cifs: fix potential buffer overrun when composing a new options string
[linux-2.6.git] / fs / fs-writeback.c
1 /*
2  * fs/fs-writeback.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  *
6  * Contains all the functions related to writing back and waiting
7  * upon dirty inodes against superblocks, and writing back dirty
8  * pages against inodes.  ie: data writeback.  Writeout of the
9  * inode itself is not handled here.
10  *
11  * 10Apr2002    Andrew Morton
12  *              Split out of fs/inode.c
13  *              Additions for address_space-based writeback
14  */
15
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
21 #include <linux/fs.h>
22 #include <linux/mm.h>
23 #include <linux/pagemap.h>
24 #include <linux/kthread.h>
25 #include <linux/freezer.h>
26 #include <linux/writeback.h>
27 #include <linux/blkdev.h>
28 #include <linux/backing-dev.h>
29 #include <linux/tracepoint.h>
30 #include "internal.h"
31
32 /*
33  * 4MB minimal write chunk size
34  */
35 #define MIN_WRITEBACK_PAGES     (4096UL >> (PAGE_CACHE_SHIFT - 10))
36
37 /*
38  * Passed into wb_writeback(), essentially a subset of writeback_control
39  */
40 struct wb_writeback_work {
41         long nr_pages;
42         struct super_block *sb;
43         unsigned long *older_than_this;
44         enum writeback_sync_modes sync_mode;
45         unsigned int tagged_writepages:1;
46         unsigned int for_kupdate:1;
47         unsigned int range_cyclic:1;
48         unsigned int for_background:1;
49         enum wb_reason reason;          /* why was writeback initiated? */
50
51         struct list_head list;          /* pending work list */
52         struct completion *done;        /* set if the caller waits */
53 };
54
55 /*
56  * We don't actually have pdflush, but this one is exported though /proc...
57  */
58 int nr_pdflush_threads;
59
60 /**
61  * writeback_in_progress - determine whether there is writeback in progress
62  * @bdi: the device's backing_dev_info structure.
63  *
64  * Determine whether there is writeback waiting to be handled against a
65  * backing device.
66  */
67 int writeback_in_progress(struct backing_dev_info *bdi)
68 {
69         return test_bit(BDI_writeback_running, &bdi->state);
70 }
71 EXPORT_SYMBOL(writeback_in_progress);
72
73 static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
74 {
75         struct super_block *sb = inode->i_sb;
76
77         if (strcmp(sb->s_type->name, "bdev") == 0)
78                 return inode->i_mapping->backing_dev_info;
79
80         return sb->s_bdi;
81 }
82
83 static inline struct inode *wb_inode(struct list_head *head)
84 {
85         return list_entry(head, struct inode, i_wb_list);
86 }
87
88 /*
89  * Include the creation of the trace points after defining the
90  * wb_writeback_work structure and inline functions so that the definition
91  * remains local to this file.
92  */
93 #define CREATE_TRACE_POINTS
94 #include <trace/events/writeback.h>
95
96 /* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
97 static void bdi_wakeup_flusher(struct backing_dev_info *bdi)
98 {
99         if (bdi->wb.task) {
100                 wake_up_process(bdi->wb.task);
101         } else {
102                 /*
103                  * The bdi thread isn't there, wake up the forker thread which
104                  * will create and run it.
105                  */
106                 wake_up_process(default_backing_dev_info.wb.task);
107         }
108 }
109
110 static void bdi_queue_work(struct backing_dev_info *bdi,
111                            struct wb_writeback_work *work)
112 {
113         trace_writeback_queue(bdi, work);
114
115         spin_lock_bh(&bdi->wb_lock);
116         list_add_tail(&work->list, &bdi->work_list);
117         if (!bdi->wb.task)
118                 trace_writeback_nothread(bdi, work);
119         bdi_wakeup_flusher(bdi);
120         spin_unlock_bh(&bdi->wb_lock);
121 }
122
123 static void
124 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
125                       bool range_cyclic, enum wb_reason reason)
126 {
127         struct wb_writeback_work *work;
128
129         /*
130          * This is WB_SYNC_NONE writeback, so if allocation fails just
131          * wakeup the thread for old dirty data writeback
132          */
133         work = kzalloc(sizeof(*work), GFP_ATOMIC);
134         if (!work) {
135                 if (bdi->wb.task) {
136                         trace_writeback_nowork(bdi);
137                         wake_up_process(bdi->wb.task);
138                 }
139                 return;
140         }
141
142         work->sync_mode = WB_SYNC_NONE;
143         work->nr_pages  = nr_pages;
144         work->range_cyclic = range_cyclic;
145         work->reason    = reason;
146
147         bdi_queue_work(bdi, work);
148 }
149
150 /**
151  * bdi_start_writeback - start writeback
152  * @bdi: the backing device to write from
153  * @nr_pages: the number of pages to write
154  * @reason: reason why some writeback work was initiated
155  *
156  * Description:
157  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
158  *   started when this function returns, we make no guarantees on
159  *   completion. Caller need not hold sb s_umount semaphore.
160  *
161  */
162 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
163                         enum wb_reason reason)
164 {
165         __bdi_start_writeback(bdi, nr_pages, true, reason);
166 }
167
168 /**
169  * bdi_start_background_writeback - start background writeback
170  * @bdi: the backing device to write from
171  *
172  * Description:
173  *   This makes sure WB_SYNC_NONE background writeback happens. When
174  *   this function returns, it is only guaranteed that for given BDI
175  *   some IO is happening if we are over background dirty threshold.
176  *   Caller need not hold sb s_umount semaphore.
177  */
178 void bdi_start_background_writeback(struct backing_dev_info *bdi)
179 {
180         /*
181          * We just wake up the flusher thread. It will perform background
182          * writeback as soon as there is no other work to do.
