Merge remote-tracking branch 'lsk/v3.10/topic/arm64-perf' into linux-linaro-lsk
[firefly-linux-kernel-4.4.55.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 <linux/timer.h>
38 #include <linux/sched/rt.h>
39 #include <trace/events/writeback.h>
40
41 /*
42  * Sleep at most 200ms at a time in balance_dirty_pages().
43  */
44 #define MAX_PAUSE               max(HZ/5, 1)
45
46 /*
47  * Try to keep balance_dirty_pages() call intervals higher than this many pages
48  * by raising pause time to max_pause when falls below it.
49  */
50 #define DIRTY_POLL_THRESH       (128 >> (PAGE_SHIFT - 10))
51
52 /*
53  * Estimate write bandwidth at 200ms intervals.
54  */
55 #define BANDWIDTH_INTERVAL      max(HZ/5, 1)
56
57 #define RATELIMIT_CALC_SHIFT    10
58
59 /*
60  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
61  * will look to see if it needs to force writeback or throttling.
62  */
63 static long ratelimit_pages = 32;
64
65 /* The following parameters are exported via /proc/sys/vm */
66
67 /*
68  * Start background writeback (via writeback threads) at this percentage
69  */
70 int dirty_background_ratio = 10;
71
72 /*
73  * dirty_background_bytes starts at 0 (disabled) so that it is a function of
74  * dirty_background_ratio * the amount of dirtyable memory
75  */
76 unsigned long dirty_background_bytes;
77
78 /*
79  * free highmem will not be subtracted from the total free memory
80  * for calculating free ratios if vm_highmem_is_dirtyable is true
81  */
82 int vm_highmem_is_dirtyable;
83
84 /*
85  * The generator of dirty data starts writeback at this percentage
86  */
87 int vm_dirty_ratio = 20;
88
89 /*
90  * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
91  * vm_dirty_ratio * the amount of dirtyable memory
92  */
93 unsigned long vm_dirty_bytes;
94
95 /*
96  * The interval between `kupdate'-style writebacks
97  */
98 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
99
100 EXPORT_SYMBOL_GPL(dirty_writeback_interval);
101
102 /*
103  * The longest time for which data is allowed to remain dirty
104  */
105 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
106
107 /*
108  * Flag that makes the machine dump writes/reads and block dirtyings.
109  */
110 int block_dump;
111
112 /*
113  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
114  * a full sync is triggered after this time elapses without any disk activity.
115  */
116 int laptop_mode;
117
118 EXPORT_SYMBOL(laptop_mode);
119
120 /* End of sysctl-exported parameters */
121
122 unsigned long global_dirty_limit;
123
124 /*
125  * Scale the writeback cache size proportional to the relative writeout speeds.
126  *
127  * We do this by keeping a floating proportion between BDIs, based on page
128  * writeback completions [end_page_writeback()]. Those devices that write out
129  * pages fastest will get the larger share, while the slower will get a smaller
130  * share.
131  *
132  * We use page writeout completions because we are interested in getting rid of
133  * dirty pages. Having them written out is the primary goal.
134  *
135  * We introduce a concept of time, a period over which we measure these events,
136  * because demand can/will vary over time. The length of this period itself is
137  * measured in page writeback completions.
138  *
139  */
140 static struct fprop_global writeout_completions;
141
142 static void writeout_period(unsigned long t);
143 /* Timer for aging of writeout_completions */
144 static struct timer_list writeout_period_timer =
145                 TIMER_DEFERRED_INITIALIZER(writeout_period, 0, 0);
146 static unsigned long writeout_period_time = 0;
147
148 /*
149  * Length of period for aging writeout fractions of bdis. This is an
150  * arbitrarily chosen number. The longer the period, the slower fractions will
151  * reflect changes in current writeout rate.
152  */
153 #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
154
155 /*
156  * Work out the current dirty-memory clamping and background writeout
157  * thresholds.
158  *
159  * The main aim here is to lower them aggressively if there is a lot of mapped
160  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
161  * pages.  It is better to clamp down on writers than to start swapping, and
162  * performing lots of scanning.
163  *
164  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
165  *
166  * We don't permit the clamping level to fall below 5% - that is getting rather
167  * excessive.
168  *
169  * We make sure that the background writeout level is below the adjusted
170  * clamping level.
171  */
172
173 /*
174  * In a memory zone, there is a certain amount of pages we consider
175  * available for the page cache, which is essentially the number of
176  * free and reclaimable pages, minus some zone reserves to protect
177  * lowmem and the ability to uphold the zone's watermarks without
178  * requiring writeback.
179  *
180  * This number of dirtyable pages is the base value of which the
181  * user-configurable dirty ratio is the effictive number of pages that
182  * are allowed to be actually dirtied.  Per individual zone, or
183  * globally by using the sum of dirtyable pages over all zones.
184  *
185  * Because the user is allowed to specify the dirty limit globally as
186  * absolute number of bytes, calculating the per-zone dirty limit can
187  * require translating the configured limit into a percentage of
188  * global dirtyable memory first.
189  */
190
191 /**
192  * zone_dirtyable_memory - number of dirtyable pages in a zone
193  * @zone: the zone
194  *
195  * Returns the zone's number of pages potentially available for dirty
196  * page cache.  This is the base value for the per-zone dirty limits.
197  */
198 static unsigned long zone_dirtyable_memory(struct zone *zone)
199 {
200         unsigned long nr_pages;
201
202         nr_pages = zone_page_state(zone, NR_FREE_PAGES);
203         nr_pages -= min(nr_pages, zone->dirty_balance_reserve);
204
205         nr_pages += zone_page_state(zone, NR_INACTIVE_FILE);
206         nr_pages += zone_page_state(zone, NR_ACTIVE_FILE);
207
208         return nr_pages;
209 }
210
211 static unsigned long highmem_dirtyable_memory(unsigned long total)
212 {
213 #ifdef CONFIG_HIGHMEM
214         int node;
215         unsigned long x = 0;
216
217         for_each_node_state(node, N_HIGH_MEMORY) {
218                 struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
219
220                 x += zone_dirtyable_memory(z);
221         }
222         /*
223          * Unreclaimable memory (kernel memory or anonymous memory
224          * without swap) can bring down the dirtyable pages below
225          * the zone's dirty balance reserve and the above calculation
226          * will underflow.  However we still want to add in nodes
227          * which are below threshold (negative values) to get a more
228          * accurate calculation but make sure that the total never
229          * underflows.
230          */
231         if ((long)x < 0)
232                 x = 0;
233
234         /*
235          * Make sure that the number of highmem pages is never larger
236          * than the number of the total dirtyable memory. This can only
237          * occur in very strange VM situations but we want to make sure
238          * that this does not occur.
239          */
240         return min(x, total);
241 #else
242         return 0;
243 #endif
244 }
245
246 /**
247  * global_dirtyable_memory - number of globally dirtyable pages
248  *
249  * Returns the global number of pages potentially available for dirty
250  * page cache.  This is the base value for the global dirty limits.
251  */
252 static unsigned long global_dirtyable_memory(void)
253 {
254         unsigned long x;
255
256         x = global_page_state(NR_FREE_PAGES);
257         x -= min(x, dirty_balance_reserve);
258
259         x += global_page_state(NR_INACTIVE_FILE);
260         x += global_page_state(NR_ACTIVE_FILE);
261
262         if (!vm_highmem_is_dirtyable)
263                 x -= highmem_dirtyable_memory(x);
264
265         /* Subtract min_free_kbytes */
266         x -= min_t(unsigned long, x, min_free_kbytes >> (PAGE_SHIFT - 10));
267
268         return x + 1;   /* Ensure that we never return 0 */
269 }
270
271 /*
272  * global_dirty_limits - background-writeback and dirty-throttling thresholds
273  *
274  * Calculate the dirty thresholds based on sysctl parameters
275  * - vm.dirty_background_ratio  or  vm.dirty_background_bytes
276  * - vm.dirty_ratio             or  vm.dirty_bytes
277  * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
278  * real-time tasks.
279  */
280 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
281 {
282         unsigned long background;
283         unsigned long dirty;
284         unsigned long uninitialized_var(available_memory);
285         struct task_struct *tsk;
286
287         if (!vm_dirty_bytes || !dirty_background_bytes)
288                 available_memory = global_dirtyable_memory();
289
290         if (vm_dirty_bytes)
291                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
292         else
293                 dirty = (vm_dirty_ratio * available_memory) / 100;
294
295         if (dirty_background_bytes)
296                 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
297         else
298                 background = (dirty_background_ratio * available_memory) / 100;
299
300         if (background >= dirty)
301                 background = dirty / 2;
302         tsk = current;
303         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
304                 background += background / 4;
305                 dirty += dirty / 4;
306         }
307         *pbackground = background;
308         *pdirty = dirty;
309         trace_global_dirty_state(background, dirty);
310 }
311
312 /**
313  * zone_dirty_limit - maximum number of dirty pages allowed in a zone
314  * @zone: the zone
315  *
316  * Returns the maximum number of dirty pages allowed in a zone, based
317  * on the zone's dirtyable memory.
318  */
319 static unsigned long zone_dirty_limit(struct zone *zone)
320 {
321         unsigned long zone_memory = zone_dirtyable_memory(zone);
322         struct task_struct *tsk = current;
323         unsigned long dirty;
324
325         if (vm_dirty_bytes)
326                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
327                         zone_memory / global_dirtyable_memory();
328         else
329                 dirty = vm_dirty_ratio * zone_memory / 100;
330
331         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
332                 dirty += dirty / 4;
333
334         return dirty;
335 }
336
337 /**
338  * zone_dirty_ok - tells whether a zone is within its dirty limits
339  * @zone: the zone to check
340  *
341  * Returns %true when the dirty pages in @zone are within the zone's
342  * dirty limit, %false if the limit is exceeded.
