Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/signal
[platform/adaptation/renesas_rcar/renesas_kernel.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 static unsigned long highmem_dirtyable_memory(unsigned long total)
192 {
193 #ifdef CONFIG_HIGHMEM
194         int node;
195         unsigned long x = 0;
196
197         for_each_node_state(node, N_HIGH_MEMORY) {
198                 struct zone *z =
199                         &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
200
201                 x += zone_page_state(z, NR_FREE_PAGES) +
202                      zone_reclaimable_pages(z) - z->dirty_balance_reserve;
203         }
204         /*
205          * Unreclaimable memory (kernel memory or anonymous memory
206          * without swap) can bring down the dirtyable pages below
207          * the zone's dirty balance reserve and the above calculation
208          * will underflow.  However we still want to add in nodes
209          * which are below threshold (negative values) to get a more
210          * accurate calculation but make sure that the total never
211          * underflows.
212          */
213         if ((long)x < 0)
214                 x = 0;
215
216         /*
217          * Make sure that the number of highmem pages is never larger
218          * than the number of the total dirtyable memory. This can only
219          * occur in very strange VM situations but we want to make sure
220          * that this does not occur.
221          */
222         return min(x, total);
223 #else
224         return 0;
225 #endif
226 }
227
228 /**
229  * global_dirtyable_memory - number of globally dirtyable pages
230  *
231  * Returns the global number of pages potentially available for dirty
232  * page cache.  This is the base value for the global dirty limits.
233  */
234 static unsigned long global_dirtyable_memory(void)
235 {
236         unsigned long x;
237
238         x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
239         x -= min(x, dirty_balance_reserve);
240
241         if (!vm_highmem_is_dirtyable)
242                 x -= highmem_dirtyable_memory(x);
243
244         /* Subtract min_free_kbytes */
245         x -= min_t(unsigned long, x, min_free_kbytes >> (PAGE_SHIFT - 10));
246
247         return x + 1;   /* Ensure that we never return 0 */
248 }
249
250 /*
251  * global_dirty_limits - background-writeback and dirty-throttling thresholds
252  *
253  * Calculate the dirty thresholds based on sysctl parameters
254  * - vm.dirty_background_ratio  or  vm.dirty_background_bytes
255  * - vm.dirty_ratio             or  vm.dirty_bytes
256  * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
257  * real-time tasks.
258  */
259 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
260 {
261         unsigned long background;
262         unsigned long dirty;
263         unsigned long uninitialized_var(available_memory);
264         struct task_struct *tsk;
265
266         if (!vm_dirty_bytes || !dirty_background_bytes)
267                 available_memory = global_dirtyable_memory();
268
269         if (vm_dirty_bytes)
270                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
271         else
272                 dirty = (vm_dirty_ratio * available_memory) / 100;
273
274         if (dirty_background_bytes)
275                 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
276         else
277                 background = (dirty_background_ratio * available_memory) / 100;
278
279         if (background >= dirty)
280                 background = dirty / 2;
281         tsk = current;
282         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
283                 background += background / 4;
284                 dirty += dirty / 4;
285         }
286         *pbackground = background;
287         *pdirty = dirty;
288         trace_global_dirty_state(background, dirty);
289 }
290
291 /**
292  * zone_dirtyable_memory - number of dirtyable pages in a zone
293  * @zone: the zone
294  *
295  * Returns the zone's number of pages potentially available for dirty
296  * page cache.  This is the base value for the per-zone dirty limits.
297  */
298 static unsigned long zone_dirtyable_memory(struct zone *zone)
299 {
300         /*
301          * The effective global number of dirtyable pages may exclude
302          * highmem as a big-picture measure to keep the ratio between
303          * dirty memory and lowmem reasonable.
304          *
305          * But this function is purely about the individual zone and a
306          * highmem zone can hold its share of dirty pages, so we don't
307          * care about vm_highmem_is_dirtyable here.
308          */
309         unsigned long nr_pages = zone_page_state(zone, NR_FREE_PAGES) +
310                 zone_reclaimable_pages(zone);
311
312         /* don't allow this to underflow */
313         nr_pages -= min(nr_pages, zone->dirty_balance_reserve);
314         return nr_pages;
315 }
316
317 /**
318  * zone_dirty_limit - maximum number of dirty pages allowed in a zone
319  * @zone: the zone
320  *
321  * Returns the maximum number of dirty pages allowed in a zone, based
322  * on the zone's dirtyable memory.
323  */
324 static unsigned long zone_dirty_limit(struct zone *zone)
325 {
326         unsigned long zone_memory = zone_dirtyable_memory(zone);
327         struct task_struct *tsk = current;
328         unsigned long dirty;
329
330         if (vm_dirty_bytes)
331                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
332                         zone_memory / global_dirtyable_memory();
333         else
334                 dirty = vm_dirty_ratio * zone_memory / 100;
335
336         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
337                 dirty += dirty / 4;
338
339         return dirty;
340 }
341
342 /**
343  * zone_dirty_ok - tells whether a zone is within its dirty limits
344  * @zone: the zone to check
345  *
346  * Returns %true when the dirty pages in @zone are within the zone's
347  * dirty limit, %false if the limit is exceeded.
348  */
349 bool zone_dirty_ok(struct zone *zone)
350 {
351         unsigned long limit = zone_dirty_limit(zone);
352
353         return zone_page_state(zone, NR_FILE_DIRTY) +
354                zone_page_state(zone, NR_UNSTABLE_NFS) +
355                zone_page_state(zone, NR_WRITEBACK) <= limit;
356 }
357
358 int dirty_background_ratio_handler(struct ctl_table *table, int write,
359                 void __user *buffer, size_t *lenp,
360                 loff_t *ppos)
361 {
362         int ret;
363
364         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
365         if (ret == 0 && write)
366                 dirty_background_bytes = 0;
367         return ret;
368 }
369
370 int dirty_background_bytes_handler(struct ctl_table *table, int write,
371                 void __user *buffer, size_t *lenp,
372                 loff_t *ppos)
373 {
374         int ret;
375
376         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
377         if (ret == 0 && write)
378                 dirty_background_ratio = 0;
379         return ret;
380 }
381
382 int dirty_ratio_handler(struct ctl_table *table, int write,
383                 void __user *buffer, size_t *lenp,
384                 loff_t *ppos)
385 {
386         int old_ratio = vm_dirty_ratio;
387         int ret;
388
389         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
390         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
391                 writeback_set_ratelimit();
392                 vm_dirty_bytes = 0;
393         }
394         return ret;
395 }
396
397 int dirty_bytes_handler(struct ctl_table *table, int write,
398                 void __user *buffer, size_t *lenp,
399                 loff_t *ppos)
400 {
401         unsigned long old_bytes = vm_dirty_bytes;
402         int ret;
403
404         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
405         if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
406                 writeback_set_ratelimit();
407                 vm_dirty_ratio = 0;
408         }
409         return ret;
410 }
411
412 static unsigned long wp_next_time(unsigned long cur_time)
413 {
414         cur_time += VM_COMPLETIONS_PERIOD_LEN;
415         /* 0 has a special meaning... */
416         if (!cur_time)
417                 return 1;
418         return cur_time;
419 }
420
421 /*
422  * Increment the BDI's writeout completion count and the global writeout
423  * completion count. Called from test_clear_page_writeback().
424  */
425 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
426 {
427         __inc_bdi_stat(bdi, BDI_WRITTEN);
428         __fprop_inc_percpu_max(&writeout_completions, &bdi->completions,
429                                bdi->max_prop_frac);
430         /* First event after period switching was turned off? */
431         if (!unlikely(writeout_period_time)) {
432                 /*
433                  * We can race with other __bdi_writeout_inc calls here but
434                  * it does not cause any harm since the resulting time when
435                  * timer will fire and what is in writeout_period_time will be
436                  * roughly the same.
437                  */
438                 writeout_period_time = wp_next_time(jiffies);
439                 mod_timer(&writeout_period_timer, writeout_period_time);
440         }
441 }
442
443 void bdi_writeout_inc(struct backing_dev_info *bdi)
444 {
445         unsigned long flags;
446
447         local_irq_save(flags);
448         __bdi_writeout_inc(bdi);
449         local_irq_restore(flags);
450 }
451 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
452
453 /*
454  * Obtain an accurate fraction of the BDI's portion.
455  */
456 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
457                 long *numerator, long *denominator)
458 {
459         fprop_fraction_percpu(&writeout_completions, &bdi->completions,
460                                 numerator, denominator);
461 }
462
463 /*
464  * On idle system, we can be called long after we scheduled because we use
465  * deferred timers so count with missed periods.
466  */
467 static void writeout_period(unsigned long t)
468 {
469         int miss_periods = (jiffies - writeout_period_time) /
470                                                  VM_COMPLETIONS_PERIOD_LEN;
471
472         if (fprop_new_period(&writeout_completions, miss_periods + 1)) {
473                 writeout_period_time = wp_next_time(writeout_period_time +
474                                 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
475                 mod_timer(&writeout_period_timer, writeout_period_time);
476         } else {
477                 /*
478                  * Aging has zeroed all fractions. Stop wasting CPU on period
479                  * updates.
