writeback: trace global_dirty_state
[linux-2.6.git] / mm / page-writeback.c
1 /*
2  * mm/page-writeback.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
6  *
7  * Contains functions related to writing back dirty pages at the
8  * address_space level.
9  *
10  * 10Apr2002    Andrew Morton
11  *              Initial version
12  */
13
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/spinlock.h>
17 #include <linux/fs.h>
18 #include <linux/mm.h>
19 #include <linux/swap.h>
20 #include <linux/slab.h>
21 #include <linux/pagemap.h>
22 #include <linux/writeback.h>
23 #include <linux/init.h>
24 #include <linux/backing-dev.h>
25 #include <linux/task_io_accounting_ops.h>
26 #include <linux/blkdev.h>
27 #include <linux/mpage.h>
28 #include <linux/rmap.h>
29 #include <linux/percpu.h>
30 #include <linux/notifier.h>
31 #include <linux/smp.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/syscalls.h>
35 #include <linux/buffer_head.h>
36 #include <linux/pagevec.h>
37 #include <trace/events/writeback.h>
38
39 /*
40  * Sleep at most 200ms at a time in balance_dirty_pages().
41  */
42 #define MAX_PAUSE               max(HZ/5, 1)
43
44 /*
45  * Estimate write bandwidth at 200ms intervals.
46  */
47 #define BANDWIDTH_INTERVAL      max(HZ/5, 1)
48
49 /*
50  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
51  * will look to see if it needs to force writeback or throttling.
52  */
53 static long ratelimit_pages = 32;
54
55 /*
56  * When balance_dirty_pages decides that the caller needs to perform some
57  * non-background writeback, this is how many pages it will attempt to write.
58  * It should be somewhat larger than dirtied pages to ensure that reasonably
59  * large amounts of I/O are submitted.
60  */
61 static inline long sync_writeback_pages(unsigned long dirtied)
62 {
63         if (dirtied < ratelimit_pages)
64                 dirtied = ratelimit_pages;
65
66         return dirtied + dirtied / 2;
67 }
68
69 /* The following parameters are exported via /proc/sys/vm */
70
71 /*
72  * Start background writeback (via writeback threads) at this percentage
73  */
74 int dirty_background_ratio = 10;
75
76 /*
77  * dirty_background_bytes starts at 0 (disabled) so that it is a function of
78  * dirty_background_ratio * the amount of dirtyable memory
79  */
80 unsigned long dirty_background_bytes;
81
82 /*
83  * free highmem will not be subtracted from the total free memory
84  * for calculating free ratios if vm_highmem_is_dirtyable is true
85  */
86 int vm_highmem_is_dirtyable;
87
88 /*
89  * The generator of dirty data starts writeback at this percentage
90  */
91 int vm_dirty_ratio = 20;
92
93 /*
94  * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
95  * vm_dirty_ratio * the amount of dirtyable memory
96  */
97 unsigned long vm_dirty_bytes;
98
99 /*
100  * The interval between `kupdate'-style writebacks
101  */
102 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
103
104 /*
105  * The longest time for which data is allowed to remain dirty
106  */
107 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
108
109 /*
110  * Flag that makes the machine dump writes/reads and block dirtyings.
111  */
112 int block_dump;
113
114 /*
115  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
116  * a full sync is triggered after this time elapses without any disk activity.
117  */
118 int laptop_mode;
119
120 EXPORT_SYMBOL(laptop_mode);
121
122 /* End of sysctl-exported parameters */
123
124 unsigned long global_dirty_limit;
125
126 /*
127  * Scale the writeback cache size proportional to the relative writeout speeds.
128  *
129  * We do this by keeping a floating proportion between BDIs, based on page
130  * writeback completions [end_page_writeback()]. Those devices that write out
131  * pages fastest will get the larger share, while the slower will get a smaller
132  * share.
133  *
134  * We use page writeout completions because we are interested in getting rid of
135  * dirty pages. Having them written out is the primary goal.
136  *
137  * We introduce a concept of time, a period over which we measure these events,
138  * because demand can/will vary over time. The length of this period itself is
139  * measured in page writeback completions.
140  *
141  */
142 static struct prop_descriptor vm_completions;
143 static struct prop_descriptor vm_dirties;
144
145 /*
146  * couple the period to the dirty_ratio:
147  *
148  *   period/2 ~ roundup_pow_of_two(dirty limit)
149  */
150 static int calc_period_shift(void)
151 {
152         unsigned long dirty_total;
153
154         if (vm_dirty_bytes)
155                 dirty_total = vm_dirty_bytes / PAGE_SIZE;
156         else
157                 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
158                                 100;
159         return 2 + ilog2(dirty_total - 1);
160 }
161
162 /*
163  * update the period when the dirty threshold changes.
164  */
165 static void update_completion_period(void)
166 {
167         int shift = calc_period_shift();
168         prop_change_shift(&vm_completions, shift);
169         prop_change_shift(&vm_dirties, shift);
170 }
171
172 int dirty_background_ratio_handler(struct ctl_table *table, int write,
173                 void __user *buffer, size_t *lenp,
174                 loff_t *ppos)
175 {
176         int ret;
177
178         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
179         if (ret == 0 && write)
180                 dirty_background_bytes = 0;
181         return ret;
182 }
183
184 int dirty_background_bytes_handler(struct ctl_table *table, int write,
185                 void __user *buffer, size_t *lenp,
186                 loff_t *ppos)
187 {
188         int ret;
189
190         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
191         if (ret == 0 && write)
192                 dirty_background_ratio = 0;
193         return ret;
194 }
195
196 int dirty_ratio_handler(struct ctl_table *table, int write,
197                 void __user *buffer, size_t *lenp,
198                 loff_t *ppos)
199 {
200         int old_ratio = vm_dirty_ratio;
201         int ret;
202
203         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
204         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
205                 update_completion_period();
206                 vm_dirty_bytes = 0;
207         }
208         return ret;
209 }
210
211
212 int dirty_bytes_handler(struct ctl_table *table, int write,
213                 void __user *buffer, size_t *lenp,
214                 loff_t *ppos)
215 {
216         unsigned long old_bytes = vm_dirty_bytes;
217         int ret;
218
219         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
220         if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
221                 update_completion_period();
222                 vm_dirty_ratio = 0;
223         }
224         return ret;
225 }
226
227 /*
228  * Increment the BDI's writeout completion count and the global writeout
229  * completion count. Called from test_clear_page_writeback().