343  */
344 bool zone_dirty_ok(struct zone *zone)
345 {
346         unsigned long limit = zone_dirty_limit(zone);
347
348         return zone_page_state(zone, NR_FILE_DIRTY) +
349                zone_page_state(zone, NR_UNSTABLE_NFS) +
350                zone_page_state(zone, NR_WRITEBACK) <= limit;
351 }
352
353 int dirty_background_ratio_handler(struct ctl_table *table, int write,
354                 void __user *buffer, size_t *lenp,
355                 loff_t *ppos)
356 {
357         int ret;
358
359         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
360         if (ret == 0 && write)
361                 dirty_background_bytes = 0;
362         return ret;
363 }
364
365 int dirty_background_bytes_handler(struct ctl_table *table, int write,
366                 void __user *buffer, size_t *lenp,
367                 loff_t *ppos)
368 {
369         int ret;
370
371         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
372         if (ret == 0 && write)
373                 dirty_background_ratio = 0;
374         return ret;
375 }
376
377 int dirty_ratio_handler(struct ctl_table *table, int write,
378                 void __user *buffer, size_t *lenp,
379                 loff_t *ppos)
380 {
381         int old_ratio = vm_dirty_ratio;
382         int ret;
383
384         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
385         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
386                 writeback_set_ratelimit();
387                 vm_dirty_bytes = 0;
388         }
389         return ret;
390 }
391
392 int dirty_bytes_handler(struct ctl_table *table, int write,
393                 void __user *buffer, size_t *lenp,
394                 loff_t *ppos)
395 {
396         unsigned long old_bytes = vm_dirty_bytes;
397         int ret;
398
399         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
400         if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
401                 writeback_set_ratelimit();
402                 vm_dirty_ratio = 0;
403         }
404         return ret;
405 }
406
407 static unsigned long wp_next_time(unsigned long cur_time)
408 {
409         cur_time += VM_COMPLETIONS_PERIOD_LEN;
410         /* 0 has a special meaning... */
411         if (!cur_time)
412                 return 1;
413         return cur_time;
414 }
415
416 /*
417  * Increment the BDI's writeout completion count and the global writeout
418  * completion count. Called from test_clear_page_writeback().
419  */
420 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
421 {
422         __inc_bdi_stat(bdi, BDI_WRITTEN);
423         __fprop_inc_percpu_max(&writeout_completions, &bdi->completions,
424                                bdi->max_prop_frac);
425         /* First event after period switching was turned off? */
426         if (!unlikely(writeout_period_time)) {
427                 /*
428                  * We can race with other __bdi_writeout_inc calls here but
429                  * it does not cause any harm since the resulting time when
430                  * timer will fire and what is in writeout_period_time will be
431                  * roughly the same.
432                  */
433                 writeout_period_time = wp_next_time(jiffies);
434                 mod_timer(&writeout_period_timer, writeout_period_time);
435         }
436 }
437
438 void bdi_writeout_inc(struct backing_dev_info *bdi)
439 {
440         unsigned long flags;
441
442         local_irq_save(flags);
443         __bdi_writeout_inc(bdi);
444         local_irq_restore(flags);
445 }
446 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
447
448 /*
449  * Obtain an accurate fraction of the BDI's portion.
450  */
451 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
452                 long *numerator, long *denominator)
453 {
454         fprop_fraction_percpu(&writeout_completions, &bdi->completions,
455                                 numerator, denominator);
456 }
457
458 /*
459  * On idle system, we can be called long after we scheduled because we use
460  * deferred timers so count with missed periods.
461  */
462 static void writeout_period(unsigned long t)
463 {
464         int miss_periods = (jiffies - writeout_period_time) /
465                                                  VM_COMPLETIONS_PERIOD_LEN;
466
467         if (fprop_new_period(&writeout_completions, miss_periods + 1)) {
468                 writeout_period_time = wp_next_time(writeout_period_time +
469                                 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
470                 mod_timer(&writeout_period_timer, writeout_period_time);
471         } else {
472                 /*
473                  * Aging has zeroed all fractions. Stop wasting CPU on period
474                  * updates.
475                  */
476                 writeout_period_time = 0;
477         }
478 }
479
480 /*
481  * bdi_min_ratio keeps the sum of the minimum dirty shares of all
482  * registered backing devices, which, for obvious reasons, can not
483  * exceed 100%.
484  */
485 static unsigned int bdi_min_ratio;
486
487 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
488 {
489         int ret = 0;
490
491         spin_lock_bh(&bdi_lock);
492         if (min_ratio > bdi->max_ratio) {
493                 ret = -EINVAL;
494         } else {
495                 min_ratio -= bdi->min_ratio;
496                 if (bdi_min_ratio + min_ratio < 100) {
497                         bdi_min_ratio += min_ratio;
498                         bdi->min_ratio += min_ratio;
499                 } else {
500                         ret = -EINVAL;
501                 }
502         }
503         spin_unlock_bh(&bdi_lock);
504
505         return ret;
506 }
507
508 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
509 {
510         int ret = 0;
511
512         if (max_ratio > 100)
513                 return -EINVAL;
514
515         spin_lock_bh(&bdi_lock);
516         if (bdi->min_ratio > max_ratio) {
517                 ret = -EINVAL;
518         } else {
519                 bdi->max_ratio = max_ratio;
520                 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
521         }
522         spin_unlock_bh(&bdi_lock);
523
524         return ret;
525 }
526 EXPORT_SYMBOL(bdi_set_max_ratio);
527
528 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
529                                            unsigned long bg_thresh)
530 {
531         return (thresh + bg_thresh) / 2;
532 }
533
534 static unsigned long hard_dirty_limit(unsigned long thresh)
535 {
536         return max(thresh, global_dirty_limit);
537 }
538
539 /**
540  * bdi_dirty_limit - @bdi's share of dirty throttling threshold
541  * @bdi: the backing_dev_info to query
542  * @dirty: global dirty limit in pages
543  *
544  * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
545  * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
546  *
547  * Note that balance_dirty_pages() will only seriously take it as a hard limit
548  * when sleeping max_pause per page is not enough to keep the dirty pages under
549  * control. For example, when the device is completely stalled due to some error
550  * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
551  * In the other normal situations, it acts more gently by throttling the tasks
552  * more (rather than completely block them) when the bdi dirty pages go high.
553  *
554  * It allocates high/low dirty limits to fast/slow devices, in order to prevent
555  * - starving fast devices
556  * - piling up dirty pages (that will take long time to sync) on slow devices
557  *
558  * The bdi's share of dirty limit will be adapting to its throughput and
559  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
560  */
561 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
562 {
563         u64 bdi_dirty;
564         long numerator, denominator;
565
566         /*
567          * Calculate this BDI's share of the dirty ratio.
568          */
569         bdi_writeout_fraction(bdi, &numerator, &denominator);
570
571         bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
572         bdi_dirty *= numerator;
573         do_div(bdi_dirty, denominator);
574
575         bdi_dirty += (dirty * bdi->min_ratio) / 100;
576         if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
577                 bdi_dirty = dirty * bdi->max_ratio / 100;
578
579         return bdi_dirty;
580 }
581
582 /*
583  * Dirty position control.
584  *
585  * (o) global/bdi setpoints
586  *
587  * We want the dirty pages be balanced around the global/bdi setpoints.
588  * When the number of dirty pages is higher/lower than the setpoint, the
589  * dirty position control ratio (and hence task dirty ratelimit) will be
590  * decreased/increased to bring the dirty pages back to the setpoint.
591  *
592  *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
593  *
594  *     if (dirty < setpoint) scale up   pos_ratio
595  *     if (dirty > setpoint) scale down pos_ratio
596  *
597  *     if (bdi_dirty < bdi_setpoint) scale up   pos_ratio
598  *     if (bdi_dirty > bdi_setpoint) scale down pos_ratio
599  *
600  *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
601  *
602  * (o) global control line
603  *
604  *     ^ pos_ratio
605  *     |
606  *     |            |<===== global dirty control scope ======>|
607  * 2.0 .............*
608  *     |            .*
609  *     |            . *
610  *     |            .   *
611  *     |            .     *
612  *     |            .        *
613  *     |            .            *
614  * 1.0 ................................*
615  *     |            .                  .     *
616  *     |            .                  .          *
617  *     |            .                  .              *
618  *     |            .                  .                 *
619  *     |            .                  .                    *
620  *   0 +------------.------------------.----------------------*------------->
621  *           freerun^          setpoint^                 limit^   dirty pages
622  *
623  * (o) bdi control line
624  *
625  *     ^ pos_ratio
626  *     |
627  *     |            *
628  *     |              *
629  *     |                *
630  *     |                  *
631  *     |                    * |<=========== span ============>|
632  * 1.0 .......................*
633  *     |                      . *
634  *     |                      .   *
635  *     |                      .     *
636  *     |                      .       *
637  *     |                      .         *
638  *     |                      .           *
639  *     |                      .             *
640  *     |                      .               *
641  *     |                      .                 *
642  *     |                      .                   *
643  *     |                      .                     *
644  * 1/4 ...............................................* * * * * * * * * * * *
645  *     |                      .                         .
646  *     |                      .                           .
647  *     |                      .                             .
648  *   0 +----------------------.-------------------------------.------------->
649  *                bdi_setpoint^                    x_intercept^
650  *
651  * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
652  * be smoothly throttled down to normal if it starts high in situations like
653  * - start writing to a slow SD card and a fast disk at the same time. The SD
654  *   card's bdi_dirty may rush to many times higher than bdi_setpoint.