183          */
184         trace_writeback_wake_background(bdi);
185         spin_lock_bh(&bdi->wb_lock);
186         bdi_wakeup_flusher(bdi);
187         spin_unlock_bh(&bdi->wb_lock);
188 }
189
190 /*
191  * Remove the inode from the writeback list it is on.
192  */
193 void inode_wb_list_del(struct inode *inode)
194 {
195         struct backing_dev_info *bdi = inode_to_bdi(inode);
196
197         spin_lock(&bdi->wb.list_lock);
198         list_del_init(&inode->i_wb_list);
199         spin_unlock(&bdi->wb.list_lock);
200 }
201
202 /*
203  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
204  * furthest end of its superblock's dirty-inode list.
205  *
206  * Before stamping the inode's ->dirtied_when, we check to see whether it is
207  * already the most-recently-dirtied inode on the b_dirty list.  If that is
208  * the case then the inode must have been redirtied while it was being written
209  * out and we don't reset its dirtied_when.
210  */
211 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
212 {
213         assert_spin_locked(&wb->list_lock);
214         if (!list_empty(&wb->b_dirty)) {
215                 struct inode *tail;
216
217                 tail = wb_inode(wb->b_dirty.next);
218                 if (time_before(inode->dirtied_when, tail->dirtied_when))
219                         inode->dirtied_when = jiffies;
220         }
221         list_move(&inode->i_wb_list, &wb->b_dirty);
222 }
223
224 /*
225  * requeue inode for re-scanning after bdi->b_io list is exhausted.
226  */
227 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
228 {
229         assert_spin_locked(&wb->list_lock);
230         list_move(&inode->i_wb_list, &wb->b_more_io);
231 }
232
233 static void inode_sync_complete(struct inode *inode)
234 {
235         /*
236          * Prevent speculative execution through
237          * spin_unlock(&wb->list_lock);
238          */
239
240         smp_mb();
241         wake_up_bit(&inode->i_state, __I_SYNC);
242 }
243
244 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
245 {
246         bool ret = time_after(inode->dirtied_when, t);
247 #ifndef CONFIG_64BIT
248         /*
249          * For inodes being constantly redirtied, dirtied_when can get stuck.
250          * It _appears_ to be in the future, but is actually in distant past.
251          * This test is necessary to prevent such wrapped-around relative times
252          * from permanently stopping the whole bdi writeback.
253          */
254         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
255 #endif
256         return ret;
257 }
258
259 /*
260  * Move expired (dirtied after work->older_than_this) dirty inodes from
261  * @delaying_queue to @dispatch_queue.
262  */
263 static int move_expired_inodes(struct list_head *delaying_queue,
264                                struct list_head *dispatch_queue,
265                                struct wb_writeback_work *work)
266 {
267         LIST_HEAD(tmp);
268         struct list_head *pos, *node;
269         struct super_block *sb = NULL;
270         struct inode *inode;
271         int do_sb_sort = 0;
272         int moved = 0;
273
274         while (!list_empty(delaying_queue)) {
275                 inode = wb_inode(delaying_queue->prev);
276                 if (work->older_than_this &&
277                     inode_dirtied_after(inode, *work->older_than_this))
278                         break;
279                 if (sb && sb != inode->i_sb)
280                         do_sb_sort = 1;
281                 sb = inode->i_sb;
282                 list_move(&inode->i_wb_list, &tmp);
283                 moved++;
284         }
285
286         /* just one sb in list, splice to dispatch_queue and we're done */
287         if (!do_sb_sort) {
288                 list_splice(&tmp, dispatch_queue);
289                 goto out;
290         }
291
292         /* Move inodes from one superblock together */
293         while (!list_empty(&tmp)) {
294                 sb = wb_inode(tmp.prev)->i_sb;
295                 list_for_each_prev_safe(pos, node, &tmp) {
296                         inode = wb_inode(pos);
297                         if (inode->i_sb == sb)
298                                 list_move(&inode->i_wb_list, dispatch_queue);
299                 }
300         }
301 out:
302         return moved;
303 }
304
305 /*
306  * Queue all expired dirty inodes for io, eldest first.
307  * Before
308  *         newly dirtied     b_dirty    b_io    b_more_io
309  *         =============>    gf         edc     BA
310  * After
311  *         newly dirtied     b_dirty    b_io    b_more_io
312  *         =============>    g          fBAedc
313  *                                           |
314  *                                           +--> dequeue for IO
315  */
316 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
317 {
318         int moved;
319         assert_spin_locked(&wb->list_lock);
320         list_splice_init(&wb->b_more_io, &wb->b_io);
321         moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, work);
322         trace_writeback_queue_io(wb, work, moved);
323 }
324
325 static int write_inode(struct inode *inode, struct writeback_control *wbc)
326 {
327         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
328                 return inode->i_sb->s_op->write_inode(inode, wbc);
329         return 0;
330 }
331
332 /*
333  * Wait for writeback on an inode to complete.
334  */
335 static void inode_wait_for_writeback(struct inode *inode,
336                                      struct bdi_writeback *wb)
337 {
338         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
339         wait_queue_head_t *wqh;
340
341         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
342         while (inode->i_state & I_SYNC) {
343                 spin_unlock(&inode->i_lock);
344                 spin_unlock(&wb->list_lock);
345                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
346                 spin_lock(&wb->list_lock);
347                 spin_lock(&inode->i_lock);
348         }
349 }
350
351 /*
352  * Write out an inode's dirty pages.  Called under wb->list_lock and
353  * inode->i_lock.  Either the caller has an active reference on the inode or
354  * the inode has I_WILL_FREE set.
355  *
356  * If `wait' is set, wait on the writeout.
