Merge tag 'armsoc-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/arm/arm-soc
[platform/kernel/linux-rpi.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
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/smp.h>
31 #include <linux/sysctl.h>
32 #include <linux/cpu.h>
33 #include <linux/syscalls.h>
34 #include <linux/buffer_head.h> /* __set_page_dirty_buffers */
35 #include <linux/pagevec.h>
36 #include <linux/timer.h>
37 #include <linux/sched/rt.h>
38 #include <linux/sched/signal.h>
39 #include <linux/mm_inline.h>
40 #include <trace/events/writeback.h>
41
42 #include "internal.h"
43
44 /*
45  * Sleep at most 200ms at a time in balance_dirty_pages().
46  */
47 #define MAX_PAUSE               max(HZ/5, 1)
48
49 /*
50  * Try to keep balance_dirty_pages() call intervals higher than this many pages
51  * by raising pause time to max_pause when falls below it.
52  */
53 #define DIRTY_POLL_THRESH       (128 >> (PAGE_SHIFT - 10))
54
55 /*
56  * Estimate write bandwidth at 200ms intervals.
57  */
58 #define BANDWIDTH_INTERVAL      max(HZ/5, 1)
59
60 #define RATELIMIT_CALC_SHIFT    10
61
62 /*
63  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
64  * will look to see if it needs to force writeback or throttling.
65  */
66 static long ratelimit_pages = 32;
67
68 /* The following parameters are exported via /proc/sys/vm */
69
70 /*
71  * Start background writeback (via writeback threads) at this percentage
72  */
73 int dirty_background_ratio = 10;
74
75 /*
76  * dirty_background_bytes starts at 0 (disabled) so that it is a function of
77  * dirty_background_ratio * the amount of dirtyable memory
78  */
79 unsigned long dirty_background_bytes;
80
81 /*
82  * free highmem will not be subtracted from the total free memory
83  * for calculating free ratios if vm_highmem_is_dirtyable is true
84  */
85 int vm_highmem_is_dirtyable;
86
87 /*
88  * The generator of dirty data starts writeback at this percentage
89  */
90 int vm_dirty_ratio = 20;
91
92 /*
93  * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
94  * vm_dirty_ratio * the amount of dirtyable memory
95  */
96 unsigned long vm_dirty_bytes;
97
98 /*
99  * The interval between `kupdate'-style writebacks
100  */
101 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
102
103 EXPORT_SYMBOL_GPL(dirty_writeback_interval);
104
105 /*
106  * The longest time for which data is allowed to remain dirty
107  */
108 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
109
110 /*
111  * Flag that makes the machine dump writes/reads and block dirtyings.
112  */
113 int block_dump;
114
115 /*
116  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
117  * a full sync is triggered after this time elapses without any disk activity.
118  */
119 int laptop_mode;
120
121 EXPORT_SYMBOL(laptop_mode);
122
123 /* End of sysctl-exported parameters */
124
125 struct wb_domain global_wb_domain;
126
127 /* consolidated parameters for balance_dirty_pages() and its subroutines */
128 struct dirty_throttle_control {
129 #ifdef CONFIG_CGROUP_WRITEBACK
130         struct wb_domain        *dom;
131         struct dirty_throttle_control *gdtc;    /* only set in memcg dtc's */
132 #endif
133         struct bdi_writeback    *wb;
134         struct fprop_local_percpu *wb_completions;
135
136         unsigned long           avail;          /* dirtyable */
137         unsigned long           dirty;          /* file_dirty + write + nfs */
138         unsigned long           thresh;         /* dirty threshold */
139         unsigned long           bg_thresh;      /* dirty background threshold */
140
141         unsigned long           wb_dirty;       /* per-wb counterparts */
142         unsigned long           wb_thresh;
143         unsigned long           wb_bg_thresh;
144
145         unsigned long           pos_ratio;
146 };
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 #ifdef CONFIG_CGROUP_WRITEBACK
156
157 #define GDTC_INIT(__wb)         .wb = (__wb),                           \
158                                 .dom = &global_wb_domain,               \
159                                 .wb_completions = &(__wb)->completions
160
161 #define GDTC_INIT_NO_WB         .dom = &global_wb_domain
162
163 #define MDTC_INIT(__wb, __gdtc) .wb = (__wb),                           \
164                                 .dom = mem_cgroup_wb_domain(__wb),      \
165                                 .wb_completions = &(__wb)->memcg_completions, \
166                                 .gdtc = __gdtc
167
168 static bool mdtc_valid(struct dirty_throttle_control *dtc)
169 {
170         return dtc->dom;
171 }
172
173 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
174 {
175         return dtc->dom;
176 }
177
178 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
179 {
180         return mdtc->gdtc;
181 }
182
183 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
184 {
185         return &wb->memcg_completions;
186 }
187
188 static void wb_min_max_ratio(struct bdi_writeback *wb,
189                              unsigned long *minp, unsigned long *maxp)
190 {
191         unsigned long this_bw = wb->avg_write_bandwidth;
192         unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
193         unsigned long long min = wb->bdi->min_ratio;
194         unsigned long long max = wb->bdi->max_ratio;
195
196         /*
197          * @wb may already be clean by the time control reaches here and
198          * the total may not include its bw.
199          */
200         if (this_bw < tot_bw) {
201                 if (min) {
202                         min *= this_bw;
203                         do_div(min, tot_bw);
204                 }
205                 if (max < 100) {
206                         max *= this_bw;
207                         do_div(max, tot_bw);
208                 }
209         }
210
211         *minp = min;
212         *maxp = max;
213 }
214
215 #else   /* CONFIG_CGROUP_WRITEBACK */
216
217 #define GDTC_INIT(__wb)         .wb = (__wb),                           \
218                                 .wb_completions = &(__wb)->completions
219 #define GDTC_INIT_NO_WB
220 #define MDTC_INIT(__wb, __gdtc)
221
222 static bool mdtc_valid(struct dirty_throttle_control *dtc)
223 {
224         return false;
225 }
226
227 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
228 {
229         return &global_wb_domain;
230 }
231
232 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
233 {
234         return NULL;
235 }
236
237 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
238 {
239         return NULL;
240 }
241
242 static void wb_min_max_ratio(struct bdi_writeback *wb,
243                              unsigned long *minp, unsigned long *maxp)
244 {
245         *minp = wb->bdi->min_ratio;
246         *maxp = wb->bdi->max_ratio;
247 }
248
249 #endif  /* CONFIG_CGROUP_WRITEBACK */
250
251 /*
252  * In a memory zone, there is a certain amount of pages we consider
253  * available for the page cache, which is essentially the number of
254  * free and reclaimable pages, minus some zone reserves to protect
255  * lowmem and the ability to uphold the zone's watermarks without
256  * requiring writeback.
257  *
258  * This number of dirtyable pages is the base value of which the
259  * user-configurable dirty ratio is the effictive number of pages that
260  * are allowed to be actually dirtied.  Per individual zone, or
261  * globally by using the sum of dirtyable pages over all zones.
262  *
263  * Because the user is allowed to specify the dirty limit globally as
264  * absolute number of bytes, calculating the per-zone dirty limit can
265  * require translating the configured limit into a percentage of
266  * global dirtyable memory first.
267  */
268
269 /**
270  * node_dirtyable_memory - number of dirtyable pages in a node
271  * @pgdat: the node
272  *
273  * Returns the node's number of pages potentially available for dirty
274  * page cache.  This is the base value for the per-node dirty limits.
275  */
276 static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
277 {
278         unsigned long nr_pages = 0;
279         int z;
280
281         for (z = 0; z < MAX_NR_ZONES; z++) {
282                 struct zone *zone = pgdat->node_zones + z;
283
284                 if (!populated_zone(zone))
285                         continue;
286
287                 nr_pages += zone_page_state(zone, NR_FREE_PAGES);
288         }
289
290         /*
291          * Pages reserved for the kernel should not be considered
292          * dirtyable, to prevent a situation where reclaim has to
293          * clean pages in order to balance the zones.
294          */
295         nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
296
297         nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
298         nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
299
300         return nr_pages;
301 }
302
303 static unsigned long highmem_dirtyable_memory(unsigned long total)
304 {
305 #ifdef CONFIG_HIGHMEM
306         int node;
307         unsigned long x = 0;
308         int i;
309
310         for_each_node_state(node, N_HIGH_MEMORY) {
311                 for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
312                         struct zone *z;
313                         unsigned long nr_pages;
314
315                         if (!is_highmem_idx(i))
316                                 continue;
317
318                         z = &NODE_DATA(node)->node_zones[i];
319                         if (!populated_zone(z))
320                                 continue;
321
322                         nr_pages = zone_page_state(z, NR_FREE_PAGES);
323                         /* watch for underflows */
324                         nr_pages -= min(nr_pages, high_wmark_pages(z));
325                         nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
326                         nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
327                         x += nr_pages;
328                 }
329         }
330
331         /*
332          * Unreclaimable memory (kernel memory or anonymous memory
333          * without swap) can bring down the dirtyable pages below
334          * the zone's dirty balance reserve and the above calculation
335          * will underflow.  However we still want to add in nodes
336          * which are below threshold (negative values) to get a more
337          * accurate calculation but make sure that the total never
338          * underflows.
339          */
340         if ((long)x < 0)
341                 x = 0;
342
343         /*
344          * Make sure that the number of highmem pages is never larger
345          * than the number of the total dirtyable memory. This can only
346          * occur in very strange VM situations but we want to make sure
347          * that this does not occur.
348          */
349         return min(x, total);
350 #else
351         return 0;
352 #endif
353 }
354
355 /**
356  * global_dirtyable_memory - number of globally dirtyable pages
357  *
358  * Returns the global number of pages potentially available for dirty
359  * page cache.  This is the base value for the global dirty limits.
360  */
361 static unsigned long global_dirtyable_memory(void)
362 {
363         unsigned long x;
364
365         x = global_zone_page_state(NR_FREE_PAGES);
366         /*
367          * Pages reserved for the kernel should not be considered
368          * dirtyable, to prevent a situation where reclaim has to
369          * clean pages in order to balance the zones.
370          */
371         x -= min(x, totalreserve_pages);
372
373         x += global_node_page_state(NR_INACTIVE_FILE);
374         x += global_node_page_state(NR_ACTIVE_FILE);
375
376         if (!vm_highmem_is_dirtyable)
377                 x -= highmem_dirtyable_memory(x);
378
379         return x + 1;   /* Ensure that we never return 0 */
380 }
381
382 /**
383  * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
384  * @dtc: dirty_throttle_control of interest
385  *
386  * Calculate @dtc->thresh and ->bg_thresh considering
387  * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}.  The caller
388  * must ensure that @dtc->avail is set before calling this function.  The
389  * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
390  * real-time tasks.