655  * - the bdi dirty thresh drops quickly due to change of JBOD workload
656  */
657 static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
658                                         unsigned long thresh,
659                                         unsigned long bg_thresh,
660                                         unsigned long dirty,
661                                         unsigned long bdi_thresh,
662                                         unsigned long bdi_dirty)
663 {
664         unsigned long write_bw = bdi->avg_write_bandwidth;
665         unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
666         unsigned long limit = hard_dirty_limit(thresh);
667         unsigned long x_intercept;
668         unsigned long setpoint;         /* dirty pages' target balance point */
669         unsigned long bdi_setpoint;
670         unsigned long span;
671         long long pos_ratio;            /* for scaling up/down the rate limit */
672         long x;
673
674         if (unlikely(dirty >= limit))
675                 return 0;
676
677         /*
678          * global setpoint
679          *
680          *                           setpoint - dirty 3
681          *        f(dirty) := 1.0 + (----------------)
682          *                           limit - setpoint
683          *
684          * it's a 3rd order polynomial that subjects to
685          *
686          * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
687          * (2) f(setpoint) = 1.0 => the balance point
688          * (3) f(limit)    = 0   => the hard limit
689          * (4) df/dx      <= 0   => negative feedback control
690          * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
691          *     => fast response on large errors; small oscillation near setpoint
692          */
693         setpoint = (freerun + limit) / 2;
694         x = div_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
695                     limit - setpoint + 1);
696         pos_ratio = x;
697         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
698         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
699         pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
700
701         /*
702          * We have computed basic pos_ratio above based on global situation. If
703          * the bdi is over/under its share of dirty pages, we want to scale
704          * pos_ratio further down/up. That is done by the following mechanism.
705          */
706
707         /*
708          * bdi setpoint
709          *
710          *        f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
711          *
712          *                        x_intercept - bdi_dirty
713          *                     := --------------------------
714          *                        x_intercept - bdi_setpoint
715          *
716          * The main bdi control line is a linear function that subjects to
717          *
718          * (1) f(bdi_setpoint) = 1.0
719          * (2) k = - 1 / (8 * write_bw)  (in single bdi case)
720          *     or equally: x_intercept = bdi_setpoint + 8 * write_bw
721          *
722          * For single bdi case, the dirty pages are observed to fluctuate
723          * regularly within range
724          *        [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
725          * for various filesystems, where (2) can yield in a reasonable 12.5%
726          * fluctuation range for pos_ratio.
727          *
728          * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
729          * own size, so move the slope over accordingly and choose a slope that
730          * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
731          */
732         if (unlikely(bdi_thresh > thresh))
733                 bdi_thresh = thresh;
734         /*
735          * It's very possible that bdi_thresh is close to 0 not because the
736          * device is slow, but that it has remained inactive for long time.
737          * Honour such devices a reasonable good (hopefully IO efficient)
738          * threshold, so that the occasional writes won't be blocked and active
739          * writes can rampup the threshold quickly.
740          */
741         bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
742         /*
743          * scale global setpoint to bdi's:
744          *      bdi_setpoint = setpoint * bdi_thresh / thresh
745          */
746         x = div_u64((u64)bdi_thresh << 16, thresh + 1);
747         bdi_setpoint = setpoint * (u64)x >> 16;
748         /*
749          * Use span=(8*write_bw) in single bdi case as indicated by
750          * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
751          *
752          *        bdi_thresh                    thresh - bdi_thresh
753          * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
754          *          thresh                            thresh
755          */
756         span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
757         x_intercept = bdi_setpoint + span;
758
759         if (bdi_dirty < x_intercept - span / 4) {
760                 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
761                                     x_intercept - bdi_setpoint + 1);
762         } else
763                 pos_ratio /= 4;
764
765         /*
766          * bdi reserve area, safeguard against dirty pool underrun and disk idle
767          * It may push the desired control point of global dirty pages higher
768          * than setpoint.
769          */
770         x_intercept = bdi_thresh / 2;
771         if (bdi_dirty < x_intercept) {
772                 if (bdi_dirty > x_intercept / 8)
773                         pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
774                 else
775                         pos_ratio *= 8;
776         }
777
778         return pos_ratio;
779 }
780
781 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
782                                        unsigned long elapsed,
783                                        unsigned long written)
784 {
785         const unsigned long period = roundup_pow_of_two(3 * HZ);
786         unsigned long avg = bdi->avg_write_bandwidth;
787         unsigned long old = bdi->write_bandwidth;
788         u64 bw;
789
790         /*
791          * bw = written * HZ / elapsed
792          *
793          *                   bw * elapsed + write_bandwidth * (period - elapsed)
794          * write_bandwidth = ---------------------------------------------------
795          *                                          period
796          */
797         bw = written - bdi->written_stamp;
798         bw *= HZ;
799         if (unlikely(elapsed > period)) {
800                 do_div(bw, elapsed);
801                 avg = bw;
802                 goto out;
803         }
804         bw += (u64)bdi->write_bandwidth * (period - elapsed);
805         bw >>= ilog2(period);
806
807         /*
808          * one more level of smoothing, for filtering out sudden spikes
809          */
810         if (avg > old && old >= (unsigned long)bw)
811                 avg -= (avg - old) >> 3;
812
813         if (avg < old && old <= (unsigned long)bw)
814                 avg += (old - avg) >> 3;
815
816 out:
817         bdi->write_bandwidth = bw;
818         bdi->avg_write_bandwidth = avg;
819 }
820
821 /*
822  * The global dirtyable memory and dirty threshold could be suddenly knocked
823  * down by a large amount (eg. on the startup of KVM in a swapless system).
824  * This may throw the system into deep dirty exceeded state and throttle
825  * heavy/light dirtiers alike. To retain good responsiveness, maintain
826  * global_dirty_limit for tracking slowly down to the knocked down dirty
827  * threshold.
828  */
829 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
830 {
831         unsigned long limit = global_dirty_limit;
832
833         /*
834          * Follow up in one step.
835          */
836         if (limit < thresh) {
837                 limit = thresh;
838                 goto update;
839         }
840
841         /*
842          * Follow down slowly. Use the higher one as the target, because thresh
843          * may drop below dirty. This is exactly the reason to introduce
844          * global_dirty_limit which is guaranteed to lie above the dirty pages.
845          */
846         thresh = max(thresh, dirty);
847         if (limit > thresh) {
848                 limit -= (limit - thresh) >> 5;
849                 goto update;
850         }
851         return;
852 update:
853         global_dirty_limit = limit;
854 }
855
856 static void global_update_bandwidth(unsigned long thresh,
857                                     unsigned long dirty,
858                                     unsigned long now)
859 {
860         static DEFINE_SPINLOCK(dirty_lock);
861         static unsigned long update_time;
862
863         /*
864          * check locklessly first to optimize away locking for the most time
865          */
866         if (time_before(now, update_time + BANDWIDTH_INTERVAL))
867                 return;
868
869         spin_lock(&dirty_lock);
870         if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
871                 update_dirty_limit(thresh, dirty);
872                 update_time = now;
873         }
874         spin_unlock(&dirty_lock);
875 }
876
877 /*
878  * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
879  *
880  * Normal bdi tasks will be curbed at or below it in long term.
881  * Obviously it should be around (write_bw / N) when there are N dd tasks.
882  */
883 static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
884                                        unsigned long thresh,
885                                        unsigned long bg_thresh,
886                                        unsigned long dirty,
887                                        unsigned long bdi_thresh,
888                                        unsigned long bdi_dirty,
889                                        unsigned long dirtied,
890                                        unsigned long elapsed)
891 {
892         unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
893         unsigned long limit = hard_dirty_limit(thresh);
894         unsigned long setpoint = (freerun + limit) / 2;
895         unsigned long write_bw = bdi->avg_write_bandwidth;
896         unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
897         unsigned long dirty_rate;
898         unsigned long task_ratelimit;
899         unsigned long balanced_dirty_ratelimit;
900         unsigned long pos_ratio;
901         unsigned long step;
902         unsigned long x;
903
904         /*
905          * The dirty rate will match the writeout rate in long term, except
906          * when dirty pages are truncated by userspace or re-dirtied by FS.
907          */
908         dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
909
910         pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
911                                        bdi_thresh, bdi_dirty);
912         /*
913          * task_ratelimit reflects each dd's dirty rate for the past 200ms.
914          */
915         task_ratelimit = (u64)dirty_ratelimit *
916                                         pos_ratio >> RATELIMIT_CALC_SHIFT;
917         task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
918
919         /*
920          * A linear estimation of the "balanced" throttle rate. The theory is,
921          * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
922          * dirty_rate will be measured to be (N * task_ratelimit). So the below
923          * formula will yield the balanced rate limit (write_bw / N).
924          *
925          * Note that the expanded form is not a pure rate feedback:
926          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate)              (1)
927          * but also takes pos_ratio into account:
928          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
929          *
930          * (1) is not realistic because pos_ratio also takes part in balancing
931          * the dirty rate.  Consider the state
932          *      pos_ratio = 0.5                                              (3)
933          *      rate = 2 * (write_bw / N)                                    (4)
934          * If (1) is used, it will stuck in that state! Because each dd will
935          * be throttled at
936          *      task_ratelimit = pos_ratio * rate = (write_bw / N)           (5)
937          * yielding
938          *      dirty_rate = N * task_ratelimit = write_bw                   (6)
939          * put (6) into (1) we get
940          *      rate_(i+1) = rate_(i)                                        (7)
941          *
942          * So we end up using (2) to always keep
943          *      rate_(i+1) ~= (write_bw / N)                                 (8)
944          * regardless of the value of pos_ratio. As long as (8) is satisfied,
945          * pos_ratio is able to drive itself to 1.0, which is not only where
946          * the dirty count meet the setpoint, but also where the slope of
947          * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
948          */
949         balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
950                                            dirty_rate | 1);
951         /*
952          * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
953          */
954         if (unlikely(balanced_dirty_ratelimit > write_bw))
955                 balanced_dirty_ratelimit = write_bw;
956
957         /*
958          * We could safely do this and return immediately:
959          *
960          *      bdi->dirty_ratelimit = balanced_dirty_ratelimit;
961          *
962          * However to get a more stable dirty_ratelimit, the below elaborated
963          * code makes use of task_ratelimit to filter out singular points and
964          * limit the step size.
