Merge branch 'for-next' of git://git.kernel.org/pub/scm/linux/kernel/git/nab/target...
[linux-3.10.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 #define TASK_LIMIT_FRACTION 8
285 static unsigned long task_dirty_limit(struct task_struct *tsk,
286                                        unsigned long bdi_dirty)
287 {
288         long numerator, denominator;
289         unsigned long dirty = bdi_dirty;
290         u64 inv = dirty / TASK_LIMIT_FRACTION;
291
292         task_dirties_fraction(tsk, &numerator, &denominator);
293         inv *= numerator;
294         do_div(inv, denominator);
295
296         dirty -= inv;
297
298         return max(dirty, bdi_dirty/2);
299 }
300
301 /* Minimum limit for any task */
302 static unsigned long task_min_dirty_limit(unsigned long bdi_dirty)
303 {
304         return bdi_dirty - bdi_dirty / TASK_LIMIT_FRACTION;
305 }
306
307 /*
308  * bdi_min_ratio keeps the sum of the minimum dirty shares of all
309  * registered backing devices, which, for obvious reasons, can not
310  * exceed 100%.
311  */
312 static unsigned int bdi_min_ratio;
313
314 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
315 {
316         int ret = 0;
317
318         spin_lock_bh(&bdi_lock);
319         if (min_ratio > bdi->max_ratio) {
320                 ret = -EINVAL;
321         } else {
322                 min_ratio -= bdi->min_ratio;
323                 if (bdi_min_ratio + min_ratio < 100) {
324                         bdi_min_ratio += min_ratio;
325                         bdi->min_ratio += min_ratio;
326                 } else {
327                         ret = -EINVAL;
328                 }
329         }
330         spin_unlock_bh(&bdi_lock);
331
332         return ret;
333 }
334
335 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
336 {
337         int ret = 0;
338
339         if (max_ratio > 100)
340                 return -EINVAL;
341
342         spin_lock_bh(&bdi_lock);
343         if (bdi->min_ratio > max_ratio) {
344                 ret = -EINVAL;
345         } else {
346                 bdi->max_ratio = max_ratio;
347                 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
348         }
349         spin_unlock_bh(&bdi_lock);
350
351         return ret;
352 }
353 EXPORT_SYMBOL(bdi_set_max_ratio);
354
355 /*
356  * Work out the current dirty-memory clamping and background writeout
357  * thresholds.
358  *
359  * The main aim here is to lower them aggressively if there is a lot of mapped
360  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
361  * pages.  It is better to clamp down on writers than to start swapping, and
362  * performing lots of scanning.
363  *
364  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
365  *
366  * We don't permit the clamping level to fall below 5% - that is getting rather
367  * excessive.
368  *
369  * We make sure that the background writeout level is below the adjusted
370  * clamping level.
371  */
372
373 static unsigned long highmem_dirtyable_memory(unsigned long total)
374 {
375 #ifdef CONFIG_HIGHMEM
376         int node;
377         unsigned long x = 0;
378
379         for_each_node_state(node, N_HIGH_MEMORY) {
380                 struct zone *z =
381                         &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
382
383                 x += zone_page_state(z, NR_FREE_PAGES) +
384                      zone_reclaimable_pages(z);
385         }
386         /*
387          * Make sure that the number of highmem pages is never larger
388          * than the number of the total dirtyable memory. This can only
389          * occur in very strange VM situations but we want to make sure
390          * that this does not occur.
391          */
392         return min(x, total);
393 #else
394         return 0;
395 #endif
396 }
397
398 /**
399  * determine_dirtyable_memory - amount of memory that may be used
400  *
401  * Returns the numebr of pages that can currently be freed and used
402  * by the kernel for direct mappings.
403  */
404 unsigned long determine_dirtyable_memory(void)
405 {
406         unsigned long x;
407
408         x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
409
410         if (!vm_highmem_is_dirtyable)
411                 x -= highmem_dirtyable_memory(x);
412
413         return x + 1;   /* Ensure that we never return 0 */
414 }
415
416 static unsigned long hard_dirty_limit(unsigned long thresh)
417 {
418         return max(thresh, global_dirty_limit);
419 }
420
421 /*
422  * global_dirty_limits - background-writeback and dirty-throttling thresholds
423  *
424  * Calculate the dirty thresholds based on sysctl parameters
425  * - vm.dirty_background_ratio  or  vm.dirty_background_bytes
426  * - vm.dirty_ratio             or  vm.dirty_bytes
427  * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
428  * real-time tasks.
429  */
430 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
431 {
432         unsigned long background;
433         unsigned long dirty;
434         unsigned long uninitialized_var(available_memory);
435         struct task_struct *tsk;
436
437         if (!vm_dirty_bytes || !dirty_background_bytes)
438                 available_memory = determine_dirtyable_memory();
439
440         if (vm_dirty_bytes)
441                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
442         else
443                 dirty = (vm_dirty_ratio * available_memory) / 100;
444
445         if (dirty_background_bytes)
446                 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
447         else
448                 background = (dirty_background_ratio * available_memory) / 100;
449
450         if (background >= dirty)
451                 background = dirty / 2;
452         tsk = current;
453         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
454                 background += background / 4;
455                 dirty += dirty / 4;
456         }
457         *pbackground = background;
458         *pdirty = dirty;
459         trace_global_dirty_state(background, dirty);
460 }
461
462 /**
463  * bdi_dirty_limit - @bdi's share of dirty throttling threshold
464  * @bdi: the backing_dev_info to query
465  * @dirty: global dirty limit in pages
466  *
467  * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
468  * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
469  * And the "limit" in the name is not seriously taken as hard limit in
470  * balance_dirty_pages().