391  */
392 static void domain_dirty_limits(struct dirty_throttle_control *dtc)
393 {
394         const unsigned long available_memory = dtc->avail;
395         struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
396         unsigned long bytes = vm_dirty_bytes;
397         unsigned long bg_bytes = dirty_background_bytes;
398         /* convert ratios to per-PAGE_SIZE for higher precision */
399         unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
400         unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
401         unsigned long thresh;
402         unsigned long bg_thresh;
403         struct task_struct *tsk;
404
405         /* gdtc is !NULL iff @dtc is for memcg domain */
406         if (gdtc) {
407                 unsigned long global_avail = gdtc->avail;
408
409                 /*
410                  * The byte settings can't be applied directly to memcg
411                  * domains.  Convert them to ratios by scaling against
412                  * globally available memory.  As the ratios are in
413                  * per-PAGE_SIZE, they can be obtained by dividing bytes by
414                  * number of pages.
415                  */
416                 if (bytes)
417                         ratio = min(DIV_ROUND_UP(bytes, global_avail),
418                                     PAGE_SIZE);
419                 if (bg_bytes)
420                         bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
421                                        PAGE_SIZE);
422                 bytes = bg_bytes = 0;
423         }
424
425         if (bytes)
426                 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
427         else
428                 thresh = (ratio * available_memory) / PAGE_SIZE;
429
430         if (bg_bytes)
431                 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
432         else
433                 bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
434
435         if (bg_thresh >= thresh)
436                 bg_thresh = thresh / 2;
437         tsk = current;
438         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
439                 bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
440                 thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
441         }
442         dtc->thresh = thresh;
443         dtc->bg_thresh = bg_thresh;
444
445         /* we should eventually report the domain in the TP */
446         if (!gdtc)
447                 trace_global_dirty_state(bg_thresh, thresh);
448 }
449
450 /**
451  * global_dirty_limits - background-writeback and dirty-throttling thresholds
452  * @pbackground: out parameter for bg_thresh
453  * @pdirty: out parameter for thresh
454  *
455  * Calculate bg_thresh and thresh for global_wb_domain.  See
456  * domain_dirty_limits() for details.
457  */
458 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
459 {
460         struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
461
462         gdtc.avail = global_dirtyable_memory();
463         domain_dirty_limits(&gdtc);
464
465         *pbackground = gdtc.bg_thresh;
466         *pdirty = gdtc.thresh;
467 }
468
469 /**
470  * node_dirty_limit - maximum number of dirty pages allowed in a node
471  * @pgdat: the node
472  *
473  * Returns the maximum number of dirty pages allowed in a node, based
474  * on the node's dirtyable memory.
475  */
476 static unsigned long node_dirty_limit(struct pglist_data *pgdat)
477 {
478         unsigned long node_memory = node_dirtyable_memory(pgdat);
479         struct task_struct *tsk = current;
480         unsigned long dirty;
481
482         if (vm_dirty_bytes)
483                 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
484                         node_memory / global_dirtyable_memory();
485         else
486                 dirty = vm_dirty_ratio * node_memory / 100;
487
488         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
489                 dirty += dirty / 4;
490
491         return dirty;
492 }
493
494 /**
495  * node_dirty_ok - tells whether a node is within its dirty limits
496  * @pgdat: the node to check
497  *
498  * Returns %true when the dirty pages in @pgdat are within the node's
499  * dirty limit, %false if the limit is exceeded.
500  */
501 bool node_dirty_ok(struct pglist_data *pgdat)
502 {
503         unsigned long limit = node_dirty_limit(pgdat);
504         unsigned long nr_pages = 0;
505
506         nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
507         nr_pages += node_page_state(pgdat, NR_UNSTABLE_NFS);
508         nr_pages += node_page_state(pgdat, NR_WRITEBACK);
509
510         return nr_pages <= limit;
511 }
512
513 int dirty_background_ratio_handler(struct ctl_table *table, int write,
514                 void __user *buffer, size_t *lenp,
515                 loff_t *ppos)
516 {
517         int ret;
518
519         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
520         if (ret == 0 && write)
521                 dirty_background_bytes = 0;
522         return ret;
523 }
524
525 int dirty_background_bytes_handler(struct ctl_table *table, int write,
526                 void __user *buffer, size_t *lenp,
527                 loff_t *ppos)
528 {
529         int ret;
530
531         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
532         if (ret == 0 && write)
533                 dirty_background_ratio = 0;
534         return ret;
535 }
536
537 int dirty_ratio_handler(struct ctl_table *table, int write,
538                 void __user *buffer, size_t *lenp,
539                 loff_t *ppos)
540 {
541         int old_ratio = vm_dirty_ratio;
542         int ret;
543
544         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
545         if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
546                 writeback_set_ratelimit();
547                 vm_dirty_bytes = 0;
548         }
549         return ret;
550 }
551
552 int dirty_bytes_handler(struct ctl_table *table, int write,
553                 void __user *buffer, size_t *lenp,
554                 loff_t *ppos)
555 {
556         unsigned long old_bytes = vm_dirty_bytes;
557         int ret;
558
559         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
560         if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
561                 writeback_set_ratelimit();
562                 vm_dirty_ratio = 0;
563         }
564         return ret;
565 }
566
567 static unsigned long wp_next_time(unsigned long cur_time)
568 {
569         cur_time += VM_COMPLETIONS_PERIOD_LEN;
570         /* 0 has a special meaning... */
571         if (!cur_time)
572                 return 1;
573         return cur_time;
574 }
575
576 static void wb_domain_writeout_inc(struct wb_domain *dom,
577                                    struct fprop_local_percpu *completions,
578                                    unsigned int max_prop_frac)
579 {
580         __fprop_inc_percpu_max(&dom->completions, completions,
581                                max_prop_frac);
582         /* First event after period switching was turned off? */
583         if (unlikely(!dom->period_time)) {
584                 /*
585                  * We can race with other __bdi_writeout_inc calls here but
586                  * it does not cause any harm since the resulting time when
587                  * timer will fire and what is in writeout_period_time will be
588                  * roughly the same.
589                  */
590                 dom->period_time = wp_next_time(jiffies);
591                 mod_timer(&dom->period_timer, dom->period_time);
592         }
593 }
594
595 /*
596  * Increment @wb's writeout completion count and the global writeout
597  * completion count. Called from test_clear_page_writeback().
598  */
599 static inline void __wb_writeout_inc(struct bdi_writeback *wb)
600 {
601         struct wb_domain *cgdom;
602
603         inc_wb_stat(wb, WB_WRITTEN);
604         wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
605                                wb->bdi->max_prop_frac);
606
607         cgdom = mem_cgroup_wb_domain(wb);
608         if (cgdom)
609                 wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
610                                        wb->bdi->max_prop_frac);
611 }
612
613 void wb_writeout_inc(struct bdi_writeback *wb)
614 {
615         unsigned long flags;
616
617         local_irq_save(flags);
618         __wb_writeout_inc(wb);
619         local_irq_restore(flags);
620 }
621 EXPORT_SYMBOL_GPL(wb_writeout_inc);
622
623 /*
624  * On idle system, we can be called long after we scheduled because we use
625  * deferred timers so count with missed periods.
626  */
627 static void writeout_period(struct timer_list *t)
628 {
629         struct wb_domain *dom = from_timer(dom, t, period_timer);
630         int miss_periods = (jiffies - dom->period_time) /
631                                                  VM_COMPLETIONS_PERIOD_LEN;
632
633         if (fprop_new_period(&dom->completions, miss_periods + 1)) {
634                 dom->period_time = wp_next_time(dom->period_time +
635                                 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
636                 mod_timer(&dom->period_timer, dom->period_time);
637         } else {
638                 /*
639                  * Aging has zeroed all fractions. Stop wasting CPU on period
640                  * updates.
641                  */
642                 dom->period_time = 0;
643         }
644 }
645
646 int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
647 {
648         memset(dom, 0, sizeof(*dom));
649
650         spin_lock_init(&dom->lock);
651
652         timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
653
654         dom->dirty_limit_tstamp = jiffies;
655
656         return fprop_global_init(&dom->completions, gfp);
657 }
658
659 #ifdef CONFIG_CGROUP_WRITEBACK
660 void wb_domain_exit(struct wb_domain *dom)
661 {
662         del_timer_sync(&dom->period_timer);
663         fprop_global_destroy(&dom->completions);
664 }
665 #endif
666
667 /*
668  * bdi_min_ratio keeps the sum of the minimum dirty shares of all
669  * registered backing devices, which, for obvious reasons, can not
670  * exceed 100%.
671  */
672 static unsigned int bdi_min_ratio;
673
674 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
675 {
676         int ret = 0;
677
678         spin_lock_bh(&bdi_lock);
679         if (min_ratio > bdi->max_ratio) {
680                 ret = -EINVAL;
681         } else {
682                 min_ratio -= bdi->min_ratio;
683                 if (bdi_min_ratio + min_ratio < 100) {
684                         bdi_min_ratio += min_ratio;
685                         bdi->min_ratio += min_ratio;
686                 } else {
687                         ret = -EINVAL;
688                 }
689         }
690         spin_unlock_bh(&bdi_lock);
691
692         return ret;
693 }
694
695 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
696 {
697         int ret = 0;
698
699         if (max_ratio > 100)
700                 return -EINVAL;
701
702         spin_lock_bh(&bdi_lock);
703         if (bdi->min_ratio > max_ratio) {
704                 ret = -EINVAL;
705         } else {
706                 bdi->max_ratio = max_ratio;
707                 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
708         }
709         spin_unlock_bh(&bdi_lock);
710
711         return ret;
712 }
713 EXPORT_SYMBOL(bdi_set_max_ratio);
714
715 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
716                                            unsigned long bg_thresh)
717 {
718         return (thresh + bg_thresh) / 2;
719 }
720
721 static unsigned long hard_dirty_limit(struct wb_domain *dom,
722                                       unsigned long thresh)
723 {
724         return max(thresh, dom->dirty_limit);
725 }
726
727 /*
728  * Memory which can be further allocated to a memcg domain is capped by
729  * system-wide clean memory excluding the amount being used in the domain.