357  *
358  * The whole writeout design is quite complex and fragile.  We want to avoid
359  * starvation of particular inodes when others are being redirtied, prevent
360  * livelocks, etc.
361  */
362 static int
363 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
364                        struct writeback_control *wbc)
365 {
366         struct address_space *mapping = inode->i_mapping;
367         long nr_to_write = wbc->nr_to_write;
368         unsigned dirty;
369         int ret;
370
371         assert_spin_locked(&wb->list_lock);
372         assert_spin_locked(&inode->i_lock);
373
374         if (!atomic_read(&inode->i_count))
375                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
376         else
377                 WARN_ON(inode->i_state & I_WILL_FREE);
378
379         if (inode->i_state & I_SYNC) {
380                 /*
381                  * If this inode is locked for writeback and we are not doing
382                  * writeback-for-data-integrity, move it to b_more_io so that
383                  * writeback can proceed with the other inodes on s_io.
384                  *
385                  * We'll have another go at writing back this inode when we
386                  * completed a full scan of b_io.
387                  */
388                 if (wbc->sync_mode != WB_SYNC_ALL) {
389                         requeue_io(inode, wb);
390                         trace_writeback_single_inode_requeue(inode, wbc,
391                                                              nr_to_write);
392                         return 0;
393                 }
394
395                 /*
396                  * It's a data-integrity sync.  We must wait.
397                  */
398                 inode_wait_for_writeback(inode, wb);
399         }
400
401         BUG_ON(inode->i_state & I_SYNC);
402
403         /* Set I_SYNC, reset I_DIRTY_PAGES */
404         inode->i_state |= I_SYNC;
405         inode->i_state &= ~I_DIRTY_PAGES;
406         spin_unlock(&inode->i_lock);
407         spin_unlock(&wb->list_lock);
408
409         ret = do_writepages(mapping, wbc);
410
411         /*
412          * Make sure to wait on the data before writing out the metadata.
413          * This is important for filesystems that modify metadata on data
414          * I/O completion.
415          */
416         if (wbc->sync_mode == WB_SYNC_ALL) {
417                 int err = filemap_fdatawait(mapping);
418                 if (ret == 0)
419                         ret = err;
420         }
421
422         /*
423          * Some filesystems may redirty the inode during the writeback
424          * due to delalloc, clear dirty metadata flags right before
425          * write_inode()
426          */
427         spin_lock(&inode->i_lock);
428         dirty = inode->i_state & I_DIRTY;
429         inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
430         spin_unlock(&inode->i_lock);
431         /* Don't write the inode if only I_DIRTY_PAGES was set */
432         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
433                 int err = write_inode(inode, wbc);
434                 if (ret == 0)
435                         ret = err;
436         }
437
438         spin_lock(&wb->list_lock);
439         spin_lock(&inode->i_lock);
440         inode->i_state &= ~I_SYNC;
441         if (!(inode->i_state & I_FREEING)) {
442                 /*
443                  * Sync livelock prevention. Each inode is tagged and synced in
444                  * one shot. If still dirty, it will be redirty_tail()'ed below.
445                  * Update the dirty time to prevent enqueue and sync it again.
446                  */
447                 if ((inode->i_state & I_DIRTY) &&
448                     (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
449                         inode->dirtied_when = jiffies;
450
451                 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
452                         /*
453                          * We didn't write back all the pages.  nfs_writepages()
454                          * sometimes bales out without doing anything.
455                          */
456                         inode->i_state |= I_DIRTY_PAGES;
457                         if (wbc->nr_to_write <= 0) {
458                                 /*
459                                  * slice used up: queue for next turn
460                                  */
461                                 requeue_io(inode, wb);
462                         } else {
463                                 /*
464                                  * Writeback blocked by something other than
465                                  * congestion. Delay the inode for some time to
466                                  * avoid spinning on the CPU (100% iowait)
467                                  * retrying writeback of the dirty page/inode
468                                  * that cannot be performed immediately.
469                                  */
470                                 redirty_tail(inode, wb);
471                         }
472                 } else if (inode->i_state & I_DIRTY) {
473                         /*
474                          * Filesystems can dirty the inode during writeback
475                          * operations, such as delayed allocation during
476                          * submission or metadata updates after data IO
477                          * completion.
478                          */
479                         redirty_tail(inode, wb);
480                 } else {
481                         /*
482                          * The inode is clean.  At this point we either have
483                          * a reference to the inode or it's on it's way out.
484                          * No need to add it back to the LRU.
485                          */
486                         list_del_init(&inode->i_wb_list);
487                 }
488         }
489         inode_sync_complete(inode);
490         trace_writeback_single_inode(inode, wbc, nr_to_write);
491         return ret;
492 }
493
494 static long writeback_chunk_size(struct backing_dev_info *bdi,
495                                  struct wb_writeback_work *work)
496 {
497         long pages;
498
499         /*
500          * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
501          * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
502          * here avoids calling into writeback_inodes_wb() more than once.
503          *
504          * The intended call sequence for WB_SYNC_ALL writeback is:
505          *
506          *      wb_writeback()
507          *          writeback_sb_inodes()       <== called only once
508          *              write_cache_pages()     <== called once for each inode
509          *                   (quickly) tag currently dirty pages
510          *                   (maybe slowly) sync all tagged pages
511          */
512         if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
513                 pages = LONG_MAX;
514         else {
515                 pages = min(bdi->avg_write_bandwidth / 2,
516                             global_dirty_limit / DIRTY_SCOPE);
517                 pages = min(pages, work->nr_pages);
518                 pages = round_down(pages + MIN_WRITEBACK_PAGES,
519                                    MIN_WRITEBACK_PAGES);
520         }
521
522         return pages;
523 }
524
525 /*
526  * Write a portion of b_io inodes which belong to @sb.