480                  */
481                 writeout_period_time = 0;
482         }
483 }
484
485 /*
486  * bdi_min_ratio keeps the sum of the minimum dirty shares of all
487  * registered backing devices, which, for obvious reasons, can not
488  * exceed 100%.
489  */
490 static unsigned int bdi_min_ratio;
491
492 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
493 {
494         int ret = 0;
495
496         spin_lock_bh(&bdi_lock);
497         if (min_ratio > bdi->max_ratio) {
498                 ret = -EINVAL;
499         } else {
500                 min_ratio -= bdi->min_ratio;
501                 if (bdi_min_ratio + min_ratio < 100) {
502                         bdi_min_ratio += min_ratio;
503                         bdi->min_ratio += min_ratio;
504                 } else {
505                         ret = -EINVAL;
506                 }
507         }
508         spin_unlock_bh(&bdi_lock);
509
510         return ret;
511 }
512
513 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
514 {
515         int ret = 0;
516
517         if (max_ratio > 100)
518                 return -EINVAL;
519
520         spin_lock_bh(&bdi_lock);
521         if (bdi->min_ratio > max_ratio) {
522                 ret = -EINVAL;
523         } else {
524                 bdi->max_ratio = max_ratio;
525                 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
526         }
527         spin_unlock_bh(&bdi_lock);
528
529         return ret;
530 }
531 EXPORT_SYMBOL(bdi_set_max_ratio);
532
533 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
534                                            unsigned long bg_thresh)
535 {
536         return (thresh + bg_thresh) / 2;
537 }
538
539 static unsigned long hard_dirty_limit(unsigned long thresh)
540 {
541         return max(thresh, global_dirty_limit);
542 }
543
544 /**
545  * bdi_dirty_limit - @bdi's share of dirty throttling threshold
546  * @bdi: the backing_dev_info to query
547  * @dirty: global dirty limit in pages
548  *
549  * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
550  * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
551  *
552  * Note that balance_dirty_pages() will only seriously take it as a hard limit
553  * when sleeping max_pause per page is not enough to keep the dirty pages under
554  * control. For example, when the device is completely stalled due to some error
555  * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
556  * In the other normal situations, it acts more gently by throttling the tasks
557  * more (rather than completely block them) when the bdi dirty pages go high.
558  *
559  * It allocates high/low dirty limits to fast/slow devices, in order to prevent
560  * - starving fast devices
561  * - piling up dirty pages (that will take long time to sync) on slow devices
562  *
563  * The bdi's share of dirty limit will be adapting to its throughput and
564  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
565  */
566 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
567 {
568         u64 bdi_dirty;
569         long numerator, denominator;
570
571         /*
572          * Calculate this BDI's share of the dirty ratio.
573          */
574         bdi_writeout_fraction(bdi, &numerator, &denominator);
575
576         bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
577         bdi_dirty *= numerator;
578         do_div(bdi_dirty, denominator);
579
580         bdi_dirty += (dirty * bdi->min_ratio) / 100;
581         if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
582                 bdi_dirty = dirty * bdi->max_ratio / 100;
583
584         return bdi_dirty;
585 }
586
587 /*
588  * Dirty position control.
589  *
590  * (o) global/bdi setpoints
591  *
592  * We want the dirty pages be balanced around the global/bdi setpoints.
593  * When the number of dirty pages is higher/lower than the setpoint, the
594  * dirty position control ratio (and hence task dirty ratelimit) will be
595  * decreased/increased to bring the dirty pages back to the setpoint.
596  *
597  *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
598  *
599  *     if (dirty < setpoint) scale up   pos_ratio
600  *     if (dirty > setpoint) scale down pos_ratio
601  *
602  *     if (bdi_dirty < bdi_setpoint) scale up   pos_ratio
603  *     if (bdi_dirty > bdi_setpoint) scale down pos_ratio
604  *
605  *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
606  *
607  * (o) global control line
608  *
609  *     ^ pos_ratio
610  *     |
611  *     |            |<===== global dirty control scope ======>|
612  * 2.0 .............*
613  *     |            .*
614  *     |            . *
615  *     |            .   *
616  *     |            .     *
617  *     |            .        *
618  *     |            .            *
619  * 1.0 ................................*
620  *     |            .                  .     *
621  *     |            .                  .          *
622  *     |            .                  .              *
623  *     |            .                  .                 *
624  *     |            .                  .                    *
625  *   0 +------------.------------------.----------------------*------------->
626  *           freerun^          setpoint^                 limit^   dirty pages
627  *
628  * (o) bdi control line
629  *
630  *     ^ pos_ratio
631  *     |
632  *     |            *
633  *     |              *
634  *     |                *
635  *     |                  *
636  *     |                    * |<=========== span ============>|
637  * 1.0 .......................*
638  *     |                      . *
639  *     |                      .   *
640  *     |                      .     *
641  *     |                      .       *
642  *     |                      .         *
643  *     |                      .           *
644  *     |                      .             *
645  *     |                      .               *
646  *     |                      .                 *
647  *     |                      .                   *
648  *     |                      .                     *
649  * 1/4 ...............................................* * * * * * * * * * * *
650  *     |                      .                         .
651  *     |                      .                           .
652  *     |                      .                             .
653  *   0 +----------------------.-------------------------------.------------->
654  *                bdi_setpoint^                    x_intercept^
655  *
656  * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
657  * be smoothly throttled down to normal if it starts high in situations like
658  * - start writing to a slow SD card and a fast disk at the same time. The SD
659  *   card's bdi_dirty may rush to many times higher than bdi_setpoint.
660  * - the bdi dirty thresh drops quickly due to change of JBOD workload
661  */
662 static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
663                                         unsigned long thresh,
664                                         unsigned long bg_thresh,
665                                         unsigned long dirty,
666                                         unsigned long bdi_thresh,
667                                         unsigned long bdi_dirty)
668 {
669         unsigned long write_bw = bdi->avg_write_bandwidth;
670         unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
671         unsigned long limit = hard_dirty_limit(thresh);
672         unsigned long x_intercept;
673         unsigned long setpoint;         /* dirty pages' target balance point */
674         unsigned long bdi_setpoint;
675         unsigned long span;
676         long long pos_ratio;            /* for scaling up/down the rate limit */
677         long x;
678
679         if (unlikely(dirty >= limit))
680                 return 0;
681
682         /*
683          * global setpoint
684          *
685          *                           setpoint - dirty 3
686          *        f(dirty) := 1.0 + (----------------)
687          *                           limit - setpoint
688          *
689          * it's a 3rd order polynomial that subjects to
690          *
691          * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
692          * (2) f(setpoint) = 1.0 => the balance point
693          * (3) f(limit)    = 0   => the hard limit
694          * (4) df/dx      <= 0   => negative feedback control
695          * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
696          *     => fast response on large errors; small oscillation near setpoint
697          */
698         setpoint = (freerun + limit) / 2;
699         x = div_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
700                     limit - setpoint + 1);
701         pos_ratio = x;
702         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
703         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
704         pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
705
706         /*
707          * We have computed basic pos_ratio above based on global situation. If
708          * the bdi is over/under its share of dirty pages, we want to scale
709          * pos_ratio further down/up. That is done by the following mechanism.
710          */
711
712         /*
713          * bdi setpoint
714          *
715          *        f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
716          *
717          *                        x_intercept - bdi_dirty
718          *                     := --------------------------
719          *                        x_intercept - bdi_setpoint
720          *
721          * The main bdi control line is a linear function that subjects to
722          *
723          * (1) f(bdi_setpoint) = 1.0
724          * (2) k = - 1 / (8 * write_bw)  (in single bdi case)
725          *     or equally: x_intercept = bdi_setpoint + 8 * write_bw
726          *
727          * For single bdi case, the dirty pages are observed to fluctuate
728          * regularly within range
729          *        [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
730          * for various filesystems, where (2) can yield in a reasonable 12.5%
731          * fluctuation range for pos_ratio.
732          *
733          * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
734          * own size, so move the slope over accordingly and choose a slope that
735          * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
736          */
737         if (unlikely(bdi_thresh > thresh))
738                 bdi_thresh = thresh;
739         /*
740          * It's very possible that bdi_thresh is close to 0 not because the
741          * device is slow, but that it has remained inactive for long time.
742          * Honour such devices a reasonable good (hopefully IO efficient)
743          * threshold, so that the occasional writes won't be blocked and active
744          * writes can rampup the threshold quickly.
745          */
746         bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
747         /*
748          * scale global setpoint to bdi's:
749          *      bdi_setpoint = setpoint * bdi_thresh / thresh
750          */
751         x = div_u64((u64)bdi_thresh << 16, thresh + 1);
752         bdi_setpoint = setpoint * (u64)x >> 16;
753         /*
754          * Use span=(8*write_bw) in single bdi case as indicated by
755          * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
756          *
757          *        bdi_thresh                    thresh - bdi_thresh
758          * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
759          *          thresh                            thresh
760          */
761         span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
762         x_intercept = bdi_setpoint + span;
763
764         if (bdi_dirty < x_intercept - span / 4) {
765                 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
766                                     x_intercept - bdi_setpoint + 1);
767         } else
768                 pos_ratio /= 4;
769
770         /*
771          * bdi reserve area, safeguard against dirty pool underrun and disk idle
772          * It may push the desired control point of global dirty pages higher
773          * than setpoint.
