Merge branch 'ucount-fixes-for-v5.15' of git://git.kernel.org/pub/scm/linux/kernel...
[linux-2.6-microblaze.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 /*
1429  * It's called from workqueue context so we need a bh_lru_lock to close
1430  * the race with preemption/irq.
1431  */
1432 void invalidate_bh_lrus_cpu(void)
1433 {
1434         struct bh_lru *b;
1435
1436         bh_lru_lock();
1437         b = this_cpu_ptr(&bh_lrus);
1438         __invalidate_bh_lrus(b);
1439         bh_lru_unlock();
1440 }
1441
1442 void set_bh_page(struct buffer_head *bh,
1443                 struct page *page, unsigned long offset)
1444 {
1445         bh->b_page = page;
1446         BUG_ON(offset >= PAGE_SIZE);
1447         if (PageHighMem(page))
1448                 /*
1449                  * This catches illegal uses and preserves the offset:
1450                  */
1451                 bh->b_data = (char *)(0 + offset);
1452         else
1453                 bh->b_data = page_address(page) + offset;
1454 }
1455 EXPORT_SYMBOL(set_bh_page);
1456
1457 /*
1458  * Called when truncating a buffer on a page completely.
1459  */
1460
1461 /* Bits that are cleared during an invalidate */
1462 #define BUFFER_FLAGS_DISCARD \
1463         (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1464          1 << BH_Delay | 1 << BH_Unwritten)
1465
1466 static void discard_buffer(struct buffer_head * bh)
1467 {
1468         unsigned long b_state, b_state_old;
1469
1470         lock_buffer(bh);
1471         clear_buffer_dirty(bh);
1472         bh->b_bdev = NULL;
1473         b_state = bh->b_state;
1474         for (;;) {
1475                 b_state_old = cmpxchg(&bh->b_state, b_state,
1476                                       (b_state & ~BUFFER_FLAGS_DISCARD));
1477                 if (b_state_old == b_state)
1478                         break;
1479                 b_state = b_state_old;
1480         }
1481         unlock_buffer(bh);
1482 }
1483
1484 /**
1485  * block_invalidatepage - invalidate part or all of a buffer-backed page
1486  *
1487  * @page: the page which is affected
1488  * @offset: start of the range to invalidate
1489  * @length: length of the range to invalidate
1490  *
1491  * block_invalidatepage() is called when all or part of the page has become
1492  * invalidated by a truncate operation.
1493  *
1494  * block_invalidatepage() does not have to release all buffers, but it must
1495  * ensure that no dirty buffer is left outside @offset and that no I/O
1496  * is underway against any of the blocks which are outside the truncation
1497  * point.  Because the caller is about to free (and possibly reuse) those
1498  * blocks on-disk.
1499  */
1500 void block_invalidatepage(struct page *page, unsigned int offset,
1501                           unsigned int length)
1502 {
1503         struct buffer_head *head, *bh, *next;
1504         unsigned int curr_off = 0;
1505         unsigned int stop = length + offset;
1506
1507         BUG_ON(!PageLocked(page));
1508         if (!page_has_buffers(page))
1509                 goto out;
1510
1511         /*
1512          * Check for overflow
1513          */
1514         BUG_ON(stop > PAGE_SIZE || stop < length);
1515
1516         head = page_buffers(page);
1517         bh = head;
1518         do {
1519                 unsigned int next_off = curr_off + bh->b_size;
1520                 next = bh->b_this_page;
1521
1522                 /*
1523                  * Are we still fully in range ?
1524                  */
1525                 if (next_off > stop)
1526                         goto out;
1527
1528                 /*
1529                  * is this block fully invalidated?
1530                  */
1531                 if (offset <= curr_off)
1532                         discard_buffer(bh);
1533                 curr_off = next_off;
1534                 bh = next;
1535         } while (bh != head);
1536
1537         /*
1538          * We release buffers only if the entire page is being invalidated.
1539          * The get_block cached value has been unconditionally invalidated,
1540          * so real IO is not possible anymore.
1541          */
1542         if (length == PAGE_SIZE)
1543                 try_to_release_page(page, 0);
1544 out:
1545         return;
1546 }
1547 EXPORT_SYMBOL(block_invalidatepage);
1548
1549
1550 /*
1551  * We attach and possibly dirty the buffers atomically wrt
1552  * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1553  * is already excluded via the page lock.
1554  */
1555 void create_empty_buffers(struct page *page,
1556                         unsigned long blocksize, unsigned long b_state)
1557 {
1558         struct buffer_head *bh, *head, *tail;
1559
1560         head = alloc_page_buffers(page, blocksize, true);
1561         bh = head;
1562         do {
1563                 bh->b_state |= b_state;
1564                 tail = bh;
1565                 bh = bh->b_this_page;
1566         } while (bh);
1567         tail->b_this_page = head;
1568
1569         spin_lock(&page->mapping->private_lock);
1570         if (PageUptodate(page) || PageDirty(page)) {
1571                 bh = head;
1572                 do {
1573                         if (PageDirty(page))
1574                                 set_buffer_dirty(bh);
1575                         if (PageUptodate(page))
1576                                 set_buffer_uptodate(bh);
1577                         bh = bh->b_this_page;
1578                 } while (bh != head);
1579         }
1580         attach_page_private(page, head);
1581         spin_unlock(&page->mapping->private_lock);
1582 }
1583 EXPORT_SYMBOL(create_empty_buffers);
1584
1585 /**
1586  * clean_bdev_aliases: clean a range of buffers in block device
1587  * @bdev: Block device to clean buffers in
1588  * @block: Start of a range of blocks to clean
1589  * @len: Number of blocks to clean
1590  *
1591  * We are taking a range of blocks for data and we don't want writeback of any
1592  * buffer-cache aliases starting from return from this function and until the
1593  * moment when something will explicitly mark the buffer dirty (hopefully that
1594  * will not happen until we will free that block ;-) We don't even need to mark
1595  * it not-uptodate - nobody can expect anything from a newly allocated buffer
1596  * anyway. We used to use unmap_buffer() for such invalidation, but that was
1597  * wrong. We definitely don't want to mark the alias unmapped, for example - it
1598  * would confuse anyone who might pick it with bread() afterwards...
1599  *
1600  * Also..  Note that bforget() doesn't lock the buffer.  So there can be
1601  * writeout I/O going on against recently-freed buffers.  We don't wait on that
1602  * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1603  * need to.  That happens here.
1604  */
1605 void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1606 {
1607         struct inode *bd_inode = bdev->bd_inode;
1608         struct address_space *bd_mapping = bd_inode->i_mapping;
1609         struct pagevec pvec;
1610         pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1611         pgoff_t end;
1612         int i, count;
1613         struct buffer_head *bh;
1614         struct buffer_head *head;
1615
1616         end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1617         pagevec_init(&pvec);
1618         while (pagevec_lookup_range(&pvec, bd_mapping, &index, end)) {
1619                 count = pagevec_count(&pvec);
1620                 for (i = 0; i < count; i++) {
1621                         struct page *page = pvec.pages[i];
1622
1623                         if (!page_has_buffers(page))
1624                                 continue;
1625                         /*
1626                          * We use page lock instead of bd_mapping->private_lock
1627                          * to pin buffers here since we can afford to sleep and
1628                          * it scales better than a global spinlock lock.
1629                          */
1630                         lock_page(page);
1631                         /* Recheck when the page is locked which pins bhs */
1632                         if (!page_has_buffers(page))
1633                                 goto unlock_page;
1634                         head = page_buffers(page);
1635                         bh = head;
1636                         do {
1637                                 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1638                                         goto next;
1639                                 if (bh->b_blocknr >= block + len)
1640                                         break;
1641                                 clear_buffer_dirty(bh);
1642                                 wait_on_buffer(bh);
1643                                 clear_buffer_req(bh);
1644 next:
1645                                 bh = bh->b_this_page;
1646                         } while (bh != head);
1647 unlock_page:
1648                         unlock_page(page);
1649                 }
1650                 pagevec_release(&pvec);
1651                 cond_resched();
1652                 /* End of range already reached? */
1653                 if (index > end || !index)
1654                         break;
1655         }
1656 }
1657 EXPORT_SYMBOL(clean_bdev_aliases);
1658
1659 /*
1660  * Size is a power-of-two in the range 512..PAGE_SIZE,
1661  * and the case we care about most is PAGE_SIZE.