471  *
472  * It allocates high/low dirty limits to fast/slow devices, in order to prevent
473  * - starving fast devices
474  * - piling up dirty pages (that will take long time to sync) on slow devices
475  *
476  * The bdi's share of dirty limit will be adapting to its throughput and
477  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
478  */
479 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
480 {
481         u64 bdi_dirty;
482         long numerator, denominator;
483
484         /*
485          * Calculate this BDI's share of the dirty ratio.
486          */
487         bdi_writeout_fraction(bdi, &numerator, &denominator);
488
489         bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
490         bdi_dirty *= numerator;
491         do_div(bdi_dirty, denominator);
492
493         bdi_dirty += (dirty * bdi->min_ratio) / 100;
494         if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
495                 bdi_dirty = dirty * bdi->max_ratio / 100;
496
497         return bdi_dirty;
498 }
499
500 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
501                                        unsigned long elapsed,
502                                        unsigned long written)
503 {
504         const unsigned long period = roundup_pow_of_two(3 * HZ);
505         unsigned long avg = bdi->avg_write_bandwidth;
506         unsigned long old = bdi->write_bandwidth;
507         u64 bw;
508
509         /*
510          * bw = written * HZ / elapsed
511          *
512          *                   bw * elapsed + write_bandwidth * (period - elapsed)
513          * write_bandwidth = ---------------------------------------------------
514          *                                          period
515          */
516         bw = written - bdi->written_stamp;
517         bw *= HZ;
518         if (unlikely(elapsed > period)) {
519                 do_div(bw, elapsed);
520                 avg = bw;
521                 goto out;
522         }
523         bw += (u64)bdi->write_bandwidth * (period - elapsed);
524         bw >>= ilog2(period);
525
526         /*
527          * one more level of smoothing, for filtering out sudden spikes
528          */
529         if (avg > old && old >= (unsigned long)bw)
530                 avg -= (avg - old) >> 3;
531
532         if (avg < old && old <= (unsigned long)bw)
533                 avg += (old - avg) >> 3;
534
535 out:
536         bdi->write_bandwidth = bw;
537         bdi->avg_write_bandwidth = avg;
538 }
539
540 /*
541  * The global dirtyable memory and dirty threshold could be suddenly knocked
542  * down by a large amount (eg. on the startup of KVM in a swapless system).
543  * This may throw the system into deep dirty exceeded state and throttle
544  * heavy/light dirtiers alike. To retain good responsiveness, maintain
545  * global_dirty_limit for tracking slowly down to the knocked down dirty
546  * threshold.
547  */
548 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
549 {
550         unsigned long limit = global_dirty_limit;
551
552         /*
553          * Follow up in one step.
554          */
555         if (limit < thresh) {
556                 limit = thresh;
557                 goto update;
558         }
559
560         /*
561          * Follow down slowly. Use the higher one as the target, because thresh
562          * may drop below dirty. This is exactly the reason to introduce
563          * global_dirty_limit which is guaranteed to lie above the dirty pages.
564          */
565         thresh = max(thresh, dirty);
566         if (limit > thresh) {
567                 limit -= (limit - thresh) >> 5;
568                 goto update;
569         }
570         return;
571 update:
572         global_dirty_limit = limit;
573 }
574
575 static void global_update_bandwidth(unsigned long thresh,
576                                     unsigned long dirty,
577                                     unsigned long now)
578 {
579         static DEFINE_SPINLOCK(dirty_lock);
580         static unsigned long update_time;
581
582         /*
583          * check locklessly first to optimize away locking for the most time
584          */
585         if (time_before(now, update_time + BANDWIDTH_INTERVAL))
586                 return;
587
588         spin_lock(&dirty_lock);
589         if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
590                 update_dirty_limit(thresh, dirty);
591                 update_time = now;
592         }
593         spin_unlock(&dirty_lock);
594 }
595
596 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
597                             unsigned long thresh,
598                             unsigned long dirty,
599                             unsigned long bdi_thresh,
600                             unsigned long bdi_dirty,
601                             unsigned long start_time)
602 {
603         unsigned long now = jiffies;
604         unsigned long elapsed = now - bdi->bw_time_stamp;
605         unsigned long written;
606
607         /*
608          * rate-limit, only update once every 200ms.
609          */
610         if (elapsed < BANDWIDTH_INTERVAL)
611                 return;
612
613         written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
614
615         /*
616          * Skip quiet periods when disk bandwidth is under-utilized.
617          * (at least 1s idle time between two flusher runs)
618          */
619         if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
620                 goto snapshot;
621
622         if (thresh)
623                 global_update_bandwidth(thresh, dirty, now);
624
625         bdi_update_write_bandwidth(bdi, elapsed, written);
626
627 snapshot:
628         bdi->written_stamp = written;
629         bdi->bw_time_stamp = now;
630 }
631
632 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
633                                  unsigned long thresh,
634                                  unsigned long dirty,
635                                  unsigned long bdi_thresh,
636                                  unsigned long bdi_dirty,
637                                  unsigned long start_time)
638 {
639         if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
640                 return;
641         spin_lock(&bdi->wb.list_lock);
642         __bdi_update_bandwidth(bdi, thresh, dirty, bdi_thresh, bdi_dirty,
643                                start_time);
644         spin_unlock(&bdi->wb.list_lock);
645 }
646
647 /*
648  * balance_dirty_pages() must be called by processes which are generating dirty
649  * data.  It looks at the number of dirty pages in the machine and will force
650  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
651  * If we're over `background_thresh' then the writeback threads are woken to
652  * perform some writeout.