774          */
775         x_intercept = bdi_thresh / 2;
776         if (bdi_dirty < x_intercept) {
777                 if (bdi_dirty > x_intercept / 8)
778                         pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
779                 else
780                         pos_ratio *= 8;
781         }
782
783         return pos_ratio;
784 }
785
786 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
787                                        unsigned long elapsed,
788                                        unsigned long written)
789 {
790         const unsigned long period = roundup_pow_of_two(3 * HZ);
791         unsigned long avg = bdi->avg_write_bandwidth;
792         unsigned long old = bdi->write_bandwidth;
793         u64 bw;
794
795         /*
796          * bw = written * HZ / elapsed
797          *
798          *                   bw * elapsed + write_bandwidth * (period - elapsed)
799          * write_bandwidth = ---------------------------------------------------
800          *                                          period
801          */
802         bw = written - bdi->written_stamp;
803         bw *= HZ;
804         if (unlikely(elapsed > period)) {
805                 do_div(bw, elapsed);
806                 avg = bw;
807                 goto out;
808         }
809         bw += (u64)bdi->write_bandwidth * (period - elapsed);
810         bw >>= ilog2(period);
811
812         /*
813          * one more level of smoothing, for filtering out sudden spikes
814          */
815         if (avg > old && old >= (unsigned long)bw)
816                 avg -= (avg - old) >> 3;
817
818         if (avg < old && old <= (unsigned long)bw)
819                 avg += (old - avg) >> 3;
820
821 out:
822         bdi->write_bandwidth = bw;
823         bdi->avg_write_bandwidth = avg;
824 }
825
826 /*
827  * The global dirtyable memory and dirty threshold could be suddenly knocked
828  * down by a large amount (eg. on the startup of KVM in a swapless system).
829  * This may throw the system into deep dirty exceeded state and throttle
830  * heavy/light dirtiers alike. To retain good responsiveness, maintain
831  * global_dirty_limit for tracking slowly down to the knocked down dirty
832  * threshold.
833  */
834 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
835 {
836         unsigned long limit = global_dirty_limit;
837
838         /*
839          * Follow up in one step.
840          */
841         if (limit < thresh) {
842                 limit = thresh;
843                 goto update;
844         }
845
846         /*
847          * Follow down slowly. Use the higher one as the target, because thresh
848          * may drop below dirty. This is exactly the reason to introduce
849          * global_dirty_limit which is guaranteed to lie above the dirty pages.
850          */
851         thresh = max(thresh, dirty);
852         if (limit > thresh) {
853                 limit -= (limit - thresh) >> 5;
854                 goto update;
855         }
856         return;
857 update:
858         global_dirty_limit = limit;
859 }
860
861 static void global_update_bandwidth(unsigned long thresh,
862                                     unsigned long dirty,
863                                     unsigned long now)
864 {
865         static DEFINE_SPINLOCK(dirty_lock);
866         static unsigned long update_time;
867
868         /*
869          * check locklessly first to optimize away locking for the most time
870          */
871         if (time_before(now, update_time + BANDWIDTH_INTERVAL))
872                 return;
873
874         spin_lock(&dirty_lock);
875         if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
876                 update_dirty_limit(thresh, dirty);
877                 update_time = now;
878         }
879         spin_unlock(&dirty_lock);
880 }
881
882 /*
883  * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
884  *
885  * Normal bdi tasks will be curbed at or below it in long term.
886  * Obviously it should be around (write_bw / N) when there are N dd tasks.
887  */
888 static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
889                                        unsigned long thresh,
890                                        unsigned long bg_thresh,
891                                        unsigned long dirty,
892                                        unsigned long bdi_thresh,
893                                        unsigned long bdi_dirty,
894                                        unsigned long dirtied,
895                                        unsigned long elapsed)
896 {
897         unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
898         unsigned long limit = hard_dirty_limit(thresh);
899         unsigned long setpoint = (freerun + limit) / 2;
900         unsigned long write_bw = bdi->avg_write_bandwidth;
901         unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
902         unsigned long dirty_rate;
903         unsigned long task_ratelimit;
904         unsigned long balanced_dirty_ratelimit;
905         unsigned long pos_ratio;
906         unsigned long step;
907         unsigned long x;
908
909         /*
910          * The dirty rate will match the writeout rate in long term, except
911          * when dirty pages are truncated by userspace or re-dirtied by FS.
912          */
913         dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
914
915         pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
916                                        bdi_thresh, bdi_dirty);
917         /*
918          * task_ratelimit reflects each dd's dirty rate for the past 200ms.
919          */
920         task_ratelimit = (u64)dirty_ratelimit *
921                                         pos_ratio >> RATELIMIT_CALC_SHIFT;
922         task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
923
924         /*
925          * A linear estimation of the "balanced" throttle rate. The theory is,
926          * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
927          * dirty_rate will be measured to be (N * task_ratelimit). So the below
928          * formula will yield the balanced rate limit (write_bw / N).
929          *
930          * Note that the expanded form is not a pure rate feedback:
931          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate)              (1)
932          * but also takes pos_ratio into account:
933          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
934          *
935          * (1) is not realistic because pos_ratio also takes part in balancing
936          * the dirty rate.  Consider the state
937          *      pos_ratio = 0.5                                              (3)
938          *      rate = 2 * (write_bw / N)                                    (4)
939          * If (1) is used, it will stuck in that state! Because each dd will
940          * be throttled at
941          *      task_ratelimit = pos_ratio * rate = (write_bw / N)           (5)
942          * yielding
943          *      dirty_rate = N * task_ratelimit = write_bw                   (6)
944          * put (6) into (1) we get
945          *      rate_(i+1) = rate_(i)                                        (7)
946          *
947          * So we end up using (2) to always keep
948          *      rate_(i+1) ~= (write_bw / N)                                 (8)
949          * regardless of the value of pos_ratio. As long as (8) is satisfied,
950          * pos_ratio is able to drive itself to 1.0, which is not only where
951          * the dirty count meet the setpoint, but also where the slope of
952          * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
953          */
954         balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
955                                            dirty_rate | 1);
956         /*
957          * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
958          */
959         if (unlikely(balanced_dirty_ratelimit > write_bw))
960                 balanced_dirty_ratelimit = write_bw;
961
962         /*
963          * We could safely do this and return immediately:
964          *
965          *      bdi->dirty_ratelimit = balanced_dirty_ratelimit;
966          *
967          * However to get a more stable dirty_ratelimit, the below elaborated
968          * code makes use of task_ratelimit to filter out singular points and
969          * limit the step size.
970          *
971          * The below code essentially only uses the relative value of
972          *
973          *      task_ratelimit - dirty_ratelimit
974          *      = (pos_ratio - 1) * dirty_ratelimit
975          *
976          * which reflects the direction and size of dirty position error.
977          */
978
979         /*
980          * dirty_ratelimit will follow balanced_dirty_ratelimit iff
981          * task_ratelimit is on the same side of dirty_ratelimit, too.
982          * For example, when
983          * - dirty_ratelimit > balanced_dirty_ratelimit
984          * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
985          * lowering dirty_ratelimit will help meet both the position and rate
986          * control targets. Otherwise, don't update dirty_ratelimit if it will
987          * only help meet the rate target. After all, what the users ultimately
988          * feel and care are stable dirty rate and small position error.
989          *
990          * |task_ratelimit - dirty_ratelimit| is used to limit the step size
991          * and filter out the singular points of balanced_dirty_ratelimit. Which
992          * keeps jumping around randomly and can even leap far away at times
993          * due to the small 200ms estimation period of dirty_rate (we want to
994          * keep that period small to reduce time lags).
995          */
996         step = 0;
997         if (dirty < setpoint) {
998                 x = min(bdi->balanced_dirty_ratelimit,
999                          min(balanced_dirty_ratelimit, task_ratelimit));
1000                 if (dirty_ratelimit < x)
1001                         step = x - dirty_ratelimit;
1002         } else {
1003                 x = max(bdi->balanced_dirty_ratelimit,
1004                          max(balanced_dirty_ratelimit, task_ratelimit));
1005                 if (dirty_ratelimit > x)
1006                         step = dirty_ratelimit - x;
1007         }
1008
1009         /*
1010          * Don't pursue 100% rate matching. It's impossible since the balanced
1011          * rate itself is constantly fluctuating. So decrease the track speed
1012          * when it gets close to the target. Helps eliminate pointless tremors.
1013          */
1014         step >>= dirty_ratelimit / (2 * step + 1);
1015         /*
1016          * Limit the tracking speed to avoid overshooting.