730  */
731 static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
732                             unsigned long filepages, unsigned long headroom)
733 {
734         struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
735         unsigned long clean = filepages - min(filepages, mdtc->dirty);
736         unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
737         unsigned long other_clean = global_clean - min(global_clean, clean);
738
739         mdtc->avail = filepages + min(headroom, other_clean);
740 }
741
742 /**
743  * __wb_calc_thresh - @wb's share of dirty throttling threshold
744  * @dtc: dirty_throttle_context of interest
745  *
746  * Returns @wb's dirty limit in pages. The term "dirty" in the context of
747  * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
748  *
749  * Note that balance_dirty_pages() will only seriously take it as a hard limit
750  * when sleeping max_pause per page is not enough to keep the dirty pages under
751  * control. For example, when the device is completely stalled due to some error
752  * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
753  * In the other normal situations, it acts more gently by throttling the tasks
754  * more (rather than completely block them) when the wb dirty pages go high.
755  *
756  * It allocates high/low dirty limits to fast/slow devices, in order to prevent
757  * - starving fast devices
758  * - piling up dirty pages (that will take long time to sync) on slow devices
759  *
760  * The wb's share of dirty limit will be adapting to its throughput and
761  * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
762  */
763 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
764 {
765         struct wb_domain *dom = dtc_dom(dtc);
766         unsigned long thresh = dtc->thresh;
767         u64 wb_thresh;
768         long numerator, denominator;
769         unsigned long wb_min_ratio, wb_max_ratio;
770
771         /*
772          * Calculate this BDI's share of the thresh ratio.
773          */
774         fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
775                               &numerator, &denominator);
776
777         wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
778         wb_thresh *= numerator;
779         do_div(wb_thresh, denominator);
780
781         wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
782
783         wb_thresh += (thresh * wb_min_ratio) / 100;
784         if (wb_thresh > (thresh * wb_max_ratio) / 100)
785                 wb_thresh = thresh * wb_max_ratio / 100;
786
787         return wb_thresh;
788 }
789
790 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
791 {
792         struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
793                                                .thresh = thresh };
794         return __wb_calc_thresh(&gdtc);
795 }
796
797 /*
798  *                           setpoint - dirty 3
799  *        f(dirty) := 1.0 + (----------------)
800  *                           limit - setpoint
801  *
802  * it's a 3rd order polynomial that subjects to
803  *
804  * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
805  * (2) f(setpoint) = 1.0 => the balance point
806  * (3) f(limit)    = 0   => the hard limit
807  * (4) df/dx      <= 0   => negative feedback control
808  * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
809  *     => fast response on large errors; small oscillation near setpoint
810  */
811 static long long pos_ratio_polynom(unsigned long setpoint,
812                                           unsigned long dirty,
813                                           unsigned long limit)
814 {
815         long long pos_ratio;
816         long x;
817
818         x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
819                       (limit - setpoint) | 1);
820         pos_ratio = x;
821         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
822         pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
823         pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
824
825         return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
826 }
827
828 /*
829  * Dirty position control.
830  *
831  * (o) global/bdi setpoints
832  *
833  * We want the dirty pages be balanced around the global/wb setpoints.
834  * When the number of dirty pages is higher/lower than the setpoint, the
835  * dirty position control ratio (and hence task dirty ratelimit) will be
836  * decreased/increased to bring the dirty pages back to the setpoint.
837  *
838  *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
839  *
840  *     if (dirty < setpoint) scale up   pos_ratio
841  *     if (dirty > setpoint) scale down pos_ratio
842  *
843  *     if (wb_dirty < wb_setpoint) scale up   pos_ratio
844  *     if (wb_dirty > wb_setpoint) scale down pos_ratio
845  *
846  *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
847  *
848  * (o) global control line
849  *
850  *     ^ pos_ratio
851  *     |
852  *     |            |<===== global dirty control scope ======>|
853  * 2.0 .............*
854  *     |            .*
855  *     |            . *
856  *     |            .   *
857  *     |            .     *
858  *     |            .        *
859  *     |            .            *
860  * 1.0 ................................*
861  *     |            .                  .     *
862  *     |            .                  .          *
863  *     |            .                  .              *
864  *     |            .                  .                 *
865  *     |            .                  .                    *
866  *   0 +------------.------------------.----------------------*------------->
867  *           freerun^          setpoint^                 limit^   dirty pages
868  *
869  * (o) wb control line
870  *
871  *     ^ pos_ratio
872  *     |
873  *     |            *
874  *     |              *
875  *     |                *
876  *     |                  *
877  *     |                    * |<=========== span ============>|
878  * 1.0 .......................*
879  *     |                      . *
880  *     |                      .   *
881  *     |                      .     *
882  *     |                      .       *
883  *     |                      .         *
884  *     |                      .           *
885  *     |                      .             *
886  *     |                      .               *
887  *     |                      .                 *
888  *     |                      .                   *
889  *     |                      .                     *
890  * 1/4 ...............................................* * * * * * * * * * * *
891  *     |                      .                         .
892  *     |                      .                           .
893  *     |                      .                             .
894  *   0 +----------------------.-------------------------------.------------->
895  *                wb_setpoint^                    x_intercept^
896  *
897  * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
898  * be smoothly throttled down to normal if it starts high in situations like
899  * - start writing to a slow SD card and a fast disk at the same time. The SD
900  *   card's wb_dirty may rush to many times higher than wb_setpoint.
901  * - the wb dirty thresh drops quickly due to change of JBOD workload
902  */
903 static void wb_position_ratio(struct dirty_throttle_control *dtc)
904 {
905         struct bdi_writeback *wb = dtc->wb;
906         unsigned long write_bw = wb->avg_write_bandwidth;
907         unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
908         unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
909         unsigned long wb_thresh = dtc->wb_thresh;
910         unsigned long x_intercept;
911         unsigned long setpoint;         /* dirty pages' target balance point */
912         unsigned long wb_setpoint;
913         unsigned long span;
914         long long pos_ratio;            /* for scaling up/down the rate limit */
915         long x;
916
917         dtc->pos_ratio = 0;
918
919         if (unlikely(dtc->dirty >= limit))
920                 return;
921
922         /*
923          * global setpoint
924          *
925          * See comment for pos_ratio_polynom().
926          */
927         setpoint = (freerun + limit) / 2;
928         pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
929
930         /*
931          * The strictlimit feature is a tool preventing mistrusted filesystems
932          * from growing a large number of dirty pages before throttling. For
933          * such filesystems balance_dirty_pages always checks wb counters
934          * against wb limits. Even if global "nr_dirty" is under "freerun".
935          * This is especially important for fuse which sets bdi->max_ratio to
936          * 1% by default. Without strictlimit feature, fuse writeback may
937          * consume arbitrary amount of RAM because it is accounted in
938          * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
939          *
940          * Here, in wb_position_ratio(), we calculate pos_ratio based on
941          * two values: wb_dirty and wb_thresh. Let's consider an example:
942          * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
943          * limits are set by default to 10% and 20% (background and throttle).
944          * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
945          * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
946          * about ~6K pages (as the average of background and throttle wb
947          * limits). The 3rd order polynomial will provide positive feedback if
948          * wb_dirty is under wb_setpoint and vice versa.
949          *
950          * Note, that we cannot use global counters in these calculations
951          * because we want to throttle process writing to a strictlimit wb
952          * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
953          * in the example above).
954          */
955         if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
956                 long long wb_pos_ratio;
957
958                 if (dtc->wb_dirty < 8) {
959                         dtc->pos_ratio = min_t(long long, pos_ratio * 2,
960                                            2 << RATELIMIT_CALC_SHIFT);
961                         return;
962                 }
963
964                 if (dtc->wb_dirty >= wb_thresh)
965                         return;
966
967                 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
968                                                     dtc->wb_bg_thresh);
969
970                 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
971                         return;
972
973                 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
974                                                  wb_thresh);
975
976                 /*
977                  * Typically, for strictlimit case, wb_setpoint << setpoint
978                  * and pos_ratio >> wb_pos_ratio. In the other words global
979                  * state ("dirty") is not limiting factor and we have to
980                  * make decision based on wb counters. But there is an
981                  * important case when global pos_ratio should get precedence:
982                  * global limits are exceeded (e.g. due to activities on other
983                  * wb's) while given strictlimit wb is below limit.
984                  *
985                  * "pos_ratio * wb_pos_ratio" would work for the case above,
986                  * but it would look too non-natural for the case of all
987                  * activity in the system coming from a single strictlimit wb
988                  * with bdi->max_ratio == 100%.
989                  *
990                  * Note that min() below somewhat changes the dynamics of the
991                  * control system. Normally, pos_ratio value can be well over 3
992                  * (when globally we are at freerun and wb is well below wb
993                  * setpoint). Now the maximum pos_ratio in the same situation
994                  * is 2. We might want to tweak this if we observe the control
995                  * system is too slow to adapt.
996                  */
997                 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
998                 return;
999         }
1000
1001         /*
1002          * We have computed basic pos_ratio above based on global situation. If
1003          * the wb is over/under its share of dirty pages, we want to scale
1004          * pos_ratio further down/up. That is done by the following mechanism.
1005          */
1006
1007         /*
1008          * wb setpoint
1009          *
1010          *        f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1011          *
1012          *                        x_intercept - wb_dirty
1013          *                     := --------------------------
1014          *                        x_intercept - wb_setpoint
1015          *
1016          * The main wb control line is a linear function that subjects to
1017          *
1018          * (1) f(wb_setpoint) = 1.0
1019          * (2) k = - 1 / (8 * write_bw)  (in single wb case)
1020          *     or equally: x_intercept = wb_setpoint + 8 * write_bw
1021          *
1022          * For single wb case, the dirty pages are observed to fluctuate
1023          * regularly within range
1024          *        [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1025          * for various filesystems, where (2) can yield in a reasonable 12.5%
1026          * fluctuation range for pos_ratio.
1027          *
1028          * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1029          * own size, so move the slope over accordingly and choose a slope that
1030          * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1031          */
1032         if (unlikely(wb_thresh > dtc->thresh))
1033                 wb_thresh = dtc->thresh;
1034         /*
1035          * It's very possible that wb_thresh is close to 0 not because the
1036          * device is slow, but that it has remained inactive for long time.
1037          * Honour such devices a reasonable good (hopefully IO efficient)
1038          * threshold, so that the occasional writes won't be blocked and active
1039          * writes can rampup the threshold quickly.