1662  *
1663  * So this *could* possibly be written with those
1664  * constraints in mind (relevant mostly if some
1665  * architecture has a slow bit-scan instruction)
1666  */
1667 static inline int block_size_bits(unsigned int blocksize)
1668 {
1669         return ilog2(blocksize);
1670 }
1671
1672 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1673 {
1674         BUG_ON(!PageLocked(page));
1675
1676         if (!page_has_buffers(page))
1677                 create_empty_buffers(page, 1 << READ_ONCE(inode->i_blkbits),
1678                                      b_state);
1679         return page_buffers(page);
1680 }
1681
1682 /*
1683  * NOTE! All mapped/uptodate combinations are valid:
1684  *
1685  *      Mapped  Uptodate        Meaning
1686  *
1687  *      No      No              "unknown" - must do get_block()
1688  *      No      Yes             "hole" - zero-filled
1689  *      Yes     No              "allocated" - allocated on disk, not read in
1690  *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1691  *
1692  * "Dirty" is valid only with the last case (mapped+uptodate).
1693  */
1694
1695 /*
1696  * While block_write_full_page is writing back the dirty buffers under
1697  * the page lock, whoever dirtied the buffers may decide to clean them
1698  * again at any time.  We handle that by only looking at the buffer
1699  * state inside lock_buffer().
1700  *
1701  * If block_write_full_page() is called for regular writeback
1702  * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1703  * locked buffer.   This only can happen if someone has written the buffer
1704  * directly, with submit_bh().  At the address_space level PageWriteback
1705  * prevents this contention from occurring.
1706  *
1707  * If block_write_full_page() is called with wbc->sync_mode ==
1708  * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1709  * causes the writes to be flagged as synchronous writes.
1710  */
1711 int __block_write_full_page(struct inode *inode, struct page *page,
1712                         get_block_t *get_block, struct writeback_control *wbc,
1713                         bh_end_io_t *handler)
1714 {
1715         int err;
1716         sector_t block;
1717         sector_t last_block;
1718         struct buffer_head *bh, *head;
1719         unsigned int blocksize, bbits;
1720         int nr_underway = 0;
1721         int write_flags = wbc_to_write_flags(wbc);
1722
1723         head = create_page_buffers(page, inode,
1724                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
1725
1726         /*
1727          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1728          * here, and the (potentially unmapped) buffers may become dirty at
1729          * any time.  If a buffer becomes dirty here after we've inspected it
1730          * then we just miss that fact, and the page stays dirty.
1731          *
1732          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1733          * handle that here by just cleaning them.
1734          */
1735
1736         bh = head;
1737         blocksize = bh->b_size;
1738         bbits = block_size_bits(blocksize);
1739
1740         block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1741         last_block = (i_size_read(inode) - 1) >> bbits;
1742
1743         /*
1744          * Get all the dirty buffers mapped to disk addresses and
1745          * handle any aliases from the underlying blockdev's mapping.
1746          */
1747         do {
1748                 if (block > last_block) {
1749                         /*
1750                          * mapped buffers outside i_size will occur, because
1751                          * this page can be outside i_size when there is a
1752                          * truncate in progress.
1753                          */
1754                         /*
1755                          * The buffer was zeroed by block_write_full_page()
1756                          */
1757                         clear_buffer_dirty(bh);
1758                         set_buffer_uptodate(bh);
1759                 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1760                            buffer_dirty(bh)) {
1761                         WARN_ON(bh->b_size != blocksize);
1762                         err = get_block(inode, block, bh, 1);
1763                         if (err)
1764                                 goto recover;
1765                         clear_buffer_delay(bh);
1766                         if (buffer_new(bh)) {
1767                                 /* blockdev mappings never come here */
1768                                 clear_buffer_new(bh);
1769                                 clean_bdev_bh_alias(bh);
1770                         }
1771                 }
1772                 bh = bh->b_this_page;
1773                 block++;
1774         } while (bh != head);
1775
1776         do {
1777                 if (!buffer_mapped(bh))
1778                         continue;
1779                 /*
1780                  * If it's a fully non-blocking write attempt and we cannot
1781                  * lock the buffer then redirty the page.  Note that this can
1782                  * potentially cause a busy-wait loop from writeback threads
1783                  * and kswapd activity, but those code paths have their own
1784                  * higher-level throttling.
1785                  */
1786                 if (wbc->sync_mode != WB_SYNC_NONE) {
1787                         lock_buffer(bh);
1788                 } else if (!trylock_buffer(bh)) {
1789                         redirty_page_for_writepage(wbc, page);
1790                         continue;
1791                 }
1792                 if (test_clear_buffer_dirty(bh)) {
1793                         mark_buffer_async_write_endio(bh, handler);
1794                 } else {
1795                         unlock_buffer(bh);
1796                 }
1797         } while ((bh = bh->b_this_page) != head);
1798
1799         /*
1800          * The page and its buffers are protected by PageWriteback(), so we can
1801          * drop the bh refcounts early.
1802          */
1803         BUG_ON(PageWriteback(page));
1804         set_page_writeback(page);
1805
1806         do {
1807                 struct buffer_head *next = bh->b_this_page;
1808                 if (buffer_async_write(bh)) {
1809                         submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1810                                         inode->i_write_hint, wbc);
1811                         nr_underway++;
1812                 }
1813                 bh = next;
1814         } while (bh != head);
1815         unlock_page(page);
1816
1817         err = 0;
1818 done:
1819         if (nr_underway == 0) {
1820                 /*
1821                  * The page was marked dirty, but the buffers were
1822                  * clean.  Someone wrote them back by hand with
1823                  * ll_rw_block/submit_bh.  A rare case.
1824                  */
1825                 end_page_writeback(page);
1826
1827                 /*
1828                  * The page and buffer_heads can be released at any time from
1829                  * here on.
1830                  */
1831         }
1832         return err;
1833
1834 recover:
1835         /*
1836          * ENOSPC, or some other error.  We may already have added some
1837          * blocks to the file, so we need to write these out to avoid
1838          * exposing stale data.
1839          * The page is currently locked and not marked for writeback
1840          */
1841         bh = head;
1842         /* Recovery: lock and submit the mapped buffers */
1843         do {
1844                 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1845                     !buffer_delay(bh)) {
1846                         lock_buffer(bh);
1847                         mark_buffer_async_write_endio(bh, handler);
1848                 } else {
1849                         /*
1850                          * The buffer may have been set dirty during
1851                          * attachment to a dirty page.
1852                          */
1853                         clear_buffer_dirty(bh);
1854                 }
1855         } while ((bh = bh->b_this_page) != head);
1856         SetPageError(page);
1857         BUG_ON(PageWriteback(page));
1858         mapping_set_error(page->mapping, err);
1859         set_page_writeback(page);
1860         do {
1861                 struct buffer_head *next = bh->b_this_page;
1862                 if (buffer_async_write(bh)) {
1863                         clear_buffer_dirty(bh);
1864                         submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1865                                         inode->i_write_hint, wbc);
1866                         nr_underway++;
1867                 }
1868                 bh = next;
1869         } while (bh != head);
1870         unlock_page(page);
1871         goto done;
1872 }
1873 EXPORT_SYMBOL(__block_write_full_page);
1874
1875 /*
1876  * If a page has any new buffers, zero them out here, and mark them uptodate
1877  * and dirty so they'll be written out (in order to prevent uninitialised
1878  * block data from leaking). And clear the new bit.
1879  */
1880 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1881 {
1882         unsigned int block_start, block_end;
1883         struct buffer_head *head, *bh;
1884
1885         BUG_ON(!PageLocked(page));
1886         if (!page_has_buffers(page))
1887                 return;
1888
1889         bh = head = page_buffers(page);
1890         block_start = 0;
1891         do {
1892                 block_end = block_start + bh->b_size;
1893
1894                 if (buffer_new(bh)) {
1895                         if (block_end > from && block_start < to) {
1896                                 if (!PageUptodate(page)) {
1897                                         unsigned start, size;
1898
1899                                         start = max(from, block_start);
1900                                         size = min(to, block_end) - start;
1901
1902                                         zero_user(page, start, size);
1903                                         set_buffer_uptodate(bh);
1904                                 }
1905
1906                                 clear_buffer_new(bh);
1907                                 mark_buffer_dirty(bh);
1908                         }
1909                 }
1910
1911                 block_start = block_end;
1912                 bh = bh->b_this_page;
1913         } while (bh != head);
1914 }
1915 EXPORT_SYMBOL(page_zero_new_buffers);
1916
1917 static void
1918 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1919                 const struct iomap *iomap)
1920 {
1921         loff_t offset = block << inode->i_blkbits;
1922
1923         bh->b_bdev = iomap->bdev;
1924
1925         /*
1926          * Block points to offset in file we need to map, iomap contains
1927          * the offset at which the map starts. If the map ends before the
1928          * current block, then do not map the buffer and let the caller
1929          * handle it.