230  */
231 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
232 {
233         __inc_bdi_stat(bdi, BDI_WRITTEN);
234         __prop_inc_percpu_max(&vm_completions, &bdi->completions,
235                               bdi->max_prop_frac);
236 }
237
238 void bdi_writeout_inc(struct backing_dev_info *bdi)
239 {
240         unsigned long flags;
241
242         local_irq_save(flags);
243         __bdi_writeout_inc(bdi);
244         local_irq_restore(flags);
245 }
246 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
247
248 void task_dirty_inc(struct task_struct *tsk)
249 {
250         prop_inc_single(&vm_dirties, &tsk->dirties);
251 }
252
253 /*
254  * Obtain an accurate fraction of the BDI's portion.
255  */
256 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
257                 long *numerator, long *denominator)
258 {
259         prop_fraction_percpu(&vm_completions, &bdi->completions,
260                                 numerator, denominator);
261 }
262
263 static inline void task_dirties_fraction(struct task_struct *tsk,
264                 long *numerator, long *denominator)
265 {
266         prop_fraction_single(&vm_dirties, &tsk->dirties,
267                                 numerator, denominator);
268 }
269
270 /*
271  * task_dirty_limit - scale down dirty throttling threshold for one task
272  *
273  * task specific dirty limit:
274  *
275  *   dirty -= (dirty/8) * p_{t}
276  *
277  * To protect light/slow dirtying tasks from heavier/fast ones, we start
278  * throttling individual tasks before reaching the bdi dirty limit.
279  * Relatively low thresholds will be allocated to heavy dirtiers. So when
280  * dirty pages grow large, heavy dirtiers will be throttled first, which will
281  * effectively curb the growth of dirty pages. Light dirtiers with high enough
282  * dirty threshold may never get throttled.
283  */
284 static unsigned long task_dirty_limit(struct task_struct *tsk,
285                                        unsigned long bdi_dirty)
286 {
287         long numerator, denominator;
288         unsigned long dirty = bdi_dirty;
289         u64 inv = dirty >> 3;
290
291         task_dirties_fraction(tsk, &numerator, &denominator);
292         inv *= numerator;
293         do_div(inv, denominator);
294
295         dirty -= inv;
296
297         return max(dirty, bdi_dirty/2);
298 }
299
300 /*
301  *
302  */
303 static unsigned int bdi_min_ratio;
304
305 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
306 {
307         int ret = 0;
308
309         spin_lock_bh(&bdi_lock);
310         if (min_ratio > bdi->max_ratio) {
311                 ret = -EINVAL;
312         } else {
313                 min_ratio -= bdi->min_ratio;
314                 if (bdi_min_ratio + min_ratio < 100) {
315                         bdi_min_ratio += min_ratio;
316                         bdi->min_ratio += min_ratio;
317                 } else {
318                         ret = -EINVAL;
319                 }
320         }
321         spin_unlock_bh(&bdi_lock);
322
323         return ret;
324 }
325
326 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
327 {
328         int ret = 0;
329
330         if (max_ratio > 100)
331                 return -EINVAL;
332
333         spin_lock_bh(&bdi_lock);
334         if (bdi->min_ratio > max_ratio) {
335                 ret = -EINVAL;
336         } else {
337                 bdi->max_ratio = max_ratio;
338                 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
339         }
340         spin_unlock_bh(&bdi_lock);
341
342         return ret;
343 }
344 EXPORT_SYMBOL(bdi_set_max_ratio);
345
346 /*
347  * Work out the current dirty-memory clamping and background writeout
348  * thresholds.
349  *
350  * The main aim here is to lower them aggressively if there is a lot of mapped
351  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
352  * pages.  It is better to clamp down on writers than to start swapping, and
353  * performing lots of scanning.
354  *
355  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
356  *
357  * We don't permit the clamping level to fall below 5% - that is getting rather
358  * excessive.
359  *
360  * We make sure that the background writeout level is below the adjusted
361  * clamping level.
362  */
363
364 static unsigned long highmem_dirtyable_memory(unsigned long total)
365 {
366 #ifdef CONFIG_HIGHMEM
367         int node;
368         unsigned long x = 0;
369
370         for_each_node_state(node, N_HIGH_MEMORY) {
371                 struct zone *z =
372                         &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
373
374                 x += zone_page_state(z, NR_FREE_PAGES) +
375                      zone_reclaimable_pages(z);
376         }
377         /*
378          * Make sure that the number of highmem pages is never larger
379          * than the number of the total dirtyable memory. This can only
380          * occur in very strange VM situations but we want to make sure
381          * that this does not occur.
382          */
383         return min(x, total);
384 #else
385         return 0;
386 #endif
387 }
388
389 /**
390  * determine_dirtyable_memory - amount of memory that may be used
391  *
392  * Returns the numebr of pages that can currently be freed and used
393  * by the kernel for direct mappings.
394  */
395 unsigned long determine_dirtyable_memory(void)
396 {
397         unsigned long x;
398
399         x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
400
401         if (!vm_highmem_is_dirtyable)
402                 x -= highmem_dirtyable_memory(x);
403
404         return x + 1;   /* Ensure that we never return 0 */
405 }
406
407 static unsigned long hard_dirty_limit(unsigned long thresh)
408 {
409         return max(thresh, global_dirty_limit);
410 }
411
412 /*
413  * global_dirty_limits - background-writeback and dirty-throttling thresholds
414  *
415  * Calculate the dirty thresholds based on sysctl parameters
416  * - vm.dirty_background_ratio  or  vm.dirty_background_bytes
417  * - vm.dirty_ratio             or  vm.dirty_bytes
418  * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
419  * real-time tasks.
420  */
421 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
422 {
423         unsigned long background;
424         unsigned long dirty;
425         unsigned long uninitialized_var(available_memory);
426         struct task_struct *tsk;
427
428         if (!vm_dirty_bytes || !dirty_background_bytes)
429                 available_memory = determine_dirtyable_memory();
430
431         if (vm_dirty_bytes)
432                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
433         else
434                 dirty = (vm_dirty_ratio * available_memory) / 100;
435
436         if (dirty_background_bytes)
437                 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
438         else
439                 background = (dirty_background_ratio * available_memory) / 100;
440
441         if (background >= dirty)
442                 background = dirty / 2;
443         tsk = current;
444         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
445                 background += background / 4;
446                 dirty += dirty / 4;
447         }
448         *pbackground = background;
449         *pdirty = dirty;
450         trace_global_dirty_state(background, dirty);
451 }
452
453 /**
454  * bdi_dirty_limit - @bdi's share of dirty throttling threshold
455  * @bdi: the backing_dev_info to query
456  * @dirty: global dirty limit in pages
457  *
458  * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
459  * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
460  * And the "limit" in the name is not seriously taken as hard limit in
461  * balance_dirty_pages().