1040          */
1041         wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1042         /*
1043          * scale global setpoint to wb's:
1044          *      wb_setpoint = setpoint * wb_thresh / thresh
1045          */
1046         x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1047         wb_setpoint = setpoint * (u64)x >> 16;
1048         /*
1049          * Use span=(8*write_bw) in single wb case as indicated by
1050          * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1051          *
1052          *        wb_thresh                    thresh - wb_thresh
1053          * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1054          *         thresh                           thresh
1055          */
1056         span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1057         x_intercept = wb_setpoint + span;
1058
1059         if (dtc->wb_dirty < x_intercept - span / 4) {
1060                 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1061                                       (x_intercept - wb_setpoint) | 1);
1062         } else
1063                 pos_ratio /= 4;
1064
1065         /*
1066          * wb reserve area, safeguard against dirty pool underrun and disk idle
1067          * It may push the desired control point of global dirty pages higher
1068          * than setpoint.
1069          */
1070         x_intercept = wb_thresh / 2;
1071         if (dtc->wb_dirty < x_intercept) {
1072                 if (dtc->wb_dirty > x_intercept / 8)
1073                         pos_ratio = div_u64(pos_ratio * x_intercept,
1074                                             dtc->wb_dirty);
1075                 else
1076                         pos_ratio *= 8;
1077         }
1078
1079         dtc->pos_ratio = pos_ratio;
1080 }
1081
1082 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1083                                       unsigned long elapsed,
1084                                       unsigned long written)
1085 {
1086         const unsigned long period = roundup_pow_of_two(3 * HZ);
1087         unsigned long avg = wb->avg_write_bandwidth;
1088         unsigned long old = wb->write_bandwidth;
1089         u64 bw;
1090
1091         /*
1092          * bw = written * HZ / elapsed
1093          *
1094          *                   bw * elapsed + write_bandwidth * (period - elapsed)
1095          * write_bandwidth = ---------------------------------------------------
1096          *                                          period
1097          *
1098          * @written may have decreased due to account_page_redirty().
1099          * Avoid underflowing @bw calculation.
1100          */
1101         bw = written - min(written, wb->written_stamp);
1102         bw *= HZ;
1103         if (unlikely(elapsed > period)) {
1104                 do_div(bw, elapsed);
1105                 avg = bw;
1106                 goto out;
1107         }
1108         bw += (u64)wb->write_bandwidth * (period - elapsed);
1109         bw >>= ilog2(period);
1110
1111         /*
1112          * one more level of smoothing, for filtering out sudden spikes
1113          */
1114         if (avg > old && old >= (unsigned long)bw)
1115                 avg -= (avg - old) >> 3;
1116
1117         if (avg < old && old <= (unsigned long)bw)
1118                 avg += (old - avg) >> 3;
1119
1120 out:
1121         /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1122         avg = max(avg, 1LU);
1123         if (wb_has_dirty_io(wb)) {
1124                 long delta = avg - wb->avg_write_bandwidth;
1125                 WARN_ON_ONCE(atomic_long_add_return(delta,
1126                                         &wb->bdi->tot_write_bandwidth) <= 0);
1127         }
1128         wb->write_bandwidth = bw;
1129         wb->avg_write_bandwidth = avg;
1130 }
1131
1132 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1133 {
1134         struct wb_domain *dom = dtc_dom(dtc);
1135         unsigned long thresh = dtc->thresh;
1136         unsigned long limit = dom->dirty_limit;
1137
1138         /*
1139          * Follow up in one step.
1140          */
1141         if (limit < thresh) {
1142                 limit = thresh;
1143                 goto update;
1144         }
1145
1146         /*
1147          * Follow down slowly. Use the higher one as the target, because thresh
1148          * may drop below dirty. This is exactly the reason to introduce
1149          * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1150          */
1151         thresh = max(thresh, dtc->dirty);
1152         if (limit > thresh) {
1153                 limit -= (limit - thresh) >> 5;
1154                 goto update;
1155         }
1156         return;
1157 update:
1158         dom->dirty_limit = limit;
1159 }
1160
1161 static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
1162                                     unsigned long now)
1163 {
1164         struct wb_domain *dom = dtc_dom(dtc);
1165
1166         /*
1167          * check locklessly first to optimize away locking for the most time
1168          */
1169         if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1170                 return;
1171
1172         spin_lock(&dom->lock);
1173         if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1174                 update_dirty_limit(dtc);
1175                 dom->dirty_limit_tstamp = now;
1176         }
1177         spin_unlock(&dom->lock);
1178 }
1179
1180 /*
1181  * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1182  *
1183  * Normal wb tasks will be curbed at or below it in long term.
1184  * Obviously it should be around (write_bw / N) when there are N dd tasks.
1185  */
1186 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1187                                       unsigned long dirtied,
1188                                       unsigned long elapsed)
1189 {
1190         struct bdi_writeback *wb = dtc->wb;
1191         unsigned long dirty = dtc->dirty;
1192         unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1193         unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1194         unsigned long setpoint = (freerun + limit) / 2;
1195         unsigned long write_bw = wb->avg_write_bandwidth;
1196         unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1197         unsigned long dirty_rate;
1198         unsigned long task_ratelimit;
1199         unsigned long balanced_dirty_ratelimit;
1200         unsigned long step;
1201         unsigned long x;
1202         unsigned long shift;
1203
1204         /*
1205          * The dirty rate will match the writeout rate in long term, except
1206          * when dirty pages are truncated by userspace or re-dirtied by FS.
1207          */
1208         dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1209
1210         /*
1211          * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1212          */
1213         task_ratelimit = (u64)dirty_ratelimit *
1214                                         dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1215         task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1216
1217         /*
1218          * A linear estimation of the "balanced" throttle rate. The theory is,
1219          * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1220          * dirty_rate will be measured to be (N * task_ratelimit). So the below
1221          * formula will yield the balanced rate limit (write_bw / N).
1222          *
1223          * Note that the expanded form is not a pure rate feedback:
1224          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate)              (1)
1225          * but also takes pos_ratio into account:
1226          *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
1227          *
1228          * (1) is not realistic because pos_ratio also takes part in balancing
1229          * the dirty rate.  Consider the state
1230          *      pos_ratio = 0.5                                              (3)
1231          *      rate = 2 * (write_bw / N)                                    (4)
1232          * If (1) is used, it will stuck in that state! Because each dd will
1233          * be throttled at
1234          *      task_ratelimit = pos_ratio * rate = (write_bw / N)           (5)
1235          * yielding
1236          *      dirty_rate = N * task_ratelimit = write_bw                   (6)
1237          * put (6) into (1) we get
1238          *      rate_(i+1) = rate_(i)                                        (7)
1239          *
1240          * So we end up using (2) to always keep
1241          *      rate_(i+1) ~= (write_bw / N)                                 (8)
1242          * regardless of the value of pos_ratio. As long as (8) is satisfied,
1243          * pos_ratio is able to drive itself to 1.0, which is not only where
1244          * the dirty count meet the setpoint, but also where the slope of
1245          * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1246          */
1247         balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1248                                            dirty_rate | 1);
1249         /*
1250          * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1251          */
1252         if (unlikely(balanced_dirty_ratelimit > write_bw))
1253                 balanced_dirty_ratelimit = write_bw;
1254
1255         /*
1256          * We could safely do this and return immediately:
1257          *
1258          *      wb->dirty_ratelimit = balanced_dirty_ratelimit;
1259          *
1260          * However to get a more stable dirty_ratelimit, the below elaborated
1261          * code makes use of task_ratelimit to filter out singular points and
1262          * limit the step size.
1263          *
1264          * The below code essentially only uses the relative value of
1265          *
1266          *      task_ratelimit - dirty_ratelimit
1267          *      = (pos_ratio - 1) * dirty_ratelimit
1268          *
1269          * which reflects the direction and size of dirty position error.
1270          */
1271
1272         /*
1273          * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1274          * task_ratelimit is on the same side of dirty_ratelimit, too.
1275          * For example, when
1276          * - dirty_ratelimit > balanced_dirty_ratelimit
1277          * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1278          * lowering dirty_ratelimit will help meet both the position and rate
1279          * control targets. Otherwise, don't update dirty_ratelimit if it will
1280          * only help meet the rate target. After all, what the users ultimately
1281          * feel and care are stable dirty rate and small position error.
1282          *
1283          * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1284          * and filter out the singular points of balanced_dirty_ratelimit. Which
1285          * keeps jumping around randomly and can even leap far away at times
1286          * due to the small 200ms estimation period of dirty_rate (we want to
1287          * keep that period small to reduce time lags).
1288          */
1289         step = 0;
1290
1291         /*
1292          * For strictlimit case, calculations above were based on wb counters
1293          * and limits (starting from pos_ratio = wb_position_ratio() and up to
1294          * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1295          * Hence, to calculate "step" properly, we have to use wb_dirty as
1296          * "dirty" and wb_setpoint as "setpoint".
1297          *
1298          * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1299          * it's possible that wb_thresh is close to zero due to inactivity
1300          * of backing device.
1301          */
1302         if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1303                 dirty = dtc->wb_dirty;
1304                 if (dtc->wb_dirty < 8)
1305                         setpoint = dtc->wb_dirty + 1;
1306                 else
1307                         setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1308         }
1309
1310         if (dirty < setpoint) {
1311                 x = min3(wb->balanced_dirty_ratelimit,
1312                          balanced_dirty_ratelimit, task_ratelimit);
1313                 if (dirty_ratelimit < x)
1314                         step = x - dirty_ratelimit;
1315         } else {
1316                 x = max3(wb->balanced_dirty_ratelimit,
1317                          balanced_dirty_ratelimit, task_ratelimit);
1318                 if (dirty_ratelimit > x)
1319                         step = dirty_ratelimit - x;
1320         }
1321
1322         /*
1323          * Don't pursue 100% rate matching. It's impossible since the balanced
1324          * rate itself is constantly fluctuating. So decrease the track speed
1325          * when it gets close to the target. Helps eliminate pointless tremors.
1326          */
1327         shift = dirty_ratelimit / (2 * step + 1);
1328         if (shift < BITS_PER_LONG)
1329                 step = DIV_ROUND_UP(step >> shift, 8);
1330         else
1331                 step = 0;
1332
1333         if (dirty_ratelimit < balanced_dirty_ratelimit)
1334                 dirty_ratelimit += step;
1335         else
1336                 dirty_ratelimit -= step;
1337
1338         wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1339         wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1340
1341         trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1342 }
1343
1344 static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1345                                   struct dirty_throttle_control *mdtc,
1346                                   unsigned long start_time,
1347                                   bool update_ratelimit)
1348 {
1349         struct bdi_writeback *wb = gdtc->wb;
1350         unsigned long now = jiffies;
1351         unsigned long elapsed = now - wb->bw_time_stamp;
1352         unsigned long dirtied;
1353         unsigned long written;
1354
1355         lockdep_assert_held(&wb->list_lock);
1356
1357         /*
1358          * rate-limit, only update once every 200ms.
