= (blocksize-1);
2421                         (*bytes)++;
2422                 }
2423                 len = offset - zerofrom;
2424
2425                 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2426                                             &page, &fsdata);
2427                 if (err)
2428                         goto out;
2429                 zero_user(page, zerofrom, len);
2430                 err = pagecache_write_end(file, mapping, curpos, len, len,
2431                                                 page, fsdata);
2432                 if (err < 0)
2433                         goto out;
2434                 BUG_ON(err != len);
2435                 err = 0;
2436         }
2437 out:
2438         return err;
2439 }
2440
2441 /*
2442  * For moronic filesystems that do not allow holes in file.
2443  * We may have to extend the file.
2444  */
2445 int cont_write_begin(struct file *file, struct address_space *mapping,
2446                         loff_t pos, unsigned len, unsigned flags,
2447                         struct page **pagep, void **fsdata,
2448                         get_block_t *get_block, loff_t *bytes)
2449 {
2450         struct inode *inode = mapping->host;
2451         unsigned int blocksize = i_blocksize(inode);
2452         unsigned int zerofrom;
2453         int err;
2454
2455         err = cont_expand_zero(file, mapping, pos, bytes);
2456         if (err)
2457                 return err;
2458
2459         zerofrom = *bytes & ~PAGE_MASK;
2460         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2461                 *bytes |= (blocksize-1);
2462                 (*bytes)++;
2463         }
2464
2465         return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2466 }
2467 EXPORT_SYMBOL(cont_write_begin);
2468
2469 int block_commit_write(struct page *page, unsigned from, unsigned to)
2470 {
2471         struct inode *inode = page->mapping->host;
2472         __block_commit_write(inode,page,from,to);
2473         return 0;
2474 }
2475 EXPORT_SYMBOL(block_commit_write);
2476
2477 /*
2478  * block_page_mkwrite() is not allowed to change the file size as it gets
2479  * called from a page fault handler when a page is first dirtied. Hence we must
2480  * be careful to check for EOF conditions here. We set the page up correctly
2481  * for a written page which means we get ENOSPC checking when writing into
2482  * holes and correct delalloc and unwritten extent mapping on filesystems that
2483  * support these features.
2484  *
2485  * We are not allowed to take the i_mutex here so we have to play games to
2486  * protect against truncate races as the page could now be beyond EOF.  Because
2487  * truncate writes the inode size before removing pages, once we have the
2488  * page lock we can determine safely if the page is beyond EOF. If it is not
2489  * beyond EOF, then the page is guaranteed safe against truncation until we
2490  * unlock the page.
2491  *
2492  * Direct callers of this function should protect against filesystem freezing
2493  * using sb_start_pagefault() - sb_end_pagefault() functions.
2494  */
2495 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2496                          get_block_t get_block)
2497 {
2498         struct page *page = vmf->page;
2499         struct inode *inode = file_inode(vma->vm_file);
2500         unsigned long end;
2501         loff_t size;
2502         int ret;
2503
2504         lock_page(page);
2505         size = i_size_read(inode);
2506         if ((page->mapping != inode->i_mapping) ||
2507             (page_offset(page) > size)) {
2508                 /* We overload EFAULT to mean page got truncated */
2509                 ret = -EFAULT;
2510                 goto out_unlock;
2511         }
2512
2513         /* page is wholly or partially inside EOF */
2514         if (((page->index + 1) << PAGE_SHIFT) > size)
2515                 end = size & ~PAGE_MASK;
2516         else
2517                 end = PAGE_SIZE;
2518
2519         ret = __block_write_begin(page, 0, end, get_block);
2520         if (!ret)
2521                 ret = block_commit_write(page, 0, end);
2522
2523         if (unlikely(ret < 0))
2524                 goto out_unlock;
2525         set_page_dirty(page);
2526         wait_for_stable_page(page);
2527         return 0;
2528 out_unlock:
2529         unlock_page(page);
2530         return ret;
2531 }
2532 EXPORT_SYMBOL(block_page_mkwrite);
2533
2534 /*
2535  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2536  * immediately, while under the page lock.  So it needs a special end_io
2537  * handler which does not touch the bh after unlocking it.