1017          */
1018         step = (step + 7) / 8;
1019
1020         if (dirty_ratelimit < balanced_dirty_ratelimit)
1021                 dirty_ratelimit += step;
1022         else
1023                 dirty_ratelimit -= step;
1024
1025         bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1026         bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1027
1028         trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
1029 }
1030
1031 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
1032                             unsigned long thresh,
1033                             unsigned long bg_thresh,
1034                             unsigned long dirty,
1035                             unsigned long bdi_thresh,
1036                             unsigned long bdi_dirty,
1037                             unsigned long start_time)
1038 {
1039         unsigned long now = jiffies;
1040         unsigned long elapsed = now - bdi->bw_time_stamp;
1041         unsigned long dirtied;
1042         unsigned long written;
1043
1044         /*
1045          * rate-limit, only update once every 200ms.
1046          */
1047         if (elapsed < BANDWIDTH_INTERVAL)
1048                 return;
1049
1050         dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
1051         written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
1052
1053         /*
1054          * Skip quiet periods when disk bandwidth is under-utilized.
1055          * (at least 1s idle time between two flusher runs)
1056          */
1057         if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
1058                 goto snapshot;
1059
1060         if (thresh) {
1061                 global_update_bandwidth(thresh, dirty, now);
1062                 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1063                                            bdi_thresh, bdi_dirty,
1064                                            dirtied, elapsed);
1065         }
1066         bdi_update_write_bandwidth(bdi, elapsed, written);
1067
1068 snapshot:
1069         bdi->dirtied_stamp = dirtied;
1070         bdi->written_stamp = written;
1071         bdi->bw_time_stamp = now;
1072 }
1073
1074 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
1075                                  unsigned long thresh,
1076                                  unsigned long bg_thresh,
1077                                  unsigned long dirty,
1078                                  unsigned long bdi_thresh,
1079                                  unsigned long bdi_dirty,
1080                                  unsigned long start_time)
1081 {
1082         if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1083                 return;
1084         spin_lock(&bdi->wb.list_lock);
1085         __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1086                                bdi_thresh, bdi_dirty, start_time);
1087         spin_unlock(&bdi->wb.list_lock);
1088 }
1089
1090 /*
1091  * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1092  * will look to see if it needs to start dirty throttling.
1093  *
1094  * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1095  * global_page_state() too often. So scale it near-sqrt to the safety margin
1096  * (the number of pages we may dirty without exceeding the dirty limits).
1097  */
1098 static unsigned long dirty_poll_interval(unsigned long dirty,
1099                                          unsigned long thresh)
1100 {
1101         if (thresh > dirty)
1102                 return 1UL << (ilog2(thresh - dirty) >> 1);
1103
1104         return 1;
1105 }
1106
1107 static long bdi_max_pause(struct backing_dev_info *bdi,
1108                           unsigned long bdi_dirty)
1109 {
1110         long bw = bdi->avg_write_bandwidth;
1111         long t;
1112
1113         /*
1114          * Limit pause time for small memory systems. If sleeping for too long
1115          * time, a small pool of dirty/writeback pages may go empty and disk go
1116          * idle.
1117          *
1118          * 8 serves as the safety ratio.
1119          */
1120         t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1121         t++;
1122
1123         return min_t(long, t, MAX_PAUSE);
1124 }
1125
1126 static long bdi_min_pause(struct backing_dev_info *bdi,
1127                           long max_pause,
1128                           unsigned long task_ratelimit,
1129                           unsigned long dirty_ratelimit,
1130                           int *nr_dirtied_pause)
1131 {
1132         long hi = ilog2(bdi->avg_write_bandwidth);
1133         long lo = ilog2(bdi->dirty_ratelimit);
1134         long t;         /* target pause */
1135         long pause;     /* estimated next pause */
1136         int pages;      /* target nr_dirtied_pause */
1137
1138         /* target for 10ms pause on 1-dd case */
1139         t = max(1, HZ / 100);
1140
1141         /*
1142          * Scale up pause time for concurrent dirtiers in order to reduce CPU
1143          * overheads.
1144          *
1145          * (N * 10ms) on 2^N concurrent tasks.
1146          */
1147         if (hi > lo)
1148                 t += (hi - lo) * (10 * HZ) / 1024;
1149
1150         /*
1151          * This is a bit convoluted. We try to base the next nr_dirtied_pause
1152          * on the much more stable dirty_ratelimit. However the next pause time
1153          * will be computed based on task_ratelimit and the two rate limits may
1154          * depart considerably at some time. Especially if task_ratelimit goes
1155          * below dirty_ratelimit/2 and the target pause is max_pause, the next
1156          * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1157          * result task_ratelimit won't be executed faithfully, which could
1158          * eventually bring down dirty_ratelimit.
1159          *
1160          * We apply two rules to fix it up:
1161          * 1) try to estimate the next pause time and if necessary, use a lower
1162          *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1163          *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1164          * 2) limit the target pause time to max_pause/2, so that the normal
1165          *    small fluctuations of task_ratelimit won't trigger rule (1) and
1166          *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1167          */
1168         t = min(t, 1 + max_pause / 2);
1169         pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1170
1171         /*
1172          * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1173          * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1174          * When the 16 consecutive reads are often interrupted by some dirty
1175          * throttling pause during the async writes, cfq will go into idles
1176          * (deadline is fine). So push nr_dirtied_pause as high as possible
1177          * until reaches DIRTY_POLL_THRESH=32 pages.
1178          */
1179         if (pages < DIRTY_POLL_THRESH) {
1180                 t = max_pause;
1181                 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1182                 if (pages > DIRTY_POLL_THRESH) {
1183                         pages = DIRTY_POLL_THRESH;
1184                         t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1185                 }
1186         }
1187
1188         pause = HZ * pages / (task_ratelimit + 1);
1189         if (pause > max_pause) {
1190                 t = max_pause;
1191                 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1192         }
1193
1194         *nr_dirtied_pause = pages;
1195         /*
1196          * The minimal pause time will normally be half the target pause time.
1197          */
1198         return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1199 }
1200
1201 /*
1202  * balance_dirty_pages() must be called by processes which are generating dirty
1203  * data.  It looks at the number of dirty pages in the machine and will force
1204  * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1205  * If we're over `background_thresh' then the writeback threads are woken to
1206  * perform some writeout.
1207  */
1208 static void balance_dirty_pages(struct address_space *mapping,
1209                                 unsigned long pages_dirtied)
1210 {
1211         unsigned long nr_reclaimable;   /* = file_dirty + unstable_nfs */
1212         unsigned long bdi_reclaimable;
1213         unsigned long nr_dirty;  /* = file_dirty + writeback + unstable_nfs */
1214         unsigned long bdi_dirty;
1215         unsigned long freerun;
1216         unsigned long background_thresh;
1217         unsigned long dirty_thresh;
1218         unsigned long bdi_thresh;
1219         long period;
1220         long pause;
1221         long max_pause;
1222         long min_pause;
1223         int nr_dirtied_pause;
1224         bool dirty_exceeded = false;
1225         unsigned long task_ratelimit;
1226         unsigned long dirty_ratelimit;
1227         unsigned long pos_ratio;
1228         struct backing_dev_info *bdi = mapping->backing_dev_info;
1229         unsigned long start_time = jiffies;
1230
1231         for (;;) {
1232                 unsigned long now = jiffies;
1233
1234                 /*
1235                  * Unstable writes are a feature of certain networked
1236                  * filesystems (i.e. NFS) in which data may have been
1237                  * written to the server's write cache, but has not yet
1238                  * been flushed to permanent storage.
1239                  */
1240                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1241                                         global_page_state(NR_UNSTABLE_NFS);
1242                 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1243
1244                 global_dirty_limits(&background_thresh, &dirty_thresh);
1245
1246                 /*
1247                  * Throttle it only when the background writeback cannot
1248                  * catch-up. This avoids (excessively) small writeouts
1249                  * when the bdi limits are ramping up.
1250                  */
1251                 freerun = dirty_freerun_ceiling(dirty_thresh,
1252                                                 background_thresh);
1253                 if (nr_dirty <= freerun) {
1254                         current->dirty_paused_when = now;
1255                         current->nr_dirtied = 0;
1256                         current->nr_dirtied_pause =
1257                                 dirty_poll_interval(nr_dirty, dirty_thresh);
1258                         break;
1259                 }
1260
1261                 if (unlikely(!writeback_in_progress(bdi)))
1262                         bdi_start_background_writeback(bdi);
1263
1264                 /*
1265                  * bdi_thresh is not treated as some limiting factor as
1266                  * dirty_thresh, due to reasons
1267                  * - in JBOD setup, bdi_thresh can fluctuate a lot
1268                  * - in a system with HDD and USB key, the USB key may somehow
1269                  *   go into state (bdi_dirty >> bdi_thresh) either because
1270                  *   bdi_dirty starts high, or because bdi_thresh drops low.
1271                  *   In this case we don't want to hard throttle the USB key
1272                  *   dirtiers for 100 seconds until bdi_dirty drops under
1273                  *   bdi_thresh. Instead the auxiliary bdi control line in
1274                  *   bdi_position_ratio() will let the dirtier task progress
1275                  *   at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1276                  */
1277                 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1278
1279                 /*
1280                  * In order to avoid the stacked BDI deadlock we need
1281                  * to ensure we accurately count the 'dirty' pages when
1282                  * the threshold is low.
1283                  *
1284                  * Otherwise it would be possible to get thresh+n pages
1285                  * reported dirty, even though there are thresh-m pages
1286                  * actually dirty; with m+n sitting in the percpu
1287                  * deltas.