653  */
654 static void balance_dirty_pages(struct address_space *mapping,
655                                 unsigned long write_chunk)
656 {
657         unsigned long nr_reclaimable, bdi_nr_reclaimable;
658         unsigned long nr_dirty;  /* = file_dirty + writeback + unstable_nfs */
659         unsigned long bdi_dirty;
660         unsigned long background_thresh;
661         unsigned long dirty_thresh;
662         unsigned long bdi_thresh;
663         unsigned long task_bdi_thresh;
664         unsigned long min_task_bdi_thresh;
665         unsigned long pages_written = 0;
666         unsigned long pause = 1;
667         bool dirty_exceeded = false;
668         bool clear_dirty_exceeded = true;
669         struct backing_dev_info *bdi = mapping->backing_dev_info;
670         unsigned long start_time = jiffies;
671
672         for (;;) {
673                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
674                                         global_page_state(NR_UNSTABLE_NFS);
675                 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
676
677                 global_dirty_limits(&background_thresh, &dirty_thresh);
678
679                 /*
680                  * Throttle it only when the background writeback cannot
681                  * catch-up. This avoids (excessively) small writeouts
682                  * when the bdi limits are ramping up.
683                  */
684                 if (nr_dirty <= (background_thresh + dirty_thresh) / 2)
685                         break;
686
687                 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
688                 min_task_bdi_thresh = task_min_dirty_limit(bdi_thresh);
689                 task_bdi_thresh = task_dirty_limit(current, bdi_thresh);
690
691                 /*
692                  * In order to avoid the stacked BDI deadlock we need
693                  * to ensure we accurately count the 'dirty' pages when
694                  * the threshold is low.
695                  *
696                  * Otherwise it would be possible to get thresh+n pages
697                  * reported dirty, even though there are thresh-m pages
698                  * actually dirty; with m+n sitting in the percpu
699                  * deltas.
700                  */
701                 if (task_bdi_thresh < 2 * bdi_stat_error(bdi)) {
702                         bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
703                         bdi_dirty = bdi_nr_reclaimable +
704                                     bdi_stat_sum(bdi, BDI_WRITEBACK);
705                 } else {
706                         bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
707                         bdi_dirty = bdi_nr_reclaimable +
708                                     bdi_stat(bdi, BDI_WRITEBACK);
709                 }
710
711                 /*
712                  * The bdi thresh is somehow "soft" limit derived from the
713                  * global "hard" limit. The former helps to prevent heavy IO
714                  * bdi or process from holding back light ones; The latter is
715                  * the last resort safeguard.
716                  */
717                 dirty_exceeded = (bdi_dirty > task_bdi_thresh) ||
718                                   (nr_dirty > dirty_thresh);
719                 clear_dirty_exceeded = (bdi_dirty <= min_task_bdi_thresh) &&
720                                         (nr_dirty <= dirty_thresh);
721
722                 if (!dirty_exceeded)
723                         break;
724
725                 if (!bdi->dirty_exceeded)
726                         bdi->dirty_exceeded = 1;
727
728                 bdi_update_bandwidth(bdi, dirty_thresh, nr_dirty,
729                                      bdi_thresh, bdi_dirty, start_time);
730
731                 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
732                  * Unstable writes are a feature of certain networked
733                  * filesystems (i.e. NFS) in which data may have been
734                  * written to the server's write cache, but has not yet
735                  * been flushed to permanent storage.
736                  * Only move pages to writeback if this bdi is over its
737                  * threshold otherwise wait until the disk writes catch
738                  * up.
739                  */
740                 trace_balance_dirty_start(bdi);
741                 if (bdi_nr_reclaimable > task_bdi_thresh) {
742                         pages_written += writeback_inodes_wb(&bdi->wb,
743                                                              write_chunk);
744                         trace_balance_dirty_written(bdi, pages_written);
745                         if (pages_written >= write_chunk)
746                                 break;          /* We've done our duty */
747                 }
748                 __set_current_state(TASK_UNINTERRUPTIBLE);
749                 io_schedule_timeout(pause);
750                 trace_balance_dirty_wait(bdi);
751
752                 dirty_thresh = hard_dirty_limit(dirty_thresh);
753                 /*
754                  * max-pause area. If dirty exceeded but still within this
755                  * area, no need to sleep for more than 200ms: (a) 8 pages per
756                  * 200ms is typically more than enough to curb heavy dirtiers;
757                  * (b) the pause time limit makes the dirtiers more responsive.
758                  */
759                 if (nr_dirty < dirty_thresh &&
760                     bdi_dirty < (task_bdi_thresh + bdi_thresh) / 2 &&
761                     time_after(jiffies, start_time + MAX_PAUSE))
762                         break;
763
764                 /*
765                  * Increase the delay for each loop, up to our previous
766                  * default of taking a 100ms nap.
767                  */
768                 pause <<= 1;
769                 if (pause > HZ / 10)
770                         pause = HZ / 10;
771         }
772
773         /* Clear dirty_exceeded flag only when no task can exceed the limit */
774         if (clear_dirty_exceeded && bdi->dirty_exceeded)
775                 bdi->dirty_exceeded = 0;
776
777         if (writeback_in_progress(bdi))
778                 return;
779
780         /*
781          * In laptop mode, we wait until hitting the higher threshold before
782          * starting background writeout, and then write out all the way down
783          * to the lower threshold.  So slow writers cause minimal disk activity.
784          *
785          * In normal mode, we start background writeout at the lower
786          * background_thresh, to keep the amount of dirty memory low.