1930          */
1931         BUG_ON(offset >= iomap->offset + iomap->length);
1932
1933         switch (iomap->type) {
1934         case IOMAP_HOLE:
1935                 /*
1936                  * If the buffer is not up to date or beyond the current EOF,
1937                  * we need to mark it as new to ensure sub-block zeroing is
1938                  * executed if necessary.
1939                  */
1940                 if (!buffer_uptodate(bh) ||
1941                     (offset >= i_size_read(inode)))
1942                         set_buffer_new(bh);
1943                 break;
1944         case IOMAP_DELALLOC:
1945                 if (!buffer_uptodate(bh) ||
1946                     (offset >= i_size_read(inode)))
1947                         set_buffer_new(bh);
1948                 set_buffer_uptodate(bh);
1949                 set_buffer_mapped(bh);
1950                 set_buffer_delay(bh);
1951                 break;
1952         case IOMAP_UNWRITTEN:
1953                 /*
1954                  * For unwritten regions, we always need to ensure that regions
1955                  * in the block we are not writing to are zeroed. Mark the
1956                  * buffer as new to ensure this.
1957                  */
1958                 set_buffer_new(bh);
1959                 set_buffer_unwritten(bh);
1960                 fallthrough;
1961         case IOMAP_MAPPED:
1962                 if ((iomap->flags & IOMAP_F_NEW) ||
1963                     offset >= i_size_read(inode))
1964                         set_buffer_new(bh);
1965                 bh->b_blocknr = (iomap->addr + offset - iomap->offset) >>
1966                                 inode->i_blkbits;
1967                 set_buffer_mapped(bh);
1968                 break;
1969         }
1970 }
1971
1972 int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
1973                 get_block_t *get_block, const struct iomap *iomap)
1974 {
1975         unsigned from = pos & (PAGE_SIZE - 1);
1976         unsigned to = from + len;
1977         struct inode *inode = page->mapping->host;
1978         unsigned block_start, block_end;
1979         sector_t block;
1980         int err = 0;
1981         unsigned blocksize, bbits;
1982         struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1983
1984         BUG_ON(!PageLocked(page));
1985         BUG_ON(from > PAGE_SIZE);
1986         BUG_ON(to > PAGE_SIZE);
1987         BUG_ON(from > to);
1988
1989         head = create_page_buffers(page, inode, 0);
1990         blocksize = head->b_size;
1991         bbits = block_size_bits(blocksize);
1992
1993         block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1994
1995         for(bh = head, block_start = 0; bh != head || !block_start;
1996             block++, block_start=block_end, bh = bh->b_this_page) {
1997                 block_end = block_start + blocksize;
1998                 if (block_end <= from || block_start >= to) {
1999                         if (PageUptodate(page)) {
2000                                 if (!buffer_uptodate(bh))
2001                                         set_buffer_uptodate(bh);
2002                         }
2003                         continue;
2004                 }
2005                 if (buffer_new(bh))
2006                         clear_buffer_new(bh);
2007                 if (!buffer_mapped(bh)) {
2008                         WARN_ON(bh->b_size != blocksize);
2009                         if (get_block) {
2010                                 err = get_block(inode, block, bh, 1);
2011                                 if (err)
2012                                         break;
2013                         } else {
2014                                 iomap_to_bh(inode, block, bh, iomap);
2015                         }
2016
2017                         if (buffer_new(bh)) {
2018                                 clean_bdev_bh_alias(bh);
2019                                 if (PageUptodate(page)) {
2020                                         clear_buffer_new(bh);
2021                                         set_buffer_uptodate(bh);
2022                                         mark_buffer_dirty(bh);
2023                                         continue;
2024                                 }
2025                                 if (block_end > to || block_start < from)
2026                                         zero_user_segments(page,
2027                                                 to, block_end,
2028                                                 block_start, from);
2029                                 continue;
2030                         }
2031                 }
2032                 if (PageUptodate(page)) {
2033                         if (!buffer_uptodate(bh))
2034                                 set_buffer_uptodate(bh);
2035                         continue; 
2036                 }
2037                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2038                     !buffer_unwritten(bh) &&
2039                      (block_start < from || block_end > to)) {
2040                         ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2041                         *wait_bh++=bh;
2042                 }
2043         }
2044         /*
2045          * If we issued read requests - let them complete.
2046          */
2047         while(wait_bh > wait) {
2048                 wait_on_buffer(*--wait_bh);
2049                 if (!buffer_uptodate(*wait_bh))
2050                         err = -EIO;
2051         }
2052         if (unlikely(err))
2053                 page_zero_new_buffers(page, from, to);
2054         return err;
2055 }
2056
2057 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2058                 get_block_t *get_block)
2059 {
2060         return __block_write_begin_int(page, pos, len, get_block, NULL);
2061 }
2062 EXPORT_SYMBOL(__block_write_begin);
2063
2064 static int __block_commit_write(struct inode *inode, struct page *page,
2065                 unsigned from, unsigned to)
2066 {
2067         unsigned block_start, block_end;
2068         int partial = 0;
2069         unsigned blocksize;
2070         struct buffer_head *bh, *head;
2071
2072         bh = head = page_buffers(page);
2073         blocksize = bh->b_size;
2074
2075         block_start = 0;
2076         do {
2077                 block_end = block_start + blocksize;
2078                 if (block_end <= from || block_start >= to) {
2079                         if (!buffer_uptodate(bh))
2080                                 partial = 1;
2081                 } else {
2082                         set_buffer_uptodate(bh);
2083                         mark_buffer_dirty(bh);
2084                 }
2085                 if (buffer_new(bh))
2086                         clear_buffer_new(bh);
2087
2088                 block_start = block_end;
2089                 bh = bh->b_this_page;
2090         } while (bh != head);
2091
2092         /*
2093          * If this is a partial write which happened to make all buffers
2094          * uptodate then we can optimize away a bogus readpage() for
2095          * the next read(). Here we 'discover' whether the page went
2096          * uptodate as a result of this (potentially partial) write.
2097          */
2098         if (!partial)
2099                 SetPageUptodate(page);
2100         return 0;
2101 }
2102
2103 /*
2104  * block_write_begin takes care of the basic task of block allocation and
2105  * bringing partial write blocks uptodate first.
2106  *
2107  * The filesystem needs to handle block truncation upon failure.
2108  */
2109 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2110                 unsigned flags, struct page **pagep, get_block_t *get_block)
2111 {
2112         pgoff_t index = pos >> PAGE_SHIFT;
2113         struct page *page;
2114         int status;
2115
2116         page = grab_cache_page_write_begin(mapping, index, flags);
2117         if (!page)
2118                 return -ENOMEM;
2119
2120         status = __block_write_begin(page, pos, len, get_block);
2121         if (unlikely(status)) {
2122                 unlock_page(page);
2123                 put_page(page);
2124                 page = NULL;
2125         }
2126
2127         *pagep = page;
2128         return status;
2129 }
2130 EXPORT_SYMBOL(block_write_begin);
2131
2132 int block_write_end(struct file *file, struct address_space *mapping,
2133                         loff_t pos, unsigned len, unsigned copied,
2134                         struct page *page, void *fsdata)
2135 {
2136         struct inode *inode = mapping->host;
2137         unsigned start;
2138
2139         start = pos & (PAGE_SIZE - 1);
2140
2141         if (unlikely(copied < len)) {
2142                 /*
2143                  * The buffers that were written will now be uptodate, so we
2144                  * don't have to worry about a readpage reading them and
2145                  * overwriting a partial write. However if we have encountered
2146                  * a short write and only partially written into a buffer, it
2147                  * will not be marked uptodate, so a readpage might come in and
2148                  * destroy our partial write.
2149                  *
2150                  * Do the simplest thing, and just treat any short write to a
2151                  * non uptodate page as a zero-length write, and force the
2152                  * caller to redo the whole thing.
