thp: add tlb_remove_pmd_tlb_entry
[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/export.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> /* __set_page_dirty_buffers */
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  * Try to keep balance_dirty_pages() call intervals higher than this many pages
46  * by raising pause time to max_pause when falls below it.
47  */
48 #define DIRTY_POLL_THRESH       (128 >> (PAGE_SHIFT - 10))
49
50 /*
51  * Estimate write bandwidth at 200ms intervals.
52  */
53 #define BANDWIDTH_INTERVAL      max(HZ/5, 1)
54
55 #define RATELIMIT_CALC_SHIFT    10
56
57 /*
58  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
59  * will look to see if it needs to force writeback or throttling.
60  */
61 static long ratelimit_pages = 32;
62
63 /* The following parameters are exported via /proc/sys/vm */
64
65 /*
66  * Start background writeback (via writeback threads) at this percentage
67  */
68 int dirty_background_ratio = 10;
69
70 /*
71  * dirty_background_bytes starts at 0 (disabled) so that it is a function of
72  * dirty_background_ratio * the amount of dirtyable memory
73  */
74 unsigned long dirty_background_bytes;
75
76 /*
77  * free highmem will not be subtracted from the total free memory
78  * for calculating free ratios if vm_highmem_is_dirtyable is true
79  */
80 int vm_highmem_is_dirtyable;
81
82 /*
83  * The generator of dirty data starts writeback at this percentage
84  */
85 int vm_dirty_ratio = 20;
86
87 /*
88  * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
89  * vm_dirty_ratio * the amount of dirtyable memory
90  */
91 unsigned long vm_dirty_bytes;
92
93 /*
94  * The interval between `kupdate'-style writebacks
95  */
96 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
97
98 /*
99  * The longest time for which data is allowed to remain dirty
100  */
101 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
102
103 /*
104  * Flag that makes the machine dump writes/reads and block dirtyings.
105  */
106 int block_dump;
107
108 /*
109  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
110  * a full sync is triggered after this time elapses without any disk activity.
111  */
112 int laptop_mode;
113
114 EXPORT_SYMBOL(laptop_mode);
115
116 /* End of sysctl-exported parameters */
117
118 unsigned long global_dirty_limit;
119
120 /*
121  * Scale the writeback cache size proportional to the relative writeout speeds.
122  *
123  * We do this by keeping a floating proportion between BDIs, based on page
124  * writeback completions [end_page_writeback()]. Those devices that write out
125  * pages fastest will get the larger share, while the slower will get a smaller
126  * share.
127  *
128  * We use page writeout completions because we are interested in getting rid of
129  * dirty pages. Having them written out is the primary goal.
130  *
131  * We introduce a concept of time, a period over which we measure these events,
132  * because demand can/will vary over time. The length of this period itself is
133  * measured in page writeback completions.
134  *
135  */
136 static struct prop_descriptor vm_completions;
137
138 /*
139  * Work out the current dirty-memory clamping and background writeout
140  * thresholds.
141  *
142  * The main aim here is to lower them aggressively if there is a lot of mapped
143  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
144  * pages.  It is better to clamp down on writers than to start swapping, and
145  * performing lots of scanning.
146  *
147  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
148  *
149  * We don't permit the clamping level to fall below 5% - that is getting rather
150  * excessive.
151  *
152  * We make sure that the background writeout level is below the adjusted
153  * clamping level.
154  */
155
156 /*
157  * In a memory zone, there is a certain amount of pages we consider
158  * available for the page cache, which is essentially the number of
159  * free and reclaimable pages, minus some zone reserves to protect
160  * lowmem and the ability to uphold the zone's watermarks without
161  * requiring writeback.
162  *
163  * This number of dirtyable pages is the base value of which the
164  * user-configurable dirty ratio is the effictive number of pages that
165  * are allowed to be actually dirtied.  Per individual zone, or
166  * globally by using the sum of dirtyable pages over all zones.
167  *
168  * Because the user is allowed to specify the dirty limit globally as
169  * absolute number of bytes, calculating the per-zone dirty limit can
170  * require translating the configured limit into a percentage of
171  * global dirtyable memory first.
172  */
173
174 static unsigned long highmem_dirtyable_memory(unsigned long total)
175 {
176 #ifdef CONFIG_HIGHMEM
177         int node;
178         unsigned long x = 0;
179
180         for_each_node_state(node, N_HIGH_MEMORY) {
181                 struct zone *z =
182                         &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
183
184                 x += zone_page_state(z, NR_FREE_PAGES) +
185                      zone_reclaimable_pages(z) - z->dirty_balance_reserve;
186         }
187         /*
188          * Make sure that the number of highmem pages is never larger
189          * than the number of the total dirtyable memory. This can only
190          * occur in very strange VM situations but we want to make sure
191          * that this does not occur.
192          */
193         return min(x, total);
194 #else
195         return 0;
196 #endif
197 }
198
199 /**
200  * global_dirtyable_memory - number of globally dirtyable pages
201  *
202  * Returns the global number of pages potentially available for dirty
203  * page cache.  This is the base value for the global dirty limits.
204  */
205 unsigned long global_dirtyable_memory(void)
206 {
207         unsigned long x;
208
209         x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages() -
210             dirty_balance_reserve;
211
212         if (!vm_highmem_is_dirtyable)
213                 x -= highmem_dirtyable_memory(x);
214
215         return x + 1;   /* Ensure that we never return 0 */
216 }
217
218 /*
219  * global_dirty_limits - background-writeback and dirty-throttling thresholds
220  *
221  * Calculate the dirty thresholds based on sysctl parameters
222  * - vm.dirty_background_ratio  or  vm.dirty_background_bytes
223  * - vm.dirty_ratio             or  vm.dirty_bytes
224  * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
225  * real-time tasks.
226  */
227 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
228 {
229         unsigned long background;
230         unsigned long dirty;
231         unsigned long uninitialized_var(available_memory);
232         struct task_struct *tsk;
233
234         if (!vm_dirty_bytes || !dirty_background_bytes)
235                 available_memory = global_dirtyable_memory();
236
237         if (vm_dirty_bytes)
238                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
239         else
240                 dirty = (vm_dirty_ratio * available_memory) / 100;
241
242         if (dirty_background_bytes)
243                 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
244         else
245                 background = (dirty_background_ratio * available_memory) / 100;
246
247         if (background >= dirty)
248                 background = dirty / 2;
249         tsk = current;
250         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
251                 background += background / 4;
252                 dirty += dirty / 4;
253         }
254         *pbackground = background;
255         *pdirty = dirty;
256         trace_global_dirty_state(background, dirty);
257 }
258
259 /**
260  * zone_dirtyable_memory - number of dirtyable pages in a zone
261  * @zone: the zone
262  *
263  * Returns the zone's number of pages potentially available for dirty
264  * page cache.  This is the base value for the per-zone dirty limits.
265  */
266 static unsigned long zone_dirtyable_memory(struct zone *zone)
267 {
268         /*
269          * The effective global number of dirtyable pages may exclude
270          * highmem as a big-picture measure to keep the ratio between
271          * dirty memory and lowmem reasonable.
272          *
273          * But this function is purely about the individual zone and a
274          * highmem zone can hold its share of dirty pages, so we don't
275          * care about vm_highmem_is_dirtyable here.
276          */
277         return zone_page_state(zone, NR_FREE_PAGES) +
278                zone_reclaimable_pages(zone) -
279                zone->dirty_balance_reserve;
280 }
281
282 /**
283  * zone_dirty_limit - maximum number of dirty pages allowed in a zone
284  * @zone: the zone
285  *
286  * Returns the maximum number of dirty pages allowed in a zone, based
287  * on the zone's dirtyable memory.
288  */
289 static unsigned long zone_dirty_limit(struct zone *zone)
290 {
291         unsigned long zone_memory = zone_dirtyable_memory(zone);
292         struct task_struct *tsk = current;
293         unsigned long dirty;
294
295         if (vm_dirty_bytes)
296                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
297                         zone_memory / global_dirtyable_memory();
298         else
299                 dirty = vm_dirty_ratio * zone_memory / 100;
300
301         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
302                 dirty += dirty / 4;
303
304         return dirty;
305 }
306
307 /**
308  * zone_dirty_ok - tells whether a zone is within its dirty limits
309  * @zone: the zone to check
310  *
311  * Returns %true when the dirty pages in @zone are within the zone's
312  * dirty limit, %false if the limit is exceeded.
313  */
314 bool zone_dirty_ok(struct zone *zone)
315 {
316         unsigned long limit = zone_dirty_limit(zone);
317
318         return zone_page_state(zone, NR_FILE_DIRTY) +
319                zone_page_state(zone, NR_UNSTABLE_NFS) +
320                zone_page_state(zone, NR_WRITEBACK) <= limit;
321 }
322
323 /*
324  * couple the period to the dirty_ratio:
325  *
326  *   period/2 ~ roundup_pow_of_two(dirty limit)
327  */
328 static int calc_period_shift(void)
329 {
330         unsigned long dirty_total;
331
332         if (vm_dirty_bytes)
333                 dirty_total = vm_dirty_bytes / PAGE_SIZE;
334         else
335                 dirty_total = (vm_dirty_ratio * global_dirtyable_memory()) /
336                                 100;
337         return 2 + ilog2(dirty_total - 1);
338 }
339
340 /*
341  * update the period when the dirty threshold changes.