2538  */
2539 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2540 {
2541         __end_buffer_read_notouch(bh, uptodate);
2542 }
2543
2544 /*
2545  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2546  * the page (converting it to circular linked list and taking care of page
2547  * dirty races).
2548  */
2549 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2550 {
2551         struct buffer_head *bh;
2552
2553         BUG_ON(!PageLocked(page));
2554
2555         spin_lock(&page->mapping->private_lock);
2556         bh = head;
2557         do {
2558                 if (PageDirty(page))
2559                         set_buffer_dirty(bh);
2560                 if (!bh->b_this_page)
2561                         bh->b_this_page = head;
2562                 bh = bh->b_this_page;
2563         } while (bh != head);
2564         attach_page_private(page, head);
2565         spin_unlock(&page->mapping->private_lock);
2566 }
2567
2568 /*
2569  * On entry, the page is fully not uptodate.
2570  * On exit the page is fully uptodate in the areas outside (from,to)
2571  * The filesystem needs to handle block truncation upon failure.
2572  */
2573 int nobh_write_begin(struct address_space *mapping,
2574                         loff_t pos, unsigned len, unsigned flags,
2575                         struct page **pagep, void **fsdata,
2576                         get_block_t *get_block)
2577 {
2578         struct inode *inode = mapping->host;
2579         const unsigned blkbits = inode->i_blkbits;
2580         const unsigned blocksize = 1 << blkbits;
2581         struct buffer_head *head, *bh;
2582         struct page *page;
2583         pgoff_t index;
2584         unsigned from, to;
2585         unsigned block_in_page;
2586         unsigned block_start, block_end;
2587         sector_t block_in_file;
2588         int nr_reads = 0;
2589         int ret = 0;
2590         int is_mapped_to_disk = 1;
2591
2592         index = pos >> PAGE_SHIFT;
2593         from = pos & (PAGE_SIZE - 1);
2594         to = from + len;
2595
2596         page = grab_cache_page_write_begin(mapping, index, flags);
2597         if (!page)
2598                 return -ENOMEM;
2599         *pagep = page;
2600         *fsdata = NULL;
2601
2602         if (page_has_buffers(page)) {
2603                 ret = __block_write_begin(page, pos, len, get_block);
2604                 if (unlikely(ret))
2605                         goto out_release;
2606                 return ret;
2607         }
2608
2609         if (PageMappedToDisk(page))
2610                 return 0;
2611
2612         /*
2613          * Allocate buffers so that we can keep track of state, and potentially
2614          * attach them to the page if an error occurs. In the common case of
2615          * no error, they will just be freed again without ever being attached
2616          * to the page (which is all OK, because we're under the page lock).
2617          *
2618          * Be careful: the buffer linked list is a NULL terminated one, rather
2619          * than the circular one we're used to.
2620          */
2621         head = alloc_page_buffers(page, blocksize, false);
2622         if (!head) {
2623                 ret = -ENOMEM;
2624                 goto out_release;
2625         }
2626
2627         block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2628
2629         /*
2630          * We loop across all blocks in the page, whether or not they are
2631          * part of the affected region.  This is so we can discover if the
2632          * page is fully mapped-to-disk.
2633          */
2634         for (block_start = 0, block_in_page = 0, bh = head;
2635                   block_start < PAGE_SIZE;
2636                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2637                 int create;
2638
2639                 block_end = block_start + blocksize;
2640                 bh->b_state = 0;
2641                 create = 1;
2642                 if (block_start >= to)
2643                         create = 0;
2644                 ret = get_block(inode, block_in_file + block_in_page,
2645                                         bh, create);
2646                 if (ret)
2647                         goto failed;
2648                 if (!buffer_mapped(bh))
2649                         is_mapped_to_disk = 0;
2650                 if (buffer_new(bh))
2651                         clean_bdev_bh_alias(bh);
2652                 if (PageUptodate(page)) {
2653                         set_buffer_uptodate(bh);
2654                         continue;
2655                 }
2656                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2657                         zero_user_segments(page, block_start, from,
2658                                                         to, block_end);
2659                         continue;
2660                 }
2661                 if (buffer_uptodate(bh))
2662                         continue;       /* reiserfs does this */
2663                 if (block_start < from || block_end > to) {
2664                         lock_buffer(bh);
2665                         bh->b_end_io = end_buffer_read_nobh;
2666                         submit_bh(REQ_OP_READ, 0, bh);
2667                         nr_reads++;
2668                 }
2669         }
2670
2671         if (nr_reads) {
2672                 /*
2673                  * The page is locked, so these buffers are protected from
2674                  * any VM or truncate activity.  Hence we don't need to care
2675                  * for the buffer_head refcounts.