2153                  */
2154                 if (!PageUptodate(page))
2155                         copied = 0;
2156
2157                 page_zero_new_buffers(page, start+copied, start+len);
2158         }
2159         flush_dcache_page(page);
2160
2161         /* This could be a short (even 0-length) commit */
2162         __block_commit_write(inode, page, start, start+copied);
2163
2164         return copied;
2165 }
2166 EXPORT_SYMBOL(block_write_end);
2167
2168 int generic_write_end(struct file *file, struct address_space *mapping,
2169                         loff_t pos, unsigned len, unsigned copied,
2170                         struct page *page, void *fsdata)
2171 {
2172         struct inode *inode = mapping->host;
2173         loff_t old_size = inode->i_size;
2174         bool i_size_changed = false;
2175
2176         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2177
2178         /*
2179          * No need to use i_size_read() here, the i_size cannot change under us
2180          * because we hold i_rwsem.
2181          *
2182          * But it's important to update i_size while still holding page lock:
2183          * page writeout could otherwise come in and zero beyond i_size.
2184          */
2185         if (pos + copied > inode->i_size) {
2186                 i_size_write(inode, pos + copied);
2187                 i_size_changed = true;
2188         }
2189
2190         unlock_page(page);
2191         put_page(page);
2192
2193         if (old_size < pos)
2194                 pagecache_isize_extended(inode, old_size, pos);
2195         /*
2196          * Don't mark the inode dirty under page lock. First, it unnecessarily
2197          * makes the holding time of page lock longer. Second, it forces lock
2198          * ordering of page lock and transaction start for journaling
2199          * filesystems.
2200          */
2201         if (i_size_changed)
2202                 mark_inode_dirty(inode);
2203         return copied;
2204 }
2205 EXPORT_SYMBOL(generic_write_end);
2206
2207 /*
2208  * block_is_partially_uptodate checks whether buffers within a page are
2209  * uptodate or not.
2210  *
2211  * Returns true if all buffers which correspond to a file portion
2212  * we want to read are uptodate.
2213  */
2214 int block_is_partially_uptodate(struct page *page, unsigned long from,
2215                                         unsigned long count)
2216 {
2217         unsigned block_start, block_end, blocksize;
2218         unsigned to;
2219         struct buffer_head *bh, *head;
2220         int ret = 1;
2221
2222         if (!page_has_buffers(page))
2223                 return 0;
2224
2225         head = page_buffers(page);
2226         blocksize = head->b_size;
2227         to = min_t(unsigned, PAGE_SIZE - from, count);
2228         to = from + to;
2229         if (from < blocksize && to > PAGE_SIZE - blocksize)
2230                 return 0;
2231
2232         bh = head;
2233         block_start = 0;
2234         do {
2235                 block_end = block_start + blocksize;
2236                 if (block_end > from && block_start < to) {
2237                         if (!buffer_uptodate(bh)) {
2238                                 ret = 0;
2239                                 break;
2240                         }
2241                         if (block_end >= to)
2242                                 break;
2243                 }
2244                 block_start = block_end;
2245                 bh = bh->b_this_page;
2246         } while (bh != head);
2247
2248         return ret;
2249 }
2250 EXPORT_SYMBOL(block_is_partially_uptodate);
2251
2252 /*
2253  * Generic "read page" function for block devices that have the normal
2254  * get_block functionality. This is most of the block device filesystems.
2255  * Reads the page asynchronously --- the unlock_buffer() and
2256  * set/clear_buffer_uptodate() functions propagate buffer state into the
2257  * page struct once IO has completed.
2258  */
2259 int block_read_full_page(struct page *page, get_block_t *get_block)
2260 {
2261         struct inode *inode = page->mapping->host;
2262         sector_t iblock, lblock;
2263         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2264         unsigned int blocksize, bbits;
2265         int nr, i;
2266         int fully_mapped = 1;
2267
2268         head = create_page_buffers(page, inode, 0);
2269         blocksize = head->b_size;
2270         bbits = block_size_bits(blocksize);
2271
2272         iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2273         lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2274         bh = head;
2275         nr = 0;
2276         i = 0;
2277
2278         do {
2279                 if (buffer_uptodate(bh))
2280                         continue;
2281
2282                 if (!buffer_mapped(bh)) {
2283                         int err = 0;
2284
2285                         fully_mapped = 0;
2286                         if (iblock < lblock) {
2287                                 WARN_ON(bh->b_size != blocksize);
2288                                 err = get_block(inode, iblock, bh, 0);
2289                                 if (err)
2290                                         SetPageError(page);
2291                         }
2292                         if (!buffer_mapped(bh)) {
2293                                 zero_user(page, i * blocksize, blocksize);
2294                                 if (!err)
2295                                         set_buffer_uptodate(bh);
2296                                 continue;
2297                         }
2298                         /*
2299                          * get_block() might have updated the buffer
2300                          * synchronously
2301                          */
2302                         if (buffer_uptodate(bh))
2303                                 continue;
2304                 }
2305                 arr[nr++] = bh;
2306         } while (i++, iblock++, (bh = bh->b_this_page) != head);
2307
2308         if (fully_mapped)
2309                 SetPageMappedToDisk(page);
2310
2311         if (!nr) {
2312                 /*
2313                  * All buffers are uptodate - we can set the page uptodate
2314                  * as well. But not if get_block() returned an error.
2315                  */
2316                 if (!PageError(page))
2317                         SetPageUptodate(page);
2318                 unlock_page(page);
2319                 return 0;
2320         }
2321
2322         /* Stage two: lock the buffers */
2323         for (i = 0; i < nr; i++) {
2324                 bh = arr[i];
2325                 lock_buffer(bh);
2326                 mark_buffer_async_read(bh);
2327         }
2328
2329         /*
2330          * Stage 3: start the IO.  Check for uptodateness
2331          * inside the buffer lock in case another process reading
2332          * the underlying blockdev brought it uptodate (the sct fix).
2333          */
2334         for (i = 0; i < nr; i++) {
2335                 bh = arr[i];
2336                 if (buffer_uptodate(bh))
2337                         end_buffer_async_read(bh, 1);
2338                 else
2339                         submit_bh(REQ_OP_READ, 0, bh);
2340         }
2341         return 0;
2342 }
2343 EXPORT_SYMBOL(block_read_full_page);
2344
2345 /* utility function for filesystems that need to do work on expanding
2346  * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2347  * deal with the hole.  
2348  */
2349 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2350 {
2351         struct address_space *mapping = inode->i_mapping;
2352         struct page *page;
2353         void *fsdata;
2354         int err;
2355
2356         err = inode_newsize_ok(inode, size);
2357         if (err)
2358                 goto out;
2359
2360         err = pagecache_write_begin(NULL, mapping, size, 0,
2361                                     AOP_FLAG_CONT_EXPAND, &page, &fsdata);
2362         if (err)
2363                 goto out;
2364
2365         err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2366         BUG_ON(err > 0);
2367
2368 out:
2369         return err;
2370 }
2371 EXPORT_SYMBOL(generic_cont_expand_simple);
2372
2373 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2374                             loff_t pos, loff_t *bytes)
2375 {
2376         struct inode *inode = mapping->host;
2377         unsigned int blocksize = i_blocksize(inode);
2378         struct page *page;
2379         void *fsdata;
2380         pgoff_t index, curidx;
2381         loff_t curpos;
2382         unsigned zerofrom, offset, len;
2383         int err = 0;
2384
2385         index = pos >> PAGE_SHIFT;
2386         offset = pos & ~PAGE_MASK;
2387
2388         while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2389                 zerofrom = curpos & ~PAGE_MASK;
2390                 if (zerofrom & (blocksize-1)) {
2391                         *bytes |= (blocksize-1);
2392                         (*bytes)++;
2393                 }
2394                 len = PAGE_SIZE - zerofrom;
2395
2396                 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2397                                             &page, &fsdata);
2398                 if (err)
2399                         goto out;
2400                 zero_user(page, zerofrom, len);
2401                 err = pagecache_write_end(file, mapping, curpos, len, len,
2402                                                 page, fsdata);
2403                 if (err < 0)
2404                         goto out;
2405                 BUG_ON(err != len);
2406                 err = 0;
2407
2408                 balance_dirty_pages_ratelimited(mapping);
2409
2410                 if (fatal_signal_pending(current)) {
2411                         err = -EINTR;
2412                         goto out;
2413                 }
2414         }
2415
2416         /* page covers the boundary, find the boundary offset */
2417         if (index == curidx) {
2418                 zerofrom = curpos & ~PAGE_MASK;
2419                 /* if we will expand the thing last block will be filled */
2420                 if (offset <= zerofrom) {
2421                         goto out;
2422                 }
2423                 if (zerofrom & (blocksize-1)) {
2424                         *bytes |= (blocksize-1);
2425                         (*bytes)++;
2426                 }
2427                 len = offset - zerofrom;
2428
2429                 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2430                                             &page, &fsdata);
2431                 if (err)
2432                         goto out;
2433                 zero_user(page, zerofrom, len);
2434                 err = pagecache_write_end(file, mapping, curpos, len, len,
2435                                                 page, fsdata);
2436                 if (err < 0)
2437                         goto out;
2438                 BUG_ON(err != len);
2439                 err = 0;
2440         }
2441 out:
2442         return err;
2443 }
2444
2445 /*
2446  * For moronic filesystems that do not allow holes in file.