965          *
966          * The below code essentially only uses the relative value of
967          *
968          *      task_ratelimit - dirty_ratelimit
969          *      = (pos_ratio - 1) * dirty_ratelimit
970          *
971          * which reflects the direction and size of dirty position error.
972          */
973
974         /*
975          * dirty_ratelimit will follow balanced_dirty_ratelimit iff
976          * task_ratelimit is on the same side of dirty_ratelimit, too.
977          * For example, when
978          * - dirty_ratelimit > balanced_dirty_ratelimit
979          * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
980          * lowering dirty_ratelimit will help meet both the position and rate
981          * control targets. Otherwise, don't update dirty_ratelimit if it will
982          * only help meet the rate target. After all, what the users ultimately
983          * feel and care are stable dirty rate and small position error.
984          *
985          * |task_ratelimit - dirty_ratelimit| is used to limit the step size
986          * and filter out the singular points of balanced_dirty_ratelimit. Which
987          * keeps jumping around randomly and can even leap far away at times
988          * due to the small 200ms estimation period of dirty_rate (we want to
989          * keep that period small to reduce time lags).
990          */
991         step = 0;
992         if (dirty < setpoint) {
993                 x = min(bdi->balanced_dirty_ratelimit,
994                          min(balanced_dirty_ratelimit, task_ratelimit));
995                 if (dirty_ratelimit < x)
996                         step = x - dirty_ratelimit;
997         } else {
998                 x = max(bdi->balanced_dirty_ratelimit,
999                          max(balanced_dirty_ratelimit, task_ratelimit));
1000                 if (dirty_ratelimit > x)
1001                         step = dirty_ratelimit - x;
1002         }
1003
1004         /*
1005          * Don't pursue 100% rate matching. It's impossible since the balanced
1006          * rate itself is constantly fluctuating. So decrease the track speed
1007          * when it gets close to the target. Helps eliminate pointless tremors.
1008          */
1009         step >>= dirty_ratelimit / (2 * step + 1);
1010         /*
1011          * Limit the tracking speed to avoid overshooting.
1012          */
1013         step = (step + 7) / 8;
1014
1015         if (dirty_ratelimit < balanced_dirty_ratelimit)
1016                 dirty_ratelimit += step;
1017         else
1018                 dirty_ratelimit -= step;
1019
1020         bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1021         bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1022
1023         trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
1024 }
1025
1026 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
1027                             unsigned long thresh,
1028                             unsigned long bg_thresh,
1029                             unsigned long dirty,
1030                             unsigned long bdi_thresh,
1031                             unsigned long bdi_dirty,
1032                             unsigned long start_time)
1033 {
1034         unsigned long now = jiffies;
1035         unsigned long elapsed = now - bdi->bw_time_stamp;
1036         unsigned long dirtied;
1037         unsigned long written;
1038
1039         /*
1040          * rate-limit, only update once every 200ms.
1041          */
1042         if (elapsed < BANDWIDTH_INTERVAL)
1043                 return;
1044
1045         dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
1046         written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
1047
1048         /*
1049          * Skip quiet periods when disk bandwidth is under-utilized.
1050          * (at least 1s idle time between two flusher runs)
1051          */
1052         if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
1053                 goto snapshot;
1054
1055         if (thresh) {
1056                 global_update_bandwidth(thresh, dirty, now);
1057                 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1058                                            bdi_thresh, bdi_dirty,
1059                                            dirtied, elapsed);
1060         }
1061         bdi_update_write_bandwidth(bdi, elapsed, written);
1062
1063 snapshot:
1064         bdi->dirtied_stamp = dirtied;
1065         bdi->written_stamp = written;
1066         bdi->bw_time_stamp = now;
1067 }
1068
1069 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
1070                                  unsigned long thresh,
1071                                  unsigned long bg_thresh,
1072                                  unsigned long dirty,
1073                                  unsigned long bdi_thresh,
1074                                  unsigned long bdi_dirty,
1075                                  unsigned long start_time)
1076 {
1077         if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1078                 return;
1079         spin_lock(&bdi->wb.list_lock);
1080         __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1081                                bdi_thresh, bdi_dirty, start_time);
1082         spin_unlock(&bdi->wb.list_lock);
1083 }
1084
1085 /*
1086  * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1087  * will look to see if it needs to start dirty throttling.
1088  *
1089  * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1090  * global_page_state() too often. So scale it near-sqrt to the safety margin
1091  * (the number of pages we may dirty without exceeding the dirty limits).
1092  */
1093 static unsigned long dirty_poll_interval(unsigned long dirty,
1094                                          unsigned long thresh)
1095 {
1096         if (thresh > dirty)
1097                 return 1UL << (ilog2(thresh - dirty) >> 1);
1098
1099         return 1;
1100 }
1101
1102 static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
1103                                    unsigned long bdi_dirty)
1104 {
1105         unsigned long bw = bdi->avg_write_bandwidth;
1106         unsigned long t;
1107
1108         /*
1109          * Limit pause time for small memory systems. If sleeping for too long
1110          * time, a small pool of dirty/writeback pages may go empty and disk go
1111          * idle.
1112          *
1113          * 8 serves as the safety ratio.
1114          */
1115         t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1116         t++;
1117
1118         return min_t(unsigned long, t, MAX_PAUSE);
1119 }
1120
1121 static long bdi_min_pause(struct backing_dev_info *bdi,
1122                           long max_pause,
1123                           unsigned long task_ratelimit,
1124                           unsigned long dirty_ratelimit,
1125                           int *nr_dirtied_pause)
1126 {
1127         long hi = ilog2(bdi->avg_write_bandwidth);
1128         long lo = ilog2(bdi->dirty_ratelimit);
1129         long t;         /* target pause */
1130         long pause;     /* estimated next pause */
1131         int pages;      /* target nr_dirtied_pause */
1132
1133         /* target for 10ms pause on 1-dd case */
1134         t = max(1, HZ / 100);
1135
1136         /*
1137          * Scale up pause time for concurrent dirtiers in order to reduce CPU
1138          * overheads.
1139          *
1140          * (N * 10ms) on 2^N concurrent tasks.
1141          */
1142         if (hi > lo)
1143                 t += (hi - lo) * (10 * HZ) / 1024;
1144
1145         /*
1146          * This is a bit convoluted. We try to base the next nr_dirtied_pause
1147          * on the much more stable dirty_ratelimit. However the next pause time
1148          * will be computed based on task_ratelimit and the two rate limits may
1149          * depart considerably at some time. Especially if task_ratelimit goes
1150          * below dirty_ratelimit/2 and the target pause is max_pause, the next
1151          * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1152          * result task_ratelimit won't be executed faithfully, which could
1153          * eventually bring down dirty_ratelimit.
1154          *
1155          * We apply two rules to fix it up:
1156          * 1) try to estimate the next pause time and if necessary, use a lower
1157          *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1158          *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1159          * 2) limit the target pause time to max_pause/2, so that the normal
1160          *    small fluctuations of task_ratelimit won't trigger rule (1) and
1161          *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1162          */
1163         t = min(t, 1 + max_pause / 2);
1164         pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1165
1166         /*
1167          * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1168          * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1169          * When the 16 consecutive reads are often interrupted by some dirty
1170          * throttling pause during the async writes, cfq will go into idles
1171          * (deadline is fine). So push nr_dirtied_pause as high as possible
1172          * until reaches DIRTY_POLL_THRESH=32 pages.
1173          */
1174         if (pages < DIRTY_POLL_THRESH) {
1175                 t = max_pause;
1176                 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1177                 if (pages > DIRTY_POLL_THRESH) {
1178                         pages = DIRTY_POLL_THRESH;
1179                         t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1180                 }
1181         }
1182
1183         pause = HZ * pages / (task_ratelimit + 1);
1184         if (pause > max_pause) {
1185                 t = max_pause;
1186                 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1187         }
1188
1189         *nr_dirtied_pause = pages;
1190         /*
1191          * The minimal pause time will normally be half the target pause time.
1192          */
1193         return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1194 }
1195
1196 /*
1197  * balance_dirty_pages() must be called by processes which are generating dirty
1198  * data.  It looks at the number of dirty pages in the machine and will force
1199  * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1200  * If we're over `background_thresh' then the writeback threads are woken to
1201  * perform some writeout.
1202  */
1203 static void balance_dirty_pages(struct address_space *mapping,
1204                                 unsigned long pages_dirtied)
1205 {
1206         unsigned long nr_reclaimable;   /* = file_dirty + unstable_nfs */
1207         unsigned long bdi_reclaimable;
1208         unsigned long nr_dirty;  /* = file_dirty + writeback + unstable_nfs */
1209         unsigned long bdi_dirty;
1210         unsigned long freerun;
1211         unsigned long background_thresh;
1212         unsigned long dirty_thresh;
1213         unsigned long bdi_thresh;
1214         long period;
1215         long pause;
1216         long max_pause;
1217         long min_pause;
1218         int nr_dirtied_pause;
1219         bool dirty_exceeded = false;
1220         unsigned long task_ratelimit;
1221         unsigned long dirty_ratelimit;
1222         unsigned long pos_ratio;
1223         struct backing_dev_info *bdi = mapping->backing_dev_info;
1224         unsigned long start_time = jiffies;
1225
1226         for (;;) {
1227                 unsigned long now = jiffies;
1228
1229                 /*
1230                  * Unstable writes are a feature of certain networked
1231                  * filesystems (i.e. NFS) in which data may have been
1232                  * written to the server's write cache, but has not yet
1233                  * been flushed to permanent storage.