2676                  */
2677                 for (bh = head; bh; bh = bh->b_this_page) {
2678                         wait_on_buffer(bh);
2679                         if (!buffer_uptodate(bh))
2680                                 ret = -EIO;
2681                 }
2682                 if (ret)
2683                         goto failed;
2684         }
2685
2686         if (is_mapped_to_disk)
2687                 SetPageMappedToDisk(page);
2688
2689         *fsdata = head; /* to be released by nobh_write_end */
2690
2691         return 0;
2692
2693 failed:
2694         BUG_ON(!ret);
2695         /*
2696          * Error recovery is a bit difficult. We need to zero out blocks that
2697          * were newly allocated, and dirty them to ensure they get written out.
2698          * Buffers need to be attached to the page at this point, otherwise
2699          * the handling of potential IO errors during writeout would be hard
2700          * (could try doing synchronous writeout, but what if that fails too?)
2701          */
2702         attach_nobh_buffers(page, head);
2703         page_zero_new_buffers(page, from, to);
2704
2705 out_release:
2706         unlock_page(page);
2707         put_page(page);
2708         *pagep = NULL;
2709
2710         return ret;
2711 }
2712 EXPORT_SYMBOL(nobh_write_begin);
2713
2714 int nobh_write_end(struct file *file, struct address_space *mapping,
2715                         loff_t pos, unsigned len, unsigned copied,
2716                         struct page *page, void *fsdata)
2717 {
2718         struct inode *inode = page->mapping->host;
2719         struct buffer_head *head = fsdata;
2720         struct buffer_head *bh;
2721         BUG_ON(fsdata != NULL && page_has_buffers(page));
2722
2723         if (unlikely(copied < len) && head)
2724                 attach_nobh_buffers(page, head);
2725         if (page_has_buffers(page))
2726                 return generic_write_end(file, mapping, pos, len,
2727                                         copied, page, fsdata);
2728
2729         SetPageUptodate(page);
2730         set_page_dirty(page);
2731         if (pos+copied > inode->i_size) {
2732                 i_size_write(inode, pos+copied);
2733                 mark_inode_dirty(inode);
2734         }
2735
2736         unlock_page(page);
2737         put_page(page);
2738
2739         while (head) {
2740                 bh = head;
2741                 head = head->b_this_page;
2742                 free_buffer_head(bh);
2743         }
2744
2745         return copied;
2746 }
2747 EXPORT_SYMBOL(nobh_write_end);
2748
2749 /*
2750  * nobh_writepage() - based on block_full_write_page() except
2751  * that it tries to operate without attaching bufferheads to
2752  * the page.
2753  */
2754 int nobh_writepage(struct page *page, get_block_t *get_block,
2755                         struct writeback_control *wbc)
2756 {
2757         struct inode * const inode = page->mapping->host;
2758         loff_t i_size = i_size_read(inode);
2759         const pgoff_t end_index = i_size >> PAGE_SHIFT;
2760         unsigned offset;
2761         int ret;
2762
2763         /* Is the page fully inside i_size? */
2764         if (page->index < end_index)
2765                 goto out;
2766
2767         /* Is the page fully outside i_size? (truncate in progress) */
2768         offset = i_size & (PAGE_SIZE-1);
2769         if (page->index >= end_index+1 || !offset) {
2770                 unlock_page(page);
2771                 return 0; /* don't care */
2772         }
2773
2774         /*
2775          * The page straddles i_size.  It must be zeroed out on each and every
2776          * writepage invocation because it may be mmapped.  "A file is mapped
2777          * in multiples of the page size.  For a file that is not a multiple of
2778          * the  page size, the remaining memory is zeroed when mapped, and
2779          * writes to that region are not written out to the file."