462  *
463  * It allocates high/low dirty limits to fast/slow devices, in order to prevent
464  * - starving fast devices
465  * - piling up dirty pages (that will take long time to sync) on slow devices
466  *
467  * The bdi's share of dirty limit will be adapting to its throughput and
468  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
469  */
470 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
471 {
472         u64 bdi_dirty;
473         long numerator, denominator;
474
475         /*
476          * Calculate this BDI's share of the dirty ratio.
477          */
478         bdi_writeout_fraction(bdi, &numerator, &denominator);
479
480         bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
481         bdi_dirty *= numerator;
482         do_div(bdi_dirty, denominator);
483
484         bdi_dirty += (dirty * bdi->min_ratio) / 100;
485         if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
486                 bdi_dirty = dirty * bdi->max_ratio / 100;
487
488         return bdi_dirty;
489 }
490
491 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
492                                        unsigned long elapsed,
493                                        unsigned long written)
494 {
495         const unsigned long period = roundup_pow_of_two(3 * HZ);
496         unsigned long avg = bdi->avg_write_bandwidth;
497         unsigned long old = bdi->write_bandwidth;
498         u64 bw;
499
500         /*
501          * bw = written * HZ / elapsed
502          *
503          *                   bw * elapsed + write_bandwidth * (period - elapsed)
504          * write_bandwidth = ---------------------------------------------------
505          *                                          period
506          */
507         bw = written - bdi->written_stamp;
508         bw *= HZ;
509         if (unlikely(elapsed > period)) {
510                 do_div(bw, elapsed);
511                 avg = bw;
512                 goto out;
513         }
514         bw += (u64)bdi->write_bandwidth * (period - elapsed);
515         bw >>= ilog2(period);
516
517         /*
518          * one more level of smoothing, for filtering out sudden spikes
519          */
520         if (avg > old && old >= (unsigned long)bw)
521                 avg -= (avg - old) >> 3;
522
523         if (avg < old && old <= (unsigned long)bw)
524                 avg += (old - avg) >> 3;
525
526 out:
527         bdi->write_bandwidth = bw;
528         bdi->avg_write_bandwidth = avg;
529 }
530
531 /*
532  * The global dirtyable memory and dirty threshold could be suddenly knocked
533  * down by a large amount (eg. on the startup of KVM in a swapless system).
534  * This may throw the system into deep dirty exceeded state and throttle
535  * heavy/light dirtiers alike. To retain good responsiveness, maintain
536  * global_dirty_limit for tracking slowly down to the knocked down dirty
537  * threshold.
538  */
539 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
540 {
541         unsigned long limit = global_dirty_limit;
542
543         /*
544          * Follow up in one step.
545          */
546         if (limit < thresh) {
547                 limit = thresh;
548                 goto update;
549         }
550
551         /*
552          * Follow down slowly. Use the higher one as the target, because thresh
553          * may drop below dirty. This is exactly the reason to introduce
554          * global_dirty_limit which is guaranteed to lie above the dirty pages.
555          */
556         thresh = max(thresh, dirty);
557         if (limit > thresh) {
558                 limit -= (limit - thresh) >> 5;
559                 goto update;
560         }
561         return;
562 update:
563         global_dirty_limit = limit;
564 }
565
566 static void global_update_bandwidth(unsigned long thresh,
567                                     unsigned long dirty,
568                                     unsigned long now)
569 {
570         static DEFINE_SPINLOCK(dirty_lock);
571         static unsigned long update_time;
572
573         /*
574          * check locklessly first to optimize away locking for the most time
575          */
576         if (time_before(now, update_time + BANDWIDTH_INTERVAL))
577                 return;
578
579         spin_lock(&dirty_lock);
580         if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
581                 update_dirty_limit(thresh, dirty);
582                 update_time = now;
583         }
584         spin_unlock(&dirty_lock);
585 }
586
587 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
588                             unsigned long thresh,
589                             unsigned long dirty,
590                             unsigned long bdi_thresh,
591                             unsigned long bdi_dirty,
592                             unsigned long start_time)
593 {
594         unsigned long now = jiffies;
595         unsigned long elapsed = now - bdi->bw_time_stamp;
596         unsigned long written;
597
598         /*
599          * rate-limit, only update once every 200ms.
600          */
601         if (elapsed < BANDWIDTH_INTERVAL)
602                 return;
603
604         written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
605
606         /*
607          * Skip quiet periods when disk bandwidth is under-utilized.
608          * (at least 1s idle time between two flusher runs)
609          */
610         if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
611                 goto snapshot;
612
613         if (thresh)
614                 global_update_bandwidth(thresh, dirty, now);
615
616         bdi_update_write_bandwidth(bdi, elapsed, written);
617
618 snapshot:
619         bdi->written_stamp = written;
620         bdi->bw_time_stamp = now;
621 }
622
623 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
624                                  unsigned long thresh,
625                                  unsigned long dirty,
626                                  unsigned long bdi_thresh,
627                                  unsigned long bdi_dirty,
628                                  unsigned long start_time)
629 {
630         if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
631                 return;
632         spin_lock(&bdi->wb.list_lock);
633         __bdi_update_bandwidth(bdi, thresh, dirty, bdi_thresh, bdi_dirty,
634                                start_time);
635         spin_unlock(&bdi->wb.list_lock);
636 }
637
638 /*
639  * balance_dirty_pages() must be called by processes which are generating dirty
640  * data.  It looks at the number of dirty pages in the machine and will force
641  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
642  * If we're over `background_thresh' then the writeback threads are woken to
643  * perform some writeout.
644  */
645 static void balance_dirty_pages(struct address_space *mapping,
646                                 unsigned long write_chunk)
647 {
648         unsigned long nr_reclaimable, bdi_nr_reclaimable;
649         unsigned long nr_dirty;  /* = file_dirty + writeback + unstable_nfs */
650         unsigned long bdi_dirty;
651         unsigned long background_thresh;
652         unsigned long dirty_thresh;
653         unsigned long bdi_thresh;
654         unsigned long pages_written = 0;
655         unsigned long pause = 1;
656         bool dirty_exceeded = false;
657         struct backing_dev_info *bdi = mapping->backing_dev_info;
658         unsigned long start_time = jiffies;
659
660         for (;;) {
661                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
662                                         global_page_state(NR_UNSTABLE_NFS);
663                 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
664
665                 global_dirty_limits(&background_thresh, &dirty_thresh);
666
667                 /*
668                  * Throttle it only when the background writeback cannot
669                  * catch-up. This avoids (excessively) small writeouts
670                  * when the bdi limits are ramping up.