2447  * We may have to extend the file.
2448  */
2449 int cont_write_begin(struct file *file, struct address_space *mapping,
2450                         loff_t pos, unsigned len, unsigned flags,
2451                         struct page **pagep, void **fsdata,
2452                         get_block_t *get_block, loff_t *bytes)
2453 {
2454         struct inode *inode = mapping->host;
2455         unsigned int blocksize = i_blocksize(inode);
2456         unsigned int zerofrom;
2457         int err;
2458
2459         err = cont_expand_zero(file, mapping, pos, bytes);
2460         if (err)
2461                 return err;
2462
2463         zerofrom = *bytes & ~PAGE_MASK;
2464         if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2465                 *bytes |= (blocksize-1);
2466                 (*bytes)++;
2467         }
2468
2469         return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2470 }
2471 EXPORT_SYMBOL(cont_write_begin);
2472
2473 int block_commit_write(struct page *page, unsigned from, unsigned to)
2474 {
2475         struct inode *inode = page->mapping->host;
2476         __block_commit_write(inode,page,from,to);
2477         return 0;
2478 }
2479 EXPORT_SYMBOL(block_commit_write);
2480
2481 /*
2482  * block_page_mkwrite() is not allowed to change the file size as it gets
2483  * called from a page fault handler when a page is first dirtied. Hence we must
2484  * be careful to check for EOF conditions here. We set the page up correctly
2485  * for a written page which means we get ENOSPC checking when writing into
2486  * holes and correct delalloc and unwritten extent mapping on filesystems that
2487  * support these features.
2488  *
2489  * We are not allowed to take the i_mutex here so we have to play games to
2490  * protect against truncate races as the page could now be beyond EOF.  Because
2491  * truncate writes the inode size before removing pages, once we have the
2492  * page lock we can determine safely if the page is beyond EOF. If it is not
2493  * beyond EOF, then the page is guaranteed safe against truncation until we
2494  * unlock the page.
2495  *
2496  * Direct callers of this function should protect against filesystem freezing
2497  * using sb_start_pagefault() - sb_end_pagefault() functions.
2498  */
2499 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2500                          get_block_t get_block)
2501 {
2502         struct page *page = vmf->page;
2503         struct inode *inode = file_inode(vma->vm_file);
2504         unsigned long end;
2505         loff_t size;
2506         int ret;
2507
2508         lock_page(page);
2509         size = i_size_read(inode);
2510         if ((page->mapping != inode->i_mapping) ||
2511             (page_offset(page) > size)) {
2512                 /* We overload EFAULT to mean page got truncated */
2513                 ret = -EFAULT;
2514                 goto out_unlock;
2515         }
2516
2517         /* page is wholly or partially inside EOF */
2518         if (((page->index + 1) << PAGE_SHIFT) > size)
2519                 end = size & ~PAGE_MASK;
2520         else
2521                 end = PAGE_SIZE;
2522
2523         ret = __block_write_begin(page, 0, end, get_block);
2524         if (!ret)
2525                 ret = block_commit_write(page, 0, end);
2526
2527         if (unlikely(ret < 0))
2528                 goto out_unlock;
2529         set_page_dirty(page);
2530         wait_for_stable_page(page);
2531         return 0;
2532 out_unlock:
2533         unlock_page(page);
2534         return ret;
2535 }
2536 EXPORT_SYMBOL(block_page_mkwrite);
2537
2538 /*
2539  * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2540  * immediately, while under the page lock.  So it needs a special end_io
2541  * handler which does not touch the bh after unlocking it.
2542  */
2543 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2544 {
2545         __end_buffer_read_notouch(bh, uptodate);
2546 }
2547
2548 /*
2549  * Attach the singly-linked list of buffers created by nobh_write_begin, to
2550  * the page (converting it to circular linked list and taking care of page
2551  * dirty races).
2552  */
2553 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2554 {
2555         struct buffer_head *bh;
2556
2557         BUG_ON(!PageLocked(page));
2558
2559         spin_lock(&page->mapping->private_lock);
2560         bh = head;
2561         do {
2562                 if (PageDirty(page))
2563                         set_buffer_dirty(bh);
2564                 if (!bh->b_this_page)
2565                         bh->b_this_page = head;
2566                 bh = bh->b_this_page;
2567         } while (bh != head);
2568         attach_page_private(page, head);
2569         spin_unlock(&page->mapping->private_lock);
2570 }
2571
2572 /*
2573  * On entry, the page is fully not uptodate.
2574  * On exit the page is fully uptodate in the areas outside (from,to)
2575  * The filesystem needs to handle block truncation upon failure.
2576  */
2577 int nobh_write_begin(struct address_space *mapping,
2578                         loff_t pos, unsigned len, unsigned flags,
2579                         struct page **pagep, void **fsdata,
2580                         get_block_t *get_block)
2581 {
2582         struct inode *inode = mapping->host;
2583         const unsigned blkbits = inode->i_blkbits;
2584         const unsigned blocksize = 1 << blkbits;
2585         struct buffer_head *head, *bh;
2586         struct page *page;
2587         pgoff_t index;
2588         unsigned from, to;
2589         unsigned block_in_page;
2590         unsigned block_start, block_end;
2591         sector_t block_in_file;
2592         int nr_reads = 0;
2593         int ret = 0;
2594         int is_mapped_to_disk = 1;
2595
2596         index = pos >> PAGE_SHIFT;
2597         from = pos & (PAGE_SIZE - 1);
2598         to = from + len;
2599
2600         page = grab_cache_page_write_begin(mapping, index, flags);
2601         if (!page)
2602                 return -ENOMEM;
2603         *pagep = page;
2604         *fsdata = NULL;
2605
2606         if (page_has_buffers(page)) {
2607                 ret = __block_write_begin(page, pos, len, get_block);
2608                 if (unlikely(ret))
2609                         goto out_release;
2610                 return ret;
2611         }
2612
2613         if (PageMappedToDisk(page))
2614                 return 0;
2615
2616         /*
2617          * Allocate buffers so that we can keep track of state, and potentially
2618          * attach them to the page if an error occurs. In the common case of
2619          * no error, they will just be freed again without ever being attached
2620          * to the page (which is all OK, because we're under the page lock).
2621          *
2622          * Be careful: the buffer linked list is a NULL terminated one, rather
2623          * than the circular one we're used to.
2624          */
2625         head = alloc_page_buffers(page, blocksize, false);
2626         if (!head) {
2627                 ret = -ENOMEM;
2628                 goto out_release;
2629         }
2630
2631         block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2632
2633         /*
2634          * We loop across all blocks in the page, whether or not they are
2635          * part of the affected region.  This is so we can discover if the
2636          * page is fully mapped-to-disk.
2637          */
2638         for (block_start = 0, block_in_page = 0, bh = head;
2639                   block_start < PAGE_SIZE;
2640                   block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2641                 int create;
2642
2643                 block_end = block_start + blocksize;
2644                 bh->b_state = 0;
2645                 create = 1;
2646                 if (block_start >= to)
2647                         create = 0;
2648                 ret = get_block(inode, block_in_file + block_in_page,
2649                                         bh, create);
2650                 if (ret)
2651                         goto failed;
2652                 if (!buffer_mapped(bh))
2653                         is_mapped_to_disk = 0;
2654                 if (buffer_new(bh))
2655                         clean_bdev_bh_alias(bh);
2656                 if (PageUptodate(page)) {
2657                         set_buffer_uptodate(bh);
2658                         continue;
2659                 }
2660                 if (buffer_new(bh) || !buffer_mapped(bh)) {
2661                         zero_user_segments(page, block_start, from,
2662                                                         to, block_end);
2663                         continue;
2664                 }
2665                 if (buffer_uptodate(bh))
2666                         continue;       /* reiserfs does this */
2667                 if (block_start < from || block_end > to) {
2668                         lock_buffer(bh);
2669                         bh->b_end_io = end_buffer_read_nobh;
2670                         submit_bh(REQ_OP_READ, 0, bh);
2671                         nr_reads++;
2672                 }
2673         }
2674
2675         if (nr_reads) {
2676                 /*
2677                  * The page is locked, so these buffers are protected from
2678                  * any VM or truncate activity.  Hence we don't need to care
2679                  * for the buffer_head refcounts.