1288                  */
1289                 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1290                         bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1291                         bdi_dirty = bdi_reclaimable +
1292                                     bdi_stat_sum(bdi, BDI_WRITEBACK);
1293                 } else {
1294                         bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1295                         bdi_dirty = bdi_reclaimable +
1296                                     bdi_stat(bdi, BDI_WRITEBACK);
1297                 }
1298
1299                 dirty_exceeded = (bdi_dirty > bdi_thresh) &&
1300                                   (nr_dirty > dirty_thresh);
1301                 if (dirty_exceeded && !bdi->dirty_exceeded)
1302                         bdi->dirty_exceeded = 1;
1303
1304                 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1305                                      nr_dirty, bdi_thresh, bdi_dirty,
1306                                      start_time);
1307
1308                 dirty_ratelimit = bdi->dirty_ratelimit;
1309                 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1310                                                background_thresh, nr_dirty,
1311                                                bdi_thresh, bdi_dirty);
1312                 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1313                                                         RATELIMIT_CALC_SHIFT;
1314                 max_pause = bdi_max_pause(bdi, bdi_dirty);
1315                 min_pause = bdi_min_pause(bdi, max_pause,
1316                                           task_ratelimit, dirty_ratelimit,
1317                                           &nr_dirtied_pause);
1318
1319                 if (unlikely(task_ratelimit == 0)) {
1320                         period = max_pause;
1321                         pause = max_pause;
1322                         goto pause;
1323                 }
1324                 period = HZ * pages_dirtied / task_ratelimit;
1325                 pause = period;
1326                 if (current->dirty_paused_when)
1327                         pause -= now - current->dirty_paused_when;
1328                 /*
1329                  * For less than 1s think time (ext3/4 may block the dirtier
1330                  * for up to 800ms from time to time on 1-HDD; so does xfs,
1331                  * however at much less frequency), try to compensate it in
1332                  * future periods by updating the virtual time; otherwise just
1333                  * do a reset, as it may be a light dirtier.
1334                  */
1335                 if (pause < min_pause) {
1336                         trace_balance_dirty_pages(bdi,
1337                                                   dirty_thresh,
1338                                                   background_thresh,
1339                                                   nr_dirty,
1340                                                   bdi_thresh,
1341                                                   bdi_dirty,
1342                                                   dirty_ratelimit,
1343                                                   task_ratelimit,
1344                                                   pages_dirtied,
1345                                                   period,
1346                                                   min(pause, 0L),
1347                                                   start_time);
1348                         if (pause < -HZ) {
1349                                 current->dirty_paused_when = now;
1350                                 current->nr_dirtied = 0;
1351                         } else if (period) {
1352                                 current->dirty_paused_when += period;
1353                                 current->nr_dirtied = 0;
1354                         } else if (current->nr_dirtied_pause <= pages_dirtied)
1355                                 current->nr_dirtied_pause += pages_dirtied;
1356                         break;
1357                 }
1358                 if (unlikely(pause > max_pause)) {
1359                         /* for occasional dropped task_ratelimit */
1360                         now += min(pause - max_pause, max_pause);
1361                         pause = max_pause;
1362                 }
1363
1364 pause:
1365                 trace_balance_dirty_pages(bdi,
1366                                           dirty_thresh,
1367                                           background_thresh,
1368                                           nr_dirty,
1369                                           bdi_thresh,
1370                                           bdi_dirty,
1371                                           dirty_ratelimit,
1372                                           task_ratelimit,
1373                                           pages_dirtied,
1374                                           period,
1375                                           pause,
1376                                           start_time);
1377                 __set_current_state(TASK_KILLABLE);
1378                 io_schedule_timeout(pause);
1379
1380                 current->dirty_paused_when = now + pause;
1381                 current->nr_dirtied = 0;
1382                 current->nr_dirtied_pause = nr_dirtied_pause;
1383
1384                 /*
1385                  * This is typically equal to (nr_dirty < dirty_thresh) and can
1386                  * also keep "1000+ dd on a slow USB stick" under control.
1387                  */
1388                 if (task_ratelimit)
1389                         break;
1390
1391                 /*
1392                  * In the case of an unresponding NFS server and the NFS dirty
1393                  * pages exceeds dirty_thresh, give the other good bdi's a pipe
1394                  * to go through, so that tasks on them still remain responsive.
1395                  *
1396                  * In theory 1 page is enough to keep the comsumer-producer
1397                  * pipe going: the flusher cleans 1 page => the task dirties 1
1398                  * more page. However bdi_dirty has accounting errors.  So use
1399                  * the larger and more IO friendly bdi_stat_error.
1400                  */
1401                 if (bdi_dirty <= bdi_stat_error(bdi))
1402                         break;
1403
1404                 if (fatal_signal_pending(current))
1405                         break;
1406         }
1407
1408         if (!dirty_exceeded && bdi->dirty_exceeded)
1409                 bdi->dirty_exceeded = 0;
1410
1411         if (writeback_in_progress(bdi))
1412                 return;
1413
1414         /*
1415          * In laptop mode, we wait until hitting the higher threshold before
1416          * starting background writeout, and then write out all the way down
1417          * to the lower threshold.  So slow writers cause minimal disk activity.
1418          *
1419          * In normal mode, we start background writeout at the lower
1420          * background_thresh, to keep the amount of dirty memory low.
1421          */
1422         if (laptop_mode)
1423                 return;
1424
1425         if (nr_reclaimable > background_thresh)
1426                 bdi_start_background_writeback(bdi);
1427 }
1428
1429 void set_page_dirty_balance(struct page *page, int page_mkwrite)
1430 {
1431         if (set_page_dirty(page) || page_mkwrite) {
1432                 struct address_space *mapping = page_mapping(page);
1433
1434                 if (mapping)
1435                         balance_dirty_pages_ratelimited(mapping);
1436         }
1437 }
1438
1439 static DEFINE_PER_CPU(int, bdp_ratelimits);
1440
1441 /*
1442  * Normal tasks are throttled by
1443  *      loop {
1444  *              dirty tsk->nr_dirtied_pause pages;
1445  *              take a snap in balance_dirty_pages();
1446  *      }
1447  * However there is a worst case. If every task exit immediately when dirtied
1448  * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1449  * called to throttle the page dirties. The solution is to save the not yet
1450  * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1451  * randomly into the running tasks. This works well for the above worst case,
1452  * as the new task will pick up and accumulate the old task's leaked dirty
1453  * count and eventually get throttled.
1454  */
1455 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1456
1457 /**
1458  * balance_dirty_pages_ratelimited - balance dirty memory state
1459  * @mapping: address_space which was dirtied
1460  *
1461  * Processes which are dirtying memory should call in here once for each page
1462  * which was newly dirtied.  The function will periodically check the system's
1463  * dirty state and will initiate writeback if needed.
1464  *
1465  * On really big machines, get_writeback_state is expensive, so try to avoid
1466  * calling it too often (ratelimiting).  But once we're over the dirty memory
1467  * limit we decrease the ratelimiting by a lot, to prevent individual processes
1468  * from overshooting the limit by (ratelimit_pages) each.
1469  */
1470 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1471 {
1472         struct backing_dev_info *bdi = mapping->backing_dev_info;
1473         int ratelimit;
1474         int *p;
1475
1476         if (!bdi_cap_account_dirty(bdi))
1477                 return;
1478
1479         ratelimit = current->nr_dirtied_pause;
1480         if (bdi->dirty_exceeded)
1481                 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1482
1483         preempt_disable();
1484         /*
1485          * This prevents one CPU to accumulate too many dirtied pages without
1486          * calling into balance_dirty_pages(), which can happen when there are
1487          * 1000+ tasks, all of them start dirtying pages at exactly the same
1488          * time, hence all honoured too large initial task->nr_dirtied_pause.
1489          */
1490         p =  &__get_cpu_var(bdp_ratelimits);
1491         if (unlikely(current->nr_dirtied >= ratelimit))
1492                 *p = 0;
1493         else if (unlikely(*p >= ratelimit_pages)) {
1494                 *p = 0;
1495                 ratelimit = 0;
1496         }
1497         /*
1498          * Pick up the dirtied pages by the exited tasks. This avoids lots of
1499          * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1500          * the dirty throttling and livelock other long-run dirtiers.
1501          */
1502         p = &__get_cpu_var(dirty_throttle_leaks);
1503         if (*p > 0 && current->nr_dirtied < ratelimit) {
1504                 unsigned long nr_pages_dirtied;
1505                 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1506                 *p -= nr_pages_dirtied;
1507                 current->nr_dirtied += nr_pages_dirtied;
1508         }
1509         preempt_enable();
1510
1511         if (unlikely(current->nr_dirtied >= ratelimit))
1512                 balance_dirty_pages(mapping, current->nr_dirtied);
1513 }
1514 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1515
1516 void throttle_vm_writeout(gfp_t gfp_mask)
1517 {
1518         unsigned long background_thresh;
1519         unsigned long dirty_thresh;
1520
1521         for ( ; ; ) {
1522                 global_dirty_limits(&background_thresh, &dirty_thresh);
1523                 dirty_thresh = hard_dirty_limit(dirty_thresh);
1524
1525                 /*
1526                  * Boost the allowable dirty threshold a bit for page
1527                  * allocators so they don't get DoS'ed by heavy writers
1528                  */
1529                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
1530
1531                 if (global_page_state(NR_UNSTABLE_NFS) +
1532                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
1533                                 break;
1534                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1535
1536                 /*
1537                  * The caller might hold locks which can prevent IO completion
1538                  * or progress in the filesystem.  So we cannot just sit here
1539                  * waiting for IO to complete.