787          */
788         if ((laptop_mode && pages_written) ||
789             (!laptop_mode && (nr_reclaimable > background_thresh)))
790                 bdi_start_background_writeback(bdi);
791 }
792
793 void set_page_dirty_balance(struct page *page, int page_mkwrite)
794 {
795         if (set_page_dirty(page) || page_mkwrite) {
796                 struct address_space *mapping = page_mapping(page);
797
798                 if (mapping)
799                         balance_dirty_pages_ratelimited(mapping);
800         }
801 }
802
803 static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
804
805 /**
806  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
807  * @mapping: address_space which was dirtied
808  * @nr_pages_dirtied: number of pages which the caller has just dirtied
809  *
810  * Processes which are dirtying memory should call in here once for each page
811  * which was newly dirtied.  The function will periodically check the system's
812  * dirty state and will initiate writeback if needed.
813  *
814  * On really big machines, get_writeback_state is expensive, so try to avoid
815  * calling it too often (ratelimiting).  But once we're over the dirty memory
816  * limit we decrease the ratelimiting by a lot, to prevent individual processes
817  * from overshooting the limit by (ratelimit_pages) each.
818  */
819 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
820                                         unsigned long nr_pages_dirtied)
821 {
822         struct backing_dev_info *bdi = mapping->backing_dev_info;
823         unsigned long ratelimit;
824         unsigned long *p;
825
826         if (!bdi_cap_account_dirty(bdi))
827                 return;
828
829         ratelimit = ratelimit_pages;
830         if (mapping->backing_dev_info->dirty_exceeded)
831                 ratelimit = 8;
832
833         /*
834          * Check the rate limiting. Also, we do not want to throttle real-time
835          * tasks in balance_dirty_pages(). Period.
836          */
837         preempt_disable();
838         p =  &__get_cpu_var(bdp_ratelimits);
839         *p += nr_pages_dirtied;
840         if (unlikely(*p >= ratelimit)) {
841                 ratelimit = sync_writeback_pages(*p);
842                 *p = 0;
843                 preempt_enable();
844                 balance_dirty_pages(mapping, ratelimit);
845                 return;
846         }
847         preempt_enable();
848 }
849 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
850
851 void throttle_vm_writeout(gfp_t gfp_mask)
852 {
853         unsigned long background_thresh;
854         unsigned long dirty_thresh;
855
856         for ( ; ; ) {
857                 global_dirty_limits(&background_thresh, &dirty_thresh);
858
859                 /*
860                  * Boost the allowable dirty threshold a bit for page
861                  * allocators so they don't get DoS'ed by heavy writers
862                  */
863                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
864
865                 if (global_page_state(NR_UNSTABLE_NFS) +
866                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
867                                 break;
868                 congestion_wait(BLK_RW_ASYNC, HZ/10);
869
870                 /*
871                  * The caller might hold locks which can prevent IO completion
872                  * or progress in the filesystem.  So we cannot just sit here
873                  * waiting for IO to complete.
874                  */
875                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
876                         break;
877         }
878 }
879
880 /*
881  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
882  */
883 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
884         void __user *buffer, size_t *length, loff_t *ppos)
885 {
886         proc_dointvec(table, write, buffer, length, ppos);
887         bdi_arm_supers_timer();
888         return 0;
889 }
890
891 #ifdef CONFIG_BLOCK
892 void laptop_mode_timer_fn(unsigned long data)
893 {
894         struct request_queue *q = (struct request_queue *)data;
895         int nr_pages = global_page_state(NR_FILE_DIRTY) +
896                 global_page_state(NR_UNSTABLE_NFS);
897
898         /*
899          * We want to write everything out, not just down to the dirty
900          * threshold
901          */
902         if (bdi_has_dirty_io(&q->backing_dev_info))
903                 bdi_start_writeback(&q->backing_dev_info, nr_pages);
904 }
905
906 /*
907  * We've spun up the disk and we're in laptop mode: schedule writeback
908  * of all dirty data a few seconds from now.  If the flush is already scheduled
909  * then push it back - the user is still using the disk.
910  */
911 void laptop_io_completion(struct backing_dev_info *info)
912 {
913         mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
914 }
915
916 /*
917  * We're in laptop mode and we've just synced. The sync's writes will have
918  * caused another writeback to be scheduled by laptop_io_completion.
919  * Nothing needs to be written back anymore, so we unschedule the writeback.
920  */
921 void laptop_sync_completion(void)
922 {
923         struct backing_dev_info *bdi;
924
925         rcu_read_lock();
926
927         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
928                 del_timer(&bdi->laptop_mode_wb_timer);
929
930         rcu_read_unlock();
931 }
932 #endif
933
934 /*
935  * If ratelimit_pages is too high then we can get into dirty-data overload
936  * if a large number of processes all perform writes at the same time.
937  * If it is too low then SMP machines will call the (expensive)
938  * get_writeback_state too often.
939  *
940  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
941  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
942  * thresholds before writeback cuts in.
943  *
944  * But the limit should not be set too high.  Because it also controls the
945  * amount of memory which the balance_dirty_pages() caller has to write back.
946  * If this is too large then the caller will block on the IO queue all the
947  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
948  * will write six megabyte chunks, max.
949  */
950
951 void writeback_set_ratelimit(void)
952 {
953         ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
954         if (ratelimit_pages < 16)
955                 ratelimit_pages = 16;
956         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
957                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
958 }
959
960 static int __cpuinit
961 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
962 {
963         writeback_set_ratelimit();
964         return NOTIFY_DONE;
965 }
966
967 static struct notifier_block __cpuinitdata ratelimit_nb = {
968         .notifier_call  = ratelimit_handler,
969         .next           = NULL,
970 };
971
972 /*
973  * Called early on to tune the page writeback dirty limits.