2680                  */
2681                 for (bh = head; bh; bh = bh->b_this_page) {
2682                         wait_on_buffer(bh);
2683                         if (!buffer_uptodate(bh))
2684                                 ret = -EIO;
2685                 }
2686                 if (ret)
2687                         goto failed;
2688         }
2689
2690         if (is_mapped_to_disk)
2691                 SetPageMappedToDisk(page);
2692
2693         *fsdata = head; /* to be released by nobh_write_end */
2694
2695         return 0;
2696
2697 failed:
2698         BUG_ON(!ret);
2699         /*
2700          * Error recovery is a bit difficult. We need to zero out blocks that
2701          * were newly allocated, and dirty them to ensure they get written out.
2702          * Buffers need to be attached to the page at this point, otherwise
2703          * the handling of potential IO errors during writeout would be hard
2704          * (could try doing synchronous writeout, but what if that fails too?)
2705          */
2706         attach_nobh_buffers(page, head);
2707         page_zero_new_buffers(page, from, to);
2708
2709 out_release:
2710         unlock_page(page);
2711         put_page(page);
2712         *pagep = NULL;
2713
2714         return ret;
2715 }
2716 EXPORT_SYMBOL(nobh_write_begin);
2717
2718 int nobh_write_end(struct file *file, struct address_space *mapping,
2719                         loff_t pos, unsigned len, unsigned copied,
2720                         struct page *page, void *fsdata)
2721 {
2722         struct inode *inode = page->mapping->host;
2723         struct buffer_head *head = fsdata;
2724         struct buffer_head *bh;
2725         BUG_ON(fsdata != NULL && page_has_buffers(page));
2726
2727         if (unlikely(copied < len) && head)
2728                 attach_nobh_buffers(page, head);
2729         if (page_has_buffers(page))
2730                 return generic_write_end(file, mapping, pos, len,
2731                                         copied, page, fsdata);
2732
2733         SetPageUptodate(page);
2734         set_page_dirty(page);
2735         if (pos+copied > inode->i_size) {
2736                 i_size_write(inode, pos+copied);
2737                 mark_inode_dirty(inode);
2738         }
2739
2740         unlock_page(page);
2741         put_page(page);
2742
2743         while (head) {
2744                 bh = head;
2745                 head = head->b_this_page;
2746                 free_buffer_head(bh);
2747         }
2748
2749         return copied;
2750 }
2751 EXPORT_SYMBOL(nobh_write_end);
2752
2753 /*
2754  * nobh_writepage() - based on block_full_write_page() except
2755  * that it tries to operate without attaching bufferheads to
2756  * the page.
2757  */
2758 int nobh_writepage(struct page *page, get_block_t *get_block,
2759                         struct writeback_control *wbc)
2760 {
2761         struct inode * const inode = page->mapping->host;
2762         loff_t i_size = i_size_read(inode);
2763         const pgoff_t end_index = i_size >> PAGE_SHIFT;
2764         unsigned offset;
2765         int ret;
2766
2767         /* Is the page fully inside i_size? */
2768         if (page->index < end_index)
2769                 goto out;
2770
2771         /* Is the page fully outside i_size? (truncate in progress) */
2772         offset = i_size & (PAGE_SIZE-1);
2773         if (page->index >= end_index+1 || !offset) {
2774                 unlock_page(page);
2775                 return 0; /* don't care */
2776         }
2777
2778         /*
2779          * The page straddles i_size.  It must be zeroed out on each and every
2780          * writepage invocation because it may be mmapped.  "A file is mapped
2781          * in multiples of the page size.  For a file that is not a multiple of
2782          * the  page size, the remaining memory is zeroed when mapped, and
2783          * writes to that region are not written out to the file."
2784          */
2785         zero_user_segment(page, offset, PAGE_SIZE);
2786 out:
2787         ret = mpage_writepage(page, get_block, wbc);
2788         if (ret == -EAGAIN)
2789                 ret = __block_write_full_page(inode, page, get_block, wbc,
2790                                               end_buffer_async_write);
2791         return ret;
2792 }
2793 EXPORT_SYMBOL(nobh_writepage);
2794
2795 int nobh_truncate_page(struct address_space *mapping,
2796                         loff_t from, get_block_t *get_block)
2797 {
2798         pgoff_t index = from >> PAGE_SHIFT;
2799         unsigned offset = from & (PAGE_SIZE-1);
2800         unsigned blocksize;
2801         sector_t iblock;
2802         unsigned length, pos;
2803         struct inode *inode = mapping->host;
2804         struct page *page;
2805         struct buffer_head map_bh;
2806         int err;
2807
2808         blocksize = i_blocksize(inode);
2809         length = offset & (blocksize - 1);
2810
2811         /* Block boundary? Nothing to do */
2812         if (!length)
2813                 return 0;
2814
2815         length = blocksize - length;
2816         iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2817
2818         page = grab_cache_page(mapping, index);
2819         err = -ENOMEM;
2820         if (!page)
2821                 goto out;
2822
2823         if (page_has_buffers(page)) {
2824 has_buffers:
2825                 unlock_page(page);
2826                 put_page(page);
2827                 return block_truncate_page(mapping, from, get_block);
2828         }
2829
2830         /* Find the buffer that contains "offset" */
2831         pos = blocksize;
2832         while (offset >= pos) {
2833                 iblock++;
2834                 pos += blocksize;
2835         }
2836
2837         map_bh.b_size = blocksize;
2838         map_bh.b_state = 0;
2839         err = get_block(inode, iblock, &map_bh, 0);
2840         if (err)
2841                 goto unlock;
2842         /* unmapped? It's a hole - nothing to do */
2843         if (!buffer_mapped(&map_bh))
2844                 goto unlock;
2845
2846         /* Ok, it's mapped. Make sure it's up-to-date */
2847         if (!PageUptodate(page)) {
2848                 err = mapping->a_ops->readpage(NULL, page);
2849                 if (err) {
2850                         put_page(page);
2851                         goto out;
2852                 }
2853                 lock_page(page);
2854                 if (!PageUptodate(page)) {
2855                         err = -EIO;
2856                         goto unlock;
2857                 }
2858                 if (page_has_buffers(page))
2859                         goto has_buffers;
2860         }
2861         zero_user(page, offset, length);
2862         set_page_dirty(page);
2863         err = 0;
2864
2865 unlock:
2866         unlock_page(page);
2867         put_page(page);
2868 out:
2869         return err;
2870 }
2871 EXPORT_SYMBOL(nobh_truncate_page);
2872
2873 int block_truncate_page(struct address_space *mapping,
2874                         loff_t from, get_block_t *get_block)
2875 {
2876         pgoff_t index = from >> PAGE_SHIFT;
2877         unsigned offset = from & (PAGE_SIZE-1);
2878         unsigned blocksize;
2879         sector_t iblock;
2880         unsigned length, pos;
2881         struct inode *inode = mapping->host;
2882         struct page *page;
2883         struct buffer_head *bh;
2884         int err;
2885
2886         blocksize = i_blocksize(inode);
2887         length = offset & (blocksize - 1);
2888
2889         /* Block boundary? Nothing to do */
2890         if (!length)
2891                 return 0;
2892
2893         length = blocksize - length;
2894         iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2895         
2896         page = grab_cache_page(mapping, index);
2897         err = -ENOMEM;
2898         if (!page)
2899                 goto out;
2900
2901         if (!page_has_buffers(page))
2902                 create_empty_buffers(page, blocksize, 0);
2903
2904         /* Find the buffer that contains "offset" */
2905         bh = page_buffers(page);
2906         pos = blocksize;
2907         while (offset >= pos) {
2908                 bh = bh->b_this_page;
2909                 iblock++;
2910                 pos += blocksize;
2911         }
2912
2913         err = 0;
2914         if (!buffer_mapped(bh)) {
2915                 WARN_ON(bh->b_size != blocksize);
2916                 err = get_block(inode, iblock, bh, 0);
2917                 if (err)
2918                         goto unlock;
2919                 /* unmapped? It's a hole - nothing to do */
2920                 if (!buffer_mapped(bh))
2921                         goto unlock;
2922         }
2923
2924         /* Ok, it's mapped. Make sure it's up-to-date */
2925         if (PageUptodate(page))
2926                 set_buffer_uptodate(bh);
2927
2928         if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2929                 err = -EIO;
2930                 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2931                 wait_on_buffer(bh);
2932                 /* Uhhuh. Read error. Complain and punt. */
2933                 if (!buffer_uptodate(bh))
2934                         goto unlock;
2935         }
2936
2937         zero_user(page, offset, length);
2938         mark_buffer_dirty(bh);
2939         err = 0;
2940
2941 unlock:
2942         unlock_page(page);
2943         put_page(page);
2944 out:
2945         return err;
2946 }
2947 EXPORT_SYMBOL(block_truncate_page);
2948
2949 /*
2950  * The generic ->writepage function for buffer-backed address_spaces
2951  */
2952 int block_write_full_page(struct page *page, get_block_t *get_block,
2953                         struct writeback_control *wbc)
2954 {
2955         struct inode * const inode = page->mapping->host;
2956         loff_t i_size = i_size_read(inode);
2957         const pgoff_t end_index = i_size >> PAGE_SHIFT;
2958         unsigned offset;
2959
2960         /* Is the page fully inside i_size? */
2961         if (page->index < end_index)
2962                 return __block_write_full_page(inode, page, get_block, wbc,
2963                                                end_buffer_async_write);
2964
2965         /* Is the page fully outside i_size? (truncate in progress) */
2966         offset = i_size & (PAGE_SIZE-1);
2967         if (page->index >= end_index+1 || !offset) {
2968                 unlock_page(page);
2969                 return 0; /* don't care */
2970         }
2971
2972         /*
2973          * The page straddles i_size.  It must be zeroed out on each and every
2974          * writepage invocation because it may be mmapped.  "A file is mapped
2975          * in multiples of the page size.  For a file that is not a multiple of
2976          * the  page size, the remaining memory is zeroed when mapped, and
2977          * writes to that region are not written out to the file."