671                  */
672                 if (nr_dirty <= (background_thresh + dirty_thresh) / 2)
673                         break;
674
675                 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
676                 bdi_thresh = task_dirty_limit(current, bdi_thresh);
677
678                 /*
679                  * In order to avoid the stacked BDI deadlock we need
680                  * to ensure we accurately count the 'dirty' pages when
681                  * the threshold is low.
682                  *
683                  * Otherwise it would be possible to get thresh+n pages
684                  * reported dirty, even though there are thresh-m pages
685                  * actually dirty; with m+n sitting in the percpu
686                  * deltas.
687                  */
688                 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
689                         bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
690                         bdi_dirty = bdi_nr_reclaimable +
691                                     bdi_stat_sum(bdi, BDI_WRITEBACK);
692                 } else {
693                         bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
694                         bdi_dirty = bdi_nr_reclaimable +
695                                     bdi_stat(bdi, BDI_WRITEBACK);
696                 }
697
698                 /*
699                  * The bdi thresh is somehow "soft" limit derived from the
700                  * global "hard" limit. The former helps to prevent heavy IO
701                  * bdi or process from holding back light ones; The latter is
702                  * the last resort safeguard.
703                  */
704                 dirty_exceeded = (bdi_dirty > bdi_thresh) ||
705                                   (nr_dirty > dirty_thresh);
706
707                 if (!dirty_exceeded)
708                         break;
709
710                 if (!bdi->dirty_exceeded)
711                         bdi->dirty_exceeded = 1;
712
713                 bdi_update_bandwidth(bdi, dirty_thresh, nr_dirty,
714                                      bdi_thresh, bdi_dirty, start_time);
715
716                 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
717                  * Unstable writes are a feature of certain networked
718                  * filesystems (i.e. NFS) in which data may have been
719                  * written to the server's write cache, but has not yet
720                  * been flushed to permanent storage.
721                  * Only move pages to writeback if this bdi is over its
722                  * threshold otherwise wait until the disk writes catch
723                  * up.
724                  */
725                 trace_balance_dirty_start(bdi);
726                 if (bdi_nr_reclaimable > bdi_thresh) {
727                         pages_written += writeback_inodes_wb(&bdi->wb,
728                                                              write_chunk);
729                         trace_balance_dirty_written(bdi, pages_written);
730                         if (pages_written >= write_chunk)
731                                 break;          /* We've done our duty */
732                 }
733                 __set_current_state(TASK_UNINTERRUPTIBLE);
734                 io_schedule_timeout(pause);
735                 trace_balance_dirty_wait(bdi);
736
737                 dirty_thresh = hard_dirty_limit(dirty_thresh);
738                 /*
739                  * max-pause area. If dirty exceeded but still within this
740                  * area, no need to sleep for more than 200ms: (a) 8 pages per
741                  * 200ms is typically more than enough to curb heavy dirtiers;
742                  * (b) the pause time limit makes the dirtiers more responsive.
743                  */
744                 if (nr_dirty < dirty_thresh +
745                                dirty_thresh / DIRTY_MAXPAUSE_AREA &&
746                     time_after(jiffies, start_time + MAX_PAUSE))
747                         break;
748                 /*
749                  * pass-good area. When some bdi gets blocked (eg. NFS server
750                  * not responding), or write bandwidth dropped dramatically due
751                  * to concurrent reads, or dirty threshold suddenly dropped and
752                  * the dirty pages cannot be brought down anytime soon (eg. on
753                  * slow USB stick), at least let go of the good bdi's.
754                  */
755                 if (nr_dirty < dirty_thresh +
756                                dirty_thresh / DIRTY_PASSGOOD_AREA &&
757                     bdi_dirty < bdi_thresh)
758                         break;
759
760                 /*
761                  * Increase the delay for each loop, up to our previous
762                  * default of taking a 100ms nap.
763                  */
764                 pause <<= 1;
765                 if (pause > HZ / 10)
766                         pause = HZ / 10;
767         }
768
769         if (!dirty_exceeded && bdi->dirty_exceeded)
770                 bdi->dirty_exceeded = 0;
771
772         if (writeback_in_progress(bdi))
773                 return;
774
775         /*
776          * In laptop mode, we wait until hitting the higher threshold before
777          * starting background writeout, and then write out all the way down
778          * to the lower threshold.  So slow writers cause minimal disk activity.
779          *
780          * In normal mode, we start background writeout at the lower
781          * background_thresh, to keep the amount of dirty memory low.
782          */
783         if ((laptop_mode && pages_written) ||
784             (!laptop_mode && (nr_reclaimable > background_thresh)))
785                 bdi_start_background_writeback(bdi);
786 }
787
788 void set_page_dirty_balance(struct page *page, int page_mkwrite)
789 {
790         if (set_page_dirty(page) || page_mkwrite) {
791                 struct address_space *mapping = page_mapping(page);
792
793                 if (mapping)
794                         balance_dirty_pages_ratelimited(mapping);
795         }
796 }
797
798 static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
799
800 /**
801  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
802  * @mapping: address_space which was dirtied
803  * @nr_pages_dirtied: number of pages which the caller has just dirtied
804  *
805  * Processes which are dirtying memory should call in here once for each page
806  * which was newly dirtied.  The function will periodically check the system's
807  * dirty state and will initiate writeback if needed.
808  *
809  * On really big machines, get_writeback_state is expensive, so try to avoid
810  * calling it too often (ratelimiting).  But once we're over the dirty memory
811  * limit we decrease the ratelimiting by a lot, to prevent individual processes
812  * from overshooting the limit by (ratelimit_pages) each.
813  */
814 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
815                                         unsigned long nr_pages_dirtied)
816 {
817         struct backing_dev_info *bdi = mapping->backing_dev_info;
818         unsigned long ratelimit;
819         unsigned long *p;
820
821         if (!bdi_cap_account_dirty(bdi))
822                 return;
823
824         ratelimit = ratelimit_pages;
825         if (mapping->backing_dev_info->dirty_exceeded)
826                 ratelimit = 8;
827
828         /*
829          * Check the rate limiting. Also, we do not want to throttle real-time
830          * tasks in balance_dirty_pages(). Period.