527  *
528  * If @only_this_sb is true, then find and write all such
529  * inodes. Otherwise write only ones which go sequentially
530  * in reverse order.
531  *
532  * Return the number of pages and/or inodes written.
533  */
534 static long writeback_sb_inodes(struct super_block *sb,
535                                 struct bdi_writeback *wb,
536                                 struct wb_writeback_work *work)
537 {
538         struct writeback_control wbc = {
539                 .sync_mode              = work->sync_mode,
540                 .tagged_writepages      = work->tagged_writepages,
541                 .for_kupdate            = work->for_kupdate,
542                 .for_background         = work->for_background,
543                 .range_cyclic           = work->range_cyclic,
544                 .range_start            = 0,
545                 .range_end              = LLONG_MAX,
546         };
547         unsigned long start_time = jiffies;
548         long write_chunk;
549         long wrote = 0;  /* count both pages and inodes */
550
551         while (!list_empty(&wb->b_io)) {
552                 struct inode *inode = wb_inode(wb->b_io.prev);
553
554                 if (inode->i_sb != sb) {
555                         if (work->sb) {
556                                 /*
557                                  * We only want to write back data for this
558                                  * superblock, move all inodes not belonging
559                                  * to it back onto the dirty list.
560                                  */
561                                 redirty_tail(inode, wb);
562                                 continue;
563                         }
564
565                         /*
566                          * The inode belongs to a different superblock.
567                          * Bounce back to the caller to unpin this and
568                          * pin the next superblock.
569                          */
570                         break;
571                 }
572
573                 /*
574                  * Don't bother with new inodes or inodes beeing freed, first
575                  * kind does not need peridic writeout yet, and for the latter
576                  * kind writeout is handled by the freer.
577                  */
578                 spin_lock(&inode->i_lock);
579                 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
580                         spin_unlock(&inode->i_lock);
581                         redirty_tail(inode, wb);
582                         continue;
583                 }
584                 __iget(inode);
585                 write_chunk = writeback_chunk_size(wb->bdi, work);
586                 wbc.nr_to_write = write_chunk;
587                 wbc.pages_skipped = 0;
588
589                 writeback_single_inode(inode, wb, &wbc);
590
591                 work->nr_pages -= write_chunk - wbc.nr_to_write;
592                 wrote += write_chunk - wbc.nr_to_write;
593                 if (!(inode->i_state & I_DIRTY))
594                         wrote++;
595                 if (wbc.pages_skipped) {
596                         /*
597                          * writeback is not making progress due to locked
598                          * buffers.  Skip this inode for now.
599                          */
600                         redirty_tail(inode, wb);
601                 }
602                 spin_unlock(&inode->i_lock);
603                 spin_unlock(&wb->list_lock);
604                 iput(inode);
605                 cond_resched();
606                 spin_lock(&wb->list_lock);
607                 /*
608                  * bail out to wb_writeback() often enough to check
609                  * background threshold and other termination conditions.
610                  */
611                 if (wrote) {
612                         if (time_is_before_jiffies(start_time + HZ / 10UL))
613                                 break;
614                         if (work->nr_pages <= 0)
615                                 break;
616                 }
617         }
618         return wrote;
619 }
620
621 static long __writeback_inodes_wb(struct bdi_writeback *wb,
622                                   struct wb_writeback_work *work)
623 {
624         unsigned long start_time = jiffies;
625         long wrote = 0;
626
627         while (!list_empty(&wb->b_io)) {
628                 struct inode *inode = wb_inode(wb->b_io.prev);
629                 struct super_block *sb = inode->i_sb;
630
631                 if (!grab_super_passive(sb)) {
632                         /*
633                          * grab_super_passive() may fail consistently due to
634                          * s_umount being grabbed by someone else. Don't use
635                          * requeue_io() to avoid busy retrying the inode/sb.
636                          */
637                         redirty_tail(inode, wb);
638                         continue;
639                 }
640                 wrote += writeback_sb_inodes(sb, wb, work);
641                 drop_super(sb);
642
643                 /* refer to the same tests at the end of writeback_sb_inodes */
644                 if (wrote) {
645                         if (time_is_before_jiffies(start_time + HZ / 10UL))
646                                 break;
647                         if (work->nr_pages <= 0)
648                                 break;
649                 }
650         }
651         /* Leave any unwritten inodes on b_io */
652         return wrote;
653 }
654
655 long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
656                                 enum wb_reason reason)
657 {
658         struct wb_writeback_work work = {
659                 .nr_pages       = nr_pages,
660                 .sync_mode      = WB_SYNC_NONE,
661                 .range_cyclic   = 1,
662                 .reason         = reason,
663         };
664
665         spin_lock(&wb->list_lock);
666         if (list_empty(&wb->b_io))
667                 queue_io(wb, &work);
668         __writeback_inodes_wb(wb, &work);
669         spin_unlock(&wb->list_lock);
670
671         return nr_pages - work.nr_pages;
672 }
673
674 static bool over_bground_thresh(struct backing_dev_info *bdi)
675 {
676         unsigned long background_thresh, dirty_thresh;
677
678         global_dirty_limits(&background_thresh, &dirty_thresh);
679
680         if (global_page_state(NR_FILE_DIRTY) +
681             global_page_state(NR_UNSTABLE_NFS) > background_thresh)
682                 return true;
683
684         if (bdi_stat(bdi, BDI_RECLAIMABLE) >
685                                 bdi_dirty_limit(bdi, background_thresh))
686                 return true;
687
688         return false;
689 }
690
691 /*
692  * Called under wb->list_lock. If there are multiple wb per bdi,
693  * only the flusher working on the first wb should do it.