342  */
343 static void update_completion_period(void)
344 {
345         int shift = calc_period_shift();
346         prop_change_shift(&vm_completions, shift);
347
348         writeback_set_ratelimit();
349 }
350
351 int dirty_background_ratio_handler(struct ctl_table *table, int write,
352                 void __user *buffer, size_t *lenp,
353                 loff_t *ppos)
354 {
355         int ret;
356
357         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
358         if (ret == 0 && write)
359                 dirty_background_bytes = 0;
360         return ret;
361 }
362
363 int dirty_background_bytes_handler(struct ctl_table *table, int write,
364                 void __user *buffer, size_t *lenp,
365                 loff_t *ppos)
366 {
367         int ret;
368
369         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
370         if (ret == 0 && write)
371                 dirty_background_ratio = 0;
372         return ret;
373 }
374
375 int dirty_ratio_handler(struct ctl_table *table, int write,
376                 void __user *buffer, size_t *lenp,
377                 loff_t *ppos)
378 {
379         int old_ratio = vm_dirty_ratio;
380         int ret;
381
382         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
383         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
384                 update_completion_period();
385                 vm_dirty_bytes = 0;
386         }
387         return ret;
388 }
389
390 int dirty_bytes_handler(struct ctl_table *table, int write,
391                 void __user *buffer, size_t *lenp,
392                 loff_t *ppos)
393 {
394         unsigned long old_bytes = vm_dirty_bytes;
395         int ret;
396
397         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
398         if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
399                 update_completion_period();
400                 vm_dirty_ratio = 0;
401         }
402         return ret;
403 }
404
405 /*
406  * Increment the BDI's writeout completion count and the global writeout
407  * completion count. Called from test_clear_page_writeback().
408  */
409 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
410 {
411         __inc_bdi_stat(bdi, BDI_WRITTEN);
412         __prop_inc_percpu_max(&vm_completions, &bdi->completions,
413                               bdi->max_prop_frac);
414 }
415
416 void bdi_writeout_inc(struct backing_dev_info *bdi)
417 {
418         unsigned long flags;
419
420         local_irq_save(flags);
421         __bdi_writeout_inc(bdi);
422         local_irq_restore(flags);
423 }
424 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
425
426 /*
427  * Obtain an accurate fraction of the BDI's portion.
428  */
429 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
430                 long *numerator, long *denominator)
431 {
432         prop_fraction_percpu(&vm_completions, &bdi->completions,
433                                 numerator, denominator);
434 }
435
436 /*
437  * bdi_min_ratio keeps the sum of the minimum dirty shares of all
438  * registered backing devices, which, for obvious reasons, can not
439  * exceed 100%.
440  */
441 static unsigned int bdi_min_ratio;
442
443 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
444 {
445         int ret = 0;
446
447         spin_lock_bh(&bdi_lock);
448         if (min_ratio > bdi->max_ratio) {
449                 ret = -EINVAL;
450         } else {
451                 min_ratio -= bdi->min_ratio;
452                 if (bdi_min_ratio + min_ratio < 100) {
453                         bdi_min_ratio += min_ratio;
454                         bdi->min_ratio += min_ratio;
455                 } else {
456                         ret = -EINVAL;
457                 }
458         }
459         spin_unlock_bh(&bdi_lock);
460
461         return ret;
462 }
463
464 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
465 {
466         int ret = 0;
467
468         if (max_ratio > 100)
469                 return -EINVAL;
470
471         spin_lock_bh(&bdi_lock);
472         if (bdi->min_ratio > max_ratio) {
473                 ret = -EINVAL;
474         } else {
475                 bdi->max_ratio = max_ratio;
476                 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
477         }
478         spin_unlock_bh(&bdi_lock);
479
480         return ret;
481 }
482 EXPORT_SYMBOL(bdi_set_max_ratio);
483
484 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
485                                            unsigned long bg_thresh)
486 {
487         return (thresh + bg_thresh) / 2;
488 }
489
490 static unsigned long hard_dirty_limit(unsigned long thresh)
491 {
492         return max(thresh, global_dirty_limit);
493 }
494
495 /**
496  * bdi_dirty_limit - @bdi's share of dirty throttling threshold
497  * @bdi: the backing_dev_info to query
498  * @dirty: global dirty limit in pages
499  *
500  * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
501  * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
502  *
503  * Note that balance_dirty_pages() will only seriously take it as a hard limit
504  * when sleeping max_pause per page is not enough to keep the dirty pages under
505  * control. For example, when the device is completely stalled due to some error
506  * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
507  * In the other normal situations, it acts more gently by throttling the tasks
508  * more (rather than completely block them) when the bdi dirty pages go high.
509  *
510  * It allocates high/low dirty limits to fast/slow devices, in order to prevent
511  * - starving fast devices
512  * - piling up dirty pages (that will take long time to sync) on slow devices
513  *
514  * The bdi's share of dirty limit will be adapting to its throughput and
515  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
516  */
517 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
518 {
519         u64 bdi_dirty;
520         long numerator, denominator;
521
522         /*
523          * Calculate this BDI's share of the dirty ratio.
524          */
525         bdi_writeout_fraction(bdi, &numerator, &denominator);
526
527         bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
528         bdi_dirty *= numerator;
529         do_div(bdi_dirty, denominator);
530
531         bdi_dirty += (dirty * bdi->min_ratio) / 100;
532         if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
533                 bdi_dirty = dirty * bdi->max_ratio / 100;
534
535         return bdi_dirty;
536 }
537
538 /*
539  * Dirty position control.
540  *
541  * (o) global/bdi setpoints
542  *
543  * We want the dirty pages be balanced around the global/bdi setpoints.
544  * When the number of dirty pages is higher/lower than the setpoint, the
545  * dirty position control ratio (and hence task dirty ratelimit) will be
546  * decreased/increased to bring the dirty pages back to the setpoint.
547  *
548  *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
549  *
550  *     if (dirty < setpoint) scale up   pos_ratio
551  *     if (dirty > setpoint) scale down pos_ratio
552  *
553  *     if (bdi_dirty < bdi_setpoint) scale up   pos_ratio
554  *     if (bdi_dirty > bdi_setpoint) scale down pos_ratio
555  *
556  *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
557  *
558  * (o) global control line
559  *
560  *     ^ pos_ratio
561  *     |
562  *     |            |<===== global dirty control scope ======>|
563  * 2.0 .............*
564  *     |            .*
565  *     |            . *
566  *     |            .   *
567  *     |            .     *
568  *     |            .        *
569  *     |            .            *
570  * 1.0 ................................*
571  *     |            .                  .     *
572  *     |            .                  .          *
573  *     |            .                  .              *
574  *     |            .                  .                 *
575  *     |            .                  .                    *
576  *   0 +------------.------------------.----------------------*------------->
577  *           freerun^          setpoint^                 limit^   dirty pages
578  *
579  * (o) bdi control line
580  *
581  *     ^ pos_ratio
582  *     |
583  *     |            *
584  *     |              *
585  *     |                *
586  *     |                  *
587  *     |                    * |<=========== span ============>|
588  * 1.0 .......................*
589  *     |                      . *
590  *     |                      .   *
591  *     |                      .     *
592  *     |                      .       *
593  *     |                      .         *
594  *     |                      .           *
595  *     |                      .             *
596  *     |                      .               *
597  *     |                      .                 *
598  *     |                      .                   *
599  *     |                      .                     *
600  * 1/4 ...............................................* * * * * * * * * * * *
601  *     |                      .                         .
602  *     |                      .                           .
603  *     |                      .                             .
604  *   0 +----------------------.-------------------------------.------------->
605  *                bdi_setpoint^                    x_intercept^
606  *
607  * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
608  * be smoothly throttled down to normal if it starts high in situations like
609  * - start writing to a slow SD card and a fast disk at the same time. The SD
610  *   card's bdi_dirty may rush to many times higher than bdi_setpoint.
611  * - the bdi dirty thresh drops quickly due to change of JBOD workload
612  */
613 static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
614                                         unsigned long thresh,
615                                         unsigned long bg_thresh,
616                                         unsigned long dirty,
617                                         unsigned long bdi_thresh,
618                                         unsigned long bdi_dirty)
619 {
620         unsigned long write_bw = bdi->avg_write_bandwidth;
621         unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
622         unsigned long limit = hard_dirty_limit(thresh);
623         unsigned long x_intercept;
624         unsigned long setpoint;         /* dirty pages' target balance point */
625         unsigned long bdi_setpoint;
626         unsigned long span;
627         long long pos_ratio;            /* for scaling up/down the rate limit */
628         long x;
629
630         if (unlikely(dirty >= limit))
631                 return 0;
632
633         /*
634          * global setpoint
635          *
636          *                           setpoint - dirty 3
637          *        f(dirty) := 1.0 + (----------------)
638          *                           limit - setpoint
639          *
640          * it's a 3rd order polynomial that subjects to
641          *
642          * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
643          * (2) f(setpoint) = 1.0 => the balance point
644          * (3) f(limit)    = 0   => the hard limit
645          * (4) df/dx      <= 0   => negative feedback control
646          * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
647          *     => fast response on large errors; small oscillation near setpoint
648          */
649         setpoint = (freerun + limit) / 2;
650         x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
651                     limit - setpoint + 1);
652         pos_ratio = x;
653         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
654         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
655         pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
656
657         /*
658          * We have computed basic pos_ratio above based on global situation. If
659          * the bdi is over/under its share of dirty pages, we want to scale
660          * pos_ratio further down/up. That is done by the following mechanism.