1359          */
1360         if (elapsed < BANDWIDTH_INTERVAL)
1361                 return;
1362
1363         dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1364         written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1365
1366         /*
1367          * Skip quiet periods when disk bandwidth is under-utilized.
1368          * (at least 1s idle time between two flusher runs)
1369          */
1370         if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1371                 goto snapshot;
1372
1373         if (update_ratelimit) {
1374                 domain_update_bandwidth(gdtc, now);
1375                 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1376
1377                 /*
1378                  * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1379                  * compiler has no way to figure that out.  Help it.
1380                  */
1381                 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1382                         domain_update_bandwidth(mdtc, now);
1383                         wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1384                 }
1385         }
1386         wb_update_write_bandwidth(wb, elapsed, written);
1387
1388 snapshot:
1389         wb->dirtied_stamp = dirtied;
1390         wb->written_stamp = written;
1391         wb->bw_time_stamp = now;
1392 }
1393
1394 void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1395 {
1396         struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1397
1398         __wb_update_bandwidth(&gdtc, NULL, start_time, false);
1399 }
1400
1401 /*
1402  * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1403  * will look to see if it needs to start dirty throttling.
1404  *
1405  * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1406  * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1407  * (the number of pages we may dirty without exceeding the dirty limits).
1408  */
1409 static unsigned long dirty_poll_interval(unsigned long dirty,
1410                                          unsigned long thresh)
1411 {
1412         if (thresh > dirty)
1413                 return 1UL << (ilog2(thresh - dirty) >> 1);
1414
1415         return 1;
1416 }
1417
1418 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1419                                   unsigned long wb_dirty)
1420 {
1421         unsigned long bw = wb->avg_write_bandwidth;
1422         unsigned long t;
1423
1424         /*
1425          * Limit pause time for small memory systems. If sleeping for too long
1426          * time, a small pool of dirty/writeback pages may go empty and disk go
1427          * idle.
1428          *
1429          * 8 serves as the safety ratio.
1430          */
1431         t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1432         t++;
1433
1434         return min_t(unsigned long, t, MAX_PAUSE);
1435 }
1436
1437 static long wb_min_pause(struct bdi_writeback *wb,
1438                          long max_pause,
1439                          unsigned long task_ratelimit,
1440                          unsigned long dirty_ratelimit,
1441                          int *nr_dirtied_pause)
1442 {
1443         long hi = ilog2(wb->avg_write_bandwidth);
1444         long lo = ilog2(wb->dirty_ratelimit);
1445         long t;         /* target pause */
1446         long pause;     /* estimated next pause */
1447         int pages;      /* target nr_dirtied_pause */
1448
1449         /* target for 10ms pause on 1-dd case */
1450         t = max(1, HZ / 100);
1451
1452         /*
1453          * Scale up pause time for concurrent dirtiers in order to reduce CPU
1454          * overheads.
1455          *
1456          * (N * 10ms) on 2^N concurrent tasks.
1457          */
1458         if (hi > lo)
1459                 t += (hi - lo) * (10 * HZ) / 1024;
1460
1461         /*
1462          * This is a bit convoluted. We try to base the next nr_dirtied_pause
1463          * on the much more stable dirty_ratelimit. However the next pause time
1464          * will be computed based on task_ratelimit and the two rate limits may
1465          * depart considerably at some time. Especially if task_ratelimit goes
1466          * below dirty_ratelimit/2 and the target pause is max_pause, the next
1467          * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1468          * result task_ratelimit won't be executed faithfully, which could
1469          * eventually bring down dirty_ratelimit.
1470          *
1471          * We apply two rules to fix it up:
1472          * 1) try to estimate the next pause time and if necessary, use a lower
1473          *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1474          *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1475          * 2) limit the target pause time to max_pause/2, so that the normal
1476          *    small fluctuations of task_ratelimit won't trigger rule (1) and
1477          *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1478          */
1479         t = min(t, 1 + max_pause / 2);
1480         pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1481
1482         /*
1483          * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1484          * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1485          * When the 16 consecutive reads are often interrupted by some dirty
1486          * throttling pause during the async writes, cfq will go into idles
1487          * (deadline is fine). So push nr_dirtied_pause as high as possible
1488          * until reaches DIRTY_POLL_THRESH=32 pages.
1489          */
1490         if (pages < DIRTY_POLL_THRESH) {
1491                 t = max_pause;
1492                 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1493                 if (pages > DIRTY_POLL_THRESH) {
1494                         pages = DIRTY_POLL_THRESH;
1495                         t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1496                 }
1497         }
1498
1499         pause = HZ * pages / (task_ratelimit + 1);
1500         if (pause > max_pause) {
1501                 t = max_pause;
1502                 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1503         }
1504
1505         *nr_dirtied_pause = pages;
1506         /*
1507          * The minimal pause time will normally be half the target pause time.
1508          */
1509         return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1510 }
1511
1512 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1513 {
1514         struct bdi_writeback *wb = dtc->wb;
1515         unsigned long wb_reclaimable;
1516
1517         /*
1518          * wb_thresh is not treated as some limiting factor as
1519          * dirty_thresh, due to reasons
1520          * - in JBOD setup, wb_thresh can fluctuate a lot
1521          * - in a system with HDD and USB key, the USB key may somehow
1522          *   go into state (wb_dirty >> wb_thresh) either because
1523          *   wb_dirty starts high, or because wb_thresh drops low.
1524          *   In this case we don't want to hard throttle the USB key
1525          *   dirtiers for 100 seconds until wb_dirty drops under
1526          *   wb_thresh. Instead the auxiliary wb control line in
1527          *   wb_position_ratio() will let the dirtier task progress
1528          *   at some rate <= (write_bw / 2) for bringing down wb_dirty.
1529          */
1530         dtc->wb_thresh = __wb_calc_thresh(dtc);
1531         dtc->wb_bg_thresh = dtc->thresh ?
1532                 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1533
1534         /*
1535          * In order to avoid the stacked BDI deadlock we need
1536          * to ensure we accurately count the 'dirty' pages when
1537          * the threshold is low.
1538          *
1539          * Otherwise it would be possible to get thresh+n pages
1540          * reported dirty, even though there are thresh-m pages
1541          * actually dirty; with m+n sitting in the percpu
1542          * deltas.
1543          */
1544         if (dtc->wb_thresh < 2 * wb_stat_error()) {
1545                 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1546                 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1547         } else {
1548                 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1549                 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1550         }
1551 }
1552
1553 /*
1554  * balance_dirty_pages() must be called by processes which are generating dirty
1555  * data.  It looks at the number of dirty pages in the machine and will force
1556  * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1557  * If we're over `background_thresh' then the writeback threads are woken to
1558  * perform some writeout.
1559  */
1560 static void balance_dirty_pages(struct bdi_writeback *wb,
1561                                 unsigned long pages_dirtied)
1562 {
1563         struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1564         struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1565         struct dirty_throttle_control * const gdtc = &gdtc_stor;
1566         struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1567                                                      &mdtc_stor : NULL;
1568         struct dirty_throttle_control *sdtc;
1569         unsigned long nr_reclaimable;   /* = file_dirty + unstable_nfs */
1570         long period;
1571         long pause;
1572         long max_pause;
1573         long min_pause;
1574         int nr_dirtied_pause;
1575         bool dirty_exceeded = false;
1576         unsigned long task_ratelimit;
1577         unsigned long dirty_ratelimit;
1578         struct backing_dev_info *bdi = wb->bdi;
1579         bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1580         unsigned long start_time = jiffies;
1581
1582         for (;;) {
1583                 unsigned long now = jiffies;
1584                 unsigned long dirty, thresh, bg_thresh;
1585                 unsigned long m_dirty = 0;      /* stop bogus uninit warnings */
1586                 unsigned long m_thresh = 0;
1587                 unsigned long m_bg_thresh = 0;
1588
1589                 /*
1590                  * Unstable writes are a feature of certain networked
1591                  * filesystems (i.e. NFS) in which data may have been
1592                  * written to the server's write cache, but has not yet
1593                  * been flushed to permanent storage.
1594                  */
1595                 nr_reclaimable = global_node_page_state(NR_FILE_DIRTY) +
1596                                         global_node_page_state(NR_UNSTABLE_NFS);
1597                 gdtc->avail = global_dirtyable_memory();
1598                 gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1599
1600                 domain_dirty_limits(gdtc);
1601
1602                 if (unlikely(strictlimit)) {
1603                         wb_dirty_limits(gdtc);
1604
1605                         dirty = gdtc->wb_dirty;
1606                         thresh = gdtc->wb_thresh;
1607                         bg_thresh = gdtc->wb_bg_thresh;
1608                 } else {
1609                         dirty = gdtc->dirty;
1610                         thresh = gdtc->thresh;
1611                         bg_thresh = gdtc->bg_thresh;
1612                 }
1613
1614                 if (mdtc) {
1615                         unsigned long filepages, headroom, writeback;
1616
1617                         /*
1618                          * If @wb belongs to !root memcg, repeat the same
1619                          * basic calculations for the memcg domain.
1620                          */
1621                         mem_cgroup_wb_stats(wb, &filepages, &headroom,
1622                                             &mdtc->dirty, &writeback);
1623                         mdtc->dirty += writeback;
1624                         mdtc_calc_avail(mdtc, filepages, headroom);
1625
1626                         domain_dirty_limits(mdtc);
1627
1628                         if (unlikely(strictlimit)) {
1629                                 wb_dirty_limits(mdtc);
1630                                 m_dirty = mdtc->wb_dirty;
1631                                 m_thresh = mdtc->wb_thresh;
1632                                 m_bg_thresh = mdtc->wb_bg_thresh;
1633                         } else {
1634                                 m_dirty = mdtc->dirty;
1635                                 m_thresh = mdtc->thresh;
1636                                 m_bg_thresh = mdtc->bg_thresh;
1637                         }
1638                 }
1639
1640                 /*
1641                  * Throttle it only when the background writeback cannot
1642                  * catch-up. This avoids (excessively) small writeouts
1643                  * when the wb limits are ramping up in case of !strictlimit.