1234                  */
1235                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1236                                         global_page_state(NR_UNSTABLE_NFS);
1237                 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1238
1239                 global_dirty_limits(&background_thresh, &dirty_thresh);
1240
1241                 /*
1242                  * Throttle it only when the background writeback cannot
1243                  * catch-up. This avoids (excessively) small writeouts
1244                  * when the bdi limits are ramping up.
1245                  */
1246                 freerun = dirty_freerun_ceiling(dirty_thresh,
1247                                                 background_thresh);
1248                 if (nr_dirty <= freerun) {
1249                         current->dirty_paused_when = now;
1250                         current->nr_dirtied = 0;
1251                         current->nr_dirtied_pause =
1252                                 dirty_poll_interval(nr_dirty, dirty_thresh);
1253                         break;
1254                 }
1255
1256                 if (unlikely(!writeback_in_progress(bdi)))
1257                         bdi_start_background_writeback(bdi);
1258
1259                 /*
1260                  * bdi_thresh is not treated as some limiting factor as
1261                  * dirty_thresh, due to reasons
1262                  * - in JBOD setup, bdi_thresh can fluctuate a lot
1263                  * - in a system with HDD and USB key, the USB key may somehow
1264                  *   go into state (bdi_dirty >> bdi_thresh) either because
1265                  *   bdi_dirty starts high, or because bdi_thresh drops low.
1266                  *   In this case we don't want to hard throttle the USB key
1267                  *   dirtiers for 100 seconds until bdi_dirty drops under
1268                  *   bdi_thresh. Instead the auxiliary bdi control line in
1269                  *   bdi_position_ratio() will let the dirtier task progress
1270                  *   at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1271                  */
1272                 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1273
1274                 /*
1275                  * In order to avoid the stacked BDI deadlock we need
1276                  * to ensure we accurately count the 'dirty' pages when
1277                  * the threshold is low.
1278                  *
1279                  * Otherwise it would be possible to get thresh+n pages
1280                  * reported dirty, even though there are thresh-m pages
1281                  * actually dirty; with m+n sitting in the percpu
1282                  * deltas.
1283                  */
1284                 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1285                         bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1286                         bdi_dirty = bdi_reclaimable +
1287                                     bdi_stat_sum(bdi, BDI_WRITEBACK);
1288                 } else {
1289                         bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1290                         bdi_dirty = bdi_reclaimable +
1291                                     bdi_stat(bdi, BDI_WRITEBACK);
1292                 }
1293
1294                 dirty_exceeded = (bdi_dirty > bdi_thresh) &&
1295                                   (nr_dirty > dirty_thresh);
1296                 if (dirty_exceeded && !bdi->dirty_exceeded)
1297                         bdi->dirty_exceeded = 1;
1298
1299                 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1300                                      nr_dirty, bdi_thresh, bdi_dirty,
1301                                      start_time);
1302
1303                 dirty_ratelimit = bdi->dirty_ratelimit;
1304                 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1305                                                background_thresh, nr_dirty,
1306                                                bdi_thresh, bdi_dirty);
1307                 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1308                                                         RATELIMIT_CALC_SHIFT;
1309                 max_pause = bdi_max_pause(bdi, bdi_dirty);
1310                 min_pause = bdi_min_pause(bdi, max_pause,
1311                                           task_ratelimit, dirty_ratelimit,
1312                                           &nr_dirtied_pause);
1313
1314                 if (unlikely(task_ratelimit == 0)) {
1315                         period = max_pause;
1316                         pause = max_pause;
1317                         goto pause;
1318                 }
1319                 period = HZ * pages_dirtied / task_ratelimit;
1320                 pause = period;
1321                 if (current->dirty_paused_when)
1322                         pause -= now - current->dirty_paused_when;
1323                 /*
1324                  * For less than 1s think time (ext3/4 may block the dirtier
1325                  * for up to 800ms from time to time on 1-HDD; so does xfs,
1326                  * however at much less frequency), try to compensate it in
1327                  * future periods by updating the virtual time; otherwise just
1328                  * do a reset, as it may be a light dirtier.
1329                  */
1330                 if (pause < min_pause) {
1331                         trace_balance_dirty_pages(bdi,
1332                                                   dirty_thresh,
1333                                                   background_thresh,
1334                                                   nr_dirty,
1335                                                   bdi_thresh,
1336                                                   bdi_dirty,
1337                                                   dirty_ratelimit,
1338                                                   task_ratelimit,
1339                                                   pages_dirtied,
1340                                                   period,
1341                                                   min(pause, 0L),
1342                                                   start_time);
1343                         if (pause < -HZ) {
1344                                 current->dirty_paused_when = now;
1345                                 current->nr_dirtied = 0;
1346                         } else if (period) {
1347                                 current->dirty_paused_when += period;
1348                                 current->nr_dirtied = 0;
1349                         } else if (current->nr_dirtied_pause <= pages_dirtied)
1350                                 current->nr_dirtied_pause += pages_dirtied;
1351                         break;
1352                 }
1353                 if (unlikely(pause > max_pause)) {
1354                         /* for occasional dropped task_ratelimit */
1355                         now += min(pause - max_pause, max_pause);
1356                         pause = max_pause;
1357                 }
1358
1359 pause:
1360                 trace_balance_dirty_pages(bdi,
1361                                           dirty_thresh,
1362                                           background_thresh,
1363                                           nr_dirty,
1364                                           bdi_thresh,
1365                                           bdi_dirty,
1366                                           dirty_ratelimit,
1367                                           task_ratelimit,
1368                                           pages_dirtied,
1369                                           period,
1370                                           pause,
1371                                           start_time);
1372                 __set_current_state(TASK_KILLABLE);
1373                 io_schedule_timeout(pause);
1374
1375                 current->dirty_paused_when = now + pause;
1376                 current->nr_dirtied = 0;
1377                 current->nr_dirtied_pause = nr_dirtied_pause;
1378
1379                 /*
1380                  * This is typically equal to (nr_dirty < dirty_thresh) and can
1381                  * also keep "1000+ dd on a slow USB stick" under control.
1382                  */
1383                 if (task_ratelimit)
1384                         break;
1385
1386                 /*
1387                  * In the case of an unresponding NFS server and the NFS dirty
1388                  * pages exceeds dirty_thresh, give the other good bdi's a pipe
1389                  * to go through, so that tasks on them still remain responsive.
1390                  *
1391                  * In theory 1 page is enough to keep the comsumer-producer
1392                  * pipe going: the flusher cleans 1 page => the task dirties 1
1393                  * more page. However bdi_dirty has accounting errors.  So use
1394                  * the larger and more IO friendly bdi_stat_error.
1395                  */
1396                 if (bdi_dirty <= bdi_stat_error(bdi))
1397                         break;
1398
1399                 if (fatal_signal_pending(current))
1400                         break;
1401         }
1402
1403         if (!dirty_exceeded && bdi->dirty_exceeded)
1404                 bdi->dirty_exceeded = 0;
1405
1406         if (writeback_in_progress(bdi))
1407                 return;
1408
1409         /*
1410          * In laptop mode, we wait until hitting the higher threshold before
1411          * starting background writeout, and then write out all the way down
1412          * to the lower threshold.  So slow writers cause minimal disk activity.
1413          *
1414          * In normal mode, we start background writeout at the lower
1415          * background_thresh, to keep the amount of dirty memory low.
1416          */
1417         if (laptop_mode)
1418                 return;
1419
1420         if (nr_reclaimable > background_thresh)
1421                 bdi_start_background_writeback(bdi);
1422 }
1423
1424 void set_page_dirty_balance(struct page *page, int page_mkwrite)
1425 {
1426         if (set_page_dirty(page) || page_mkwrite) {
1427                 struct address_space *mapping = page_mapping(page);
1428
1429                 if (mapping)
1430                         balance_dirty_pages_ratelimited(mapping);
1431         }
1432 }
1433
1434 static DEFINE_PER_CPU(int, bdp_ratelimits);
1435
1436 /*
1437  * Normal tasks are throttled by
1438  *      loop {
1439  *              dirty tsk->nr_dirtied_pause pages;
1440  *              take a snap in balance_dirty_pages();
1441  *      }
1442  * However there is a worst case. If every task exit immediately when dirtied
1443  * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1444  * called to throttle the page dirties. The solution is to save the not yet
1445  * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1446  * randomly into the running tasks. This works well for the above worst case,
1447  * as the new task will pick up and accumulate the old task's leaked dirty
1448  * count and eventually get throttled.
1449  */
1450 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1451
1452 /**
1453  * balance_dirty_pages_ratelimited - balance dirty memory state
1454  * @mapping: address_space which was dirtied
1455  *
1456  * Processes which are dirtying memory should call in here once for each page
1457  * which was newly dirtied.  The function will periodically check the system's
1458  * dirty state and will initiate writeback if needed.
1459  *
1460  * On really big machines, get_writeback_state is expensive, so try to avoid
1461  * calling it too often (ratelimiting).  But once we're over the dirty memory
1462  * limit we decrease the ratelimiting by a lot, to prevent individual processes
1463  * from overshooting the limit by (ratelimit_pages) each.