1540                  */
1541                 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1542                         break;
1543         }
1544 }
1545
1546 /*
1547  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1548  */
1549 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1550         void __user *buffer, size_t *length, loff_t *ppos)
1551 {
1552         proc_dointvec(table, write, buffer, length, ppos);
1553         return 0;
1554 }
1555
1556 #ifdef CONFIG_BLOCK
1557 void laptop_mode_timer_fn(unsigned long data)
1558 {
1559         struct request_queue *q = (struct request_queue *)data;
1560         int nr_pages = global_page_state(NR_FILE_DIRTY) +
1561                 global_page_state(NR_UNSTABLE_NFS);
1562
1563         /*
1564          * We want to write everything out, not just down to the dirty
1565          * threshold
1566          */
1567         if (bdi_has_dirty_io(&q->backing_dev_info))
1568                 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1569                                         WB_REASON_LAPTOP_TIMER);
1570 }
1571
1572 /*
1573  * We've spun up the disk and we're in laptop mode: schedule writeback
1574  * of all dirty data a few seconds from now.  If the flush is already scheduled
1575  * then push it back - the user is still using the disk.
1576  */
1577 void laptop_io_completion(struct backing_dev_info *info)
1578 {
1579         mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1580 }
1581
1582 /*
1583  * We're in laptop mode and we've just synced. The sync's writes will have
1584  * caused another writeback to be scheduled by laptop_io_completion.
1585  * Nothing needs to be written back anymore, so we unschedule the writeback.
1586  */
1587 void laptop_sync_completion(void)
1588 {
1589         struct backing_dev_info *bdi;
1590
1591         rcu_read_lock();
1592
1593         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1594                 del_timer(&bdi->laptop_mode_wb_timer);
1595
1596         rcu_read_unlock();
1597 }
1598 #endif
1599
1600 /*
1601  * If ratelimit_pages is too high then we can get into dirty-data overload
1602  * if a large number of processes all perform writes at the same time.
1603  * If it is too low then SMP machines will call the (expensive)
1604  * get_writeback_state too often.
1605  *
1606  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1607  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1608  * thresholds.
1609  */
1610
1611 void writeback_set_ratelimit(void)
1612 {
1613         unsigned long background_thresh;
1614         unsigned long dirty_thresh;
1615         global_dirty_limits(&background_thresh, &dirty_thresh);
1616         global_dirty_limit = dirty_thresh;
1617         ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1618         if (ratelimit_pages < 16)
1619                 ratelimit_pages = 16;
1620 }
1621
1622 static int __cpuinit
1623 ratelimit_handler(struct notifier_block *self, unsigned long action,
1624                   void *hcpu)
1625 {
1626
1627         switch (action & ~CPU_TASKS_FROZEN) {
1628         case CPU_ONLINE:
1629         case CPU_DEAD:
1630                 writeback_set_ratelimit();
1631                 return NOTIFY_OK;
1632         default:
1633                 return NOTIFY_DONE;
1634         }
1635 }
1636
1637 static struct notifier_block __cpuinitdata ratelimit_nb = {
1638         .notifier_call  = ratelimit_handler,
1639         .next           = NULL,
1640 };
1641
1642 /*
1643  * Called early on to tune the page writeback dirty limits.
1644  *
1645  * We used to scale dirty pages according to how total memory
1646  * related to pages that could be allocated for buffers (by
1647  * comparing nr_free_buffer_pages() to vm_total_pages.
1648  *
1649  * However, that was when we used "dirty_ratio" to scale with
1650  * all memory, and we don't do that any more. "dirty_ratio"
1651  * is now applied to total non-HIGHPAGE memory (by subtracting
1652  * totalhigh_pages from vm_total_pages), and as such we can't
1653  * get into the old insane situation any more where we had
1654  * large amounts of dirty pages compared to a small amount of
1655  * non-HIGHMEM memory.
1656  *
1657  * But we might still want to scale the dirty_ratio by how
1658  * much memory the box has..
1659  */
1660 void __init page_writeback_init(void)
1661 {
1662         writeback_set_ratelimit();
1663         register_cpu_notifier(&ratelimit_nb);
1664
1665         fprop_global_init(&writeout_completions);
1666 }
1667
1668 /**
1669  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1670  * @mapping: address space structure to write
1671  * @start: starting page index
1672  * @end: ending page index (inclusive)
1673  *
1674  * This function scans the page range from @start to @end (inclusive) and tags
1675  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1676  * that write_cache_pages (or whoever calls this function) will then use
1677  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
1678  * used to avoid livelocking of writeback by a process steadily creating new
1679  * dirty pages in the file (thus it is important for this function to be quick
1680  * so that it can tag pages faster than a dirtying process can create them).
1681  */
1682 /*
1683  * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1684  */
1685 void tag_pages_for_writeback(struct address_space *mapping,
1686                              pgoff_t start, pgoff_t end)
1687 {
1688 #define WRITEBACK_TAG_BATCH 4096
1689         unsigned long tagged;
1690
1691         do {
1692                 spin_lock_irq(&mapping->tree_lock);
1693                 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1694                                 &start, end, WRITEBACK_TAG_BATCH,
1695                                 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1696                 spin_unlock_irq(&mapping->tree_lock);
1697                 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1698                 cond_resched();
1699                 /* We check 'start' to handle wrapping when end == ~0UL */
1700         } while (tagged >= WRITEBACK_TAG_BATCH && start);
1701 }
1702 EXPORT_SYMBOL(tag_pages_for_writeback);
1703
1704 /**
1705  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1706  * @mapping: address space structure to write
1707  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1708  * @writepage: function called for each page
1709  * @data: data passed to writepage function
1710  *
1711  * If a page is already under I/O, write_cache_pages() skips it, even
1712  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
1713  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
1714  * and msync() need to guarantee that all the data which was dirty at the time
1715  * the call was made get new I/O started against them.  If wbc->sync_mode is
1716  * WB_SYNC_ALL then we were called for data integrity and we must wait for
1717  * existing IO to complete.
1718  *
1719  * To avoid livelocks (when other process dirties new pages), we first tag
1720  * pages which should be written back with TOWRITE tag and only then start
1721  * writing them. For data-integrity sync we have to be careful so that we do
1722  * not miss some pages (e.g., because some other process has cleared TOWRITE
1723  * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1724  * by the process clearing the DIRTY tag (and submitting the page for IO).
1725  */
1726 int write_cache_pages(struct address_space *mapping,
1727                       struct writeback_control *wbc, writepage_t writepage,
1728                       void *data)
1729 {
1730         int ret = 0;
1731         int done = 0;
1732         struct pagevec pvec;
1733         int nr_pages;
1734         pgoff_t uninitialized_var(writeback_index);
1735         pgoff_t index;
1736         pgoff_t end;            /* Inclusive */
1737         pgoff_t done_index;
1738         int cycled;
1739         int range_whole = 0;
1740         int tag;
1741
1742         pagevec_init(&pvec, 0);
1743         if (wbc->range_cyclic) {
1744                 writeback_index = mapping->writeback_index; /* prev offset */
1745                 index = writeback_index;
1746                 if (index == 0)
1747                         cycled = 1;
1748                 else
1749                         cycled = 0;
1750                 end = -1;
1751         } else {
1752                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1753                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1754                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1755                         range_whole = 1;
1756                 cycled = 1; /* ignore range_cyclic tests */
1757         }
1758         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1759                 tag = PAGECACHE_TAG_TOWRITE;
1760         else
1761                 tag = PAGECACHE_TAG_DIRTY;
1762 retry:
1763         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1764                 tag_pages_for_writeback(mapping, index, end);
1765         done_index = index;
1766         while (!done && (index <= end)) {
1767                 int i;
1768
1769                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1770                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1771                 if (nr_pages == 0)
1772                         break;
1773
1774                 for (i = 0; i < nr_pages; i++) {
1775                         struct page *page = pvec.pages[i];
1776
1777                         /*
1778                          * At this point, the page may be truncated or
1779                          * invalidated (changing page->mapping to NULL), or
1780                          * even swizzled back from swapper_space to tmpfs file
1781                          * mapping. However, page->index will not change
1782                          * because we have a reference on the page.
1783                          */
1784                         if (page->index > end) {
1785                                 /*
1786                                  * can't be range_cyclic (1st pass) because
1787                                  * end == -1 in that case.
1788                                  */
1789                                 done = 1;
1790                                 break;
1791                         }
1792
1793                         done_index = page->index;
1794
1795                         lock_page(page);
1796
1797                         /*
1798                          * Page truncated or invalidated. We can freely skip it
1799                          * then, even for data integrity operations: the page
1800                          * has disappeared concurrently, so there could be no
1801                          * real expectation of this data interity operation
1802                          * even if there is now a new, dirty page at the same
1803                          * pagecache address.