831          */
832         preempt_disable();
833         p =  &__get_cpu_var(bdp_ratelimits);
834         *p += nr_pages_dirtied;
835         if (unlikely(*p >= ratelimit)) {
836                 ratelimit = sync_writeback_pages(*p);
837                 *p = 0;
838                 preempt_enable();
839                 balance_dirty_pages(mapping, ratelimit);
840                 return;
841         }
842         preempt_enable();
843 }
844 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
845
846 void throttle_vm_writeout(gfp_t gfp_mask)
847 {
848         unsigned long background_thresh;
849         unsigned long dirty_thresh;
850
851         for ( ; ; ) {
852                 global_dirty_limits(&background_thresh, &dirty_thresh);
853
854                 /*
855                  * Boost the allowable dirty threshold a bit for page
856                  * allocators so they don't get DoS'ed by heavy writers
857                  */
858                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
859
860                 if (global_page_state(NR_UNSTABLE_NFS) +
861                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
862                                 break;
863                 congestion_wait(BLK_RW_ASYNC, HZ/10);
864
865                 /*
866                  * The caller might hold locks which can prevent IO completion
867                  * or progress in the filesystem.  So we cannot just sit here
868                  * waiting for IO to complete.
869                  */
870                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
871                         break;
872         }
873 }
874
875 /*
876  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
877  */
878 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
879         void __user *buffer, size_t *length, loff_t *ppos)
880 {
881         proc_dointvec(table, write, buffer, length, ppos);
882         bdi_arm_supers_timer();
883         return 0;
884 }
885
886 #ifdef CONFIG_BLOCK
887 void laptop_mode_timer_fn(unsigned long data)
888 {
889         struct request_queue *q = (struct request_queue *)data;
890         int nr_pages = global_page_state(NR_FILE_DIRTY) +
891                 global_page_state(NR_UNSTABLE_NFS);
892
893         /*
894          * We want to write everything out, not just down to the dirty
895          * threshold
896          */
897         if (bdi_has_dirty_io(&q->backing_dev_info))
898                 bdi_start_writeback(&q->backing_dev_info, nr_pages);
899 }
900
901 /*
902  * We've spun up the disk and we're in laptop mode: schedule writeback
903  * of all dirty data a few seconds from now.  If the flush is already scheduled
904  * then push it back - the user is still using the disk.
905  */
906 void laptop_io_completion(struct backing_dev_info *info)
907 {
908         mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
909 }
910
911 /*
912  * We're in laptop mode and we've just synced. The sync's writes will have
913  * caused another writeback to be scheduled by laptop_io_completion.
914  * Nothing needs to be written back anymore, so we unschedule the writeback.
915  */
916 void laptop_sync_completion(void)
917 {
918         struct backing_dev_info *bdi;
919
920         rcu_read_lock();
921
922         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
923                 del_timer(&bdi->laptop_mode_wb_timer);
924
925         rcu_read_unlock();
926 }
927 #endif
928
929 /*
930  * If ratelimit_pages is too high then we can get into dirty-data overload
931  * if a large number of processes all perform writes at the same time.
932  * If it is too low then SMP machines will call the (expensive)
933  * get_writeback_state too often.
934  *
935  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
936  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
937  * thresholds before writeback cuts in.
938  *
939  * But the limit should not be set too high.  Because it also controls the
940  * amount of memory which the balance_dirty_pages() caller has to write back.
941  * If this is too large then the caller will block on the IO queue all the
942  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
943  * will write six megabyte chunks, max.
944  */
945
946 void writeback_set_ratelimit(void)
947 {
948         ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
949         if (ratelimit_pages < 16)
950                 ratelimit_pages = 16;
951         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
952                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
953 }
954
955 static int __cpuinit
956 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
957 {
958         writeback_set_ratelimit();
959         return NOTIFY_DONE;
960 }
961
962 static struct notifier_block __cpuinitdata ratelimit_nb = {
963         .notifier_call  = ratelimit_handler,
964         .next           = NULL,
965 };
966
967 /*
968  * Called early on to tune the page writeback dirty limits.
969  *
970  * We used to scale dirty pages according to how total memory
971  * related to pages that could be allocated for buffers (by
972  * comparing nr_free_buffer_pages() to vm_total_pages.
973  *
974  * However, that was when we used "dirty_ratio" to scale with
975  * all memory, and we don't do that any more. "dirty_ratio"
976  * is now applied to total non-HIGHPAGE memory (by subtracting
977  * totalhigh_pages from vm_total_pages), and as such we can't
978  * get into the old insane situation any more where we had
979  * large amounts of dirty pages compared to a small amount of
980  * non-HIGHMEM memory.
981  *
982  * But we might still want to scale the dirty_ratio by how
983  * much memory the box has..
984  */
985 void __init page_writeback_init(void)
986 {
987         int shift;
988
989         writeback_set_ratelimit();
990         register_cpu_notifier(&ratelimit_nb);
991
992         shift = calc_period_shift();
993         prop_descriptor_init(&vm_completions, shift);
994         prop_descriptor_init(&vm_dirties, shift);
995 }
996
997 /**
998  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
999  * @mapping: address space structure to write
1000  * @start: starting page index
1001  * @end: ending page index (inclusive)
1002  *
1003  * This function scans the page range from @start to @end (inclusive) and tags
1004  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1005  * that write_cache_pages (or whoever calls this function) will then use
1006  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
1007  * used to avoid livelocking of writeback by a process steadily creating new
1008  * dirty pages in the file (thus it is important for this function to be quick
1009  * so that it can tag pages faster than a dirtying process can create them).
1010  */
1011 /*
1012  * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1013  */
1014 void tag_pages_for_writeback(struct address_space *mapping,
1015                              pgoff_t start, pgoff_t end)
1016 {
1017 #define WRITEBACK_TAG_BATCH 4096
1018         unsigned long tagged;
1019
1020         do {
1021                 spin_lock_irq(&mapping->tree_lock);
1022                 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1023                                 &start, end, WRITEBACK_TAG_BATCH,
1024                                 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1025                 spin_unlock_irq(&mapping->tree_lock);
1026                 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1027                 cond_resched();
1028                 /* We check 'start' to handle wrapping when end == ~0UL */
1029         } while (tagged >= WRITEBACK_TAG_BATCH && start);
1030 }
1031 EXPORT_SYMBOL(tag_pages_for_writeback);
1032
1033 /**
1034  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1035  * @mapping: address space structure to write
1036  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1037  * @writepage: function called for each page
1038  * @data: data passed to writepage function
1039  *
1040  * If a page is already under I/O, write_cache_pages() skips it, even
1041  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
1042  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
1043  * and msync() need to guarantee that all the data which was dirty at the time
1044  * the call was made get new I/O started against them.  If wbc->sync_mode is
1045  * WB_SYNC_ALL then we were called for data integrity and we must wait for
1046  * existing IO to complete.