1464  */
1465 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1466 {
1467         struct backing_dev_info *bdi = mapping->backing_dev_info;
1468         int ratelimit;
1469         int *p;
1470
1471         if (!bdi_cap_account_dirty(bdi))
1472                 return;
1473
1474         ratelimit = current->nr_dirtied_pause;
1475         if (bdi->dirty_exceeded)
1476                 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1477
1478         preempt_disable();
1479         /*
1480          * This prevents one CPU to accumulate too many dirtied pages without
1481          * calling into balance_dirty_pages(), which can happen when there are
1482          * 1000+ tasks, all of them start dirtying pages at exactly the same
1483          * time, hence all honoured too large initial task->nr_dirtied_pause.
1484          */
1485         p =  &__get_cpu_var(bdp_ratelimits);
1486         if (unlikely(current->nr_dirtied >= ratelimit))
1487                 *p = 0;
1488         else if (unlikely(*p >= ratelimit_pages)) {
1489                 *p = 0;
1490                 ratelimit = 0;
1491         }
1492         /*
1493          * Pick up the dirtied pages by the exited tasks. This avoids lots of
1494          * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1495          * the dirty throttling and livelock other long-run dirtiers.
1496          */
1497         p = &__get_cpu_var(dirty_throttle_leaks);
1498         if (*p > 0 && current->nr_dirtied < ratelimit) {
1499                 unsigned long nr_pages_dirtied;
1500                 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1501                 *p -= nr_pages_dirtied;
1502                 current->nr_dirtied += nr_pages_dirtied;
1503         }
1504         preempt_enable();
1505
1506         if (unlikely(current->nr_dirtied >= ratelimit))
1507                 balance_dirty_pages(mapping, current->nr_dirtied);
1508 }
1509 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1510
1511 void throttle_vm_writeout(gfp_t gfp_mask)
1512 {
1513         unsigned long background_thresh;
1514         unsigned long dirty_thresh;
1515
1516         for ( ; ; ) {
1517                 global_dirty_limits(&background_thresh, &dirty_thresh);
1518                 dirty_thresh = hard_dirty_limit(dirty_thresh);
1519
1520                 /*
1521                  * Boost the allowable dirty threshold a bit for page
1522                  * allocators so they don't get DoS'ed by heavy writers
1523                  */
1524                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
1525
1526                 if (global_page_state(NR_UNSTABLE_NFS) +
1527                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
1528                                 break;
1529                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1530
1531                 /*
1532                  * The caller might hold locks which can prevent IO completion
1533                  * or progress in the filesystem.  So we cannot just sit here
1534                  * waiting for IO to complete.
1535                  */
1536                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1537                         break;
1538         }
1539 }
1540
1541 /*
1542  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1543  */
1544 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1545         void __user *buffer, size_t *length, loff_t *ppos)
1546 {
1547         proc_dointvec(table, write, buffer, length, ppos);
1548         return 0;
1549 }
1550
1551 #ifdef CONFIG_BLOCK
1552 void laptop_mode_timer_fn(unsigned long data)
1553 {
1554         struct request_queue *q = (struct request_queue *)data;
1555         int nr_pages = global_page_state(NR_FILE_DIRTY) +
1556                 global_page_state(NR_UNSTABLE_NFS);
1557
1558         /*
1559          * We want to write everything out, not just down to the dirty
1560          * threshold
1561          */
1562         if (bdi_has_dirty_io(&q->backing_dev_info))
1563                 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1564                                         WB_REASON_LAPTOP_TIMER);
1565 }
1566
1567 /*
1568  * We've spun up the disk and we're in laptop mode: schedule writeback
1569  * of all dirty data a few seconds from now.  If the flush is already scheduled
1570  * then push it back - the user is still using the disk.
1571  */
1572 void laptop_io_completion(struct backing_dev_info *info)
1573 {
1574         mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1575 }
1576
1577 /*
1578  * We're in laptop mode and we've just synced. The sync's writes will have
1579  * caused another writeback to be scheduled by laptop_io_completion.
1580  * Nothing needs to be written back anymore, so we unschedule the writeback.
1581  */
1582 void laptop_sync_completion(void)
1583 {
1584         struct backing_dev_info *bdi;
1585
1586         rcu_read_lock();
1587
1588         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1589                 del_timer(&bdi->laptop_mode_wb_timer);
1590
1591         rcu_read_unlock();
1592 }
1593 #endif
1594
1595 /*
1596  * If ratelimit_pages is too high then we can get into dirty-data overload
1597  * if a large number of processes all perform writes at the same time.
1598  * If it is too low then SMP machines will call the (expensive)
1599  * get_writeback_state too often.
1600  *
1601  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1602  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1603  * thresholds.
1604  */
1605
1606 void writeback_set_ratelimit(void)
1607 {
1608         unsigned long background_thresh;
1609         unsigned long dirty_thresh;
1610         global_dirty_limits(&background_thresh, &dirty_thresh);
1611         global_dirty_limit = dirty_thresh;
1612         ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1613         if (ratelimit_pages < 16)
1614                 ratelimit_pages = 16;
1615 }
1616
1617 static int __cpuinit
1618 ratelimit_handler(struct notifier_block *self, unsigned long action,
1619                   void *hcpu)
1620 {
1621
1622         switch (action & ~CPU_TASKS_FROZEN) {
1623         case CPU_ONLINE:
1624         case CPU_DEAD:
1625                 writeback_set_ratelimit();
1626                 return NOTIFY_OK;
1627         default:
1628                 return NOTIFY_DONE;
1629         }
1630 }
1631
1632 static struct notifier_block __cpuinitdata ratelimit_nb = {
1633         .notifier_call  = ratelimit_handler,
1634         .next           = NULL,
1635 };
1636
1637 /*
1638  * Called early on to tune the page writeback dirty limits.
1639  *
1640  * We used to scale dirty pages according to how total memory
1641  * related to pages that could be allocated for buffers (by
1642  * comparing nr_free_buffer_pages() to vm_total_pages.
1643  *
1644  * However, that was when we used "dirty_ratio" to scale with
1645  * all memory, and we don't do that any more. "dirty_ratio"
1646  * is now applied to total non-HIGHPAGE memory (by subtracting
1647  * totalhigh_pages from vm_total_pages), and as such we can't
1648  * get into the old insane situation any more where we had
1649  * large amounts of dirty pages compared to a small amount of
1650  * non-HIGHMEM memory.
1651  *
1652  * But we might still want to scale the dirty_ratio by how
1653  * much memory the box has..
1654  */
1655 void __init page_writeback_init(void)
1656 {
1657         writeback_set_ratelimit();
1658         register_cpu_notifier(&ratelimit_nb);
1659
1660         fprop_global_init(&writeout_completions);
1661 }
1662
1663 /**
1664  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1665  * @mapping: address space structure to write
1666  * @start: starting page index
1667  * @end: ending page index (inclusive)
1668  *
1669  * This function scans the page range from @start to @end (inclusive) and tags
1670  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1671  * that write_cache_pages (or whoever calls this function) will then use
1672  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
1673  * used to avoid livelocking of writeback by a process steadily creating new
1674  * dirty pages in the file (thus it is important for this function to be quick
1675  * so that it can tag pages faster than a dirtying process can create them).
1676  */
1677 /*
1678  * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1679  */
1680 void tag_pages_for_writeback(struct address_space *mapping,
1681                              pgoff_t start, pgoff_t end)
1682 {
1683 #define WRITEBACK_TAG_BATCH 4096
1684         unsigned long tagged;
1685
1686         do {
1687                 spin_lock_irq(&mapping->tree_lock);
1688                 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1689                                 &start, end, WRITEBACK_TAG_BATCH,
1690                                 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1691                 spin_unlock_irq(&mapping->tree_lock);
1692                 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1693                 cond_resched();
1694                 /* We check 'start' to handle wrapping when end == ~0UL */
1695         } while (tagged >= WRITEBACK_TAG_BATCH && start);
1696 }
1697 EXPORT_SYMBOL(tag_pages_for_writeback);
1698
1699 /**
1700  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1701  * @mapping: address space structure to write
1702  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1703  * @writepage: function called for each page
1704  * @data: data passed to writepage function
1705  *
1706  * If a page is already under I/O, write_cache_pages() skips it, even
1707  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
1708  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
1709  * and msync() need to guarantee that all the data which was dirty at the time
1710  * the call was made get new I/O started against them.  If wbc->sync_mode is
1711  * WB_SYNC_ALL then we were called for data integrity and we must wait for
1712  * existing IO to complete.
1713  *
1714  * To avoid livelocks (when other process dirties new pages), we first tag
1715  * pages which should be written back with TOWRITE tag and only then start
1716  * writing them. For data-integrity sync we have to be careful so that we do
1717  * not miss some pages (e.g., because some other process has cleared TOWRITE
1718  * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1719  * by the process clearing the DIRTY tag (and submitting the page for IO).