2780          */
2781         zero_user_segment(page, offset, PAGE_SIZE);
2782 out:
2783         ret = mpage_writepage(page, get_block, wbc);
2784         if (ret == -EAGAIN)
2785                 ret = __block_write_full_page(inode, page, get_block, wbc,
2786                                               end_buffer_async_write);
2787         return ret;
2788 }
2789 EXPORT_SYMBOL(nobh_writepage);
2790
2791 int nobh_truncate_page(struct address_space *mapping,
2792                         loff_t from, get_block_t *get_block)
2793 {
2794         pgoff_t index = from >> PAGE_SHIFT;
2795         unsigned offset = from & (PAGE_SIZE-1);
2796         unsigned blocksize;
2797         sector_t iblock;
2798         unsigned length, pos;
2799         struct inode *inode = mapping->host;
2800         struct page *page;
2801         struct buffer_head map_bh;
2802         int err;
2803
2804         blocksize = i_blocksize(inode);
2805         length = offset & (blocksize - 1);
2806
2807         /* Block boundary? Nothing to do */
2808         if (!length)
2809                 return 0;
2810
2811         length = blocksize - length;
2812         iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2813
2814         page = grab_cache_page(mapping, index);
2815         err = -ENOMEM;
2816         if (!page)
2817                 goto out;
2818
2819         if (page_has_buffers(page)) {
2820 has_buffers:
2821                 unlock_page(page);
2822                 put_page(page);
2823                 return block_truncate_page(mapping, from, get_block);
2824         }
2825
2826         /* Find the buffer that contains "offset" */
2827         pos = blocksize;
2828         while (offset >= pos) {
2829                 iblock++;
2830                 pos += blocksize;
2831         }
2832
2833         map_bh.b_size = blocksize;
2834         map_bh.b_state = 0;
2835         err = get_block(inode, iblock, &map_bh, 0);
2836         if (err)
2837                 goto unlock;
2838         /* unmapped? It's a hole - nothing to do */
2839         if (!buffer_mapped(&map_bh))
2840                 goto unlock;
2841
2842         /* Ok, it's mapped. Make sure it's up-to-date */
2843         if (!PageUptodate(page)) {
2844                 err = mapping->a_ops->readpage(NULL, page);
2845                 if (err) {
2846                         put_page(page);
2847                         goto out;
2848                 }
2849                 lock_page(page);
2850                 if (!PageUptodate(page)) {
2851                         err = -EIO;
2852                         goto unlock;
2853                 }
2854                 if (page_has_buffers(page))
2855                         goto has_buffers;
2856         }
2857         zero_user(page, offset, length);
2858         set_page_dirty(page);
2859         err = 0;
2860
2861 unlock:
2862         unlock_page(page);
2863         put_page(page);
2864 out:
2865         return err;
2866 }
2867 EXPORT_SYMBOL(nobh_truncate_page);
2868
2869 int block_truncate_page(struct address_space *mapping,
2870                         loff_t from, get_block_t *get_block)
2871 {
2872         pgoff_t index = from >> PAGE_SHIFT;
2873         unsigned offset = from & (PAGE_SIZE-1);
2874         unsigned blocksize;
2875         sector_t iblock;
2876         unsigned length, pos;
2877         struct inode *inode = mapping->host;
2878         struct page *page;
2879         struct buffer_head *bh;
2880         int err;
2881
2882         blocksize = i_blocksize(inode);
2883         length = offset & (blocksize - 1);
2884
2885         /* Block boundary? Nothing to do */
2886         if (!length)
2887                 return 0;
2888
2889         length = blocksize - length;
2890         iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2891         
2892         page = grab_cache_page(mapping, index);
2893         err = -ENOMEM;
2894         if (!page)
2895                 goto out;
2896
2897         if (!page_has_buffers(page))
2898                 create_empty_buffers(page, blocksize, 0);
2899
2900         /* Find the buffer that contains "offset" */
2901         bh = page_buffers(page);
2902         pos = blocksize;
2903         while (offset >= pos) {
2904                 bh = bh->b_this_page;
2905                 iblock++;
2906                 pos += blocksize;
2907         }
2908
2909         err = 0;
2910         if (!buffer_mapped(bh)) {