2978          */
2979         zero_user_segment(page, offset, PAGE_SIZE);
2980         return __block_write_full_page(inode, page, get_block, wbc,
2981                                                         end_buffer_async_write);
2982 }
2983 EXPORT_SYMBOL(block_write_full_page);
2984
2985 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2986                             get_block_t *get_block)
2987 {
2988         struct inode *inode = mapping->host;
2989         struct buffer_head tmp = {
2990                 .b_size = i_blocksize(inode),
2991         };
2992
2993         get_block(inode, block, &tmp, 0);
2994         return tmp.b_blocknr;
2995 }
2996 EXPORT_SYMBOL(generic_block_bmap);
2997
2998 static void end_bio_bh_io_sync(struct bio *bio)
2999 {
3000         struct buffer_head *bh = bio->bi_private;
3001
3002         if (unlikely(bio_flagged(bio, BIO_QUIET)))
3003                 set_bit(BH_Quiet, &bh->b_state);
3004
3005         bh->b_end_io(bh, !bio->bi_status);
3006         bio_put(bio);
3007 }
3008
3009 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3010                          enum rw_hint write_hint, struct writeback_control *wbc)
3011 {
3012         struct bio *bio;
3013
3014         BUG_ON(!buffer_locked(bh));
3015         BUG_ON(!buffer_mapped(bh));
3016         BUG_ON(!bh->b_end_io);
3017         BUG_ON(buffer_delay(bh));
3018         BUG_ON(buffer_unwritten(bh));
3019
3020         /*
3021          * Only clear out a write error when rewriting
3022          */
3023         if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3024                 clear_buffer_write_io_error(bh);
3025
3026         bio = bio_alloc(GFP_NOIO, 1);
3027
3028         fscrypt_set_bio_crypt_ctx_bh(bio, bh, GFP_NOIO);
3029
3030         bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3031         bio_set_dev(bio, bh->b_bdev);
3032         bio->bi_write_hint = write_hint;
3033
3034         bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3035         BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3036
3037         bio->bi_end_io = end_bio_bh_io_sync;
3038         bio->bi_private = bh;
3039
3040         if (buffer_meta(bh))
3041                 op_flags |= REQ_META;
3042         if (buffer_prio(bh))
3043                 op_flags |= REQ_PRIO;
3044         bio_set_op_attrs(bio, op, op_flags);
3045
3046         /* Take care of bh's that straddle the end of the device */
3047         guard_bio_eod(bio);
3048
3049         if (wbc) {
3050                 wbc_init_bio(wbc, bio);
3051                 wbc_account_cgroup_owner(wbc, bh->b_page, bh->b_size);
3052         }
3053
3054         submit_bio(bio);
3055         return 0;
3056 }
3057
3058 int submit_bh(int op, int op_flags, struct buffer_head *bh)
3059 {
3060         return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3061 }
3062 EXPORT_SYMBOL(submit_bh);
3063
3064 /**
3065  * ll_rw_block: low-level access to block devices (DEPRECATED)
3066  * @op: whether to %READ or %WRITE
3067  * @op_flags: req_flag_bits
3068  * @nr: number of &struct buffer_heads in the array
3069  * @bhs: array of pointers to &struct buffer_head
3070  *
3071  * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3072  * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3073  * @op_flags contains flags modifying the detailed I/O behavior, most notably
3074  * %REQ_RAHEAD.
3075  *
3076  * This function drops any buffer that it cannot get a lock on (with the
3077  * BH_Lock state bit), any buffer that appears to be clean when doing a write
3078  * request, and any buffer that appears to be up-to-date when doing read
3079  * request.  Further it marks as clean buffers that are processed for
3080  * writing (the buffer cache won't assume that they are actually clean
3081  * until the buffer gets unlocked).
3082  *
3083  * ll_rw_block sets b_end_io to simple completion handler that marks
3084  * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3085  * any waiters. 
3086  *
3087  * All of the buffers must be for the same device, and must also be a
3088  * multiple of the current approved size for the device.
3089  */
3090 void ll_rw_block(int op, int op_flags,  int nr, struct buffer_head *bhs[])
3091 {
3092         int i;
3093
3094         for (i = 0; i < nr; i++) {
3095                 struct buffer_head *bh = bhs[i];
3096
3097                 if (!trylock_buffer(bh))
3098                         continue;
3099                 if (op == WRITE) {
3100                         if (test_clear_buffer_dirty(bh)) {
3101                                 bh->b_end_io = end_buffer_write_sync;
3102                                 get_bh(bh);
3103                                 submit_bh(op, op_flags, bh);
3104                                 continue;
3105                         }
3106                 } else {
3107                         if (!buffer_uptodate(bh)) {
3108                                 bh->b_end_io = end_buffer_read_sync;
3109                                 get_bh(bh);
3110                                 submit_bh(op, op_flags, bh);
3111                                 continue;
3112                         }
3113                 }
3114                 unlock_buffer(bh);
3115         }
3116 }
3117 EXPORT_SYMBOL(ll_rw_block);
3118
3119 void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3120 {
3121         lock_buffer(bh);
3122         if (!test_clear_buffer_dirty(bh)) {
3123                 unlock_buffer(bh);
3124                 return;
3125         }
3126         bh->b_end_io = end_buffer_write_sync;
3127         get_bh(bh);
3128         submit_bh(REQ_OP_WRITE, op_flags, bh);
3129 }
3130 EXPORT_SYMBOL(write_dirty_buffer);
3131
3132 /*
3133  * For a data-integrity writeout, we need to wait upon any in-progress I/O
3134  * and then start new I/O and then wait upon it.  The caller must have a ref on
3135  * the buffer_head.
3136  */
3137 int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3138 {
3139         int ret = 0;
3140
3141         WARN_ON(atomic_read(&bh->b_count) < 1);
3142         lock_buffer(bh);
3143         if (test_clear_buffer_dirty(bh)) {
3144                 /*
3145                  * The bh should be mapped, but it might not be if the
3146                  * device was hot-removed. Not much we can do but fail the I/O.