661          */
662
663         /*
664          * bdi setpoint
665          *
666          *        f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
667          *
668          *                        x_intercept - bdi_dirty
669          *                     := --------------------------
670          *                        x_intercept - bdi_setpoint
671          *
672          * The main bdi control line is a linear function that subjects to
673          *
674          * (1) f(bdi_setpoint) = 1.0
675          * (2) k = - 1 / (8 * write_bw)  (in single bdi case)
676          *     or equally: x_intercept = bdi_setpoint + 8 * write_bw
677          *
678          * For single bdi case, the dirty pages are observed to fluctuate
679          * regularly within range
680          *        [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
681          * for various filesystems, where (2) can yield in a reasonable 12.5%
682          * fluctuation range for pos_ratio.
683          *
684          * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
685          * own size, so move the slope over accordingly and choose a slope that
686          * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
687          */
688         if (unlikely(bdi_thresh > thresh))
689                 bdi_thresh = thresh;
690         /*
691          * It's very possible that bdi_thresh is close to 0 not because the
692          * device is slow, but that it has remained inactive for long time.
693          * Honour such devices a reasonable good (hopefully IO efficient)
694          * threshold, so that the occasional writes won't be blocked and active
695          * writes can rampup the threshold quickly.
696          */
697         bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
698         /*
699          * scale global setpoint to bdi's:
700          *      bdi_setpoint = setpoint * bdi_thresh / thresh
701          */
702         x = div_u64((u64)bdi_thresh << 16, thresh + 1);
703         bdi_setpoint = setpoint * (u64)x >> 16;
704         /*
705          * Use span=(8*write_bw) in single bdi case as indicated by
706          * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
707          *
708          *        bdi_thresh                    thresh - bdi_thresh
709          * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
710          *          thresh                            thresh
711          */
712         span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
713         x_intercept = bdi_setpoint + span;
714
715         if (bdi_dirty < x_intercept - span / 4) {
716                 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
717                                     x_intercept - bdi_setpoint + 1);
718         } else
719                 pos_ratio /= 4;
720
721         /*
722          * bdi reserve area, safeguard against dirty pool underrun and disk idle
723          * It may push the desired control point of global dirty pages higher
724          * than setpoint.
725          */
726         x_intercept = bdi_thresh / 2;
727         if (bdi_dirty < x_intercept) {
728                 if (bdi_dirty > x_intercept / 8)
729                         pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
730                 else
731                         pos_ratio *= 8;
732         }
733
734         return pos_ratio;
735 }
736
737 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
738                                        unsigned long elapsed,
739                                        unsigned long written)
740 {
741         const unsigned long period = roundup_pow_of_two(3 * HZ);
742         unsigned long avg = bdi->avg_write_bandwidth;
743         unsigned long old = bdi->write_bandwidth;
744         u64 bw;
745
746         /*
747          * bw = written * HZ / elapsed
748          *
749          *                   bw * elapsed + write_bandwidth * (period - elapsed)
750          * write_bandwidth = ---------------------------------------------------
751          *                                          period
752          */
753         bw = written - bdi->written_stamp;
754         bw *= HZ;
755         if (unlikely(elapsed > period)) {
756                 do_div(bw, elapsed);
757                 avg = bw;
758                 goto out;
759         }
760         bw += (u64)bdi->write_bandwidth * (period - elapsed);
761         bw >>= ilog2(period);
762
763         /*
764          * one more level of smoothing, for filtering out sudden spikes
765          */
766         if (avg > old && old >= (unsigned long)bw)
767                 avg -= (avg - old) >> 3;
768
769         if (avg < old && old <= (unsigned long)bw)
770                 avg += (old - avg) >> 3;
771
772 out:
773         bdi->write_bandwidth = bw;
774         bdi->avg_write_bandwidth = avg;
775 }
776
777 /*
778  * The global dirtyable memory and dirty threshold could be suddenly knocked
779  * down by a large amount (eg. on the startup of KVM in a swapless system).
780  * This may throw the system into deep dirty exceeded state and throttle
781  * heavy/light dirtiers alike. To retain good responsiveness, maintain
782  * global_dirty_limit for tracking slowly down to the knocked down dirty
783  * threshold.
784  */
785 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
786 {
787         unsigned long limit = global_dirty_limit;
788
789         /*
790          * Follow up in one step.
791          */
792         if (limit < thresh) {
793                 limit = thresh;
794                 goto update;
795         }
796
797         /*
798          * Follow down slowly. Use the higher one as the target, because thresh
799          * may drop below dirty. This is exactly the reason to introduce
800          * global_dirty_limit which is guaranteed to lie above the dirty pages.
801          */
802         thresh = max(thresh, dirty);
803         if (limit > thresh) {
804                 limit -= (limit - thresh) >> 5;
805                 goto update;
806         }
807         return;
808 update:
809         global_dirty_limit = limit;
810 }
811
812 static void global_update_bandwidth(unsigned long thresh,
813                                     unsigned long dirty,
814                                     unsigned long now)
815 {
816         static DEFINE_SPINLOCK(dirty_lock);
817         static unsigned long update_time;
818
819         /*
820          * check locklessly first to optimize away locking for the most time
821          */
822         if (time_before(now, update_time + BANDWIDTH_INTERVAL))
823                 return;
824
825         spin_lock(&dirty_lock);
826         if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
827                 update_dirty_limit(thresh, dirty);
828                 update_time = now;
829         }
830         spin_unlock(&dirty_lock);
831 }
832
833 /*
834  * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
835  *
836  * Normal bdi tasks will be curbed at or below it in long term.
837  * Obviously it should be around (write_bw / N) when there are N dd tasks.
838  */
839 static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
840                                        unsigned long thresh,
841                                        unsigned long bg_thresh,
842                                        unsigned long dirty,
843                                        unsigned long bdi_thresh,
844                                        unsigned long bdi_dirty,
845                                        unsigned long dirtied,
846                                        unsigned long elapsed)
847 {
848         unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
849         unsigned long limit = hard_dirty_limit(thresh);
850         unsigned long setpoint = (freerun + limit) / 2;
851         unsigned long write_bw = bdi->avg_write_bandwidth;
852         unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
853         unsigned long dirty_rate;
854         unsigned long task_ratelimit;
855         unsigned long balanced_dirty_ratelimit;
856         unsigned long pos_ratio;
857         unsigned long step;
858         unsigned long x;
859
860         /*
861          * The dirty rate will match the writeout rate in long term, except
862          * when dirty pages are truncated by userspace or re-dirtied by FS.
863          */
864         dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
865
866         pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
867                                        bdi_thresh, bdi_dirty);
868         /*
869          * task_ratelimit reflects each dd's dirty rate for the past 200ms.
870          */
871         task_ratelimit = (u64)dirty_ratelimit *
872                                         pos_ratio >> RATELIMIT_CALC_SHIFT;
873         task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
874
875         /*
876          * A linear estimation of the "balanced" throttle rate. The theory is,
877          * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
878          * dirty_rate will be measured to be (N * task_ratelimit). So the below
879          * formula will yield the balanced rate limit (write_bw / N).
880          *
881          * Note that the expanded form is not a pure rate feedback:
882          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate)              (1)
883          * but also takes pos_ratio into account:
884          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
885          *
886          * (1) is not realistic because pos_ratio also takes part in balancing
887          * the dirty rate.  Consider the state
888          *      pos_ratio = 0.5                                              (3)
889          *      rate = 2 * (write_bw / N)                                    (4)
890          * If (1) is used, it will stuck in that state! Because each dd will
891          * be throttled at
892          *      task_ratelimit = pos_ratio * rate = (write_bw / N)           (5)
893          * yielding
894          *      dirty_rate = N * task_ratelimit = write_bw                   (6)
895          * put (6) into (1) we get
896          *      rate_(i+1) = rate_(i)                                        (7)
897          *
898          * So we end up using (2) to always keep
899          *      rate_(i+1) ~= (write_bw / N)                                 (8)
900          * regardless of the value of pos_ratio. As long as (8) is satisfied,
901          * pos_ratio is able to drive itself to 1.0, which is not only where
902          * the dirty count meet the setpoint, but also where the slope of
903          * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
904          */
905         balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
906                                            dirty_rate | 1);
907         /*
908          * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
909          */
910         if (unlikely(balanced_dirty_ratelimit > write_bw))
911                 balanced_dirty_ratelimit = write_bw;
912
913         /*
914          * We could safely do this and return immediately:
915          *
916          *      bdi->dirty_ratelimit = balanced_dirty_ratelimit;
917          *
918          * However to get a more stable dirty_ratelimit, the below elaborated
919          * code makes use of task_ratelimit to filter out sigular points and
920          * limit the step size.
921          *
922          * The below code essentially only uses the relative value of
923          *
924          *      task_ratelimit - dirty_ratelimit
925          *      = (pos_ratio - 1) * dirty_ratelimit
926          *
927          * which reflects the direction and size of dirty position error.