1644                  *
1645                  * In strictlimit case make decision based on the wb counters
1646                  * and limits. Small writeouts when the wb limits are ramping
1647                  * up are the price we consciously pay for strictlimit-ing.
1648                  *
1649                  * If memcg domain is in effect, @dirty should be under
1650                  * both global and memcg freerun ceilings.
1651                  */
1652                 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1653                     (!mdtc ||
1654                      m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1655                         unsigned long intv = dirty_poll_interval(dirty, thresh);
1656                         unsigned long m_intv = ULONG_MAX;
1657
1658                         current->dirty_paused_when = now;
1659                         current->nr_dirtied = 0;
1660                         if (mdtc)
1661                                 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1662                         current->nr_dirtied_pause = min(intv, m_intv);
1663                         break;
1664                 }
1665
1666                 if (unlikely(!writeback_in_progress(wb)))
1667                         wb_start_background_writeback(wb);
1668
1669                 /*
1670                  * Calculate global domain's pos_ratio and select the
1671                  * global dtc by default.
1672                  */
1673                 if (!strictlimit)
1674                         wb_dirty_limits(gdtc);
1675
1676                 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1677                         ((gdtc->dirty > gdtc->thresh) || strictlimit);
1678
1679                 wb_position_ratio(gdtc);
1680                 sdtc = gdtc;
1681
1682                 if (mdtc) {
1683                         /*
1684                          * If memcg domain is in effect, calculate its
1685                          * pos_ratio.  @wb should satisfy constraints from
1686                          * both global and memcg domains.  Choose the one
1687                          * w/ lower pos_ratio.
1688                          */
1689                         if (!strictlimit)
1690                                 wb_dirty_limits(mdtc);
1691
1692                         dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1693                                 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1694
1695                         wb_position_ratio(mdtc);
1696                         if (mdtc->pos_ratio < gdtc->pos_ratio)
1697                                 sdtc = mdtc;
1698                 }
1699
1700                 if (dirty_exceeded && !wb->dirty_exceeded)
1701                         wb->dirty_exceeded = 1;
1702
1703                 if (time_is_before_jiffies(wb->bw_time_stamp +
1704                                            BANDWIDTH_INTERVAL)) {
1705                         spin_lock(&wb->list_lock);
1706                         __wb_update_bandwidth(gdtc, mdtc, start_time, true);
1707                         spin_unlock(&wb->list_lock);
1708                 }
1709
1710                 /* throttle according to the chosen dtc */
1711                 dirty_ratelimit = wb->dirty_ratelimit;
1712                 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1713                                                         RATELIMIT_CALC_SHIFT;
1714                 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1715                 min_pause = wb_min_pause(wb, max_pause,
1716                                          task_ratelimit, dirty_ratelimit,
1717                                          &nr_dirtied_pause);
1718
1719                 if (unlikely(task_ratelimit == 0)) {
1720                         period = max_pause;
1721                         pause = max_pause;
1722                         goto pause;
1723                 }
1724                 period = HZ * pages_dirtied / task_ratelimit;
1725                 pause = period;
1726                 if (current->dirty_paused_when)
1727                         pause -= now - current->dirty_paused_when;
1728                 /*
1729                  * For less than 1s think time (ext3/4 may block the dirtier
1730                  * for up to 800ms from time to time on 1-HDD; so does xfs,
1731                  * however at much less frequency), try to compensate it in
1732                  * future periods by updating the virtual time; otherwise just
1733                  * do a reset, as it may be a light dirtier.
1734                  */
1735                 if (pause < min_pause) {
1736                         trace_balance_dirty_pages(wb,
1737                                                   sdtc->thresh,
1738                                                   sdtc->bg_thresh,
1739                                                   sdtc->dirty,
1740                                                   sdtc->wb_thresh,
1741                                                   sdtc->wb_dirty,
1742                                                   dirty_ratelimit,
1743                                                   task_ratelimit,
1744                                                   pages_dirtied,
1745                                                   period,
1746                                                   min(pause, 0L),
1747                                                   start_time);
1748                         if (pause < -HZ) {
1749                                 current->dirty_paused_when = now;
1750                                 current->nr_dirtied = 0;
1751                         } else if (period) {
1752                                 current->dirty_paused_when += period;
1753                                 current->nr_dirtied = 0;
1754                         } else if (current->nr_dirtied_pause <= pages_dirtied)
1755                                 current->nr_dirtied_pause += pages_dirtied;
1756                         break;
1757                 }
1758                 if (unlikely(pause > max_pause)) {
1759                         /* for occasional dropped task_ratelimit */
1760                         now += min(pause - max_pause, max_pause);
1761                         pause = max_pause;
1762                 }
1763
1764 pause:
1765                 trace_balance_dirty_pages(wb,
1766                                           sdtc->thresh,
1767                                           sdtc->bg_thresh,
1768                                           sdtc->dirty,
1769                                           sdtc->wb_thresh,
1770                                           sdtc->wb_dirty,
1771                                           dirty_ratelimit,
1772                                           task_ratelimit,
1773                                           pages_dirtied,
1774                                           period,
1775                                           pause,
1776                                           start_time);
1777                 __set_current_state(TASK_KILLABLE);
1778                 wb->dirty_sleep = now;
1779                 io_schedule_timeout(pause);
1780
1781                 current->dirty_paused_when = now + pause;
1782                 current->nr_dirtied = 0;
1783                 current->nr_dirtied_pause = nr_dirtied_pause;
1784
1785                 /*
1786                  * This is typically equal to (dirty < thresh) and can also
1787                  * keep "1000+ dd on a slow USB stick" under control.
1788                  */
1789                 if (task_ratelimit)
1790                         break;
1791
1792                 /*
1793                  * In the case of an unresponding NFS server and the NFS dirty
1794                  * pages exceeds dirty_thresh, give the other good wb's a pipe
1795                  * to go through, so that tasks on them still remain responsive.
1796                  *
1797                  * In theory 1 page is enough to keep the consumer-producer
1798                  * pipe going: the flusher cleans 1 page => the task dirties 1
1799                  * more page. However wb_dirty has accounting errors.  So use
1800                  * the larger and more IO friendly wb_stat_error.
1801                  */
1802                 if (sdtc->wb_dirty <= wb_stat_error())
1803                         break;
1804
1805                 if (fatal_signal_pending(current))
1806                         break;
1807         }
1808
1809         if (!dirty_exceeded && wb->dirty_exceeded)
1810                 wb->dirty_exceeded = 0;
1811
1812         if (writeback_in_progress(wb))
1813                 return;
1814
1815         /*
1816          * In laptop mode, we wait until hitting the higher threshold before
1817          * starting background writeout, and then write out all the way down
1818          * to the lower threshold.  So slow writers cause minimal disk activity.
1819          *
1820          * In normal mode, we start background writeout at the lower
1821          * background_thresh, to keep the amount of dirty memory low.
1822          */
1823         if (laptop_mode)
1824                 return;
1825
1826         if (nr_reclaimable > gdtc->bg_thresh)
1827                 wb_start_background_writeback(wb);
1828 }
1829
1830 static DEFINE_PER_CPU(int, bdp_ratelimits);
1831
1832 /*
1833  * Normal tasks are throttled by
1834  *      loop {
1835  *              dirty tsk->nr_dirtied_pause pages;
1836  *              take a snap in balance_dirty_pages();
1837  *      }
1838  * However there is a worst case. If every task exit immediately when dirtied
1839  * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1840  * called to throttle the page dirties. The solution is to save the not yet
1841  * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1842  * randomly into the running tasks. This works well for the above worst case,
1843  * as the new task will pick up and accumulate the old task's leaked dirty
1844  * count and eventually get throttled.
1845  */
1846 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1847
1848 /**
1849  * balance_dirty_pages_ratelimited - balance dirty memory state
1850  * @mapping: address_space which was dirtied
1851  *
1852  * Processes which are dirtying memory should call in here once for each page
1853  * which was newly dirtied.  The function will periodically check the system's
1854  * dirty state and will initiate writeback if needed.
1855  *
1856  * On really big machines, get_writeback_state is expensive, so try to avoid
1857  * calling it too often (ratelimiting).  But once we're over the dirty memory
1858  * limit we decrease the ratelimiting by a lot, to prevent individual processes
1859  * from overshooting the limit by (ratelimit_pages) each.
1860  */
1861 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1862 {
1863         struct inode *inode = mapping->host;
1864         struct backing_dev_info *bdi = inode_to_bdi(inode);
1865         struct bdi_writeback *wb = NULL;
1866         int ratelimit;
1867         int *p;
1868
1869         if (!bdi_cap_account_dirty(bdi))
1870                 return;
1871
1872         if (inode_cgwb_enabled(inode))
1873                 wb = wb_get_create_current(bdi, GFP_KERNEL);
1874         if (!wb)
1875                 wb = &bdi->wb;
1876
1877         ratelimit = current->nr_dirtied_pause;
1878         if (wb->dirty_exceeded)
1879                 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1880
1881         preempt_disable();
1882         /*
1883          * This prevents one CPU to accumulate too many dirtied pages without
1884          * calling into balance_dirty_pages(), which can happen when there are
1885          * 1000+ tasks, all of them start dirtying pages at exactly the same
1886          * time, hence all honoured too large initial task->nr_dirtied_pause.
1887          */
1888         p =  this_cpu_ptr(&bdp_ratelimits);
1889         if (unlikely(current->nr_dirtied >= ratelimit))
1890                 *p = 0;
1891         else if (unlikely(*p >= ratelimit_pages)) {
1892                 *p = 0;
1893                 ratelimit = 0;
1894         }
1895         /*
1896          * Pick up the dirtied pages by the exited tasks. This avoids lots of
1897          * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1898          * the dirty throttling and livelock other long-run dirtiers.
1899          */
1900         p = this_cpu_ptr(&dirty_throttle_leaks);
1901         if (*p > 0 && current->nr_dirtied < ratelimit) {
1902                 unsigned long nr_pages_dirtied;
1903                 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1904                 *p -= nr_pages_dirtied;
1905                 current->nr_dirtied += nr_pages_dirtied;
1906         }
1907         preempt_enable();
1908
1909         if (unlikely(current->nr_dirtied >= ratelimit))
1910                 balance_dirty_pages(wb, current->nr_dirtied);
1911
1912         wb_put(wb);
1913 }
1914 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1915
1916 /**
1917  * wb_over_bg_thresh - does @wb need to be written back?