974  *
975  * We used to scale dirty pages according to how total memory
976  * related to pages that could be allocated for buffers (by
977  * comparing nr_free_buffer_pages() to vm_total_pages.
978  *
979  * However, that was when we used "dirty_ratio" to scale with
980  * all memory, and we don't do that any more. "dirty_ratio"
981  * is now applied to total non-HIGHPAGE memory (by subtracting
982  * totalhigh_pages from vm_total_pages), and as such we can't
983  * get into the old insane situation any more where we had
984  * large amounts of dirty pages compared to a small amount of
985  * non-HIGHMEM memory.
986  *
987  * But we might still want to scale the dirty_ratio by how
988  * much memory the box has..
989  */
990 void __init page_writeback_init(void)
991 {
992         int shift;
993
994         writeback_set_ratelimit();
995         register_cpu_notifier(&ratelimit_nb);
996
997         shift = calc_period_shift();
998         prop_descriptor_init(&vm_completions, shift);
999         prop_descriptor_init(&vm_dirties, shift);
1000 }
1001
1002 /**
1003  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1004  * @mapping: address space structure to write
1005  * @start: starting page index
1006  * @end: ending page index (inclusive)
1007  *
1008  * This function scans the page range from @start to @end (inclusive) and tags
1009  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1010  * that write_cache_pages (or whoever calls this function) will then use
1011  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
1012  * used to avoid livelocking of writeback by a process steadily creating new
1013  * dirty pages in the file (thus it is important for this function to be quick
1014  * so that it can tag pages faster than a dirtying process can create them).
1015  */
1016 /*
1017  * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1018  */
1019 void tag_pages_for_writeback(struct address_space *mapping,
1020                              pgoff_t start, pgoff_t end)
1021 {
1022 #define WRITEBACK_TAG_BATCH 4096
1023         unsigned long tagged;
1024
1025         do {
1026                 spin_lock_irq(&mapping->tree_lock);
1027                 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1028                                 &start, end, WRITEBACK_TAG_BATCH,
1029                                 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1030                 spin_unlock_irq(&mapping->tree_lock);
1031                 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1032                 cond_resched();
1033                 /* We check 'start' to handle wrapping when end == ~0UL */
1034         } while (tagged >= WRITEBACK_TAG_BATCH && start);
1035 }
1036 EXPORT_SYMBOL(tag_pages_for_writeback);
1037
1038 /**
1039  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1040  * @mapping: address space structure to write
1041  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1042  * @writepage: function called for each page
1043  * @data: data passed to writepage function
1044  *
1045  * If a page is already under I/O, write_cache_pages() skips it, even
1046  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
1047  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
1048  * and msync() need to guarantee that all the data which was dirty at the time
1049  * the call was made get new I/O started against them.  If wbc->sync_mode is
1050  * WB_SYNC_ALL then we were called for data integrity and we must wait for
1051  * existing IO to complete.
1052  *
1053  * To avoid livelocks (when other process dirties new pages), we first tag
1054  * pages which should be written back with TOWRITE tag and only then start
1055  * writing them. For data-integrity sync we have to be careful so that we do
1056  * not miss some pages (e.g., because some other process has cleared TOWRITE
1057  * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1058  * by the process clearing the DIRTY tag (and submitting the page for IO).
1059  */
1060 int write_cache_pages(struct address_space *mapping,
1061                       struct writeback_control *wbc, writepage_t writepage,
1062                       void *data)
1063 {
1064         int ret = 0;
1065         int done = 0;
1066         struct pagevec pvec;
1067         int nr_pages;
1068         pgoff_t uninitialized_var(writeback_index);
1069         pgoff_t index;
1070         pgoff_t end;            /* Inclusive */
1071         pgoff_t done_index;
1072         int cycled;
1073         int range_whole = 0;
1074         int tag;
1075
1076         pagevec_init(&pvec, 0);
1077         if (wbc->range_cyclic) {
1078                 writeback_index = mapping->writeback_index; /* prev offset */
1079                 index = writeback_index;
1080                 if (index == 0)
1081                         cycled = 1;
1082                 else
1083                         cycled = 0;
1084                 end = -1;
1085         } else {
1086                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1087                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1088                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1089                         range_whole = 1;
1090                 cycled = 1; /* ignore range_cyclic tests */
1091         }
1092         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1093                 tag = PAGECACHE_TAG_TOWRITE;
1094         else
1095                 tag = PAGECACHE_TAG_DIRTY;
1096 retry:
1097         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1098                 tag_pages_for_writeback(mapping, index, end);
1099         done_index = index;
1100         while (!done && (index <= end)) {
1101                 int i;
1102
1103                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1104                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1105                 if (nr_pages == 0)
1106                         break;
1107
1108                 for (i = 0; i < nr_pages; i++) {
1109                         struct page *page = pvec.pages[i];
1110
1111                         /*
1112                          * At this point, the page may be truncated or
1113                          * invalidated (changing page->mapping to NULL), or
1114                          * even swizzled back from swapper_space to tmpfs file
1115                          * mapping. However, page->index will not change
1116                          * because we have a reference on the page.
1117                          */
1118                         if (page->index > end) {
1119                                 /*
1120                                  * can't be range_cyclic (1st pass) because
1121                                  * end == -1 in that case.