928          */
929
930         /*
931          * dirty_ratelimit will follow balanced_dirty_ratelimit iff
932          * task_ratelimit is on the same side of dirty_ratelimit, too.
933          * For example, when
934          * - dirty_ratelimit > balanced_dirty_ratelimit
935          * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
936          * lowering dirty_ratelimit will help meet both the position and rate
937          * control targets. Otherwise, don't update dirty_ratelimit if it will
938          * only help meet the rate target. After all, what the users ultimately
939          * feel and care are stable dirty rate and small position error.
940          *
941          * |task_ratelimit - dirty_ratelimit| is used to limit the step size
942          * and filter out the sigular points of balanced_dirty_ratelimit. Which
943          * keeps jumping around randomly and can even leap far away at times
944          * due to the small 200ms estimation period of dirty_rate (we want to
945          * keep that period small to reduce time lags).
946          */
947         step = 0;
948         if (dirty < setpoint) {
949                 x = min(bdi->balanced_dirty_ratelimit,
950                          min(balanced_dirty_ratelimit, task_ratelimit));
951                 if (dirty_ratelimit < x)
952                         step = x - dirty_ratelimit;
953         } else {
954                 x = max(bdi->balanced_dirty_ratelimit,
955                          max(balanced_dirty_ratelimit, task_ratelimit));
956                 if (dirty_ratelimit > x)
957                         step = dirty_ratelimit - x;
958         }
959
960         /*
961          * Don't pursue 100% rate matching. It's impossible since the balanced
962          * rate itself is constantly fluctuating. So decrease the track speed
963          * when it gets close to the target. Helps eliminate pointless tremors.
964          */
965         step >>= dirty_ratelimit / (2 * step + 1);
966         /*
967          * Limit the tracking speed to avoid overshooting.
968          */
969         step = (step + 7) / 8;
970
971         if (dirty_ratelimit < balanced_dirty_ratelimit)
972                 dirty_ratelimit += step;
973         else
974                 dirty_ratelimit -= step;
975
976         bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
977         bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
978
979         trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
980 }
981
982 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
983                             unsigned long thresh,
984                             unsigned long bg_thresh,
985                             unsigned long dirty,
986                             unsigned long bdi_thresh,
987                             unsigned long bdi_dirty,
988                             unsigned long start_time)
989 {
990         unsigned long now = jiffies;
991         unsigned long elapsed = now - bdi->bw_time_stamp;
992         unsigned long dirtied;
993         unsigned long written;
994
995         /*
996          * rate-limit, only update once every 200ms.
997          */
998         if (elapsed < BANDWIDTH_INTERVAL)
999                 return;
1000
1001         dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
1002         written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
1003
1004         /*
1005          * Skip quiet periods when disk bandwidth is under-utilized.
1006          * (at least 1s idle time between two flusher runs)
1007          */
1008         if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
1009                 goto snapshot;
1010
1011         if (thresh) {
1012                 global_update_bandwidth(thresh, dirty, now);
1013                 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1014                                            bdi_thresh, bdi_dirty,
1015                                            dirtied, elapsed);
1016         }
1017         bdi_update_write_bandwidth(bdi, elapsed, written);
1018
1019 snapshot:
1020         bdi->dirtied_stamp = dirtied;
1021         bdi->written_stamp = written;
1022         bdi->bw_time_stamp = now;
1023 }
1024
1025 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
1026                                  unsigned long thresh,
1027                                  unsigned long bg_thresh,
1028                                  unsigned long dirty,
1029                                  unsigned long bdi_thresh,
1030                                  unsigned long bdi_dirty,
1031                                  unsigned long start_time)
1032 {
1033         if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1034                 return;
1035         spin_lock(&bdi->wb.list_lock);
1036         __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1037                                bdi_thresh, bdi_dirty, start_time);
1038         spin_unlock(&bdi->wb.list_lock);
1039 }
1040
1041 /*
1042  * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
1043  * will look to see if it needs to start dirty throttling.
1044  *
1045  * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1046  * global_page_state() too often. So scale it near-sqrt to the safety margin
1047  * (the number of pages we may dirty without exceeding the dirty limits).
1048  */
1049 static unsigned long dirty_poll_interval(unsigned long dirty,
1050                                          unsigned long thresh)
1051 {
1052         if (thresh > dirty)
1053                 return 1UL << (ilog2(thresh - dirty) >> 1);
1054
1055         return 1;
1056 }
1057
1058 static long bdi_max_pause(struct backing_dev_info *bdi,
1059                           unsigned long bdi_dirty)
1060 {
1061         long bw = bdi->avg_write_bandwidth;
1062         long t;
1063
1064         /*
1065          * Limit pause time for small memory systems. If sleeping for too long
1066          * time, a small pool of dirty/writeback pages may go empty and disk go
1067          * idle.
1068          *
1069          * 8 serves as the safety ratio.
1070          */
1071         t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1072         t++;
1073
1074         return min_t(long, t, MAX_PAUSE);
1075 }
1076
1077 static long bdi_min_pause(struct backing_dev_info *bdi,
1078                           long max_pause,
1079                           unsigned long task_ratelimit,
1080                           unsigned long dirty_ratelimit,
1081                           int *nr_dirtied_pause)
1082 {
1083         long hi = ilog2(bdi->avg_write_bandwidth);
1084         long lo = ilog2(bdi->dirty_ratelimit);
1085         long t;         /* target pause */
1086         long pause;     /* estimated next pause */
1087         int pages;      /* target nr_dirtied_pause */
1088
1089         /* target for 10ms pause on 1-dd case */
1090         t = max(1, HZ / 100);
1091
1092         /*
1093          * Scale up pause time for concurrent dirtiers in order to reduce CPU
1094          * overheads.
1095          *
1096          * (N * 10ms) on 2^N concurrent tasks.
1097          */
1098         if (hi > lo)
1099                 t += (hi - lo) * (10 * HZ) / 1024;
1100
1101         /*
1102          * This is a bit convoluted. We try to base the next nr_dirtied_pause
1103          * on the much more stable dirty_ratelimit. However the next pause time
1104          * will be computed based on task_ratelimit and the two rate limits may
1105          * depart considerably at some time. Especially if task_ratelimit goes
1106          * below dirty_ratelimit/2 and the target pause is max_pause, the next
1107          * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1108          * result task_ratelimit won't be executed faithfully, which could
1109          * eventually bring down dirty_ratelimit.
1110          *
1111          * We apply two rules to fix it up:
1112          * 1) try to estimate the next pause time and if necessary, use a lower
1113          *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1114          *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1115          * 2) limit the target pause time to max_pause/2, so that the normal
1116          *    small fluctuations of task_ratelimit won't trigger rule (1) and
1117          *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1118          */
1119         t = min(t, 1 + max_pause / 2);
1120         pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1121
1122         /*
1123          * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1124          * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1125          * When the 16 consecutive reads are often interrupted by some dirty
1126          * throttling pause during the async writes, cfq will go into idles
1127          * (deadline is fine). So push nr_dirtied_pause as high as possible
1128          * until reaches DIRTY_POLL_THRESH=32 pages.
1129          */
1130         if (pages < DIRTY_POLL_THRESH) {
1131                 t = max_pause;
1132                 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1133                 if (pages > DIRTY_POLL_THRESH) {
1134                         pages = DIRTY_POLL_THRESH;
1135                         t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1136                 }
1137         }
1138
1139         pause = HZ * pages / (task_ratelimit + 1);
1140         if (pause > max_pause) {
1141                 t = max_pause;
1142                 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1143         }
1144
1145         *nr_dirtied_pause = pages;
1146         /*
1147          * The minimal pause time will normally be half the target pause time.
1148          */
1149         return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1150 }
1151
1152 /*
1153  * balance_dirty_pages() must be called by processes which are generating dirty
1154  * data.  It looks at the number of dirty pages in the machine and will force
1155  * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1156  * If we're over `background_thresh' then the writeback threads are woken to
1157  * perform some writeout.
1158  */
1159 static void balance_dirty_pages(struct address_space *mapping,
1160                                 unsigned long pages_dirtied)
1161 {
1162         unsigned long nr_reclaimable;   /* = file_dirty + unstable_nfs */
1163         unsigned long bdi_reclaimable;
1164         unsigned long nr_dirty;  /* = file_dirty + writeback + unstable_nfs */
1165         unsigned long bdi_dirty;
1166         unsigned long freerun;
1167         unsigned long background_thresh;
1168         unsigned long dirty_thresh;
1169         unsigned long bdi_thresh;
1170         long period;
1171         long pause;
1172         long max_pause;
1173         long min_pause;
1174         int nr_dirtied_pause;
1175         bool dirty_exceeded = false;
1176         unsigned long task_ratelimit;
1177         unsigned long dirty_ratelimit;
1178         unsigned long pos_ratio;
1179         struct backing_dev_info *bdi = mapping->backing_dev_info;
1180         unsigned long start_time = jiffies;
1181
1182         for (;;) {
1183                 unsigned long now = jiffies;
1184
1185                 /*
1186                  * Unstable writes are a feature of certain networked
1187                  * filesystems (i.e. NFS) in which data may have been
1188                  * written to the server's write cache, but has not yet
1189                  * been flushed to permanent storage.