2911                 WARN_ON(bh->b_size != blocksize);
2912                 err = get_block(inode, iblock, bh, 0);
2913                 if (err)
2914                         goto unlock;
2915                 /* unmapped? It's a hole - nothing to do */
2916                 if (!buffer_mapped(bh))
2917                         goto unlock;
2918         }
2919
2920         /* Ok, it's mapped. Make sure it's up-to-date */
2921         if (PageUptodate(page))
2922                 set_buffer_uptodate(bh);
2923
2924         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2925                 err = -EIO;
2926                 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2927                 wait_on_buffer(bh);
2928                 /* Uhhuh. Read error. Complain and punt. */
2929                 if (!buffer_uptodate(bh))
2930                         goto unlock;
2931         }
2932
2933         zero_user(page, offset, length);
2934         mark_buffer_dirty(bh);
2935         err = 0;
2936
2937 unlock:
2938         unlock_page(page);
2939         put_page(page);
2940 out:
2941         return err;
2942 }
2943 EXPORT_SYMBOL(block_truncate_page);
2944
2945 /*
2946  * The generic ->writepage function for buffer-backed address_spaces
2947  */
2948 int block_write_full_page(struct page *page, get_block_t *get_block,
2949                         struct writeback_control *wbc)
2950 {
2951         struct inode * const inode = page->mapping->host;
2952         loff_t i_size = i_size_read(inode);
2953         const pgoff_t end_index = i_size >> PAGE_SHIFT;
2954         unsigned offset;
2955
2956         /* Is the page fully inside i_size? */
2957         if (page->index < end_index)
2958                 return __block_write_full_page(inode, page, get_block, wbc,
2959                                                end_buffer_async_write);
2960
2961         /* Is the page fully outside i_size? (truncate in progress) */
2962         offset = i_size & (PAGE_SIZE-1);
2963         if (page->index >= end_index+1 || !offset) {
2964                 unlock_page(page);
2965                 return 0; /* don't care */
2966         }
2967
2968         /*
2969          * The page straddles i_size.  It must be zeroed out on each and every
2970          * writepage invocation because it may be mmapped.  "A file is mapped
2971          * in multiples of the page size.  For a file that is not a multiple of
2972          * the  page size, the remaining memory is zeroed when mapped, and
2973          * writes to that region are not written out to the file."
2974          */
2975         zero_user_segment(page, offset, PAGE_SIZE);
2976         return __block_write_full_page(inode, page, get_block, wbc,
2977                                                         end_buffer_async_write);
2978 }
2979 EXPORT_SYMBOL(block_write_full_page);
2980
2981 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2982                             get_block_t *get_block)
2983 {
2984         struct inode *inode = mapping->host;
2985         struct buffer_head tmp = {
2986                 .b_size = i_blocksize(inode),
2987         };
2988
2989         get_block(inode, block, &tmp, 0);
2990         return tmp.b_blocknr;
2991 }
2992 EXPORT_SYMBOL(generic_block_bmap);
2993
2994 static void end_bio_bh_io_sync(struct bio *bio)
2995 {
2996         struct buffer_head *bh = bio->bi_private;
2997
2998         if (unlikely(bio_flagged(bio, BIO_QUIET)))
2999                 set_bit(BH_Quiet, &bh->b_state);
3000
3001         bh->b_end_io(bh, !bio->bi_status);
3002         bio_put(bio);
3003 }
3004
3005 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3006                          enum rw_hint write_hint, struct writeback_control *wbc)
3007 {
3008         struct bio *bio;
3009
3010         BUG_ON(!buffer_locked(bh));
3011         BUG_ON(!buffer_mapped(bh));
3012         BUG_ON(!bh->b_end_io);
3013         BUG_ON(buffer_delay(bh));
3014         BUG_ON(buffer_unwritten(bh));
3015
3016         /*
3017          * Only clear out a write error when rewriting
3018          */
3019         if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3020                 clear_buffer_write_io_error(bh);
3021
3022         bio = bio_alloc(GFP_NOIO, 1);
3023
3024         fscrypt_set_bio_crypt_ctx_bh(bio, bh, GFP_NOIO);
3025
3026         bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3027         bio_set_dev(bio, bh->b_bdev);