1047  *
1048  * To avoid livelocks (when other process dirties new pages), we first tag
1049  * pages which should be written back with TOWRITE tag and only then start
1050  * writing them. For data-integrity sync we have to be careful so that we do
1051  * not miss some pages (e.g., because some other process has cleared TOWRITE
1052  * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1053  * by the process clearing the DIRTY tag (and submitting the page for IO).
1054  */
1055 int write_cache_pages(struct address_space *mapping,
1056                       struct writeback_control *wbc, writepage_t writepage,
1057                       void *data)
1058 {
1059         int ret = 0;
1060         int done = 0;
1061         struct pagevec pvec;
1062         int nr_pages;
1063         pgoff_t uninitialized_var(writeback_index);
1064         pgoff_t index;
1065         pgoff_t end;            /* Inclusive */
1066         pgoff_t done_index;
1067         int cycled;
1068         int range_whole = 0;
1069         int tag;
1070
1071         pagevec_init(&pvec, 0);
1072         if (wbc->range_cyclic) {
1073                 writeback_index = mapping->writeback_index; /* prev offset */
1074                 index = writeback_index;
1075                 if (index == 0)
1076                         cycled = 1;
1077                 else
1078                         cycled = 0;
1079                 end = -1;
1080         } else {
1081                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1082                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1083                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1084                         range_whole = 1;
1085                 cycled = 1; /* ignore range_cyclic tests */
1086         }
1087         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1088                 tag = PAGECACHE_TAG_TOWRITE;
1089         else
1090                 tag = PAGECACHE_TAG_DIRTY;
1091 retry:
1092         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1093                 tag_pages_for_writeback(mapping, index, end);
1094         done_index = index;
1095         while (!done && (index <= end)) {
1096                 int i;
1097
1098                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1099                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1100                 if (nr_pages == 0)
1101                         break;
1102
1103                 for (i = 0; i < nr_pages; i++) {
1104                         struct page *page = pvec.pages[i];
1105
1106                         /*
1107                          * At this point, the page may be truncated or
1108                          * invalidated (changing page->mapping to NULL), or
1109                          * even swizzled back from swapper_space to tmpfs file
1110                          * mapping. However, page->index will not change
1111                          * because we have a reference on the page.
1112                          */
1113                         if (page->index > end) {
1114                                 /*
1115                                  * can't be range_cyclic (1st pass) because
1116                                  * end == -1 in that case.
1117                                  */
1118                                 done = 1;
1119                                 break;
1120                         }
1121
1122                         done_index = page->index;
1123
1124                         lock_page(page);
1125
1126                         /*
1127                          * Page truncated or invalidated. We can freely skip it
1128                          * then, even for data integrity operations: the page
1129                          * has disappeared concurrently, so there could be no
1130                          * real expectation of this data interity operation
1131                          * even if there is now a new, dirty page at the same
1132                          * pagecache address.
1133                          */
1134                         if (unlikely(page->mapping != mapping)) {
1135 continue_unlock:
1136                                 unlock_page(page);
1137                                 continue;
1138                         }
1139
1140                         if (!PageDirty(page)) {
1141                                 /* someone wrote it for us */
1142                                 goto continue_unlock;
1143                         }
1144
1145                         if (PageWriteback(page)) {
1146                                 if (wbc->sync_mode != WB_SYNC_NONE)
1147                                         wait_on_page_writeback(page);
1148                                 else
1149                                         goto continue_unlock;
1150                         }
1151
1152                         BUG_ON(PageWriteback(page));
1153                         if (!clear_page_dirty_for_io(page))
1154                                 goto continue_unlock;
1155
1156                         trace_wbc_writepage(wbc, mapping->backing_dev_info);
1157                         ret = (*writepage)(page, wbc, data);
1158                         if (unlikely(ret)) {
1159                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1160                                         unlock_page(page);
1161                                         ret = 0;
1162                                 } else {
1163                                         /*
1164                                          * done_index is set past this page,
1165                                          * so media errors will not choke
1166                                          * background writeout for the entire
1167                                          * file. This has consequences for
1168                                          * range_cyclic semantics (ie. it may
1169                                          * not be suitable for data integrity
1170                                          * writeout).
1171                                          */
1172                                         done_index = page->index + 1;
1173                                         done = 1;
1174                                         break;
1175                                 }
1176                         }
1177
1178                         /*
1179                          * We stop writing back only if we are not doing
1180                          * integrity sync. In case of integrity sync we have to
1181                          * keep going until we have written all the pages
1182                          * we tagged for writeback prior to entering this loop.
1183                          */
1184                         if (--wbc->nr_to_write <= 0 &&
1185                             wbc->sync_mode == WB_SYNC_NONE) {
1186                                 done = 1;
1187                                 break;
1188                         }
1189                 }
1190                 pagevec_release(&pvec);
1191                 cond_resched();
1192         }
1193         if (!cycled && !done) {
1194                 /*
1195                  * range_cyclic:
1196                  * We hit the last page and there is more work to be done: wrap
1197                  * back to the start of the file
1198                  */
1199                 cycled = 1;
1200                 index = 0;
1201                 end = writeback_index - 1;
1202                 goto retry;
1203         }
1204         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1205                 mapping->writeback_index = done_index;
1206
1207         return ret;
1208 }
1209 EXPORT_SYMBOL(write_cache_pages);
1210
1211 /*
1212  * Function used by generic_writepages to call the real writepage
1213  * function and set the mapping flags on error
1214  */
1215 static int __writepage(struct page *page, struct writeback_control *wbc,
1216                        void *data)
1217 {
1218         struct address_space *mapping = data;
1219         int ret = mapping->a_ops->writepage(page, wbc);
1220         mapping_set_error(mapping, ret);
1221         return ret;
1222 }
1223
1224 /**
1225  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1226  * @mapping: address space structure to write
1227  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1228  *
1229  * This is a library function, which implements the writepages()
1230  * address_space_operation.