1190                  */
1191                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1192                                         global_page_state(NR_UNSTABLE_NFS);
1193                 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1194
1195                 global_dirty_limits(&background_thresh, &dirty_thresh);
1196
1197                 /*
1198                  * Throttle it only when the background writeback cannot
1199                  * catch-up. This avoids (excessively) small writeouts
1200                  * when the bdi limits are ramping up.
1201                  */
1202                 freerun = dirty_freerun_ceiling(dirty_thresh,
1203                                                 background_thresh);
1204                 if (nr_dirty <= freerun) {
1205                         current->dirty_paused_when = now;
1206                         current->nr_dirtied = 0;
1207                         current->nr_dirtied_pause =
1208                                 dirty_poll_interval(nr_dirty, dirty_thresh);
1209                         break;
1210                 }
1211
1212                 if (unlikely(!writeback_in_progress(bdi)))
1213                         bdi_start_background_writeback(bdi);
1214
1215                 /*
1216                  * bdi_thresh is not treated as some limiting factor as
1217                  * dirty_thresh, due to reasons
1218                  * - in JBOD setup, bdi_thresh can fluctuate a lot
1219                  * - in a system with HDD and USB key, the USB key may somehow
1220                  *   go into state (bdi_dirty >> bdi_thresh) either because
1221                  *   bdi_dirty starts high, or because bdi_thresh drops low.
1222                  *   In this case we don't want to hard throttle the USB key
1223                  *   dirtiers for 100 seconds until bdi_dirty drops under
1224                  *   bdi_thresh. Instead the auxiliary bdi control line in
1225                  *   bdi_position_ratio() will let the dirtier task progress
1226                  *   at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1227                  */
1228                 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1229
1230                 /*
1231                  * In order to avoid the stacked BDI deadlock we need
1232                  * to ensure we accurately count the 'dirty' pages when
1233                  * the threshold is low.
1234                  *
1235                  * Otherwise it would be possible to get thresh+n pages
1236                  * reported dirty, even though there are thresh-m pages
1237                  * actually dirty; with m+n sitting in the percpu
1238                  * deltas.
1239                  */
1240                 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1241                         bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1242                         bdi_dirty = bdi_reclaimable +
1243                                     bdi_stat_sum(bdi, BDI_WRITEBACK);
1244                 } else {
1245                         bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1246                         bdi_dirty = bdi_reclaimable +
1247                                     bdi_stat(bdi, BDI_WRITEBACK);
1248                 }
1249
1250                 dirty_exceeded = (bdi_dirty > bdi_thresh) &&
1251                                   (nr_dirty > dirty_thresh);
1252                 if (dirty_exceeded && !bdi->dirty_exceeded)
1253                         bdi->dirty_exceeded = 1;
1254
1255                 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1256                                      nr_dirty, bdi_thresh, bdi_dirty,
1257                                      start_time);
1258
1259                 dirty_ratelimit = bdi->dirty_ratelimit;
1260                 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1261                                                background_thresh, nr_dirty,
1262                                                bdi_thresh, bdi_dirty);
1263                 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1264                                                         RATELIMIT_CALC_SHIFT;
1265                 max_pause = bdi_max_pause(bdi, bdi_dirty);
1266                 min_pause = bdi_min_pause(bdi, max_pause,
1267                                           task_ratelimit, dirty_ratelimit,
1268                                           &nr_dirtied_pause);
1269
1270                 if (unlikely(task_ratelimit == 0)) {
1271                         period = max_pause;
1272                         pause = max_pause;
1273                         goto pause;
1274                 }
1275                 period = HZ * pages_dirtied / task_ratelimit;
1276                 pause = period;
1277                 if (current->dirty_paused_when)
1278                         pause -= now - current->dirty_paused_when;
1279                 /*
1280                  * For less than 1s think time (ext3/4 may block the dirtier
1281                  * for up to 800ms from time to time on 1-HDD; so does xfs,
1282                  * however at much less frequency), try to compensate it in
1283                  * future periods by updating the virtual time; otherwise just
1284                  * do a reset, as it may be a light dirtier.
1285                  */
1286                 if (pause < min_pause) {
1287                         trace_balance_dirty_pages(bdi,
1288                                                   dirty_thresh,
1289                                                   background_thresh,
1290                                                   nr_dirty,
1291                                                   bdi_thresh,
1292                                                   bdi_dirty,
1293                                                   dirty_ratelimit,
1294                                                   task_ratelimit,
1295                                                   pages_dirtied,
1296                                                   period,
1297                                                   min(pause, 0L),
1298                                                   start_time);
1299                         if (pause < -HZ) {
1300                                 current->dirty_paused_when = now;
1301                                 current->nr_dirtied = 0;
1302                         } else if (period) {
1303                                 current->dirty_paused_when += period;
1304                                 current->nr_dirtied = 0;
1305                         } else if (current->nr_dirtied_pause <= pages_dirtied)
1306                                 current->nr_dirtied_pause += pages_dirtied;
1307                         break;
1308                 }
1309                 if (unlikely(pause > max_pause)) {
1310                         /* for occasional dropped task_ratelimit */
1311                         now += min(pause - max_pause, max_pause);
1312                         pause = max_pause;
1313                 }
1314
1315 pause:
1316                 trace_balance_dirty_pages(bdi,
1317                                           dirty_thresh,
1318                                           background_thresh,
1319                                           nr_dirty,
1320                                           bdi_thresh,
1321                                           bdi_dirty,
1322                                           dirty_ratelimit,
1323                                           task_ratelimit,
1324                                           pages_dirtied,
1325                                           period,
1326                                           pause,
1327                                           start_time);
1328                 __set_current_state(TASK_KILLABLE);
1329                 io_schedule_timeout(pause);
1330
1331                 current->dirty_paused_when = now + pause;
1332                 current->nr_dirtied = 0;
1333                 current->nr_dirtied_pause = nr_dirtied_pause;
1334
1335                 /*
1336                  * This is typically equal to (nr_dirty < dirty_thresh) and can
1337                  * also keep "1000+ dd on a slow USB stick" under control.
1338                  */
1339                 if (task_ratelimit)
1340                         break;
1341
1342                 /*
1343                  * In the case of an unresponding NFS server and the NFS dirty
1344                  * pages exceeds dirty_thresh, give the other good bdi's a pipe
1345                  * to go through, so that tasks on them still remain responsive.
1346                  *
1347                  * In theory 1 page is enough to keep the comsumer-producer
1348                  * pipe going: the flusher cleans 1 page => the task dirties 1
1349                  * more page. However bdi_dirty has accounting errors.  So use
1350                  * the larger and more IO friendly bdi_stat_error.
1351                  */
1352                 if (bdi_dirty <= bdi_stat_error(bdi))
1353                         break;
1354
1355                 if (fatal_signal_pending(current))
1356                         break;
1357         }
1358
1359         if (!dirty_exceeded && bdi->dirty_exceeded)
1360                 bdi->dirty_exceeded = 0;
1361
1362         if (writeback_in_progress(bdi))
1363                 return;
1364
1365         /*
1366          * In laptop mode, we wait until hitting the higher threshold before
1367          * starting background writeout, and then write out all the way down
1368          * to the lower threshold.  So slow writers cause minimal disk activity.
1369          *
1370          * In normal mode, we start background writeout at the lower
1371          * background_thresh, to keep the amount of dirty memory low.
1372          */
1373         if (laptop_mode)
1374                 return;
1375
1376         if (nr_reclaimable > background_thresh)
1377                 bdi_start_background_writeback(bdi);
1378 }
1379
1380 void set_page_dirty_balance(struct page *page, int page_mkwrite)
1381 {
1382         if (set_page_dirty(page) || page_mkwrite) {
1383                 struct address_space *mapping = page_mapping(page);
1384
1385                 if (mapping)
1386                         balance_dirty_pages_ratelimited(mapping);
1387         }
1388 }
1389
1390 static DEFINE_PER_CPU(int, bdp_ratelimits);
1391
1392 /*
1393  * Normal tasks are throttled by
1394  *      loop {
1395  *              dirty tsk->nr_dirtied_pause pages;
1396  *              take a snap in balance_dirty_pages();
1397  *      }
1398  * However there is a worst case. If every task exit immediately when dirtied
1399  * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1400  * called to throttle the page dirties. The solution is to save the not yet
1401  * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1402  * randomly into the running tasks. This works well for the above worst case,
1403  * as the new task will pick up and accumulate the old task's leaked dirty
1404  * count and eventually get throttled.
1405  */
1406 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1407
1408 /**
1409  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
1410  * @mapping: address_space which was dirtied
1411  * @nr_pages_dirtied: number of pages which the caller has just dirtied
1412  *
1413  * Processes which are dirtying memory should call in here once for each page
1414  * which was newly dirtied.  The function will periodically check the system's
1415  * dirty state and will initiate writeback if needed.
1416  *
1417  * On really big machines, get_writeback_state is expensive, so try to avoid
1418  * calling it too often (ratelimiting).  But once we're over the dirty memory
1419  * limit we decrease the ratelimiting by a lot, to prevent individual processes
1420  * from overshooting the limit by (ratelimit_pages) each.