1720  */
1721 int write_cache_pages(struct address_space *mapping,
1722                       struct writeback_control *wbc, writepage_t writepage,
1723                       void *data)
1724 {
1725         int ret = 0;
1726         int done = 0;
1727         struct pagevec pvec;
1728         int nr_pages;
1729         pgoff_t uninitialized_var(writeback_index);
1730         pgoff_t index;
1731         pgoff_t end;            /* Inclusive */
1732         pgoff_t done_index;
1733         int cycled;
1734         int range_whole = 0;
1735         int tag;
1736
1737         pagevec_init(&pvec, 0);
1738         if (wbc->range_cyclic) {
1739                 writeback_index = mapping->writeback_index; /* prev offset */
1740                 index = writeback_index;
1741                 if (index == 0)
1742                         cycled = 1;
1743                 else
1744                         cycled = 0;
1745                 end = -1;
1746         } else {
1747                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1748                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1749                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1750                         range_whole = 1;
1751                 cycled = 1; /* ignore range_cyclic tests */
1752         }
1753         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1754                 tag = PAGECACHE_TAG_TOWRITE;
1755         else
1756                 tag = PAGECACHE_TAG_DIRTY;
1757 retry:
1758         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1759                 tag_pages_for_writeback(mapping, index, end);
1760         done_index = index;
1761         while (!done && (index <= end)) {
1762                 int i;
1763
1764                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1765                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1766                 if (nr_pages == 0)
1767                         break;
1768
1769                 for (i = 0; i < nr_pages; i++) {
1770                         struct page *page = pvec.pages[i];
1771
1772                         /*
1773                          * At this point, the page may be truncated or
1774                          * invalidated (changing page->mapping to NULL), or
1775                          * even swizzled back from swapper_space to tmpfs file
1776                          * mapping. However, page->index will not change
1777                          * because we have a reference on the page.
1778                          */
1779                         if (page->index > end) {
1780                                 /*
1781                                  * can't be range_cyclic (1st pass) because
1782                                  * end == -1 in that case.
1783                                  */
1784                                 done = 1;
1785                                 break;
1786                         }
1787
1788                         done_index = page->index;
1789
1790                         lock_page(page);
1791
1792                         /*
1793                          * Page truncated or invalidated. We can freely skip it
1794                          * then, even for data integrity operations: the page
1795                          * has disappeared concurrently, so there could be no
1796                          * real expectation of this data interity operation
1797                          * even if there is now a new, dirty page at the same
1798                          * pagecache address.
1799                          */
1800                         if (unlikely(page->mapping != mapping)) {
1801 continue_unlock:
1802                                 unlock_page(page);
1803                                 continue;
1804                         }
1805
1806                         if (!PageDirty(page)) {
1807                                 /* someone wrote it for us */
1808                                 goto continue_unlock;
1809                         }
1810
1811                         if (PageWriteback(page)) {
1812                                 if (wbc->sync_mode != WB_SYNC_NONE)
1813                                         wait_on_page_writeback(page);
1814                                 else
1815                                         goto continue_unlock;
1816                         }
1817
1818                         BUG_ON(PageWriteback(page));
1819                         if (!clear_page_dirty_for_io(page))
1820                                 goto continue_unlock;
1821
1822                         trace_wbc_writepage(wbc, mapping->backing_dev_info);
1823                         ret = (*writepage)(page, wbc, data);
1824                         if (unlikely(ret)) {
1825                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1826                                         unlock_page(page);
1827                                         ret = 0;
1828                                 } else {
1829                                         /*
1830                                          * done_index is set past this page,
1831                                          * so media errors will not choke
1832                                          * background writeout for the entire
1833                                          * file. This has consequences for
1834                                          * range_cyclic semantics (ie. it may
1835                                          * not be suitable for data integrity
1836                                          * writeout).
1837                                          */
1838                                         done_index = page->index + 1;
1839                                         done = 1;
1840                                         break;
1841                                 }
1842                         }
1843
1844                         /*
1845                          * We stop writing back only if we are not doing
1846                          * integrity sync. In case of integrity sync we have to
1847                          * keep going until we have written all the pages
1848                          * we tagged for writeback prior to entering this loop.
1849                          */
1850                         if (--wbc->nr_to_write <= 0 &&
1851                             wbc->sync_mode == WB_SYNC_NONE) {
1852                                 done = 1;
1853                                 break;
1854                         }
1855                 }
1856                 pagevec_release(&pvec);
1857                 cond_resched();
1858         }
1859         if (!cycled && !done) {
1860                 /*
1861                  * range_cyclic:
1862                  * We hit the last page and there is more work to be done: wrap
1863                  * back to the start of the file
1864                  */
1865                 cycled = 1;
1866                 index = 0;
1867                 end = writeback_index - 1;
1868                 goto retry;
1869         }
1870         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1871                 mapping->writeback_index = done_index;
1872
1873         return ret;
1874 }
1875 EXPORT_SYMBOL(write_cache_pages);
1876
1877 /*
1878  * Function used by generic_writepages to call the real writepage
1879  * function and set the mapping flags on error
1880  */
1881 static int __writepage(struct page *page, struct writeback_control *wbc,
1882                        void *data)
1883 {
1884         struct address_space *mapping = data;
1885         int ret = mapping->a_ops->writepage(page, wbc);
1886         mapping_set_error(mapping, ret);
1887         return ret;
1888 }
1889
1890 /**
1891  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1892  * @mapping: address space structure to write
1893  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1894  *
1895  * This is a library function, which implements the writepages()
1896  * address_space_operation.
1897  */
1898 int generic_writepages(struct address_space *mapping,
1899                        struct writeback_control *wbc)
1900 {
1901         struct blk_plug plug;
1902         int ret;
1903
1904         /* deal with chardevs and other special file */
1905         if (!mapping->a_ops->writepage)
1906                 return 0;
1907
1908         blk_start_plug(&plug);
1909         ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1910         blk_finish_plug(&plug);
1911         return ret;
1912 }
1913
1914 EXPORT_SYMBOL(generic_writepages);
1915
1916 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1917 {
1918         int ret;
1919
1920         if (wbc->nr_to_write <= 0)
1921                 return 0;
1922         if (mapping->a_ops->writepages)
1923                 ret = mapping->a_ops->writepages(mapping, wbc);
1924         else
1925                 ret = generic_writepages(mapping, wbc);
1926         return ret;
1927 }
1928
1929 /**
1930  * write_one_page - write out a single page and optionally wait on I/O
1931  * @page: the page to write
1932  * @wait: if true, wait on writeout
1933  *
1934  * The page must be locked by the caller and will be unlocked upon return.
1935  *
1936  * write_one_page() returns a negative error code if I/O failed.
1937  */
1938 int write_one_page(struct page *page, int wait)
1939 {
1940         struct address_space *mapping = page->mapping;
1941         int ret = 0;
1942         struct writeback_control wbc = {
1943                 .sync_mode = WB_SYNC_ALL,
1944                 .nr_to_write = 1,
1945         };
1946
1947         BUG_ON(!PageLocked(page));
1948
1949         if (wait)
1950                 wait_on_page_writeback(page);
1951
1952         if (clear_page_dirty_for_io(page)) {
1953                 page_cache_get(page);
1954                 ret = mapping->a_ops->writepage(page, &wbc);
1955                 if (ret == 0 && wait) {
1956                         wait_on_page_writeback(page);
1957                         if (PageError(page))
1958                                 ret = -EIO;
1959                 }
1960                 page_cache_release(page);
1961         } else {
1962                 unlock_page(page);
1963         }
1964         return ret;
1965 }
1966 EXPORT_SYMBOL(write_one_page);
1967
1968 /*
1969  * For address_spaces which do not use buffers nor write back.
1970  */
1971 int __set_page_dirty_no_writeback(struct page *page)
1972 {
1973         if (!PageDirty(page))
1974                 return !TestSetPageDirty(page);
1975         return 0;
1976 }
1977
1978 /*
1979  * Helper function for set_page_dirty family.
1980  * NOTE: This relies on being atomic wrt interrupts.
1981  */
1982 void account_page_dirtied(struct page *page, struct address_space *mapping)
1983 {
1984         trace_writeback_dirty_page(page, mapping);
1985
1986         if (mapping_cap_account_dirty(mapping)) {
1987                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1988                 __inc_zone_page_state(page, NR_DIRTIED);
1989                 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1990                 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1991                 task_io_account_write(PAGE_CACHE_SIZE);
1992                 current->nr_dirtied++;
1993                 this_cpu_inc(bdp_ratelimits);
1994         }
1995 }
1996 EXPORT_SYMBOL(account_page_dirtied);
1997
1998 /*
1999  * Helper function for set_page_writeback family.
2000  * NOTE: Unlike account_page_dirtied this does not rely on being atomic
2001  * wrt interrupts.
2002  */
2003 void account_page_writeback(struct page *page)
2004 {
2005         inc_zone_page_state(page, NR_WRITEBACK);
2006 }
2007 EXPORT_SYMBOL(account_page_writeback);
2008
2009 /*
2010  * For address_spaces which do not use buffers.  Just tag the page as dirty in
2011  * its radix tree.
2012  *
2013  * This is also used when a single buffer is being dirtied: we want to set the
2014  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
2015  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2016  *
2017  * Most callers have locked the page, which pins the address_space in memory.
2018  * But zap_pte_range() does not lock the page, however in that case the
2019  * mapping is pinned by the vma's ->vm_file reference.
2020  *
2021  * We take care to handle the case where the page was truncated from the
2022  * mapping by re-checking page_mapping() inside tree_lock.
2023  */
2024 int __set_page_dirty_nobuffers(struct page *page)
2025 {
2026         if (!TestSetPageDirty(page)) {
2027                 struct address_space *mapping = page_mapping(page);
2028                 struct address_space *mapping2;
2029                 unsigned long flags;
2030
2031                 if (!mapping)
2032                         return 1;
2033
2034                 spin_lock_irqsave(&mapping->tree_lock, flags);
2035                 mapping2 = page_mapping(page);
2036                 if (mapping2) { /* Race with truncate? */
2037                         BUG_ON(mapping2 != mapping);
2038                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2039                         account_page_dirtied(page, mapping);
2040                         radix_tree_tag_set(&mapping->page_tree,
2041                                 page_index(page), PAGECACHE_TAG_DIRTY);
2042                 }
2043                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2044                 if (mapping->host) {
2045                         /* !PageAnon && !swapper_space */
2046                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2047                 }
2048                 return 1;
2049         }
2050         return 0;
2051 }
2052 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2053
2054 /*
2055  * Call this whenever redirtying a page, to de-account the dirty counters
2056  * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2057  * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2058  * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2059  * control.