1918  * @wb: bdi_writeback of interest
1919  *
1920  * Determines whether background writeback should keep writing @wb or it's
1921  * clean enough.  Returns %true if writeback should continue.
1922  */
1923 bool wb_over_bg_thresh(struct bdi_writeback *wb)
1924 {
1925         struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1926         struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1927         struct dirty_throttle_control * const gdtc = &gdtc_stor;
1928         struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1929                                                      &mdtc_stor : NULL;
1930
1931         /*
1932          * Similar to balance_dirty_pages() but ignores pages being written
1933          * as we're trying to decide whether to put more under writeback.
1934          */
1935         gdtc->avail = global_dirtyable_memory();
1936         gdtc->dirty = global_node_page_state(NR_FILE_DIRTY) +
1937                       global_node_page_state(NR_UNSTABLE_NFS);
1938         domain_dirty_limits(gdtc);
1939
1940         if (gdtc->dirty > gdtc->bg_thresh)
1941                 return true;
1942
1943         if (wb_stat(wb, WB_RECLAIMABLE) >
1944             wb_calc_thresh(gdtc->wb, gdtc->bg_thresh))
1945                 return true;
1946
1947         if (mdtc) {
1948                 unsigned long filepages, headroom, writeback;
1949
1950                 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1951                                     &writeback);
1952                 mdtc_calc_avail(mdtc, filepages, headroom);
1953                 domain_dirty_limits(mdtc);      /* ditto, ignore writeback */
1954
1955                 if (mdtc->dirty > mdtc->bg_thresh)
1956                         return true;
1957
1958                 if (wb_stat(wb, WB_RECLAIMABLE) >
1959                     wb_calc_thresh(mdtc->wb, mdtc->bg_thresh))
1960                         return true;
1961         }
1962
1963         return false;
1964 }
1965
1966 /*
1967  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1968  */
1969 int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1970         void __user *buffer, size_t *length, loff_t *ppos)
1971 {
1972         unsigned int old_interval = dirty_writeback_interval;
1973         int ret;
1974
1975         ret = proc_dointvec(table, write, buffer, length, ppos);
1976
1977         /*
1978          * Writing 0 to dirty_writeback_interval will disable periodic writeback
1979          * and a different non-zero value will wakeup the writeback threads.
1980          * wb_wakeup_delayed() would be more appropriate, but it's a pain to
1981          * iterate over all bdis and wbs.
1982          * The reason we do this is to make the change take effect immediately.
1983          */
1984         if (!ret && write && dirty_writeback_interval &&
1985                 dirty_writeback_interval != old_interval)
1986                 wakeup_flusher_threads(WB_REASON_PERIODIC);
1987
1988         return ret;
1989 }
1990
1991 #ifdef CONFIG_BLOCK
1992 void laptop_mode_timer_fn(struct timer_list *t)
1993 {
1994         struct backing_dev_info *backing_dev_info =
1995                 from_timer(backing_dev_info, t, laptop_mode_wb_timer);
1996
1997         wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
1998 }
1999
2000 /*
2001  * We've spun up the disk and we're in laptop mode: schedule writeback
2002  * of all dirty data a few seconds from now.  If the flush is already scheduled
2003  * then push it back - the user is still using the disk.
2004  */
2005 void laptop_io_completion(struct backing_dev_info *info)
2006 {
2007         mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2008 }
2009
2010 /*
2011  * We're in laptop mode and we've just synced. The sync's writes will have
2012  * caused another writeback to be scheduled by laptop_io_completion.
2013  * Nothing needs to be written back anymore, so we unschedule the writeback.
2014  */
2015 void laptop_sync_completion(void)
2016 {
2017         struct backing_dev_info *bdi;
2018
2019         rcu_read_lock();
2020
2021         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2022                 del_timer(&bdi->laptop_mode_wb_timer);
2023
2024         rcu_read_unlock();
2025 }
2026 #endif
2027
2028 /*
2029  * If ratelimit_pages is too high then we can get into dirty-data overload
2030  * if a large number of processes all perform writes at the same time.
2031  * If it is too low then SMP machines will call the (expensive)
2032  * get_writeback_state too often.
2033  *
2034  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2035  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2036  * thresholds.
2037  */
2038
2039 void writeback_set_ratelimit(void)
2040 {
2041         struct wb_domain *dom = &global_wb_domain;
2042         unsigned long background_thresh;
2043         unsigned long dirty_thresh;
2044
2045         global_dirty_limits(&background_thresh, &dirty_thresh);
2046         dom->dirty_limit = dirty_thresh;
2047         ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2048         if (ratelimit_pages < 16)
2049                 ratelimit_pages = 16;
2050 }
2051
2052 static int page_writeback_cpu_online(unsigned int cpu)
2053 {
2054         writeback_set_ratelimit();
2055         return 0;
2056 }
2057
2058 /*
2059  * Called early on to tune the page writeback dirty limits.
2060  *
2061  * We used to scale dirty pages according to how total memory
2062  * related to pages that could be allocated for buffers (by
2063  * comparing nr_free_buffer_pages() to vm_total_pages.
2064  *
2065  * However, that was when we used "dirty_ratio" to scale with
2066  * all memory, and we don't do that any more. "dirty_ratio"
2067  * is now applied to total non-HIGHPAGE memory (by subtracting
2068  * totalhigh_pages from vm_total_pages), and as such we can't
2069  * get into the old insane situation any more where we had
2070  * large amounts of dirty pages compared to a small amount of
2071  * non-HIGHMEM memory.
2072  *
2073  * But we might still want to scale the dirty_ratio by how
2074  * much memory the box has..
2075  */
2076 void __init page_writeback_init(void)
2077 {
2078         BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2079
2080         cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2081                           page_writeback_cpu_online, NULL);
2082         cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2083                           page_writeback_cpu_online);
2084 }
2085
2086 /**
2087  * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2088  * @mapping: address space structure to write
2089  * @start: starting page index
2090  * @end: ending page index (inclusive)
2091  *
2092  * This function scans the page range from @start to @end (inclusive) and tags
2093  * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2094  * that write_cache_pages (or whoever calls this function) will then use
2095  * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
2096  * used to avoid livelocking of writeback by a process steadily creating new
2097  * dirty pages in the file (thus it is important for this function to be quick
2098  * so that it can tag pages faster than a dirtying process can create them).
2099  */
2100 /*
2101  * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce the i_pages lock
2102  * latency.
2103  */
2104 void tag_pages_for_writeback(struct address_space *mapping,
2105                              pgoff_t start, pgoff_t end)
2106 {
2107 #define WRITEBACK_TAG_BATCH 4096
2108         unsigned long tagged = 0;
2109         struct radix_tree_iter iter;
2110         void **slot;
2111
2112         xa_lock_irq(&mapping->i_pages);
2113         radix_tree_for_each_tagged(slot, &mapping->i_pages, &iter, start,
2114                                                         PAGECACHE_TAG_DIRTY) {
2115                 if (iter.index > end)
2116                         break;
2117                 radix_tree_iter_tag_set(&mapping->i_pages, &iter,
2118                                                         PAGECACHE_TAG_TOWRITE);
2119                 tagged++;
2120                 if ((tagged % WRITEBACK_TAG_BATCH) != 0)
2121                         continue;
2122                 slot = radix_tree_iter_resume(slot, &iter);
2123                 xa_unlock_irq(&mapping->i_pages);
2124                 cond_resched();
2125                 xa_lock_irq(&mapping->i_pages);
2126         }
2127         xa_unlock_irq(&mapping->i_pages);
2128 }
2129 EXPORT_SYMBOL(tag_pages_for_writeback);
2130
2131 /**
2132  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2133  * @mapping: address space structure to write
2134  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2135  * @writepage: function called for each page
2136  * @data: data passed to writepage function
2137  *
2138  * If a page is already under I/O, write_cache_pages() skips it, even
2139  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2140  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2141  * and msync() need to guarantee that all the data which was dirty at the time
2142  * the call was made get new I/O started against them.  If wbc->sync_mode is
2143  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2144  * existing IO to complete.
2145  *
2146  * To avoid livelocks (when other process dirties new pages), we first tag
2147  * pages which should be written back with TOWRITE tag and only then start
2148  * writing them. For data-integrity sync we have to be careful so that we do
2149  * not miss some pages (e.g., because some other process has cleared TOWRITE
2150  * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2151  * by the process clearing the DIRTY tag (and submitting the page for IO).
2152  */
2153 int write_cache_pages(struct address_space *mapping,
2154                       struct writeback_control *wbc, writepage_t writepage,
2155                       void *data)
2156 {
2157         int ret = 0;
2158         int done = 0;
2159         struct pagevec pvec;
2160         int nr_pages;
2161         pgoff_t uninitialized_var(writeback_index);
2162         pgoff_t index;
2163         pgoff_t end;            /* Inclusive */
2164         pgoff_t done_index;
2165         int cycled;
2166         int range_whole = 0;
2167         int tag;
2168
2169         pagevec_init(&pvec);
2170         if (wbc->range_cyclic) {
2171                 writeback_index = mapping->writeback_index; /* prev offset */
2172                 index = writeback_index;
2173                 if (index == 0)
2174                         cycled = 1;
2175                 else
2176                         cycled = 0;
2177                 end = -1;
2178         } else {
2179                 index = wbc->range_start >> PAGE_SHIFT;
2180                 end = wbc->range_end >> PAGE_SHIFT;
2181                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2182                         range_whole = 1;
2183                 cycled = 1; /* ignore range_cyclic tests */
2184         }
2185         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2186                 tag = PAGECACHE_TAG_TOWRITE;
2187         else
2188                 tag = PAGECACHE_TAG_DIRTY;
2189 retry:
2190         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2191                 tag_pages_for_writeback(mapping, index, end);
2192         done_index = index;
2193         while (!done && (index <= end)) {
2194                 int i;
2195
2196                 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2197                                 tag);
2198                 if (nr_pages == 0)
2199                         break;
2200
2201                 for (i = 0; i < nr_pages; i++) {
2202                         struct page *page = pvec.pages[i];
2203
2204                         done_index = page->index;
2205
2206                         lock_page(page);
2207
2208                         /*
2209                          * Page truncated or invalidated. We can freely skip it
2210                          * then, even for data integrity operations: the page
2211                          * has disappeared concurrently, so there could be no
2212                          * real expectation of this data interity operation
2213                          * even if there is now a new, dirty page at the same
2214                          * pagecache address.