1421  */
1422 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1423                                         unsigned long nr_pages_dirtied)
1424 {
1425         struct backing_dev_info *bdi = mapping->backing_dev_info;
1426         int ratelimit;
1427         int *p;
1428
1429         if (!bdi_cap_account_dirty(bdi))
1430                 return;
1431
1432         ratelimit = current->nr_dirtied_pause;
1433         if (bdi->dirty_exceeded)
1434                 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1435
1436         preempt_disable();
1437         /*
1438          * This prevents one CPU to accumulate too many dirtied pages without
1439          * calling into balance_dirty_pages(), which can happen when there are
1440          * 1000+ tasks, all of them start dirtying pages at exactly the same
1441          * time, hence all honoured too large initial task->nr_dirtied_pause.
1442          */
1443         p =  &__get_cpu_var(bdp_ratelimits);
1444         if (unlikely(current->nr_dirtied >= ratelimit))
1445                 *p = 0;
1446         else if (unlikely(*p >= ratelimit_pages)) {
1447                 *p = 0;
1448                 ratelimit = 0;
1449         }
1450         /*
1451          * Pick up the dirtied pages by the exited tasks. This avoids lots of
1452          * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1453          * the dirty throttling and livelock other long-run dirtiers.
1454          */
1455         p = &__get_cpu_var(dirty_throttle_leaks);
1456         if (*p > 0 && current->nr_dirtied < ratelimit) {
1457                 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1458                 *p -= nr_pages_dirtied;
1459                 current->nr_dirtied += nr_pages_dirtied;
1460         }
1461         preempt_enable();
1462
1463         if (unlikely(current->nr_dirtied >= ratelimit))
1464                 balance_dirty_pages(mapping, current->nr_dirtied);
1465 }
1466 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
1467
1468 void throttle_vm_writeout(gfp_t gfp_mask)
1469 {
1470         unsigned long background_thresh;
1471         unsigned long dirty_thresh;
1472
1473         for ( ; ; ) {
1474                 global_dirty_limits(&background_thresh, &dirty_thresh);
1475
1476                 /*
1477                  * Boost the allowable dirty threshold a bit for page
1478                  * allocators so they don't get DoS'ed by heavy writers
1479                  */
1480                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
1481
1482                 if (global_page_state(NR_UNSTABLE_NFS) +
1483                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
1484                                 break;
1485                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1486
1487                 /*
1488                  * The caller might hold locks which can prevent IO completion
1489                  * or progress in the filesystem.  So we cannot just sit here
1490                  * waiting for IO to complete.
1491                  */
1492                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1493                         break;
1494         }
1495 }
1496
1497 /*
1498  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1499  */
1500 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1501         void __user *buffer, size_t *length, loff_t *ppos)
1502 {
1503         proc_dointvec(table, write, buffer, length, ppos);
1504         bdi_arm_supers_timer();
1505         return 0;
1506 }
1507
1508 #ifdef CONFIG_BLOCK
1509 void laptop_mode_timer_fn(unsigned long data)
1510 {
1511         struct request_queue *q = (struct request_queue *)data;
1512         int nr_pages = global_page_state(NR_FILE_DIRTY) +
1513                 global_page_state(NR_UNSTABLE_NFS);
1514
1515         /*
1516          * We want to write everything out, not just down to the dirty
1517          * threshold
1518          */
1519         if (bdi_has_dirty_io(&q->backing_dev_info))
1520                 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1521                                         WB_REASON_LAPTOP_TIMER);
1522 }
1523
1524 /*
1525  * We've spun up the disk and we're in laptop mode: schedule writeback
1526  * of all dirty data a few seconds from now.  If the flush is already scheduled
1527  * then push it back - the user is still using the disk.
1528  */
1529 void laptop_io_completion(struct backing_dev_info *info)
1530 {
1531         mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1532 }
1533
1534 /*
1535  * We're in laptop mode and we've just synced. The sync's writes will have
1536  * caused another writeback to be scheduled by laptop_io_completion.
1537  * Nothing needs to be written back anymore, so we unschedule the writeback.
1538  */
1539 void laptop_sync_completion(void)
1540 {
1541         struct backing_dev_info *bdi;
1542
1543         rcu_read_lock();
1544
1545         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1546                 del_timer(&bdi->laptop_mode_wb_timer);
1547
1548         rcu_read_unlock();
1549 }
1550 #endif
1551
1552 /*
1553  * If ratelimit_pages is too high then we can get into dirty-data overload
1554  * if a large number of processes all perform writes at the same time.
1555  * If it is too low then SMP machines will call the (expensive)
1556  * get_writeback_state too often.
1557  *
1558  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1559  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1560  * thresholds.
1561  */
1562
1563 void writeback_set_ratelimit(void)
1564 {
1565         unsigned long background_thresh;
1566         unsigned long dirty_thresh;
1567         global_dirty_limits(&background_thresh, &dirty_thresh);
1568         ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1569         if (ratelimit_pages < 16)
1570                 ratelimit_pages = 16;
1571 }
1572
1573 static int __cpuinit
1574 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1575 {
1576         writeback_set_ratelimit();
1577         return NOTIFY_DONE;
1578 }
1579
1580 static struct notifier_block __cpuinitdata ratelimit_nb = {
1581         .notifier_call  = ratelimit_handler,
1582         .next           = NULL,
1583 };
1584
1585 /*
1586  * Called early on to tune the page writeback dirty limits.
1587  *
1588  * We used to scale dirty pages according to how total memory
1589  * related to pages that could be allocated for buffers (by
1590  * comparing nr_free_buffer_pages() to vm_total_pages.
1591  *
1592  * However, that was when we used "dirty_ratio" to scale with
1593  * all memory, and we don't do that any more. "dirty_ratio"
1594  * is now applied to total non-HIGHPAGE memory (by subtracting
1595  * totalhigh_pages from vm_total_pages), and as such we can't
1596  * get into the old insane situation any more where we had
1597  * large amounts of dirty pages compared to a small amount of
1598  * non-HIGHMEM memory.
1599  *
1600  * But we might still want to scale the dirty_ratio by how
1601  * much memory the box has..
1602  */
1603 void __init page_writeback_init(void)
1604 {
1605         int shift;
1606
1607         writeback_set_ratelimit();
1608         register_cpu_notifier(&ratelimit_nb);
1609
1610         shift = calc_period_shift();
1611         prop_descriptor_init(&vm_completions, shift);
1612 }
1613
1614 /**
1615  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1616  * @mapping: address space structure to write
1617  * @start: starting page index
1618  * @end: ending page index (inclusive)
1619  *
1620  * This function scans the page range from @start to @end (inclusive) and tags
1621  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1622  * that write_cache_pages (or whoever calls this function) will then use
1623  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
1624  * used to avoid livelocking of writeback by a process steadily creating new
1625  * dirty pages in the file (thus it is important for this function to be quick
1626  * so that it can tag pages faster than a dirtying process can create them).
1627  */
1628 /*
1629  * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1630  */
1631 void tag_pages_for_writeback(struct address_space *mapping,
1632                              pgoff_t start, pgoff_t end)
1633 {
1634 #define WRITEBACK_TAG_BATCH 4096
1635         unsigned long tagged;
1636
1637         do {
1638                 spin_lock_irq(&mapping->tree_lock);
1639                 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1640                                 &start, end, WRITEBACK_TAG_BATCH,
1641                                 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1642                 spin_unlock_irq(&mapping->tree_lock);
1643                 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1644                 cond_resched();
1645                 /* We check 'start' to handle wrapping when end == ~0UL */
1646         } while (tagged >= WRITEBACK_TAG_BATCH && start);
1647 }
1648 EXPORT_SYMBOL(tag_pages_for_writeback);
1649
1650 /**
1651  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1652  * @mapping: address space structure to write
1653  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1654  * @writepage: function called for each page
1655  * @data: data passed to writepage function
1656  *
1657  * If a page is already under I/O, write_cache_pages() skips it, even
1658  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
1659  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
1660  * and msync() need to guarantee that all the data which was dirty at the time
1661  * the call was made get new I/O started against them.  If wbc->sync_mode is
1662  * WB_SYNC_ALL then we were called for data integrity and we must wait for
1663  * existing IO to complete.
1664  *
1665  * To avoid livelocks (when other process dirties new pages), we first tag
1666  * pages which should be written back with TOWRITE tag and only then start
1667  * writing them. For data-integrity sync we have to be careful so that we do
1668  * not miss some pages (e.g., because some other process has cleared TOWRITE
1669  * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1670  * by the process clearing the DIRTY tag (and submitting the page for IO).