694  */
695 static void wb_update_bandwidth(struct bdi_writeback *wb,
696                                 unsigned long start_time)
697 {
698         __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
699 }
700
701 /*
702  * Explicit flushing or periodic writeback of "old" data.
703  *
704  * Define "old": the first time one of an inode's pages is dirtied, we mark the
705  * dirtying-time in the inode's address_space.  So this periodic writeback code
706  * just walks the superblock inode list, writing back any inodes which are
707  * older than a specific point in time.
708  *
709  * Try to run once per dirty_writeback_interval.  But if a writeback event
710  * takes longer than a dirty_writeback_interval interval, then leave a
711  * one-second gap.
712  *
713  * older_than_this takes precedence over nr_to_write.  So we'll only write back
714  * all dirty pages if they are all attached to "old" mappings.
715  */
716 static long wb_writeback(struct bdi_writeback *wb,
717                          struct wb_writeback_work *work)
718 {
719         unsigned long wb_start = jiffies;
720         long nr_pages = work->nr_pages;
721         unsigned long oldest_jif;
722         struct inode *inode;
723         long progress;
724
725         oldest_jif = jiffies;
726         work->older_than_this = &oldest_jif;
727
728         spin_lock(&wb->list_lock);
729         for (;;) {
730                 /*
731                  * Stop writeback when nr_pages has been consumed
732                  */
733                 if (work->nr_pages <= 0)
734                         break;
735
736                 /*
737                  * Background writeout and kupdate-style writeback may
738                  * run forever. Stop them if there is other work to do
739                  * so that e.g. sync can proceed. They'll be restarted
740                  * after the other works are all done.
741                  */
742                 if ((work->for_background || work->for_kupdate) &&
743                     !list_empty(&wb->bdi->work_list))
744                         break;
745
746                 /*
747                  * For background writeout, stop when we are below the
748                  * background dirty threshold
749                  */
750                 if (work->for_background && !over_bground_thresh(wb->bdi))
751                         break;
752
753                 /*
754                  * Kupdate and background works are special and we want to
755                  * include all inodes that need writing. Livelock avoidance is
756                  * handled by these works yielding to any other work so we are
757                  * safe.
758                  */
759                 if (work->for_kupdate) {
760                         oldest_jif = jiffies -
761                                 msecs_to_jiffies(dirty_expire_interval * 10);
762                 } else if (work->for_background)
763                         oldest_jif = jiffies;
764
765                 trace_writeback_start(wb->bdi, work);
766                 if (list_empty(&wb->b_io))
767                         queue_io(wb, work);
768                 if (work->sb)
769                         progress = writeback_sb_inodes(work->sb, wb, work);
770                 else
771                         progress = __writeback_inodes_wb(wb, work);
772                 trace_writeback_written(wb->bdi, work);
773
774                 wb_update_bandwidth(wb, wb_start);
775
776                 /*
777                  * Did we write something? Try for more
778                  *
779                  * Dirty inodes are moved to b_io for writeback in batches.
780                  * The completion of the current batch does not necessarily
781                  * mean the overall work is done. So we keep looping as long
782                  * as made some progress on cleaning pages or inodes.
783                  */
784                 if (progress)
785                         continue;
786                 /*
787                  * No more inodes for IO, bail
788                  */
789                 if (list_empty(&wb->b_more_io))
790                         break;
791                 /*
792                  * Nothing written. Wait for some inode to
793                  * become available for writeback. Otherwise
794                  * we'll just busyloop.
795                  */
796                 if (!list_empty(&wb->b_more_io))  {
797                         trace_writeback_wait(wb->bdi, work);
798                         inode = wb_inode(wb->b_more_io.prev);
799                         spin_lock(&inode->i_lock);
800                         inode_wait_for_writeback(inode, wb);
801                         spin_unlock(&inode->i_lock);
802                 }
803         }
804         spin_unlock(&wb->list_lock);
805
806         return nr_pages - work->nr_pages;
807 }
808
809 /*
810  * Return the next wb_writeback_work struct that hasn't been processed yet.
811  */
812 static struct wb_writeback_work *
813 get_next_work_item(struct backing_dev_info *bdi)
814 {
815         struct wb_writeback_work *work = NULL;
816
817         spin_lock_bh(&bdi->wb_lock);
818         if (!list_empty(&bdi->work_list)) {
819                 work = list_entry(bdi->work_list.next,
820                                   struct wb_writeback_work, list);
821                 list_del_init(&work->list);
822         }
823         spin_unlock_bh(&bdi->wb_lock);
824         return work;
825 }
826
827 /*
828  * Add in the number of potentially dirty inodes, because each inode
829  * write can dirty pagecache in the underlying blockdev.
830  */
831 static unsigned long get_nr_dirty_pages(void)
832 {
833         return global_page_state(NR_FILE_DIRTY) +
834                 global_page_state(NR_UNSTABLE_NFS) +
835                 get_nr_dirty_inodes();
836 }
837
838 static long wb_check_background_flush(struct bdi_writeback *wb)
839 {
840         if (over_bground_thresh(wb->bdi)) {
841
842                 struct wb_writeback_work work = {
843                         .nr_pages       = LONG_MAX,
844                         .sync_mode      = WB_SYNC_NONE,
845                         .for_background = 1,
846                         .range_cyclic   = 1,
847                         .reason         = WB_REASON_BACKGROUND,
848                 };
849
850                 return wb_writeback(wb, &work);
851         }
852
853         return 0;
854 }
855
856 static long wb_check_old_data_flush(struct bdi_writeback *wb)
857 {
858         unsigned long expired;
859         long nr_pages;
860
861         /*
862          * When set to zero, disable periodic writeback
863          */
864         if (!dirty_writeback_interval)
865                 return 0;
866
867         expired = wb->last_old_flush +
868                         msecs_to_jiffies(dirty_writeback_interval * 10);
869         if (time_before(jiffies, expired))
870                 return 0;
871
872         wb->last_old_flush = jiffies;
873         nr_pages = get_nr_dirty_pages();
874
875         if (nr_pages) {
876                 struct wb_writeback_work work = {
877                         .nr_pages       = nr_pages,
878                         .sync_mode      = WB_SYNC_NONE,
879                         .for_kupdate    = 1,
880                         .range_cyclic   = 1,
881                         .reason         = WB_REASON_PERIODIC,
882                 };
883
884                 return wb_writeback(wb, &work);
885         }
886
887         return 0;
888 }
889
890 /*
891  * Retrieve work items and do the writeback they describe
892  */
893 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
894 {
895         struct backing_dev_info *bdi = wb->bdi;
896         struct wb_writeback_work *work;
897         long wrote = 0;
898
899         set_bit(BDI_writeback_running, &wb->bdi->state);
900         while ((work = get_next_work_item(bdi)) != NULL) {
901                 /*
902                  * Override sync mode, in case we must wait for completion
903                  * because this thread is exiting now.