1231  */
1232 int generic_writepages(struct address_space *mapping,
1233                        struct writeback_control *wbc)
1234 {
1235         struct blk_plug plug;
1236         int ret;
1237
1238         /* deal with chardevs and other special file */
1239         if (!mapping->a_ops->writepage)
1240                 return 0;
1241
1242         blk_start_plug(&plug);
1243         ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1244         blk_finish_plug(&plug);
1245         return ret;
1246 }
1247
1248 EXPORT_SYMBOL(generic_writepages);
1249
1250 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1251 {
1252         int ret;
1253
1254         if (wbc->nr_to_write <= 0)
1255                 return 0;
1256         if (mapping->a_ops->writepages)
1257                 ret = mapping->a_ops->writepages(mapping, wbc);
1258         else
1259                 ret = generic_writepages(mapping, wbc);
1260         return ret;
1261 }
1262
1263 /**
1264  * write_one_page - write out a single page and optionally wait on I/O
1265  * @page: the page to write
1266  * @wait: if true, wait on writeout
1267  *
1268  * The page must be locked by the caller and will be unlocked upon return.
1269  *
1270  * write_one_page() returns a negative error code if I/O failed.
1271  */
1272 int write_one_page(struct page *page, int wait)
1273 {
1274         struct address_space *mapping = page->mapping;
1275         int ret = 0;
1276         struct writeback_control wbc = {
1277                 .sync_mode = WB_SYNC_ALL,
1278                 .nr_to_write = 1,
1279         };
1280
1281         BUG_ON(!PageLocked(page));
1282
1283         if (wait)
1284                 wait_on_page_writeback(page);
1285
1286         if (clear_page_dirty_for_io(page)) {
1287                 page_cache_get(page);
1288                 ret = mapping->a_ops->writepage(page, &wbc);
1289                 if (ret == 0 && wait) {
1290                         wait_on_page_writeback(page);
1291                         if (PageError(page))
1292                                 ret = -EIO;
1293                 }
1294                 page_cache_release(page);
1295         } else {
1296                 unlock_page(page);
1297         }
1298         return ret;
1299 }
1300 EXPORT_SYMBOL(write_one_page);
1301
1302 /*
1303  * For address_spaces which do not use buffers nor write back.
1304  */
1305 int __set_page_dirty_no_writeback(struct page *page)
1306 {
1307         if (!PageDirty(page))
1308                 return !TestSetPageDirty(page);
1309         return 0;
1310 }
1311
1312 /*
1313  * Helper function for set_page_dirty family.
1314  * NOTE: This relies on being atomic wrt interrupts.
1315  */
1316 void account_page_dirtied(struct page *page, struct address_space *mapping)
1317 {
1318         if (mapping_cap_account_dirty(mapping)) {
1319                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1320                 __inc_zone_page_state(page, NR_DIRTIED);
1321                 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1322                 task_dirty_inc(current);
1323                 task_io_account_write(PAGE_CACHE_SIZE);
1324         }
1325 }
1326 EXPORT_SYMBOL(account_page_dirtied);
1327
1328 /*
1329  * Helper function for set_page_writeback family.
1330  * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1331  * wrt interrupts.
1332  */
1333 void account_page_writeback(struct page *page)
1334 {
1335         inc_zone_page_state(page, NR_WRITEBACK);
1336         inc_zone_page_state(page, NR_WRITTEN);
1337 }
1338 EXPORT_SYMBOL(account_page_writeback);
1339
1340 /*
1341  * For address_spaces which do not use buffers.  Just tag the page as dirty in
1342  * its radix tree.
1343  *
1344  * This is also used when a single buffer is being dirtied: we want to set the
1345  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
1346  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1347  *
1348  * Most callers have locked the page, which pins the address_space in memory.
1349  * But zap_pte_range() does not lock the page, however in that case the
1350  * mapping is pinned by the vma's ->vm_file reference.
1351  *
1352  * We take care to handle the case where the page was truncated from the
1353  * mapping by re-checking page_mapping() inside tree_lock.
1354  */
1355 int __set_page_dirty_nobuffers(struct page *page)
1356 {
1357         if (!TestSetPageDirty(page)) {
1358                 struct address_space *mapping = page_mapping(page);
1359                 struct address_space *mapping2;
1360
1361                 if (!mapping)
1362                         return 1;
1363
1364                 spin_lock_irq(&mapping->tree_lock);
1365                 mapping2 = page_mapping(page);
1366                 if (mapping2) { /* Race with truncate? */
1367                         BUG_ON(mapping2 != mapping);
1368                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1369                         account_page_dirtied(page, mapping);
1370                         radix_tree_tag_set(&mapping->page_tree,
1371                                 page_index(page), PAGECACHE_TAG_DIRTY);
1372                 }
1373                 spin_unlock_irq(&mapping->tree_lock);
1374                 if (mapping->host) {
1375                         /* !PageAnon && !swapper_space */
1376                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1377                 }
1378                 return 1;
1379         }
1380         return 0;
1381 }
1382 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1383
1384 /*
1385  * When a writepage implementation decides that it doesn't want to write this
1386  * page for some reason, it should redirty the locked page via
1387  * redirty_page_for_writepage() and it should then unlock the page and return 0
1388  */
1389 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1390 {
1391         wbc->pages_skipped++;
1392         return __set_page_dirty_nobuffers(page);
1393 }
1394 EXPORT_SYMBOL(redirty_page_for_writepage);
1395
1396 /*
1397  * Dirty a page.
1398  *
1399  * For pages with a mapping this should be done under the page lock
1400  * for the benefit of asynchronous memory errors who prefer a consistent
1401  * dirty state. This rule can be broken in some special cases,
1402  * but should be better not to.
1403  *
1404  * If the mapping doesn't provide a set_page_dirty a_op, then
1405  * just fall through and assume that it wants buffer_heads.
1406  */
1407 int set_page_dirty(struct page *page)
1408 {
1409         struct address_space *mapping = page_mapping(page);
1410
1411         if (likely(mapping)) {
1412                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1413                 /*
1414                  * readahead/lru_deactivate_page could remain
1415                  * PG_readahead/PG_reclaim due to race with end_page_writeback
1416                  * About readahead, if the page is written, the flags would be
1417                  * reset. So no problem.
1418                  * About lru_deactivate_page, if the page is redirty, the flag
1419                  * will be reset. So no problem. but if the page is used by readahead
1420                  * it will confuse readahead and make it restart the size rampup
1421                  * process. But it's a trivial problem.