3028         bio->bi_write_hint = write_hint;
3029
3030         bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3031         BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3032
3033         bio->bi_end_io = end_bio_bh_io_sync;
3034         bio->bi_private = bh;
3035
3036         if (buffer_meta(bh))
3037                 op_flags |= REQ_META;
3038         if (buffer_prio(bh))
3039                 op_flags |= REQ_PRIO;
3040         bio_set_op_attrs(bio, op, op_flags);
3041
3042         /* Take care of bh's that straddle the end of the device */
3043         guard_bio_eod(bio);
3044
3045         if (wbc) {
3046                 wbc_init_bio(wbc, bio);
3047                 wbc_account_cgroup_owner(wbc, bh->b_page, bh->b_size);
3048         }
3049
3050         submit_bio(bio);
3051         return 0;
3052 }
3053
3054 int submit_bh(int op, int op_flags, struct buffer_head *bh)
3055 {
3056         return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3057 }
3058 EXPORT_SYMBOL(submit_bh);
3059
3060 /**
3061  * ll_rw_block: low-level access to block devices (DEPRECATED)
3062  * @op: whether to %READ or %WRITE
3063  * @op_flags: req_flag_bits
3064  * @nr: number of &struct buffer_heads in the array
3065  * @bhs: array of pointers to &struct buffer_head
3066  *
3067  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3068  * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3069  * @op_flags contains flags modifying the detailed I/O behavior, most notably
3070  * %REQ_RAHEAD.
3071  *
3072  * This function drops any buffer that it cannot get a lock on (with the
3073  * BH_Lock state bit), any buffer that appears to be clean when doing a write
3074  * request, and any buffer that appears to be up-to-date when doing read
3075  * request.  Further it marks as clean buffers that are processed for
3076  * writing (the buffer cache won't assume that they are actually clean
3077  * until the buffer gets unlocked).
3078  *
3079  * ll_rw_block sets b_end_io to simple completion handler that marks
3080  * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3081  * any waiters. 
3082  *
3083  * All of the buffers must be for the same device, and must also be a
3084  * multiple of the current approved size for the device.
3085  */
3086 void ll_rw_block(int op, int op_flags,  int nr, struct buffer_head *bhs[])
3087 {
3088         int i;
3089
3090         for (i = 0; i < nr; i++) {
3091                 struct buffer_head *bh = bhs[i];
3092
3093                 if (!trylock_buffer(bh))
3094                         continue;
3095                 if (op == WRITE) {
3096                         if (test_clear_buffer_dirty(bh)) {
3097                                 bh->b_end_io = end_buffer_write_sync;
3098                                 get_bh(bh);
3099                                 submit_bh(op, op_flags, bh);
3100                                 continue;
3101                         }
3102                 } else {
3103                         if (!buffer_uptodate(bh)) {
3104                                 bh->b_end_io = end_buffer_read_sync;
3105                                 get_bh(bh);
3106                                 submit_bh(op, op_flags, bh);
3107                                 continue;
3108                         }
3109                 }
3110                 unlock_buffer(bh);
3111         }
3112 }
3113 EXPORT_SYMBOL(ll_rw_block);
3114
3115 void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3116 {
3117         lock_buffer(bh);
3118         if (!test_clear_buffer_dirty(bh)) {
3119                 unlock_buffer(bh);
3120                 return;
3121         }
3122         bh->b_end_io = end_buffer_write_sync;
3123         get_bh(bh);
3124         submit_bh(REQ_OP_WRITE, op_flags, bh);
3125 }
3126 EXPORT_SYMBOL(write_dirty_buffer);
3127
3128 /*
3129  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3130  * and then start new I/O and then wait upon it.  The caller must have a ref on
3131  * the buffer_head.
3132  */
3133 int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3134 {
3135         int ret = 0;
3136
3137         WARN_ON(atomic_read(&bh->b_count) < 1);
3138         lock_buffer(bh);
3139         if (test_clear_buffer_dirty(bh)) {
3140                 /*
3141                  * The bh should be mapped, but it might not be if the
3142                  * device was hot-removed. Not much we can do but fail the I/O.