904                  */
905                 if (force_wait)
906                         work->sync_mode = WB_SYNC_ALL;
907
908                 trace_writeback_exec(bdi, work);
909
910                 wrote += wb_writeback(wb, work);
911
912                 /*
913                  * Notify the caller of completion if this is a synchronous
914                  * work item, otherwise just free it.
915                  */
916                 if (work->done)
917                         complete(work->done);
918                 else
919                         kfree(work);
920         }
921
922         /*
923          * Check for periodic writeback, kupdated() style
924          */
925         wrote += wb_check_old_data_flush(wb);
926         wrote += wb_check_background_flush(wb);
927         clear_bit(BDI_writeback_running, &wb->bdi->state);
928
929         return wrote;
930 }
931
932 /*
933  * Handle writeback of dirty data for the device backed by this bdi. Also
934  * wakes up periodically and does kupdated style flushing.
935  */
936 int bdi_writeback_thread(void *data)
937 {
938         struct bdi_writeback *wb = data;
939         struct backing_dev_info *bdi = wb->bdi;
940         long pages_written;
941
942         current->flags |= PF_SWAPWRITE;
943         set_freezable();
944         wb->last_active = jiffies;
945
946         /*
947          * Our parent may run at a different priority, just set us to normal
948          */
949         set_user_nice(current, 0);
950
951         trace_writeback_thread_start(bdi);
952
953         while (!kthread_freezable_should_stop(NULL)) {
954                 /*
955                  * Remove own delayed wake-up timer, since we are already awake
956                  * and we'll take care of the preriodic write-back.
957                  */
958                 del_timer(&wb->wakeup_timer);
959
960                 pages_written = wb_do_writeback(wb, 0);
961
962                 trace_writeback_pages_written(pages_written);
963
964                 if (pages_written)
965                         wb->last_active = jiffies;
966
967                 set_current_state(TASK_INTERRUPTIBLE);
968                 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
969                         __set_current_state(TASK_RUNNING);
970                         continue;
971                 }
972
973                 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
974                         schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
975                 else {
976                         /*
977                          * We have nothing to do, so can go sleep without any
978                          * timeout and save power. When a work is queued or
979                          * something is made dirty - we will be woken up.
980                          */
981                         schedule();
982                 }
983         }
984
985         /* Flush any work that raced with us exiting */
986         if (!list_empty(&bdi->work_list))
987                 wb_do_writeback(wb, 1);
988
989         trace_writeback_thread_stop(bdi);
990         return 0;
991 }
992
993
994 /*
995  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
996  * the whole world.
997  */
998 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
999 {
1000         struct backing_dev_info *bdi;
1001
1002         if (!nr_pages) {
1003                 nr_pages = global_page_state(NR_FILE_DIRTY) +
1004                                 global_page_state(NR_UNSTABLE_NFS);
1005         }
1006
1007         rcu_read_lock();
1008         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1009                 if (!bdi_has_dirty_io(bdi))
1010                         continue;
1011                 __bdi_start_writeback(bdi, nr_pages, false, reason);
1012         }
1013         rcu_read_unlock();
1014 }
1015
1016 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1017 {
1018         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1019                 struct dentry *dentry;
1020                 const char *name = "?";
1021
1022                 dentry = d_find_alias(inode);
1023                 if (dentry) {
1024                         spin_lock(&dentry->d_lock);
1025                         name = (const char *) dentry->d_name.name;
1026                 }
1027                 printk(KERN_DEBUG
1028                        "%s(%d): dirtied inode %lu (%s) on %s\n",
1029                        current->comm, task_pid_nr(current), inode->i_ino,
1030                        name, inode->i_sb->s_id);
1031                 if (dentry) {
1032                         spin_unlock(&dentry->d_lock);
1033                         dput(dentry);
1034                 }
1035         }
1036 }
1037
1038 /**
1039  *      __mark_inode_dirty -    internal function
1040  *      @inode: inode to mark
1041  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1042  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
1043  *      mark_inode_dirty_sync.
1044  *
1045  * Put the inode on the super block's dirty list.
1046  *
1047  * CAREFUL! We mark it dirty unconditionally, but move it onto the
1048  * dirty list only if it is hashed or if it refers to a blockdev.
1049  * If it was not hashed, it will never be added to the dirty list
1050  * even if it is later hashed, as it will have been marked dirty already.
1051  *
1052  * In short, make sure you hash any inodes _before_ you start marking
1053  * them dirty.
1054  *
1055  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1056  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1057  * the kernel-internal blockdev inode represents the dirtying time of the
1058  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1059  * page->mapping->host, so the page-dirtying time is recorded in the internal
1060  * blockdev inode.