1671  */
1672 int write_cache_pages(struct address_space *mapping,
1673                       struct writeback_control *wbc, writepage_t writepage,
1674                       void *data)
1675 {
1676         int ret = 0;
1677         int done = 0;
1678         struct pagevec pvec;
1679         int nr_pages;
1680         pgoff_t uninitialized_var(writeback_index);
1681         pgoff_t index;
1682         pgoff_t end;            /* Inclusive */
1683         pgoff_t done_index;
1684         int cycled;
1685         int range_whole = 0;
1686         int tag;
1687
1688         pagevec_init(&pvec, 0);
1689         if (wbc->range_cyclic) {
1690                 writeback_index = mapping->writeback_index; /* prev offset */
1691                 index = writeback_index;
1692                 if (index == 0)
1693                         cycled = 1;
1694                 else
1695                         cycled = 0;
1696                 end = -1;
1697         } else {
1698                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1699                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1700                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1701                         range_whole = 1;
1702                 cycled = 1; /* ignore range_cyclic tests */
1703         }
1704         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1705                 tag = PAGECACHE_TAG_TOWRITE;
1706         else
1707                 tag = PAGECACHE_TAG_DIRTY;
1708 retry:
1709         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1710                 tag_pages_for_writeback(mapping, index, end);
1711         done_index = index;
1712         while (!done && (index <= end)) {
1713                 int i;
1714
1715                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1716                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1717                 if (nr_pages == 0)
1718                         break;
1719
1720                 for (i = 0; i < nr_pages; i++) {
1721                         struct page *page = pvec.pages[i];
1722
1723                         /*
1724                          * At this point, the page may be truncated or
1725                          * invalidated (changing page->mapping to NULL), or
1726                          * even swizzled back from swapper_space to tmpfs file
1727                          * mapping. However, page->index will not change
1728                          * because we have a reference on the page.
1729                          */
1730                         if (page->index > end) {
1731                                 /*
1732                                  * can't be range_cyclic (1st pass) because
1733                                  * end == -1 in that case.
1734                                  */
1735                                 done = 1;
1736                                 break;
1737                         }
1738
1739                         done_index = page->index;
1740
1741                         lock_page(page);
1742
1743                         /*
1744                          * Page truncated or invalidated. We can freely skip it
1745                          * then, even for data integrity operations: the page
1746                          * has disappeared concurrently, so there could be no
1747                          * real expectation of this data interity operation
1748                          * even if there is now a new, dirty page at the same
1749                          * pagecache address.
1750                          */
1751                         if (unlikely(page->mapping != mapping)) {
1752 continue_unlock:
1753                                 unlock_page(page);
1754                                 continue;
1755                         }
1756
1757                         if (!PageDirty(page)) {
1758                                 /* someone wrote it for us */
1759                                 goto continue_unlock;
1760                         }
1761
1762                         if (PageWriteback(page)) {
1763                                 if (wbc->sync_mode != WB_SYNC_NONE)
1764                                         wait_on_page_writeback(page);
1765                                 else
1766                                         goto continue_unlock;
1767                         }
1768
1769                         BUG_ON(PageWriteback(page));
1770                         if (!clear_page_dirty_for_io(page))
1771                                 goto continue_unlock;
1772
1773                         trace_wbc_writepage(wbc, mapping->backing_dev_info);
1774                         ret = (*writepage)(page, wbc, data);
1775                         if (unlikely(ret)) {
1776                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1777                                         unlock_page(page);
1778                                         ret = 0;
1779                                 } else {
1780                                         /*
1781                                          * done_index is set past this page,
1782                                          * so media errors will not choke
1783                                          * background writeout for the entire
1784                                          * file. This has consequences for
1785                                          * range_cyclic semantics (ie. it may
1786                                          * not be suitable for data integrity
1787                                          * writeout).
1788                                          */
1789                                         done_index = page->index + 1;
1790                                         done = 1;
1791                                         break;
1792                                 }
1793                         }
1794
1795                         /*
1796                          * We stop writing back only if we are not doing
1797                          * integrity sync. In case of integrity sync we have to
1798                          * keep going until we have written all the pages
1799                          * we tagged for writeback prior to entering this loop.
1800                          */
1801                         if (--wbc->nr_to_write <= 0 &&
1802                             wbc->sync_mode == WB_SYNC_NONE) {
1803                                 done = 1;
1804                                 break;
1805                         }
1806                 }
1807                 pagevec_release(&pvec);
1808                 cond_resched();
1809         }
1810         if (!cycled && !done) {
1811                 /*
1812                  * range_cyclic:
1813                  * We hit the last page and there is more work to be done: wrap
1814                  * back to the start of the file
1815                  */
1816                 cycled = 1;
1817                 index = 0;
1818                 end = writeback_index - 1;
1819                 goto retry;
1820         }
1821         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1822                 mapping->writeback_index = done_index;
1823
1824         return ret;
1825 }
1826 EXPORT_SYMBOL(write_cache_pages);
1827
1828 /*
1829  * Function used by generic_writepages to call the real writepage
1830  * function and set the mapping flags on error
1831  */
1832 static int __writepage(struct page *page, struct writeback_control *wbc,
1833                        void *data)
1834 {
1835         struct address_space *mapping = data;
1836         int ret = mapping->a_ops->writepage(page, wbc);
1837         mapping_set_error(mapping, ret);
1838         return ret;
1839 }
1840
1841 /**
1842  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1843  * @mapping: address space structure to write
1844  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1845  *
1846  * This is a library function, which implements the writepages()
1847  * address_space_operation.
1848  */
1849 int generic_writepages(struct address_space *mapping,
1850                        struct writeback_control *wbc)
1851 {
1852         struct blk_plug plug;
1853         int ret;
1854
1855         /* deal with chardevs and other special file */
1856         if (!mapping->a_ops->writepage)
1857                 return 0;
1858
1859         blk_start_plug(&plug);
1860         ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1861         blk_finish_plug(&plug);
1862         return ret;
1863 }
1864
1865 EXPORT_SYMBOL(generic_writepages);
1866
1867 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1868 {
1869         int ret;
1870
1871         if (wbc->nr_to_write <= 0)
1872                 return 0;
1873         if (mapping->a_ops->writepages)
1874                 ret = mapping->a_ops->writepages(mapping, wbc);
1875         else
1876                 ret = generic_writepages(mapping, wbc);
1877         return ret;
1878 }
1879
1880 /**
1881  * write_one_page - write out a single page and optionally wait on I/O
1882  * @page: the page to write
1883  * @wait: if true, wait on writeout
1884  *
1885  * The page must be locked by the caller and will be unlocked upon return.
1886  *
1887  * write_one_page() returns a negative error code if I/O failed.
1888  */
1889 int write_one_page(struct page *page, int wait)
1890 {
1891         struct address_space *mapping = page->mapping;
1892         int ret = 0;
1893         struct writeback_control wbc = {
1894                 .sync_mode = WB_SYNC_ALL,
1895                 .nr_to_write = 1,
1896         };
1897
1898         BUG_ON(!PageLocked(page));
1899
1900         if (wait)
1901                 wait_on_page_writeback(page);
1902
1903         if (clear_page_dirty_for_io(page)) {
1904                 page_cache_get(page);
1905                 ret = mapping->a_ops->writepage(page, &wbc);
1906                 if (ret == 0 && wait) {
1907                         wait_on_page_writeback(page);
1908                         if (PageError(page))
1909                                 ret = -EIO;
1910                 }
1911                 page_cache_release(page);
1912         } else {
1913                 unlock_page(page);
1914         }
1915         return ret;
1916 }
1917 EXPORT_SYMBOL(write_one_page);
1918
1919 /*
1920  * For address_spaces which do not use buffers nor write back.
1921  */
1922 int __set_page_dirty_no_writeback(struct page *page)
1923 {
1924         if (!PageDirty(page))
1925                 return !TestSetPageDirty(page);
1926         return 0;
1927 }
1928
1929 /*
1930  * Helper function for set_page_dirty family.
1931  * NOTE: This relies on being atomic wrt interrupts.
1932  */
1933 void account_page_dirtied(struct page *page, struct address_space *mapping)
1934 {
1935         if (mapping_cap_account_dirty(mapping)) {
1936                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1937                 __inc_zone_page_state(page, NR_DIRTIED);
1938                 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1939                 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1940                 task_io_account_write(PAGE_CACHE_SIZE);
1941                 current->nr_dirtied++;
1942                 this_cpu_inc(bdp_ratelimits);
1943         }
1944 }
1945 EXPORT_SYMBOL(account_page_dirtied);
1946
1947 /*
1948  * Helper function for set_page_writeback family.
1949  * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1950  * wrt interrupts.
1951  */
1952 void account_page_writeback(struct page *page)
1953 {
1954         inc_zone_page_state(page, NR_WRITEBACK);
1955 }
1956 EXPORT_SYMBOL(account_page_writeback);
1957
1958 /*
1959  * For address_spaces which do not use buffers.  Just tag the page as dirty in
1960  * its radix tree.
1961  *
1962  * This is also used when a single buffer is being dirtied: we want to set the
1963  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
1964  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1965  *
1966  * Most callers have locked the page, which pins the address_space in memory.
1967  * But zap_pte_range() does not lock the page, however in that case the
1968  * mapping is pinned by the vma's ->vm_file reference.
1969  *
1970  * We take care to handle the case where the page was truncated from the
1971  * mapping by re-checking page_mapping() inside tree_lock.