3143                  */
3144                 if (!buffer_mapped(bh)) {
3145                         unlock_buffer(bh);
3146                         return -EIO;
3147                 }
3148
3149                 get_bh(bh);
3150                 bh->b_end_io = end_buffer_write_sync;
3151                 ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3152                 wait_on_buffer(bh);
3153                 if (!ret && !buffer_uptodate(bh))
3154                         ret = -EIO;
3155         } else {
3156                 unlock_buffer(bh);
3157         }
3158         return ret;
3159 }
3160 EXPORT_SYMBOL(__sync_dirty_buffer);
3161
3162 int sync_dirty_buffer(struct buffer_head *bh)
3163 {
3164         return __sync_dirty_buffer(bh, REQ_SYNC);
3165 }
3166 EXPORT_SYMBOL(sync_dirty_buffer);
3167
3168 /*
3169  * try_to_free_buffers() checks if all the buffers on this particular page
3170  * are unused, and releases them if so.
3171  *
3172  * Exclusion against try_to_free_buffers may be obtained by either
3173  * locking the page or by holding its mapping's private_lock.
3174  *
3175  * If the page is dirty but all the buffers are clean then we need to
3176  * be sure to mark the page clean as well.  This is because the page
3177  * may be against a block device, and a later reattachment of buffers
3178  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3179  * filesystem data on the same device.
3180  *
3181  * The same applies to regular filesystem pages: if all the buffers are
3182  * clean then we set the page clean and proceed.  To do that, we require
3183  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3184  * private_lock.
3185  *
3186  * try_to_free_buffers() is non-blocking.
3187  */
3188 static inline int buffer_busy(struct buffer_head *bh)
3189 {
3190         return atomic_read(&bh->b_count) |
3191                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3192 }
3193
3194 static int
3195 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3196 {
3197         struct buffer_head *head = page_buffers(page);
3198         struct buffer_head *bh;
3199
3200         bh = head;
3201         do {
3202                 if (buffer_busy(bh))
3203                         goto failed;
3204                 bh = bh->b_this_page;
3205         } while (bh != head);
3206
3207         do {
3208                 struct buffer_head *next = bh->b_this_page;
3209
3210                 if (bh->b_assoc_map)
3211                         __remove_assoc_queue(bh);
3212                 bh = next;
3213         } while (bh != head);
3214         *buffers_to_free = head;
3215         detach_page_private(page);
3216         return 1;
3217 failed:
3218         return 0;
3219 }
3220
3221 int try_to_free_buffers(struct page *page)
3222 {
3223         struct address_space * const mapping = page->mapping;
3224         struct buffer_head *buffers_to_free = NULL;
3225         int ret = 0;
3226
3227         BUG_ON(!PageLocked(page));
3228         if (PageWriteback(page))
3229                 return 0;
3230
3231         if (mapping == NULL) {          /* can this still happen? */
3232                 ret = drop_buffers(page, &buffers_to_free);
3233                 goto out;
3234         }
3235
3236         spin_lock(&mapping->private_lock);
3237         ret = drop_buffers(page, &buffers_to_free);
3238
3239         /*
3240          * If the filesystem writes its buffers by hand (eg ext3)
3241          * then we can have clean buffers against a dirty page.  We
3242          * clean the page here; otherwise the VM will never notice
3243          * that the filesystem did any IO at all.
3244          *
3245          * Also, during truncate, discard_buffer will have marked all
3246          * the page's buffers clean.  We discover that here and clean
3247          * the page also.
3248          *
3249          * private_lock must be held over this entire operation in order
3250          * to synchronise against __set_page_dirty_buffers and prevent the
3251          * dirty bit from being lost.
3252          */
3253         if (ret)
3254                 cancel_dirty_page(page);
3255         spin_unlock(&mapping->private_lock);
3256 out:
3257         if (buffers_to_free) {
3258                 struct buffer_head *bh = buffers_to_free;
3259
3260                 do {
3261                         struct buffer_head *next = bh->b_this_page;
3262                         free_buffer_head(bh);
3263                         bh = next;
3264                 } while (bh != buffers_to_free);
3265         }
3266         return ret;
3267 }
3268 EXPORT_SYMBOL(try_to_free_buffers);
3269
3270 /*
3271  * Buffer-head allocation
3272  */
3273 static struct kmem_cache *bh_cachep __read_mostly;
3274
3275 /*
3276  * Once the number of bh's in the machine exceeds this level, we start
3277  * stripping them in writeback.