1061  */
1062 void __mark_inode_dirty(struct inode *inode, int flags)
1063 {
1064         struct super_block *sb = inode->i_sb;
1065         struct backing_dev_info *bdi = NULL;
1066
1067         /*
1068          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1069          * dirty the inode itself
1070          */
1071         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1072                 if (sb->s_op->dirty_inode)
1073                         sb->s_op->dirty_inode(inode, flags);
1074         }
1075
1076         /*
1077          * make sure that changes are seen by all cpus before we test i_state
1078          * -- mikulas
1079          */
1080         smp_mb();
1081
1082         /* avoid the locking if we can */
1083         if ((inode->i_state & flags) == flags)
1084                 return;
1085
1086         if (unlikely(block_dump))
1087                 block_dump___mark_inode_dirty(inode);
1088
1089         spin_lock(&inode->i_lock);
1090         if ((inode->i_state & flags) != flags) {
1091                 const int was_dirty = inode->i_state & I_DIRTY;
1092
1093                 inode->i_state |= flags;
1094
1095                 /*
1096                  * If the inode is being synced, just update its dirty state.
1097                  * The unlocker will place the inode on the appropriate
1098                  * superblock list, based upon its state.
1099                  */
1100                 if (inode->i_state & I_SYNC)
1101                         goto out_unlock_inode;
1102
1103                 /*
1104                  * Only add valid (hashed) inodes to the superblock's
1105                  * dirty list.  Add blockdev inodes as well.
1106                  */
1107                 if (!S_ISBLK(inode->i_mode)) {
1108                         if (inode_unhashed(inode))
1109                                 goto out_unlock_inode;
1110                 }
1111                 if (inode->i_state & I_FREEING)
1112                         goto out_unlock_inode;
1113
1114                 /*
1115                  * If the inode was already on b_dirty/b_io/b_more_io, don't
1116                  * reposition it (that would break b_dirty time-ordering).
1117                  */
1118                 if (!was_dirty) {
1119                         bool wakeup_bdi = false;
1120                         bdi = inode_to_bdi(inode);
1121
1122                         if (bdi_cap_writeback_dirty(bdi)) {
1123                                 WARN(!test_bit(BDI_registered, &bdi->state),
1124                                      "bdi-%s not registered\n", bdi->name);
1125
1126                                 /*
1127                                  * If this is the first dirty inode for this
1128                                  * bdi, we have to wake-up the corresponding
1129                                  * bdi thread to make sure background
1130                                  * write-back happens later.
1131                                  */
1132                                 if (!wb_has_dirty_io(&bdi->wb))
1133                                         wakeup_bdi = true;
1134                         }
1135
1136                         spin_unlock(&inode->i_lock);
1137                         spin_lock(&bdi->wb.list_lock);
1138                         inode->dirtied_when = jiffies;
1139                         list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1140                         spin_unlock(&bdi->wb.list_lock);
1141
1142                         if (wakeup_bdi)
1143                                 bdi_wakeup_thread_delayed(bdi);
1144                         return;
1145                 }
1146         }
1147 out_unlock_inode:
1148         spin_unlock(&inode->i_lock);
1149
1150 }
1151 EXPORT_SYMBOL(__mark_inode_dirty);
1152
1153 static void wait_sb_inodes(struct super_block *sb)
1154 {
1155         struct inode *inode, *old_inode = NULL;
1156
1157         /*
1158          * We need to be protected against the filesystem going from
1159          * r/o to r/w or vice versa.
1160          */
1161         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1162
1163         spin_lock(&inode_sb_list_lock);
1164
1165         /*
1166          * Data integrity sync. Must wait for all pages under writeback,
1167          * because there may have been pages dirtied before our sync
1168          * call, but which had writeout started before we write it out.
1169          * In which case, the inode may not be on the dirty list, but
1170          * we still have to wait for that writeout.
1171          */
1172         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1173                 struct address_space *mapping = inode->i_mapping;
1174
1175                 spin_lock(&inode->i_lock);
1176                 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1177                     (mapping->nrpages == 0)) {
1178                         spin_unlock(&inode->i_lock);
1179                         continue;
1180                 }
1181                 __iget(inode);
1182                 spin_unlock(&inode->i_lock);
1183                 spin_unlock(&inode_sb_list_lock);
1184
1185                 /*
1186                  * We hold a reference to 'inode' so it couldn't have been
1187                  * removed from s_inodes list while we dropped the
1188                  * inode_sb_list_lock.  We cannot iput the inode now as we can
1189                  * be holding the last reference and we cannot iput it under
1190                  * inode_sb_list_lock. So we keep the reference and iput it
1191                  * later.
1192                  */
1193                 iput(old_inode);
1194                 old_inode = inode;
1195
1196                 filemap_fdatawait(mapping);
1197
1198                 cond_resched();
1199
1200                 spin_lock(&inode_sb_list_lock);
1201         }
1202         spin_unlock(&inode_sb_list_lock);
1203         iput(old_inode);
1204 }
1205
1206 /**
1207  * writeback_inodes_sb_nr -     writeback dirty inodes from given super_block
1208  * @sb: the superblock
1209  * @nr: the number of pages to write
1210  * @reason: reason why some writeback work initiated
1211  *
1212  * Start writeback on some inodes on this super_block. No guarantees are made
1213  * on how many (if any) will be written, and this function does not wait
1214  * for IO completion of submitted IO.