3147                  */
3148                 if (!buffer_mapped(bh)) {
3149                         unlock_buffer(bh);
3150                         return -EIO;
3151                 }
3152
3153                 get_bh(bh);
3154                 bh->b_end_io = end_buffer_write_sync;
3155                 ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3156                 wait_on_buffer(bh);
3157                 if (!ret && !buffer_uptodate(bh))
3158                         ret = -EIO;
3159         } else {
3160                 unlock_buffer(bh);
3161         }
3162         return ret;
3163 }
3164 EXPORT_SYMBOL(__sync_dirty_buffer);
3165
3166 int sync_dirty_buffer(struct buffer_head *bh)
3167 {
3168         return __sync_dirty_buffer(bh, REQ_SYNC);
3169 }
3170 EXPORT_SYMBOL(sync_dirty_buffer);
3171
3172 /*
3173  * try_to_free_buffers() checks if all the buffers on this particular page
3174  * are unused, and releases them if so.
3175  *
3176  * Exclusion against try_to_free_buffers may be obtained by either
3177  * locking the page or by holding its mapping's private_lock.
3178  *
3179  * If the page is dirty but all the buffers are clean then we need to
3180  * be sure to mark the page clean as well.  This is because the page
3181  * may be against a block device, and a later reattachment of buffers
3182  * to a dirty page will set *all* buffers dirty.  Which would corrupt
3183  * filesystem data on the same device.
3184  *
3185  * The same applies to regular filesystem pages: if all the buffers are
3186  * clean then we set the page clean and proceed.  To do that, we require
3187  * total exclusion from __set_page_dirty_buffers().  That is obtained with
3188  * private_lock.
3189  *
3190  * try_to_free_buffers() is non-blocking.
3191  */
3192 static inline int buffer_busy(struct buffer_head *bh)
3193 {
3194         return atomic_read(&bh->b_count) |
3195                 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3196 }
3197
3198 static int
3199 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3200 {
3201         struct buffer_head *head = page_buffers(page);
3202         struct buffer_head *bh;
3203
3204         bh = head;
3205         do {
3206                 if (buffer_busy(bh))
3207                         goto failed;
3208                 bh = bh->b_this_page;
3209         } while (bh != head);
3210
3211         do {
3212                 struct buffer_head *next = bh->b_this_page;
3213
3214                 if (bh->b_assoc_map)
3215                         __remove_assoc_queue(bh);
3216                 bh = next;
3217         } while (bh != head);
3218         *buffers_to_free = head;
3219         detach_page_private(page);
3220         return 1;
3221 failed:
3222         return 0;
3223 }
3224
3225 int try_to_free_buffers(struct page *page)
3226 {
3227         struct address_space * const mapping = page->mapping;
3228         struct buffer_head *buffers_to_free = NULL;
3229         int ret = 0;
3230
3231         BUG_ON(!PageLocked(page));
3232         if (PageWriteback(page))
3233                 return 0;
3234
3235         if (mapping == NULL) {          /* can this still happen? */
3236                 ret = drop_buffers(page, &buffers_to_free);
3237                 goto out;
3238         }
3239
3240         spin_lock(&mapping->private_lock);
3241         ret = drop_buffers(page, &buffers_to_free);
3242
3243         /*
3244          * If the filesystem writes its buffers by hand (eg ext3)
3245          * then we can have clean buffers against a dirty page.  We
3246          * clean the page here; otherwise the VM will never notice
3247          * that the filesystem did any IO at all.
3248          *
3249          * Also, during truncate, discard_buffer will have marked all
3250          * the page's buffers clean.  We discover that here and clean
3251          * the page also.
3252          *
3253          * private_lock must be held over this entire operation in order
3254          * to synchronise against __set_page_dirty_buffers and prevent the
3255          * dirty bit from being lost.
3256          */
3257         if (ret)
3258                 cancel_dirty_page(page);
3259         spin_unlock(&mapping->private_lock);
3260 out:
3261         if (buffers_to_free) {
3262                 struct buffer_head *bh = buffers_to_free;
3263
3264                 do {
3265                         struct buffer_head *next = bh->b_this_page;
3266                         free_buffer_head(bh);
3267                         bh = next;
3268                 } while (bh != buffers_to_free);
3269         }
3270         return ret;
3271 }
3272 EXPORT_SYMBOL(try_to_free_buffers);
3273
3274 /*
3275  * Buffer-head allocation
3276  */
3277 static struct kmem_cache *bh_cachep __read_mostly;
3278
3279 /*
3280  * Once the number of bh's in the machine exceeds this level, we start
3281  * stripping them in writeback.
3282  */
3283 static unsigned long max_buffer_heads;
3284
3285 int buffer_heads_over_limit;
3286
3287 struct bh_accounting {
3288         int nr;                 /* Number of live bh's */
3289         int ratelimit;          /* Limit cacheline bouncing */
3290 };
3291
3292 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3293
3294 static void recalc_bh_state(void)
3295 {
3296         int i;
3297         int tot = 0;
3298
3299         if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3300                 return;
3301         __this_cpu_write(bh_accounting.ratelimit, 0);
3302         for_each_online_cpu(i)
3303                 tot += per_cpu(bh_accounting, i).nr;
3304         buffer_heads_over_limit = (tot > max_buffer_heads);
3305 }
3306
3307 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3308 {
3309         struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3310         if (ret) {
3311                 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3312                 spin_lock_init(&ret->b_uptodate_lock);
3313                 preempt_disable();
3314                 __this_cpu_inc(bh_accounting.nr);
3315                 recalc_bh_state();
3316                 preempt_enable();
3317         }
3318         return ret;
3319 }
3320 EXPORT_SYMBOL(alloc_buffer_head);
3321
3322 void free_buffer_head(struct buffer_head *bh)
3323 {
3324         BUG_ON(!list_empty(&bh->b_assoc_buffers));
3325         kmem_cache_free(bh_cachep, bh);
3326         preempt_disable();
3327         __this_cpu_dec(bh_accounting.nr);
3328         recalc_bh_state();
3329         preempt_enable();
3330 }
3331 EXPORT_SYMBOL(free_buffer_head);
3332
3333 static int buffer_exit_cpu_dead(unsigned int cpu)
3334 {
3335         int i;
3336         struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3337
3338         for (i = 0; i < BH_LRU_SIZE; i++) {
3339                 brelse(b->bhs[i]);
3340                 b->bhs[i] = NULL;
3341         }
3342         this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3343         per_cpu(bh_accounting, cpu).nr = 0;
3344         return 0;
3345 }
3346
3347 /**
3348  * bh_uptodate_or_lock - Test whether the buffer is uptodate
3349  * @bh: struct buffer_head
3350  *
3351  * Return true if the buffer is up-to-date and false,
3352  * with the buffer locked, if not.
3353  */
3354 int bh_uptodate_or_lock(struct buffer_head *bh)
3355 {
3356         if (!buffer_uptodate(bh)) {
3357                 lock_buffer(bh);
3358                 if (!buffer_uptodate(bh))
3359                         return 0;
3360                 unlock_buffer(bh);
3361         }
3362         return 1;
3363 }
3364 EXPORT_SYMBOL(bh_uptodate_or_lock);
3365
3366 /**
3367  * bh_submit_read - Submit a locked buffer for reading
3368  * @bh: struct buffer_head
3369  *
3370  * Returns zero on success and -EIO on error.
3371  */
3372 int bh_submit_read(struct buffer_head *bh)
3373 {
3374         BUG_ON(!buffer_locked(bh));
3375
3376         if (buffer_uptodate(bh)) {
3377                 unlock_buffer(bh);
3378                 return 0;
3379         }
3380
3381         get_bh(bh);
3382         bh->b_end_io = end_buffer_read_sync;
3383         submit_bh(REQ_OP_READ, 0, bh);
3384         wait_on_buffer(bh);
3385         if (buffer_uptodate(bh))
3386                 return 0;
3387         return -EIO;
3388 }
3389 EXPORT_SYMBOL(bh_submit_read);
3390
3391 void __init buffer_init(void)
3392 {
3393         unsigned long nrpages;
3394         int ret;
3395
3396         bh_cachep = kmem_cache_create("buffer_head",
3397                         sizeof(struct buffer_head), 0,
3398                                 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3399                                 SLAB_MEM_SPREAD),
3400                                 NULL);
3401
3402         /*
3403          * Limit the bh occupancy to 10% of ZONE_NORMAL
3404          */
3405         nrpages = (nr_free_buffer_pages() * 10) / 100;
3406         max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3407         ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3408                                         NULL, buffer_exit_cpu_dead);
3409         WARN_ON(ret < 0);
3410 }