1122                                  */
1123                                 done = 1;
1124                                 break;
1125                         }
1126
1127                         done_index = page->index;
1128
1129                         lock_page(page);
1130
1131                         /*
1132                          * Page truncated or invalidated. We can freely skip it
1133                          * then, even for data integrity operations: the page
1134                          * has disappeared concurrently, so there could be no
1135                          * real expectation of this data interity operation
1136                          * even if there is now a new, dirty page at the same
1137                          * pagecache address.
1138                          */
1139                         if (unlikely(page->mapping != mapping)) {
1140 continue_unlock:
1141                                 unlock_page(page);
1142                                 continue;
1143                         }
1144
1145                         if (!PageDirty(page)) {
1146                                 /* someone wrote it for us */
1147                                 goto continue_unlock;
1148                         }
1149
1150                         if (PageWriteback(page)) {
1151                                 if (wbc->sync_mode != WB_SYNC_NONE)
1152                                         wait_on_page_writeback(page);
1153                                 else
1154                                         goto continue_unlock;
1155                         }
1156
1157                         BUG_ON(PageWriteback(page));
1158                         if (!clear_page_dirty_for_io(page))
1159                                 goto continue_unlock;
1160
1161                         trace_wbc_writepage(wbc, mapping->backing_dev_info);
1162                         ret = (*writepage)(page, wbc, data);
1163                         if (unlikely(ret)) {
1164                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1165                                         unlock_page(page);
1166                                         ret = 0;
1167                                 } else {
1168                                         /*
1169                                          * done_index is set past this page,
1170                                          * so media errors will not choke
1171                                          * background writeout for the entire
1172                                          * file. This has consequences for
1173                                          * range_cyclic semantics (ie. it may
1174                                          * not be suitable for data integrity
1175                                          * writeout).
1176                                          */
1177                                         done_index = page->index + 1;
1178                                         done = 1;
1179                                         break;
1180                                 }
1181                         }
1182
1183                         /*
1184                          * We stop writing back only if we are not doing
1185                          * integrity sync. In case of integrity sync we have to
1186                          * keep going until we have written all the pages
1187                          * we tagged for writeback prior to entering this loop.
1188                          */
1189                         if (--wbc->nr_to_write <= 0 &&
1190                             wbc->sync_mode == WB_SYNC_NONE) {
1191                                 done = 1;
1192                                 break;
1193                         }
1194                 }
1195                 pagevec_release(&pvec);
1196                 cond_resched();
1197         }
1198         if (!cycled && !done) {
1199                 /*
1200                  * range_cyclic:
1201                  * We hit the last page and there is more work to be done: wrap
1202                  * back to the start of the file
1203                  */
1204                 cycled = 1;
1205                 index = 0;
1206                 end = writeback_index - 1;
1207                 goto retry;
1208         }
1209         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1210                 mapping->writeback_index = done_index;
1211
1212         return ret;
1213 }
1214 EXPORT_SYMBOL(write_cache_pages);
1215
1216 /*
1217  * Function used by generic_writepages to call the real writepage
1218  * function and set the mapping flags on error
1219  */
1220 static int __writepage(struct page *page, struct writeback_control *wbc,
1221                        void *data)
1222 {
1223         struct address_space *mapping = data;
1224         int ret = mapping->a_ops->writepage(page, wbc);
1225         mapping_set_error(mapping, ret);
1226         return ret;
1227 }
1228
1229 /**
1230  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1231  * @mapping: address space structure to write
1232  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1233  *
1234  * This is a library function, which implements the writepages()
1235  * address_space_operation.
1236  */
1237 int generic_writepages(struct address_space *mapping,
1238                        struct writeback_control *wbc)
1239 {
1240         struct blk_plug plug;
1241         int ret;
1242
1243         /* deal with chardevs and other special file */
1244         if (!mapping->a_ops->writepage)
1245                 return 0;
1246
1247         blk_start_plug(&plug);
1248         ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1249         blk_finish_plug(&plug);
1250         return ret;
1251 }
1252
1253 EXPORT_SYMBOL(generic_writepages);
1254
1255 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1256 {
1257         int ret;
1258
1259         if (wbc->nr_to_write <= 0)
1260                 return 0;
1261         if (mapping->a_ops->writepages)
1262                 ret = mapping->a_ops->writepages(mapping, wbc);
1263         else
1264                 ret = generic_writepages(mapping, wbc);
1265         return ret;
1266 }
1267
1268 /**
1269  * write_one_page - write out a single page and optionally wait on I/O
1270  * @page: the page to write
1271  * @wait: if true, wait on writeout
1272  *
1273  * The page must be locked by the caller and will be unlocked upon return.
1274  *
1275  * write_one_page() returns a negative error code if I/O failed.
1276  */
1277 int write_one_page(struct page *page, int wait)
1278 {
1279         struct address_space *mapping = page->mapping;
1280         int ret = 0;
1281         struct writeback_control wbc = {
1282                 .sync_mode = WB_SYNC_ALL,
1283                 .nr_to_write = 1,
1284         };
1285
1286         BUG_ON(!PageLocked(page));
1287
1288         if (wait)
1289                 wait_on_page_writeback(page);
1290
1291         if (clear_page_dirty_for_io(page)) {
1292                 page_cache_get(page);
1293                 ret = mapping->a_ops->writepage(page, &wbc);
1294                 if (ret == 0 && wait) {
1295                         wait_on_page_writeback(page);
1296                         if (PageError(page))
1297                                 ret = -EIO;
1298                 }
1299                 page_cache_release(page);
1300         } else {
1301                 unlock_page(page);
1302         }
1303         return ret;
1304 }
1305 EXPORT_SYMBOL(write_one_page);
1306
1307 /*
1308  * For address_spaces which do not use buffers nor write back.