3278  */
3279 static unsigned long max_buffer_heads;
3280
3281 int buffer_heads_over_limit;
3282
3283 struct bh_accounting {
3284         int nr;                 /* Number of live bh's */
3285         int ratelimit;          /* Limit cacheline bouncing */
3286 };
3287
3288 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3289
3290 static void recalc_bh_state(void)
3291 {
3292         int i;
3293         int tot = 0;
3294
3295         if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3296                 return;
3297         __this_cpu_write(bh_accounting.ratelimit, 0);
3298         for_each_online_cpu(i)
3299                 tot += per_cpu(bh_accounting, i).nr;
3300         buffer_heads_over_limit = (tot > max_buffer_heads);
3301 }
3302
3303 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3304 {
3305         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3306         if (ret) {
3307                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3308                 spin_lock_init(&ret->b_uptodate_lock);
3309                 preempt_disable();
3310                 __this_cpu_inc(bh_accounting.nr);
3311                 recalc_bh_state();
3312                 preempt_enable();
3313         }
3314         return ret;
3315 }
3316 EXPORT_SYMBOL(alloc_buffer_head);
3317
3318 void free_buffer_head(struct buffer_head *bh)
3319 {
3320         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3321         kmem_cache_free(bh_cachep, bh);
3322         preempt_disable();
3323         __this_cpu_dec(bh_accounting.nr);
3324         recalc_bh_state();
3325         preempt_enable();
3326 }
3327 EXPORT_SYMBOL(free_buffer_head);
3328
3329 static int buffer_exit_cpu_dead(unsigned int cpu)
3330 {
3331         int i;
3332         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3333
3334         for (i = 0; i < BH_LRU_SIZE; i++) {
3335                 brelse(b->bhs[i]);
3336                 b->bhs[i] = NULL;
3337         }
3338         this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3339         per_cpu(bh_accounting, cpu).nr = 0;
3340         return 0;
3341 }
3342
3343 /**
3344  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3345  * @bh: struct buffer_head
3346  *
3347  * Return true if the buffer is up-to-date and false,
3348  * with the buffer locked, if not.
3349  */
3350 int bh_uptodate_or_lock(struct buffer_head *bh)
3351 {
3352         if (!buffer_uptodate(bh)) {
3353                 lock_buffer(bh);
3354                 if (!buffer_uptodate(bh))
3355                         return 0;
3356                 unlock_buffer(bh);
3357         }
3358         return 1;
3359 }
3360 EXPORT_SYMBOL(bh_uptodate_or_lock);
3361
3362 /**
3363  * bh_submit_read - Submit a locked buffer for reading
3364  * @bh: struct buffer_head
3365  *
3366  * Returns zero on success and -EIO on error.
3367  */
3368 int bh_submit_read(struct buffer_head *bh)
3369 {
3370         BUG_ON(!buffer_locked(bh));
3371
3372         if (buffer_uptodate(bh)) {
3373                 unlock_buffer(bh);
3374                 return 0;
3375         }
3376
3377         get_bh(bh);
3378         bh->b_end_io = end_buffer_read_sync;
3379         submit_bh(REQ_OP_READ, 0, bh);
3380         wait_on_buffer(bh);
3381         if (buffer_uptodate(bh))
3382                 return 0;
3383         return -EIO;
3384 }
3385 EXPORT_SYMBOL(bh_submit_read);
3386
3387 void __init buffer_init(void)
3388 {
3389         unsigned long nrpages;
3390         int ret;
3391
3392         bh_cachep = kmem_cache_create("buffer_head",
3393                         sizeof(struct buffer_head), 0,
3394                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3395                                 SLAB_MEM_SPREAD),
3396                                 NULL);
3397
3398         /*
3399          * Limit the bh occupancy to 10% of ZONE_NORMAL
3400          */
3401         nrpages = (nr_free_buffer_pages() * 10) / 100;
3402         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3403         ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3404                                         NULL, buffer_exit_cpu_dead);
3405         WARN_ON(ret < 0);
3406 }