1422                  */
1423                 ClearPageReclaim(page);
1424 #ifdef CONFIG_BLOCK
1425                 if (!spd)
1426                         spd = __set_page_dirty_buffers;
1427 #endif
1428                 return (*spd)(page);
1429         }
1430         if (!PageDirty(page)) {
1431                 if (!TestSetPageDirty(page))
1432                         return 1;
1433         }
1434         return 0;
1435 }
1436 EXPORT_SYMBOL(set_page_dirty);
1437
1438 /*
1439  * set_page_dirty() is racy if the caller has no reference against
1440  * page->mapping->host, and if the page is unlocked.  This is because another
1441  * CPU could truncate the page off the mapping and then free the mapping.
1442  *
1443  * Usually, the page _is_ locked, or the caller is a user-space process which
1444  * holds a reference on the inode by having an open file.
1445  *
1446  * In other cases, the page should be locked before running set_page_dirty().
1447  */
1448 int set_page_dirty_lock(struct page *page)
1449 {
1450         int ret;
1451
1452         lock_page(page);
1453         ret = set_page_dirty(page);
1454         unlock_page(page);
1455         return ret;
1456 }
1457 EXPORT_SYMBOL(set_page_dirty_lock);
1458
1459 /*
1460  * Clear a page's dirty flag, while caring for dirty memory accounting.
1461  * Returns true if the page was previously dirty.
1462  *
1463  * This is for preparing to put the page under writeout.  We leave the page
1464  * tagged as dirty in the radix tree so that a concurrent write-for-sync
1465  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
1466  * implementation will run either set_page_writeback() or set_page_dirty(),
1467  * at which stage we bring the page's dirty flag and radix-tree dirty tag
1468  * back into sync.
1469  *
1470  * This incoherency between the page's dirty flag and radix-tree tag is
1471  * unfortunate, but it only exists while the page is locked.
1472  */
1473 int clear_page_dirty_for_io(struct page *page)
1474 {
1475         struct address_space *mapping = page_mapping(page);
1476
1477         BUG_ON(!PageLocked(page));
1478
1479         if (mapping && mapping_cap_account_dirty(mapping)) {
1480                 /*
1481                  * Yes, Virginia, this is indeed insane.
1482                  *
1483                  * We use this sequence to make sure that
1484                  *  (a) we account for dirty stats properly
1485                  *  (b) we tell the low-level filesystem to
1486                  *      mark the whole page dirty if it was
1487                  *      dirty in a pagetable. Only to then
1488                  *  (c) clean the page again and return 1 to
1489                  *      cause the writeback.
1490                  *
1491                  * This way we avoid all nasty races with the
1492                  * dirty bit in multiple places and clearing
1493                  * them concurrently from different threads.
1494                  *
1495                  * Note! Normally the "set_page_dirty(page)"
1496                  * has no effect on the actual dirty bit - since
1497                  * that will already usually be set. But we
1498                  * need the side effects, and it can help us
1499                  * avoid races.
1500                  *
1501                  * We basically use the page "master dirty bit"
1502                  * as a serialization point for all the different
1503                  * threads doing their things.
1504                  */
1505                 if (page_mkclean(page))
1506                         set_page_dirty(page);
1507                 /*
1508                  * We carefully synchronise fault handlers against
1509                  * installing a dirty pte and marking the page dirty
1510                  * at this point. We do this by having them hold the
1511                  * page lock at some point after installing their
1512                  * pte, but before marking the page dirty.
1513                  * Pages are always locked coming in here, so we get
1514                  * the desired exclusion. See mm/memory.c:do_wp_page()
1515                  * for more comments.
1516                  */
1517                 if (TestClearPageDirty(page)) {
1518                         dec_zone_page_state(page, NR_FILE_DIRTY);
1519                         dec_bdi_stat(mapping->backing_dev_info,
1520                                         BDI_RECLAIMABLE);
1521                         return 1;
1522                 }
1523                 return 0;
1524         }
1525         return TestClearPageDirty(page);
1526 }
1527 EXPORT_SYMBOL(clear_page_dirty_for_io);
1528
1529 int test_clear_page_writeback(struct page *page)
1530 {
1531         struct address_space *mapping = page_mapping(page);
1532         int ret;
1533
1534         if (mapping) {
1535                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1536                 unsigned long flags;
1537
1538                 spin_lock_irqsave(&mapping->tree_lock, flags);
1539                 ret = TestClearPageWriteback(page);
1540                 if (ret) {
1541                         radix_tree_tag_clear(&mapping->page_tree,
1542                                                 page_index(page),
1543                                                 PAGECACHE_TAG_WRITEBACK);
1544                         if (bdi_cap_account_writeback(bdi)) {
1545                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1546                                 __bdi_writeout_inc(bdi);
1547                         }
1548                 }
1549                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1550         } else {
1551                 ret = TestClearPageWriteback(page);
1552         }
1553         if (ret)
1554                 dec_zone_page_state(page, NR_WRITEBACK);
1555         return ret;
1556 }
1557
1558 int test_set_page_writeback(struct page *page)
1559 {
1560         struct address_space *mapping = page_mapping(page);
1561         int ret;
1562
1563         if (mapping) {
1564                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1565                 unsigned long flags;
1566
1567                 spin_lock_irqsave(&mapping->tree_lock, flags);
1568                 ret = TestSetPageWriteback(page);
1569                 if (!ret) {
1570                         radix_tree_tag_set(&mapping->page_tree,
1571                                                 page_index(page),
1572                                                 PAGECACHE_TAG_WRITEBACK);
1573                         if (bdi_cap_account_writeback(bdi))
1574                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1575                 }
1576                 if (!PageDirty(page))
1577                         radix_tree_tag_clear(&mapping->page_tree,
1578                                                 page_index(page),
1579                                                 PAGECACHE_TAG_DIRTY);
1580                 radix_tree_tag_clear(&mapping->page_tree,
1581                                      page_index(page),
1582                                      PAGECACHE_TAG_TOWRITE);
1583                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1584         } else {
1585                 ret = TestSetPageWriteback(page);
1586         }
1587         if (!ret)
1588                 account_page_writeback(page);
1589         return ret;
1590
1591 }
1592 EXPORT_SYMBOL(test_set_page_writeback);
1593
1594 /*
1595  * Return true if any of the pages in the mapping are marked with the
1596  * passed tag.
1597  */
1598 int mapping_tagged(struct address_space *mapping, int tag)
1599 {
1600         int ret;
1601         rcu_read_lock();
1602         ret = radix_tree_tagged(&mapping->page_tree, tag);
1603         rcu_read_unlock();
1604         return ret;
1605 }
1606 EXPORT_SYMBOL(mapping_tagged);