2060  */
2061 void account_page_redirty(struct page *page)
2062 {
2063         struct address_space *mapping = page->mapping;
2064         if (mapping && mapping_cap_account_dirty(mapping)) {
2065                 current->nr_dirtied--;
2066                 dec_zone_page_state(page, NR_DIRTIED);
2067                 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2068         }
2069 }
2070 EXPORT_SYMBOL(account_page_redirty);
2071
2072 /*
2073  * When a writepage implementation decides that it doesn't want to write this
2074  * page for some reason, it should redirty the locked page via
2075  * redirty_page_for_writepage() and it should then unlock the page and return 0
2076  */
2077 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2078 {
2079         wbc->pages_skipped++;
2080         account_page_redirty(page);
2081         return __set_page_dirty_nobuffers(page);
2082 }
2083 EXPORT_SYMBOL(redirty_page_for_writepage);
2084
2085 /*
2086  * Dirty a page.
2087  *
2088  * For pages with a mapping this should be done under the page lock
2089  * for the benefit of asynchronous memory errors who prefer a consistent
2090  * dirty state. This rule can be broken in some special cases,
2091  * but should be better not to.
2092  *
2093  * If the mapping doesn't provide a set_page_dirty a_op, then
2094  * just fall through and assume that it wants buffer_heads.
2095  */
2096 int set_page_dirty(struct page *page)
2097 {
2098         struct address_space *mapping = page_mapping(page);
2099
2100         if (likely(mapping)) {
2101                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2102                 /*
2103                  * readahead/lru_deactivate_page could remain
2104                  * PG_readahead/PG_reclaim due to race with end_page_writeback
2105                  * About readahead, if the page is written, the flags would be
2106                  * reset. So no problem.
2107                  * About lru_deactivate_page, if the page is redirty, the flag
2108                  * will be reset. So no problem. but if the page is used by readahead
2109                  * it will confuse readahead and make it restart the size rampup
2110                  * process. But it's a trivial problem.
2111                  */
2112                 ClearPageReclaim(page);
2113 #ifdef CONFIG_BLOCK
2114                 if (!spd)
2115                         spd = __set_page_dirty_buffers;
2116 #endif
2117                 return (*spd)(page);
2118         }
2119         if (!PageDirty(page)) {
2120                 if (!TestSetPageDirty(page))
2121                         return 1;
2122         }
2123         return 0;
2124 }
2125 EXPORT_SYMBOL(set_page_dirty);
2126
2127 /*
2128  * set_page_dirty() is racy if the caller has no reference against
2129  * page->mapping->host, and if the page is unlocked.  This is because another
2130  * CPU could truncate the page off the mapping and then free the mapping.
2131  *
2132  * Usually, the page _is_ locked, or the caller is a user-space process which
2133  * holds a reference on the inode by having an open file.
2134  *
2135  * In other cases, the page should be locked before running set_page_dirty().
2136  */
2137 int set_page_dirty_lock(struct page *page)
2138 {
2139         int ret;
2140
2141         lock_page(page);
2142         ret = set_page_dirty(page);
2143         unlock_page(page);
2144         return ret;
2145 }
2146 EXPORT_SYMBOL(set_page_dirty_lock);
2147
2148 /*
2149  * Clear a page's dirty flag, while caring for dirty memory accounting.
2150  * Returns true if the page was previously dirty.
2151  *
2152  * This is for preparing to put the page under writeout.  We leave the page
2153  * tagged as dirty in the radix tree so that a concurrent write-for-sync
2154  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
2155  * implementation will run either set_page_writeback() or set_page_dirty(),
2156  * at which stage we bring the page's dirty flag and radix-tree dirty tag
2157  * back into sync.
2158  *
2159  * This incoherency between the page's dirty flag and radix-tree tag is
2160  * unfortunate, but it only exists while the page is locked.
2161  */
2162 int clear_page_dirty_for_io(struct page *page)
2163 {
2164         struct address_space *mapping = page_mapping(page);
2165
2166         BUG_ON(!PageLocked(page));
2167
2168         if (mapping && mapping_cap_account_dirty(mapping)) {
2169                 /*
2170                  * Yes, Virginia, this is indeed insane.
2171                  *
2172                  * We use this sequence to make sure that
2173                  *  (a) we account for dirty stats properly
2174                  *  (b) we tell the low-level filesystem to
2175                  *      mark the whole page dirty if it was
2176                  *      dirty in a pagetable. Only to then
2177                  *  (c) clean the page again and return 1 to
2178                  *      cause the writeback.
2179                  *
2180                  * This way we avoid all nasty races with the
2181                  * dirty bit in multiple places and clearing
2182                  * them concurrently from different threads.
2183                  *
2184                  * Note! Normally the "set_page_dirty(page)"
2185                  * has no effect on the actual dirty bit - since
2186                  * that will already usually be set. But we
2187                  * need the side effects, and it can help us
2188                  * avoid races.
2189                  *
2190                  * We basically use the page "master dirty bit"
2191                  * as a serialization point for all the different
2192                  * threads doing their things.
2193                  */
2194                 if (page_mkclean(page))
2195                         set_page_dirty(page);
2196                 /*
2197                  * We carefully synchronise fault handlers against
2198                  * installing a dirty pte and marking the page dirty
2199                  * at this point. We do this by having them hold the
2200                  * page lock at some point after installing their
2201                  * pte, but before marking the page dirty.
2202                  * Pages are always locked coming in here, so we get
2203                  * the desired exclusion. See mm/memory.c:do_wp_page()
2204                  * for more comments.
2205                  */
2206                 if (TestClearPageDirty(page)) {
2207                         dec_zone_page_state(page, NR_FILE_DIRTY);
2208                         dec_bdi_stat(mapping->backing_dev_info,
2209                                         BDI_RECLAIMABLE);
2210                         return 1;
2211                 }
2212                 return 0;
2213         }
2214         return TestClearPageDirty(page);
2215 }
2216 EXPORT_SYMBOL(clear_page_dirty_for_io);
2217
2218 int test_clear_page_writeback(struct page *page)
2219 {
2220         struct address_space *mapping = page_mapping(page);
2221         int ret;
2222
2223         if (mapping) {
2224                 struct backing_dev_info *bdi = mapping->backing_dev_info;
2225                 unsigned long flags;
2226
2227                 spin_lock_irqsave(&mapping->tree_lock, flags);
2228                 ret = TestClearPageWriteback(page);
2229                 if (ret) {
2230                         radix_tree_tag_clear(&mapping->page_tree,
2231                                                 page_index(page),
2232                                                 PAGECACHE_TAG_WRITEBACK);
2233                         if (bdi_cap_account_writeback(bdi)) {
2234                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
2235                                 __bdi_writeout_inc(bdi);
2236                         }
2237                 }
2238                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2239         } else {
2240                 ret = TestClearPageWriteback(page);
2241         }
2242         if (ret) {
2243                 dec_zone_page_state(page, NR_WRITEBACK);
2244                 inc_zone_page_state(page, NR_WRITTEN);
2245         }
2246         return ret;
2247 }
2248
2249 int test_set_page_writeback(struct page *page)
2250 {
2251         struct address_space *mapping = page_mapping(page);
2252         int ret;
2253
2254         if (mapping) {
2255                 struct backing_dev_info *bdi = mapping->backing_dev_info;
2256                 unsigned long flags;
2257
2258                 spin_lock_irqsave(&mapping->tree_lock, flags);
2259                 ret = TestSetPageWriteback(page);
2260                 if (!ret) {
2261                         radix_tree_tag_set(&mapping->page_tree,
2262                                                 page_index(page),
2263                                                 PAGECACHE_TAG_WRITEBACK);
2264                         if (bdi_cap_account_writeback(bdi))
2265                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2266                 }
2267                 if (!PageDirty(page))
2268                         radix_tree_tag_clear(&mapping->page_tree,
2269                                                 page_index(page),
2270                                                 PAGECACHE_TAG_DIRTY);
2271                 radix_tree_tag_clear(&mapping->page_tree,
2272                                      page_index(page),
2273                                      PAGECACHE_TAG_TOWRITE);
2274                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2275         } else {
2276                 ret = TestSetPageWriteback(page);
2277         }
2278         if (!ret)
2279                 account_page_writeback(page);
2280         return ret;
2281
2282 }
2283 EXPORT_SYMBOL(test_set_page_writeback);
2284
2285 /*
2286  * Return true if any of the pages in the mapping are marked with the
2287  * passed tag.
2288  */
2289 int mapping_tagged(struct address_space *mapping, int tag)
2290 {
2291         return radix_tree_tagged(&mapping->page_tree, tag);
2292 }
2293 EXPORT_SYMBOL(mapping_tagged);
2294
2295 /**
2296  * wait_for_stable_page() - wait for writeback to finish, if necessary.
2297  * @page:       The page to wait on.
2298  *
2299  * This function determines if the given page is related to a backing device
2300  * that requires page contents to be held stable during writeback.  If so, then
2301  * it will wait for any pending writeback to complete.
2302  */
2303 void wait_for_stable_page(struct page *page)
2304 {
2305         struct address_space *mapping = page_mapping(page);
2306         struct backing_dev_info *bdi = mapping->backing_dev_info;
2307
2308         if (!bdi_cap_stable_pages_required(bdi))
2309                 return;
2310
2311         wait_on_page_writeback(page);
2312 }
2313 EXPORT_SYMBOL_GPL(wait_for_stable_page);