1309  */
1310 int __set_page_dirty_no_writeback(struct page *page)
1311 {
1312         if (!PageDirty(page))
1313                 return !TestSetPageDirty(page);
1314         return 0;
1315 }
1316
1317 /*
1318  * Helper function for set_page_dirty family.
1319  * NOTE: This relies on being atomic wrt interrupts.
1320  */
1321 void account_page_dirtied(struct page *page, struct address_space *mapping)
1322 {
1323         if (mapping_cap_account_dirty(mapping)) {
1324                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1325                 __inc_zone_page_state(page, NR_DIRTIED);
1326                 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1327                 task_dirty_inc(current);
1328                 task_io_account_write(PAGE_CACHE_SIZE);
1329         }
1330 }
1331 EXPORT_SYMBOL(account_page_dirtied);
1332
1333 /*
1334  * Helper function for set_page_writeback family.
1335  * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1336  * wrt interrupts.
1337  */
1338 void account_page_writeback(struct page *page)
1339 {
1340         inc_zone_page_state(page, NR_WRITEBACK);
1341 }
1342 EXPORT_SYMBOL(account_page_writeback);
1343
1344 /*
1345  * For address_spaces which do not use buffers.  Just tag the page as dirty in
1346  * its radix tree.
1347  *
1348  * This is also used when a single buffer is being dirtied: we want to set the
1349  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
1350  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1351  *
1352  * Most callers have locked the page, which pins the address_space in memory.
1353  * But zap_pte_range() does not lock the page, however in that case the
1354  * mapping is pinned by the vma's ->vm_file reference.
1355  *
1356  * We take care to handle the case where the page was truncated from the
1357  * mapping by re-checking page_mapping() inside tree_lock.
1358  */
1359 int __set_page_dirty_nobuffers(struct page *page)
1360 {
1361         if (!TestSetPageDirty(page)) {
1362                 struct address_space *mapping = page_mapping(page);
1363                 struct address_space *mapping2;
1364
1365                 if (!mapping)
1366                         return 1;
1367
1368                 spin_lock_irq(&mapping->tree_lock);
1369                 mapping2 = page_mapping(page);
1370                 if (mapping2) { /* Race with truncate? */
1371                         BUG_ON(mapping2 != mapping);
1372                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1373                         account_page_dirtied(page, mapping);
1374                         radix_tree_tag_set(&mapping->page_tree,
1375                                 page_index(page), PAGECACHE_TAG_DIRTY);
1376                 }
1377                 spin_unlock_irq(&mapping->tree_lock);
1378                 if (mapping->host) {
1379                         /* !PageAnon && !swapper_space */
1380                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1381                 }
1382                 return 1;
1383         }
1384         return 0;
1385 }
1386 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1387
1388 /*
1389  * When a writepage implementation decides that it doesn't want to write this
1390  * page for some reason, it should redirty the locked page via
1391  * redirty_page_for_writepage() and it should then unlock the page and return 0
1392  */
1393 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1394 {
1395         wbc->pages_skipped++;
1396         return __set_page_dirty_nobuffers(page);
1397 }
1398 EXPORT_SYMBOL(redirty_page_for_writepage);
1399
1400 /*
1401  * Dirty a page.
1402  *
1403  * For pages with a mapping this should be done under the page lock
1404  * for the benefit of asynchronous memory errors who prefer a consistent
1405  * dirty state. This rule can be broken in some special cases,
1406  * but should be better not to.
1407  *
1408  * If the mapping doesn't provide a set_page_dirty a_op, then
1409  * just fall through and assume that it wants buffer_heads.
1410  */
1411 int set_page_dirty(struct page *page)
1412 {
1413         struct address_space *mapping = page_mapping(page);
1414
1415         if (likely(mapping)) {
1416                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1417                 /*
1418                  * readahead/lru_deactivate_page could remain
1419                  * PG_readahead/PG_reclaim due to race with end_page_writeback
1420                  * About readahead, if the page is written, the flags would be
1421                  * reset. So no problem.
1422                  * About lru_deactivate_page, if the page is redirty, the flag
1423                  * will be reset. So no problem. but if the page is used by readahead
1424                  * it will confuse readahead and make it restart the size rampup
1425                  * process. But it's a trivial problem.
1426                  */
1427                 ClearPageReclaim(page);
1428 #ifdef CONFIG_BLOCK
1429                 if (!spd)
1430                         spd = __set_page_dirty_buffers;
1431 #endif
1432                 return (*spd)(page);
1433         }
1434         if (!PageDirty(page)) {
1435                 if (!TestSetPageDirty(page))
1436                         return 1;
1437         }
1438         return 0;
1439 }
1440 EXPORT_SYMBOL(set_page_dirty);
1441
1442 /*
1443  * set_page_dirty() is racy if the caller has no reference against
1444  * page->mapping->host, and if the page is unlocked.  This is because another
1445  * CPU could truncate the page off the mapping and then free the mapping.
1446  *
1447  * Usually, the page _is_ locked, or the caller is a user-space process which
1448  * holds a reference on the inode by having an open file.
1449  *
1450  * In other cases, the page should be locked before running set_page_dirty().
1451  */
1452 int set_page_dirty_lock(struct page *page)
1453 {
1454         int ret;
1455
1456         lock_page(page);
1457         ret = set_page_dirty(page);
1458         unlock_page(page);
1459         return ret;
1460 }
1461 EXPORT_SYMBOL(set_page_dirty_lock);
1462
1463 /*
1464  * Clear a page's dirty flag, while caring for dirty memory accounting.