1804                          */
1805                         if (unlikely(page->mapping != mapping)) {
1806 continue_unlock:
1807                                 unlock_page(page);
1808                                 continue;
1809                         }
1810
1811                         if (!PageDirty(page)) {
1812                                 /* someone wrote it for us */
1813                                 goto continue_unlock;
1814                         }
1815
1816                         if (PageWriteback(page)) {
1817                                 if (wbc->sync_mode != WB_SYNC_NONE)
1818                                         wait_on_page_writeback(page);
1819                                 else
1820                                         goto continue_unlock;
1821                         }
1822
1823                         BUG_ON(PageWriteback(page));
1824                         if (!clear_page_dirty_for_io(page))
1825                                 goto continue_unlock;
1826
1827                         trace_wbc_writepage(wbc, mapping->backing_dev_info);
1828                         ret = (*writepage)(page, wbc, data);
1829                         if (unlikely(ret)) {
1830                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1831                                         unlock_page(page);
1832                                         ret = 0;
1833                                 } else {
1834                                         /*
1835                                          * done_index is set past this page,
1836                                          * so media errors will not choke
1837                                          * background writeout for the entire
1838                                          * file. This has consequences for
1839                                          * range_cyclic semantics (ie. it may
1840                                          * not be suitable for data integrity
1841                                          * writeout).
1842                                          */
1843                                         done_index = page->index + 1;
1844                                         done = 1;
1845                                         break;
1846                                 }
1847                         }
1848
1849                         /*
1850                          * We stop writing back only if we are not doing
1851                          * integrity sync. In case of integrity sync we have to
1852                          * keep going until we have written all the pages
1853                          * we tagged for writeback prior to entering this loop.
1854                          */
1855                         if (--wbc->nr_to_write <= 0 &&
1856                             wbc->sync_mode == WB_SYNC_NONE) {
1857                                 done = 1;
1858                                 break;
1859                         }
1860                 }
1861                 pagevec_release(&pvec);
1862                 cond_resched();
1863         }
1864         if (!cycled && !done) {
1865                 /*
1866                  * range_cyclic:
1867                  * We hit the last page and there is more work to be done: wrap
1868                  * back to the start of the file
1869                  */
1870                 cycled = 1;
1871                 index = 0;
1872                 end = writeback_index - 1;
1873                 goto retry;
1874         }
1875         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1876                 mapping->writeback_index = done_index;
1877
1878         return ret;
1879 }
1880 EXPORT_SYMBOL(write_cache_pages);
1881
1882 /*
1883  * Function used by generic_writepages to call the real writepage
1884  * function and set the mapping flags on error
1885  */
1886 static int __writepage(struct page *page, struct writeback_control *wbc,
1887                        void *data)
1888 {
1889         struct address_space *mapping = data;
1890         int ret = mapping->a_ops->writepage(page, wbc);
1891         mapping_set_error(mapping, ret);
1892         return ret;
1893 }
1894
1895 /**
1896  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1897  * @mapping: address space structure to write
1898  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1899  *
1900  * This is a library function, which implements the writepages()
1901  * address_space_operation.
1902  */
1903 int generic_writepages(struct address_space *mapping,
1904                        struct writeback_control *wbc)
1905 {
1906         struct blk_plug plug;
1907         int ret;
1908
1909         /* deal with chardevs and other special file */
1910         if (!mapping->a_ops->writepage)
1911                 return 0;
1912
1913         blk_start_plug(&plug);
1914         ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1915         blk_finish_plug(&plug);
1916         return ret;
1917 }
1918
1919 EXPORT_SYMBOL(generic_writepages);
1920
1921 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1922 {
1923         int ret;
1924
1925         if (wbc->nr_to_write <= 0)
1926                 return 0;
1927         if (mapping->a_ops->writepages)
1928                 ret = mapping->a_ops->writepages(mapping, wbc);
1929         else
1930                 ret = generic_writepages(mapping, wbc);
1931         return ret;
1932 }
1933
1934 /**
1935  * write_one_page - write out a single page and optionally wait on I/O
1936  * @page: the page to write
1937  * @wait: if true, wait on writeout
1938  *
1939  * The page must be locked by the caller and will be unlocked upon return.
1940  *
1941  * write_one_page() returns a negative error code if I/O failed.
1942  */
1943 int write_one_page(struct page *page, int wait)
1944 {
1945         struct address_space *mapping = page->mapping;
1946         int ret = 0;
1947         struct writeback_control wbc = {
1948                 .sync_mode = WB_SYNC_ALL,
1949                 .nr_to_write = 1,
1950         };
1951
1952         BUG_ON(!PageLocked(page));
1953
1954         if (wait)
1955                 wait_on_page_writeback(page);
1956
1957         if (clear_page_dirty_for_io(page)) {
1958                 page_cache_get(page);
1959                 ret = mapping->a_ops->writepage(page, &wbc);
1960                 if (ret == 0 && wait) {
1961                         wait_on_page_writeback(page);
1962                         if (PageError(page))
1963                                 ret = -EIO;
1964                 }
1965                 page_cache_release(page);
1966         } else {
1967                 unlock_page(page);
1968         }
1969         return ret;
1970 }
1971 EXPORT_SYMBOL(write_one_page);
1972
1973 /*
1974  * For address_spaces which do not use buffers nor write back.
1975  */
1976 int __set_page_dirty_no_writeback(struct page *page)
1977 {
1978         if (!PageDirty(page))
1979                 return !TestSetPageDirty(page);
1980         return 0;
1981 }
1982
1983 /*
1984  * Helper function for set_page_dirty family.
1985  * NOTE: This relies on being atomic wrt interrupts.
1986  */
1987 void account_page_dirtied(struct page *page, struct address_space *mapping)
1988 {
1989         trace_writeback_dirty_page(page, mapping);
1990
1991         if (mapping_cap_account_dirty(mapping)) {
1992                 __inc_zone_page_state(page, NR_FILE_DIRTY);
1993                 __inc_zone_page_state(page, NR_DIRTIED);
1994                 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1995                 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1996                 task_io_account_write(PAGE_CACHE_SIZE);
1997                 current->nr_dirtied++;
1998                 this_cpu_inc(bdp_ratelimits);
1999         }
2000 }
2001 EXPORT_SYMBOL(account_page_dirtied);
2002
2003 /*
2004  * Helper function for set_page_writeback family.
2005  * NOTE: Unlike account_page_dirtied this does not rely on being atomic
2006  * wrt interrupts.
2007  */
2008 void account_page_writeback(struct page *page)
2009 {
2010         inc_zone_page_state(page, NR_WRITEBACK);
2011 }
2012 EXPORT_SYMBOL(account_page_writeback);
2013
2014 /*
2015  * For address_spaces which do not use buffers.  Just tag the page as dirty in
2016  * its radix tree.
2017  *
2018  * This is also used when a single buffer is being dirtied: we want to set the
2019  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
2020  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2021  *
2022  * Most callers have locked the page, which pins the address_space in memory.
2023  * But zap_pte_range() does not lock the page, however in that case the
2024  * mapping is pinned by the vma's ->vm_file reference.
2025  *
2026  * We take care to handle the case where the page was truncated from the
2027  * mapping by re-checking page_mapping() inside tree_lock.
2028  */
2029 int __set_page_dirty_nobuffers(struct page *page)
2030 {
2031         if (!TestSetPageDirty(page)) {
2032                 struct address_space *mapping = page_mapping(page);
2033                 struct address_space *mapping2;
2034
2035                 if (!mapping)
2036                         return 1;
2037
2038                 spin_lock_irq(&mapping->tree_lock);
2039                 mapping2 = page_mapping(page);
2040                 if (mapping2) { /* Race with truncate? */
2041                         BUG_ON(mapping2 != mapping);
2042                         WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2043                         account_page_dirtied(page, mapping);
2044                         radix_tree_tag_set(&mapping->page_tree,
2045                                 page_index(page), PAGECACHE_TAG_DIRTY);
2046                 }
2047                 spin_unlock_irq(&mapping->tree_lock);
2048                 if (mapping->host) {
2049                         /* !PageAnon && !swapper_space */
2050                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2051                 }
2052                 return 1;
2053         }
2054         return 0;
2055 }
2056 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2057
2058 /*
2059  * Call this whenever redirtying a page, to de-account the dirty counters
2060  * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2061  * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2062  * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2063  * control.
2064  */
2065 void account_page_redirty(struct page *page)
2066 {
2067         struct address_space *mapping = page->mapping;
2068         if (mapping && mapping_cap_account_dirty(mapping)) {
2069                 current->nr_dirtied--;
2070                 dec_zone_page_state(page, NR_DIRTIED);
2071                 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2072         }
2073 }
2074 EXPORT_SYMBOL(account_page_redirty);
2075
2076 /*
2077  * When a writepage implementation decides that it doesn't want to write this
2078  * page for some reason, it should redirty the locked page via
2079  * redirty_page_for_writepage() and it should then unlock the page and return 0
2080  */
2081 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2082 {
2083         wbc->pages_skipped++;
2084         account_page_redirty(page);
2085         return __set_page_dirty_nobuffers(page);
2086 }
2087 EXPORT_SYMBOL(redirty_page_for_writepage);
2088
2089 /*
2090  * Dirty a page.