1972  */
1973 int __set_page_dirty_nobuffers(struct page *page)
1974 {
1975         if (!TestSetPageDirty(page)) {
1976                 struct address_space *mapping = page_mapping(page);
1977                 struct address_space *mapping2;
1978
1979                 if (!mapping)
1980                         return 1;
1981
1982                 spin_lock_irq(&mapping->tree_lock);
1983                 mapping2 = page_mapping(page);
1984                 if (mapping2) { /* Race with truncate? */
1985                         BUG_ON(mapping2 != mapping);
1986                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1987                         account_page_dirtied(page, mapping);
1988                         radix_tree_tag_set(&mapping->page_tree,
1989                                 page_index(page), PAGECACHE_TAG_DIRTY);
1990                 }
1991                 spin_unlock_irq(&mapping->tree_lock);
1992                 if (mapping->host) {
1993                         /* !PageAnon && !swapper_space */
1994                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1995                 }
1996                 return 1;
1997         }
1998         return 0;
1999 }
2000 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2001
2002 /*
2003  * Call this whenever redirtying a page, to de-account the dirty counters
2004  * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2005  * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2006  * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2007  * control.
2008  */
2009 void account_page_redirty(struct page *page)
2010 {
2011         struct address_space *mapping = page->mapping;
2012         if (mapping && mapping_cap_account_dirty(mapping)) {
2013                 current->nr_dirtied--;
2014                 dec_zone_page_state(page, NR_DIRTIED);
2015                 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2016         }
2017 }
2018 EXPORT_SYMBOL(account_page_redirty);
2019
2020 /*
2021  * When a writepage implementation decides that it doesn't want to write this
2022  * page for some reason, it should redirty the locked page via
2023  * redirty_page_for_writepage() and it should then unlock the page and return 0
2024  */
2025 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2026 {
2027         wbc->pages_skipped++;
2028         account_page_redirty(page);
2029         return __set_page_dirty_nobuffers(page);
2030 }
2031 EXPORT_SYMBOL(redirty_page_for_writepage);
2032
2033 /*
2034  * Dirty a page.
2035  *
2036  * For pages with a mapping this should be done under the page lock
2037  * for the benefit of asynchronous memory errors who prefer a consistent
2038  * dirty state. This rule can be broken in some special cases,
2039  * but should be better not to.
2040  *
2041  * If the mapping doesn't provide a set_page_dirty a_op, then
2042  * just fall through and assume that it wants buffer_heads.
2043  */
2044 int set_page_dirty(struct page *page)
2045 {
2046         struct address_space *mapping = page_mapping(page);
2047
2048         if (likely(mapping)) {
2049                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2050                 /*
2051                  * readahead/lru_deactivate_page could remain
2052                  * PG_readahead/PG_reclaim due to race with end_page_writeback
2053                  * About readahead, if the page is written, the flags would be
2054                  * reset. So no problem.
2055                  * About lru_deactivate_page, if the page is redirty, the flag
2056                  * will be reset. So no problem. but if the page is used by readahead
2057                  * it will confuse readahead and make it restart the size rampup
2058                  * process. But it's a trivial problem.
2059                  */
2060                 ClearPageReclaim(page);
2061 #ifdef CONFIG_BLOCK
2062                 if (!spd)
2063                         spd = __set_page_dirty_buffers;
2064 #endif
2065                 return (*spd)(page);
2066         }
2067         if (!PageDirty(page)) {
2068                 if (!TestSetPageDirty(page))
2069                         return 1;
2070         }
2071         return 0;
2072 }
2073 EXPORT_SYMBOL(set_page_dirty);
2074
2075 /*
2076  * set_page_dirty() is racy if the caller has no reference against
2077  * page->mapping->host, and if the page is unlocked.  This is because another
2078  * CPU could truncate the page off the mapping and then free the mapping.
2079  *
2080  * Usually, the page _is_ locked, or the caller is a user-space process which
2081  * holds a reference on the inode by having an open file.
2082  *
2083  * In other cases, the page should be locked before running set_page_dirty().
2084  */
2085 int set_page_dirty_lock(struct page *page)
2086 {
2087         int ret;
2088
2089         lock_page(page);
2090         ret = set_page_dirty(page);
2091         unlock_page(page);
2092         return ret;
2093 }
2094 EXPORT_SYMBOL(set_page_dirty_lock);
2095
2096 /*
2097  * Clear a page's dirty flag, while caring for dirty memory accounting.
2098  * Returns true if the page was previously dirty.
2099  *
2100  * This is for preparing to put the page under writeout.  We leave the page
2101  * tagged as dirty in the radix tree so that a concurrent write-for-sync
2102  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
2103  * implementation will run either set_page_writeback() or set_page_dirty(),
2104  * at which stage we bring the page's dirty flag and radix-tree dirty tag
2105  * back into sync.
2106  *
2107  * This incoherency between the page's dirty flag and radix-tree tag is
2108  * unfortunate, but it only exists while the page is locked.
2109  */
2110 int clear_page_dirty_for_io(struct page *page)
2111 {
2112         struct address_space *mapping = page_mapping(page);
2113
2114         BUG_ON(!PageLocked(page));
2115
2116         if (mapping && mapping_cap_account_dirty(mapping)) {
2117                 /*
2118                  * Yes, Virginia, this is indeed insane.
2119                  *
2120                  * We use this sequence to make sure that
2121                  *  (a) we account for dirty stats properly
2122                  *  (b) we tell the low-level filesystem to
2123                  *      mark the whole page dirty if it was
2124                  *      dirty in a pagetable. Only to then
2125                  *  (c) clean the page again and return 1 to
2126                  *      cause the writeback.
2127                  *
2128                  * This way we avoid all nasty races with the
2129                  * dirty bit in multiple places and clearing
2130                  * them concurrently from different threads.
2131                  *
2132                  * Note! Normally the "set_page_dirty(page)"
2133                  * has no effect on the actual dirty bit - since
2134                  * that will already usually be set. But we
2135                  * need the side effects, and it can help us
2136                  * avoid races.
2137                  *
2138                  * We basically use the page "master dirty bit"
2139                  * as a serialization point for all the different
2140                  * threads doing their things.
2141                  */
2142                 if (page_mkclean(page))
2143                         set_page_dirty(page);
2144                 /*
2145                  * We carefully synchronise fault handlers against
2146                  * installing a dirty pte and marking the page dirty
2147                  * at this point. We do this by having them hold the
2148                  * page lock at some point after installing their
2149                  * pte, but before marking the page dirty.
2150                  * Pages are always locked coming in here, so we get
2151                  * the desired exclusion. See mm/memory.c:do_wp_page()
2152                  * for more comments.
2153                  */
2154                 if (TestClearPageDirty(page)) {
2155                         dec_zone_page_state(page, NR_FILE_DIRTY);
2156                         dec_bdi_stat(mapping->backing_dev_info,
2157                                         BDI_RECLAIMABLE);
2158                         return 1;
2159                 }
2160                 return 0;
2161         }
2162         return TestClearPageDirty(page);
2163 }
2164 EXPORT_SYMBOL(clear_page_dirty_for_io);
2165
2166 int test_clear_page_writeback(struct page *page)
2167 {
2168         struct address_space *mapping = page_mapping(page);
2169         int ret;
2170
2171         if (mapping) {
2172                 struct backing_dev_info *bdi = mapping->backing_dev_info;
2173                 unsigned long flags;
2174
2175                 spin_lock_irqsave(&mapping->tree_lock, flags);
2176                 ret = TestClearPageWriteback(page);
2177                 if (ret) {
2178                         radix_tree_tag_clear(&mapping->page_tree,
2179                                                 page_index(page),
2180                                                 PAGECACHE_TAG_WRITEBACK);
2181                         if (bdi_cap_account_writeback(bdi)) {
2182                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
2183                                 __bdi_writeout_inc(bdi);
2184                         }
2185                 }
2186                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2187         } else {
2188                 ret = TestClearPageWriteback(page);
2189         }
2190         if (ret) {
2191                 dec_zone_page_state(page, NR_WRITEBACK);
2192                 inc_zone_page_state(page, NR_WRITTEN);
2193         }
2194         return ret;
2195 }
2196
2197 int test_set_page_writeback(struct page *page)
2198 {
2199         struct address_space *mapping = page_mapping(page);
2200         int ret;
2201
2202         if (mapping) {
2203                 struct backing_dev_info *bdi = mapping->backing_dev_info;
2204                 unsigned long flags;
2205
2206                 spin_lock_irqsave(&mapping->tree_lock, flags);
2207                 ret = TestSetPageWriteback(page);
2208                 if (!ret) {
2209                         radix_tree_tag_set(&mapping->page_tree,
2210                                                 page_index(page),
2211                                                 PAGECACHE_TAG_WRITEBACK);
2212                         if (bdi_cap_account_writeback(bdi))
2213                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2214                 }
2215                 if (!PageDirty(page))
2216                         radix_tree_tag_clear(&mapping->page_tree,
2217                                                 page_index(page),
2218                                                 PAGECACHE_TAG_DIRTY);
2219                 radix_tree_tag_clear(&mapping->page_tree,
2220                                      page_index(page),
2221                                      PAGECACHE_TAG_TOWRITE);
2222                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2223         } else {
2224                 ret = TestSetPageWriteback(page);
2225         }
2226         if (!ret)
2227                 account_page_writeback(page);
2228         return ret;
2229
2230 }
2231 EXPORT_SYMBOL(test_set_page_writeback);
2232
2233 /*
2234  * Return true if any of the pages in the mapping are marked with the
2235  * passed tag.
2236  */
2237 int mapping_tagged(struct address_space *mapping, int tag)
2238 {
2239         return radix_tree_tagged(&mapping->page_tree, tag);
2240 }
2241 EXPORT_SYMBOL(mapping_tagged);