1465  * Returns true if the page was previously dirty.
1466  *
1467  * This is for preparing to put the page under writeout.  We leave the page
1468  * tagged as dirty in the radix tree so that a concurrent write-for-sync
1469  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
1470  * implementation will run either set_page_writeback() or set_page_dirty(),
1471  * at which stage we bring the page's dirty flag and radix-tree dirty tag
1472  * back into sync.
1473  *
1474  * This incoherency between the page's dirty flag and radix-tree tag is
1475  * unfortunate, but it only exists while the page is locked.
1476  */
1477 int clear_page_dirty_for_io(struct page *page)
1478 {
1479         struct address_space *mapping = page_mapping(page);
1480
1481         BUG_ON(!PageLocked(page));
1482
1483         if (mapping && mapping_cap_account_dirty(mapping)) {
1484                 /*
1485                  * Yes, Virginia, this is indeed insane.
1486                  *
1487                  * We use this sequence to make sure that
1488                  *  (a) we account for dirty stats properly
1489                  *  (b) we tell the low-level filesystem to
1490                  *      mark the whole page dirty if it was
1491                  *      dirty in a pagetable. Only to then
1492                  *  (c) clean the page again and return 1 to
1493                  *      cause the writeback.
1494                  *
1495                  * This way we avoid all nasty races with the
1496                  * dirty bit in multiple places and clearing
1497                  * them concurrently from different threads.
1498                  *
1499                  * Note! Normally the "set_page_dirty(page)"
1500                  * has no effect on the actual dirty bit - since
1501                  * that will already usually be set. But we
1502                  * need the side effects, and it can help us
1503                  * avoid races.
1504                  *
1505                  * We basically use the page "master dirty bit"
1506                  * as a serialization point for all the different
1507                  * threads doing their things.
1508                  */
1509                 if (page_mkclean(page))
1510                         set_page_dirty(page);
1511                 /*
1512                  * We carefully synchronise fault handlers against
1513                  * installing a dirty pte and marking the page dirty
1514                  * at this point. We do this by having them hold the
1515                  * page lock at some point after installing their
1516                  * pte, but before marking the page dirty.
1517                  * Pages are always locked coming in here, so we get
1518                  * the desired exclusion. See mm/memory.c:do_wp_page()
1519                  * for more comments.
1520                  */
1521                 if (TestClearPageDirty(page)) {
1522                         dec_zone_page_state(page, NR_FILE_DIRTY);
1523                         dec_bdi_stat(mapping->backing_dev_info,
1524                                         BDI_RECLAIMABLE);
1525                         return 1;
1526                 }
1527                 return 0;
1528         }
1529         return TestClearPageDirty(page);
1530 }
1531 EXPORT_SYMBOL(clear_page_dirty_for_io);
1532
1533 int test_clear_page_writeback(struct page *page)
1534 {
1535         struct address_space *mapping = page_mapping(page);
1536         int ret;
1537
1538         if (mapping) {
1539                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1540                 unsigned long flags;
1541
1542                 spin_lock_irqsave(&mapping->tree_lock, flags);
1543                 ret = TestClearPageWriteback(page);
1544                 if (ret) {
1545                         radix_tree_tag_clear(&mapping->page_tree,
1546                                                 page_index(page),
1547                                                 PAGECACHE_TAG_WRITEBACK);
1548                         if (bdi_cap_account_writeback(bdi)) {
1549                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1550                                 __bdi_writeout_inc(bdi);
1551                         }
1552                 }
1553                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1554         } else {
1555                 ret = TestClearPageWriteback(page);
1556         }
1557         if (ret) {
1558                 dec_zone_page_state(page, NR_WRITEBACK);
1559                 inc_zone_page_state(page, NR_WRITTEN);
1560         }
1561         return ret;
1562 }
1563
1564 int test_set_page_writeback(struct page *page)
1565 {
1566         struct address_space *mapping = page_mapping(page);
1567         int ret;
1568
1569         if (mapping) {
1570                 struct backing_dev_info *bdi = mapping->backing_dev_info;
1571                 unsigned long flags;
1572
1573                 spin_lock_irqsave(&mapping->tree_lock, flags);
1574                 ret = TestSetPageWriteback(page);
1575                 if (!ret) {
1576                         radix_tree_tag_set(&mapping->page_tree,
1577                                                 page_index(page),
1578                                                 PAGECACHE_TAG_WRITEBACK);
1579                         if (bdi_cap_account_writeback(bdi))
1580                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1581                 }
1582                 if (!PageDirty(page))
1583                         radix_tree_tag_clear(&mapping->page_tree,
1584                                                 page_index(page),
1585                                                 PAGECACHE_TAG_DIRTY);
1586                 radix_tree_tag_clear(&mapping->page_tree,
1587                                      page_index(page),
1588                                      PAGECACHE_TAG_TOWRITE);
1589                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1590         } else {
1591                 ret = TestSetPageWriteback(page);
1592         }
1593         if (!ret)
1594                 account_page_writeback(page);
1595         return ret;
1596
1597 }
1598 EXPORT_SYMBOL(test_set_page_writeback);
1599
1600 /*
1601  * Return true if any of the pages in the mapping are marked with the
1602  * passed tag.
1603  */
1604 int mapping_tagged(struct address_space *mapping, int tag)
1605 {
1606         return radix_tree_tagged(&mapping->page_tree, tag);
1607 }
1608 EXPORT_SYMBOL(mapping_tagged);