2091  *
2092  * For pages with a mapping this should be done under the page lock
2093  * for the benefit of asynchronous memory errors who prefer a consistent
2094  * dirty state. This rule can be broken in some special cases,
2095  * but should be better not to.
2096  *
2097  * If the mapping doesn't provide a set_page_dirty a_op, then
2098  * just fall through and assume that it wants buffer_heads.
2099  */
2100 int set_page_dirty(struct page *page)
2101 {
2102         struct address_space *mapping = page_mapping(page);
2103
2104         if (likely(mapping)) {
2105                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2106                 /*
2107                  * readahead/lru_deactivate_page could remain
2108                  * PG_readahead/PG_reclaim due to race with end_page_writeback
2109                  * About readahead, if the page is written, the flags would be
2110                  * reset. So no problem.
2111                  * About lru_deactivate_page, if the page is redirty, the flag
2112                  * will be reset. So no problem. but if the page is used by readahead
2113                  * it will confuse readahead and make it restart the size rampup
2114                  * process. But it's a trivial problem.
2115                  */
2116                 ClearPageReclaim(page);
2117 #ifdef CONFIG_BLOCK
2118                 if (!spd)
2119                         spd = __set_page_dirty_buffers;
2120 #endif
2121                 return (*spd)(page);
2122         }
2123         if (!PageDirty(page)) {
2124                 if (!TestSetPageDirty(page))
2125                         return 1;
2126         }
2127         return 0;
2128 }
2129 EXPORT_SYMBOL(set_page_dirty);
2130
2131 /*
2132  * set_page_dirty() is racy if the caller has no reference against
2133  * page->mapping->host, and if the page is unlocked.  This is because another
2134  * CPU could truncate the page off the mapping and then free the mapping.
2135  *
2136  * Usually, the page _is_ locked, or the caller is a user-space process which
2137  * holds a reference on the inode by having an open file.
2138  *
2139  * In other cases, the page should be locked before running set_page_dirty().
2140  */
2141 int set_page_dirty_lock(struct page *page)
2142 {
2143         int ret;
2144
2145         lock_page(page);
2146         ret = set_page_dirty(page);
2147         unlock_page(page);
2148         return ret;
2149 }
2150 EXPORT_SYMBOL(set_page_dirty_lock);
2151
2152 /*
2153  * Clear a page's dirty flag, while caring for dirty memory accounting.
2154  * Returns true if the page was previously dirty.
2155  *
2156  * This is for preparing to put the page under writeout.  We leave the page
2157  * tagged as dirty in the radix tree so that a concurrent write-for-sync
2158  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
2159  * implementation will run either set_page_writeback() or set_page_dirty(),
2160  * at which stage we bring the page's dirty flag and radix-tree dirty tag
2161  * back into sync.
2162  *
2163  * This incoherency between the page's dirty flag and radix-tree tag is
2164  * unfortunate, but it only exists while the page is locked.
2165  */
2166 int clear_page_dirty_for_io(struct page *page)
2167 {
2168         struct address_space *mapping = page_mapping(page);
2169
2170         BUG_ON(!PageLocked(page));
2171
2172         if (mapping && mapping_cap_account_dirty(mapping)) {
2173                 /*
2174                  * Yes, Virginia, this is indeed insane.
2175                  *
2176                  * We use this sequence to make sure that
2177                  *  (a) we account for dirty stats properly
2178                  *  (b) we tell the low-level filesystem to
2179                  *      mark the whole page dirty if it was
2180                  *      dirty in a pagetable. Only to then
2181                  *  (c) clean the page again and return 1 to
2182                  *      cause the writeback.
2183                  *
2184                  * This way we avoid all nasty races with the
2185                  * dirty bit in multiple places and clearing
2186                  * them concurrently from different threads.
2187                  *
2188                  * Note! Normally the "set_page_dirty(page)"
2189                  * has no effect on the actual dirty bit - since
2190                  * that will already usually be set. But we
2191                  * need the side effects, and it can help us
2192                  * avoid races.
2193                  *
2194                  * We basically use the page "master dirty bit"
2195                  * as a serialization point for all the different
2196                  * threads doing their things.
2197                  */
2198                 if (page_mkclean(page))
2199                         set_page_dirty(page);
2200                 /*
2201                  * We carefully synchronise fault handlers against
2202                  * installing a dirty pte and marking the page dirty
2203                  * at this point. We do this by having them hold the
2204                  * page lock at some point after installing their
2205                  * pte, but before marking the page dirty.
2206                  * Pages are always locked coming in here, so we get
2207                  * the desired exclusion. See mm/memory.c:do_wp_page()
2208                  * for more comments.
2209                  */
2210                 if (TestClearPageDirty(page)) {
2211                         dec_zone_page_state(page, NR_FILE_DIRTY);
2212                         dec_bdi_stat(mapping->backing_dev_info,
2213                                         BDI_RECLAIMABLE);
2214                         return 1;
2215                 }
2216                 return 0;
2217         }
2218         return TestClearPageDirty(page);
2219 }
2220 EXPORT_SYMBOL(clear_page_dirty_for_io);
2221
2222 int test_clear_page_writeback(struct page *page)
2223 {
2224         struct address_space *mapping = page_mapping(page);
2225         int ret;
2226
2227         if (mapping) {
2228                 struct backing_dev_info *bdi = mapping->backing_dev_info;
2229                 unsigned long flags;
2230
2231                 spin_lock_irqsave(&mapping->tree_lock, flags);
2232                 ret = TestClearPageWriteback(page);
2233                 if (ret) {
2234                         radix_tree_tag_clear(&mapping->page_tree,
2235                                                 page_index(page),
2236                                                 PAGECACHE_TAG_WRITEBACK);
2237                         if (bdi_cap_account_writeback(bdi)) {
2238                                 __dec_bdi_stat(bdi, BDI_WRITEBACK);
2239                                 __bdi_writeout_inc(bdi);
2240                         }
2241                 }
2242                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2243         } else {
2244                 ret = TestClearPageWriteback(page);
2245         }
2246         if (ret) {
2247                 dec_zone_page_state(page, NR_WRITEBACK);
2248                 inc_zone_page_state(page, NR_WRITTEN);
2249         }
2250         return ret;
2251 }
2252
2253 int test_set_page_writeback(struct page *page)
2254 {
2255         struct address_space *mapping = page_mapping(page);
2256         int ret;
2257
2258         if (mapping) {
2259                 struct backing_dev_info *bdi = mapping->backing_dev_info;
2260                 unsigned long flags;
2261
2262                 spin_lock_irqsave(&mapping->tree_lock, flags);
2263                 ret = TestSetPageWriteback(page);
2264                 if (!ret) {
2265                         radix_tree_tag_set(&mapping->page_tree,
2266                                                 page_index(page),
2267                                                 PAGECACHE_TAG_WRITEBACK);
2268                         if (bdi_cap_account_writeback(bdi))
2269                                 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2270                 }
2271                 if (!PageDirty(page))
2272                         radix_tree_tag_clear(&mapping->page_tree,
2273                                                 page_index(page),
2274                                                 PAGECACHE_TAG_DIRTY);
2275                 radix_tree_tag_clear(&mapping->page_tree,
2276                                      page_index(page),
2277                                      PAGECACHE_TAG_TOWRITE);
2278                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2279         } else {
2280                 ret = TestSetPageWriteback(page);
2281         }
2282         if (!ret)
2283                 account_page_writeback(page);
2284         return ret;
2285
2286 }
2287 EXPORT_SYMBOL(test_set_page_writeback);
2288
2289 /*
2290  * Return true if any of the pages in the mapping are marked with the
2291  * passed tag.
2292  */
2293 int mapping_tagged(struct address_space *mapping, int tag)
2294 {
2295         return radix_tree_tagged(&mapping->page_tree, tag);
2296 }
2297 EXPORT_SYMBOL(mapping_tagged);
2298
2299 /**
2300  * wait_for_stable_page() - wait for writeback to finish, if necessary.
2301  * @page:       The page to wait on.
2302  *
2303  * This function determines if the given page is related to a backing device
2304  * that requires page contents to be held stable during writeback.  If so, then
2305  * it will wait for any pending writeback to complete.
2306  */
2307 void wait_for_stable_page(struct page *page)
2308 {
2309         struct address_space *mapping = page_mapping(page);
2310         struct backing_dev_info *bdi = mapping->backing_dev_info;
2311
2312         if (!bdi_cap_stable_pages_required(bdi))
2313                 return;
2314 #ifdef CONFIG_NEED_BOUNCE_POOL
2315         if (mapping->host->i_sb->s_flags & MS_SNAP_STABLE)
2316                 return;
2317 #endif /* CONFIG_NEED_BOUNCE_POOL */
2318
2319         wait_on_page_writeback(page);
2320 }
2321 EXPORT_SYMBOL_GPL(wait_for_stable_page);