1215  */
1216 void writeback_inodes_sb_nr(struct super_block *sb,
1217                             unsigned long nr,
1218                             enum wb_reason reason)
1219 {
1220         DECLARE_COMPLETION_ONSTACK(done);
1221         struct wb_writeback_work work = {
1222                 .sb                     = sb,
1223                 .sync_mode              = WB_SYNC_NONE,
1224                 .tagged_writepages      = 1,
1225                 .done                   = &done,
1226                 .nr_pages               = nr,
1227                 .reason                 = reason,
1228         };
1229
1230         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1231         bdi_queue_work(sb->s_bdi, &work);
1232         wait_for_completion(&done);
1233 }
1234 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1235
1236 /**
1237  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1238  * @sb: the superblock
1239  * @reason: reason why some writeback work was initiated
1240  *
1241  * Start writeback on some inodes on this super_block. No guarantees are made
1242  * on how many (if any) will be written, and this function does not wait
1243  * for IO completion of submitted IO.
1244  */
1245 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1246 {
1247         return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1248 }
1249 EXPORT_SYMBOL(writeback_inodes_sb);
1250
1251 /**
1252  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1253  * @sb: the superblock
1254  * @reason: reason why some writeback work was initiated
1255  *
1256  * Invoke writeback_inodes_sb if no writeback is currently underway.
1257  * Returns 1 if writeback was started, 0 if not.
1258  */
1259 int writeback_inodes_sb_if_idle(struct super_block *sb, enum wb_reason reason)
1260 {
1261         if (!writeback_in_progress(sb->s_bdi)) {
1262                 down_read(&sb->s_umount);
1263                 writeback_inodes_sb(sb, reason);
1264                 up_read(&sb->s_umount);
1265                 return 1;
1266         } else
1267                 return 0;
1268 }
1269 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1270
1271 /**
1272  * writeback_inodes_sb_nr_if_idle       -       start writeback if none underway
1273  * @sb: the superblock
1274  * @nr: the number of pages to write
1275  * @reason: reason why some writeback work was initiated
1276  *
1277  * Invoke writeback_inodes_sb if no writeback is currently underway.
1278  * Returns 1 if writeback was started, 0 if not.
1279  */
1280 int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1281                                    unsigned long nr,
1282                                    enum wb_reason reason)
1283 {
1284         if (!writeback_in_progress(sb->s_bdi)) {
1285                 down_read(&sb->s_umount);
1286                 writeback_inodes_sb_nr(sb, nr, reason);
1287                 up_read(&sb->s_umount);
1288                 return 1;
1289         } else
1290                 return 0;
1291 }
1292 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1293
1294 /**
1295  * sync_inodes_sb       -       sync sb inode pages
1296  * @sb: the superblock
1297  *
1298  * This function writes and waits on any dirty inode belonging to this
1299  * super_block.
1300  */
1301 void sync_inodes_sb(struct super_block *sb)
1302 {
1303         DECLARE_COMPLETION_ONSTACK(done);
1304         struct wb_writeback_work work = {
1305                 .sb             = sb,
1306                 .sync_mode      = WB_SYNC_ALL,
1307                 .nr_pages       = LONG_MAX,
1308                 .range_cyclic   = 0,
1309                 .done           = &done,
1310                 .reason         = WB_REASON_SYNC,
1311         };
1312
1313         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1314
1315         bdi_queue_work(sb->s_bdi, &work);
1316         wait_for_completion(&done);
1317
1318         wait_sb_inodes(sb);
1319 }
1320 EXPORT_SYMBOL(sync_inodes_sb);
1321
1322 /**
1323  * write_inode_now      -       write an inode to disk
1324  * @inode: inode to write to disk
1325  * @sync: whether the write should be synchronous or not
1326  *
1327  * This function commits an inode to disk immediately if it is dirty. This is
1328  * primarily needed by knfsd.
1329  *
1330  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1331  */
1332 int write_inode_now(struct inode *inode, int sync)
1333 {
1334         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1335         int ret;
1336         struct writeback_control wbc = {
1337                 .nr_to_write = LONG_MAX,
1338                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1339                 .range_start = 0,
1340                 .range_end = LLONG_MAX,
1341         };
1342
1343         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1344                 wbc.nr_to_write = 0;
1345
1346         might_sleep();
1347         spin_lock(&wb->list_lock);
1348         spin_lock(&inode->i_lock);
1349         ret = writeback_single_inode(inode, wb, &wbc);
1350         spin_unlock(&inode->i_lock);
1351         spin_unlock(&wb->list_lock);
1352         return ret;
1353 }
1354 EXPORT_SYMBOL(write_inode_now);
1355
1356 /**
1357  * sync_inode - write an inode and its pages to disk.
1358  * @inode: the inode to sync
1359  * @wbc: controls the writeback mode
1360  *
1361  * sync_inode() will write an inode and its pages to disk.  It will also
1362  * correctly update the inode on its superblock's dirty inode lists and will
1363  * update inode->i_state.
1364  *
1365  * The caller must have a ref on the inode.
1366  */
1367 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1368 {
1369         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1370         int ret;
1371
1372         spin_lock(&wb->list_lock);
1373         spin_lock(&inode->i_lock);
1374         ret = writeback_single_inode(inode, wb, wbc);
1375         spin_unlock(&inode->i_lock);
1376         spin_unlock(&wb->list_lock);
1377         return ret;
1378 }
1379 EXPORT_SYMBOL(sync_inode);
1380
1381 /**
1382  * sync_inode_metadata - write an inode to disk
1383  * @inode: the inode to sync
1384  * @wait: wait for I/O to complete.
1385  *
1386  * Write an inode to disk and adjust its dirty state after completion.
1387  *
1388  * Note: only writes the actual inode, no associated data or other metadata.
1389  */
1390 int sync_inode_metadata(struct inode *inode, int wait)
1391 {
1392         struct writeback_control wbc = {
1393                 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1394                 .nr_to_write = 0, /* metadata-only */
1395         };
1396
1397         return sync_inode(inode, &wbc);
1398 }
1399 EXPORT_SYMBOL(sync_inode_metadata);