2215                          */
2216                         if (unlikely(page->mapping != mapping)) {
2217 continue_unlock:
2218                                 unlock_page(page);
2219                                 continue;
2220                         }
2221
2222                         if (!PageDirty(page)) {
2223                                 /* someone wrote it for us */
2224                                 goto continue_unlock;
2225                         }
2226
2227                         if (PageWriteback(page)) {
2228                                 if (wbc->sync_mode != WB_SYNC_NONE)
2229                                         wait_on_page_writeback(page);
2230                                 else
2231                                         goto continue_unlock;
2232                         }
2233
2234                         BUG_ON(PageWriteback(page));
2235                         if (!clear_page_dirty_for_io(page))
2236                                 goto continue_unlock;
2237
2238                         trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2239                         ret = (*writepage)(page, wbc, data);
2240                         if (unlikely(ret)) {
2241                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2242                                         unlock_page(page);
2243                                         ret = 0;
2244                                 } else {
2245                                         /*
2246                                          * done_index is set past this page,
2247                                          * so media errors will not choke
2248                                          * background writeout for the entire
2249                                          * file. This has consequences for
2250                                          * range_cyclic semantics (ie. it may
2251                                          * not be suitable for data integrity
2252                                          * writeout).
2253                                          */
2254                                         done_index = page->index + 1;
2255                                         done = 1;
2256                                         break;
2257                                 }
2258                         }
2259
2260                         /*
2261                          * We stop writing back only if we are not doing
2262                          * integrity sync. In case of integrity sync we have to
2263                          * keep going until we have written all the pages
2264                          * we tagged for writeback prior to entering this loop.
2265                          */
2266                         if (--wbc->nr_to_write <= 0 &&
2267                             wbc->sync_mode == WB_SYNC_NONE) {
2268                                 done = 1;
2269                                 break;
2270                         }
2271                 }
2272                 pagevec_release(&pvec);
2273                 cond_resched();
2274         }
2275         if (!cycled && !done) {
2276                 /*
2277                  * range_cyclic:
2278                  * We hit the last page and there is more work to be done: wrap
2279                  * back to the start of the file
2280                  */
2281                 cycled = 1;
2282                 index = 0;
2283                 end = writeback_index - 1;
2284                 goto retry;
2285         }
2286         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2287                 mapping->writeback_index = done_index;
2288
2289         return ret;
2290 }
2291 EXPORT_SYMBOL(write_cache_pages);
2292
2293 /*
2294  * Function used by generic_writepages to call the real writepage
2295  * function and set the mapping flags on error
2296  */
2297 static int __writepage(struct page *page, struct writeback_control *wbc,
2298                        void *data)
2299 {
2300         struct address_space *mapping = data;
2301         int ret = mapping->a_ops->writepage(page, wbc);
2302         mapping_set_error(mapping, ret);
2303         return ret;
2304 }
2305
2306 /**
2307  * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2308  * @mapping: address space structure to write
2309  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2310  *
2311  * This is a library function, which implements the writepages()
2312  * address_space_operation.
2313  */
2314 int generic_writepages(struct address_space *mapping,
2315                        struct writeback_control *wbc)
2316 {
2317         struct blk_plug plug;
2318         int ret;
2319
2320         /* deal with chardevs and other special file */
2321         if (!mapping->a_ops->writepage)
2322                 return 0;
2323
2324         blk_start_plug(&plug);
2325         ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2326         blk_finish_plug(&plug);
2327         return ret;
2328 }
2329
2330 EXPORT_SYMBOL(generic_writepages);
2331
2332 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2333 {
2334         int ret;
2335
2336         if (wbc->nr_to_write <= 0)
2337                 return 0;
2338         while (1) {
2339                 if (mapping->a_ops->writepages)
2340                         ret = mapping->a_ops->writepages(mapping, wbc);
2341                 else
2342                         ret = generic_writepages(mapping, wbc);
2343                 if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2344                         break;
2345                 cond_resched();
2346                 congestion_wait(BLK_RW_ASYNC, HZ/50);
2347         }
2348         return ret;
2349 }
2350
2351 /**
2352  * write_one_page - write out a single page and wait on I/O
2353  * @page: the page to write
2354  *
2355  * The page must be locked by the caller and will be unlocked upon return.
2356  *
2357  * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2358  * function returns.
2359  */
2360 int write_one_page(struct page *page)
2361 {
2362         struct address_space *mapping = page->mapping;
2363         int ret = 0;
2364         struct writeback_control wbc = {
2365                 .sync_mode = WB_SYNC_ALL,
2366                 .nr_to_write = 1,
2367         };
2368
2369         BUG_ON(!PageLocked(page));
2370
2371         wait_on_page_writeback(page);
2372
2373         if (clear_page_dirty_for_io(page)) {
2374                 get_page(page);
2375                 ret = mapping->a_ops->writepage(page, &wbc);
2376                 if (ret == 0)
2377                         wait_on_page_writeback(page);
2378                 put_page(page);
2379         } else {
2380                 unlock_page(page);
2381         }
2382
2383         if (!ret)
2384                 ret = filemap_check_errors(mapping);
2385         return ret;
2386 }
2387 EXPORT_SYMBOL(write_one_page);
2388
2389 /*
2390  * For address_spaces which do not use buffers nor write back.
2391  */
2392 int __set_page_dirty_no_writeback(struct page *page)
2393 {
2394         if (!PageDirty(page))
2395                 return !TestSetPageDirty(page);
2396         return 0;
2397 }
2398
2399 /*
2400  * Helper function for set_page_dirty family.
2401  *
2402  * Caller must hold lock_page_memcg().
2403  *
2404  * NOTE: This relies on being atomic wrt interrupts.
2405  */
2406 void account_page_dirtied(struct page *page, struct address_space *mapping)
2407 {
2408         struct inode *inode = mapping->host;
2409
2410         trace_writeback_dirty_page(page, mapping);
2411
2412         if (mapping_cap_account_dirty(mapping)) {
2413                 struct bdi_writeback *wb;
2414
2415                 inode_attach_wb(inode, page);
2416                 wb = inode_to_wb(inode);
2417
2418                 __inc_lruvec_page_state(page, NR_FILE_DIRTY);
2419                 __inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2420                 __inc_node_page_state(page, NR_DIRTIED);
2421                 inc_wb_stat(wb, WB_RECLAIMABLE);
2422                 inc_wb_stat(wb, WB_DIRTIED);
2423                 task_io_account_write(PAGE_SIZE);
2424                 current->nr_dirtied++;
2425                 this_cpu_inc(bdp_ratelimits);
2426         }
2427 }
2428 EXPORT_SYMBOL(account_page_dirtied);
2429
2430 /*
2431  * Helper function for deaccounting dirty page without writeback.
2432  *
2433  * Caller must hold lock_page_memcg().
2434  */
2435 void account_page_cleaned(struct page *page, struct address_space *mapping,
2436                           struct bdi_writeback *wb)
2437 {
2438         if (mapping_cap_account_dirty(mapping)) {
2439                 dec_lruvec_page_state(page, NR_FILE_DIRTY);
2440                 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2441                 dec_wb_stat(wb, WB_RECLAIMABLE);
2442                 task_io_account_cancelled_write(PAGE_SIZE);
2443         }
2444 }
2445
2446 /*
2447  * For address_spaces which do not use buffers.  Just tag the page as dirty in
2448  * its radix tree.
2449  *
2450  * This is also used when a single buffer is being dirtied: we want to set the
2451  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
2452  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2453  *
2454  * The caller must ensure this doesn't race with truncation.  Most will simply
2455  * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2456  * the pte lock held, which also locks out truncation.
2457  */
2458 int __set_page_dirty_nobuffers(struct page *page)
2459 {
2460         lock_page_memcg(page);
2461         if (!TestSetPageDirty(page)) {
2462                 struct address_space *mapping = page_mapping(page);
2463                 unsigned long flags;
2464
2465                 if (!mapping) {
2466                         unlock_page_memcg(page);
2467                         return 1;
2468                 }
2469
2470                 xa_lock_irqsave(&mapping->i_pages, flags);
2471                 BUG_ON(page_mapping(page) != mapping);
2472                 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2473                 account_page_dirtied(page, mapping);
2474                 radix_tree_tag_set(&mapping->i_pages, page_index(page),
2475                                    PAGECACHE_TAG_DIRTY);
2476                 xa_unlock_irqrestore(&mapping->i_pages, flags);
2477                 unlock_page_memcg(page);
2478
2479                 if (mapping->host) {
2480                         /* !PageAnon && !swapper_space */
2481                         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2482                 }
2483                 return 1;
2484         }
2485         unlock_page_memcg(page);
2486         return 0;
2487 }
2488 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2489
2490 /*
2491  * Call this whenever redirtying a page, to de-account the dirty counters
2492  * (NR_DIRTIED, WB_DIRTIED, tsk->nr_dirtied), so that they match the written
2493  * counters (NR_WRITTEN, WB_WRITTEN) in long term. The mismatches will lead to
2494  * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2495  * control.
2496  */
2497 void account_page_redirty(struct page *page)
2498 {
2499         struct address_space *mapping = page->mapping;
2500
2501         if (mapping && mapping_cap_account_dirty(mapping)) {
2502                 struct inode *inode = mapping->host;
2503                 struct bdi_writeback *wb;
2504                 struct wb_lock_cookie cookie = {};
2505
2506                 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2507                 current->nr_dirtied--;
2508                 dec_node_page_state(page, NR_DIRTIED);
2509                 dec_wb_stat(wb, WB_DIRTIED);
2510                 unlocked_inode_to_wb_end(inode, &cookie);
2511         }
2512 }
2513 EXPORT_SYMBOL(account_page_redirty);
2514
2515 /*
2516  * When a writepage implementation decides that it doesn't want to write this
2517  * page for some reason, it should redirty the locked page via
2518  * redirty_page_for_writepage() and it should then unlock the page and return 0
2519  */
2520 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2521 {
2522         int ret;
2523
2524         wbc->pages_skipped++;
2525         ret = __set_page_dirty_nobuffers(page);
2526         account_page_redirty(page);
2527         return ret;
2528 }
2529 EXPORT_SYMBOL(redirty_page_for_writepage);
2530
2531 /*
2532  * Dirty a page.
2533  *
2534  * For pages with a mapping this should be done under the page lock
2535  * for the benefit of asynchronous memory errors who prefer a consistent
2536  * dirty state. This rule can be broken in some special cases,
2537  * but should be better not to.
2538  *
2539  * If the mapping doesn't provide a set_page_dirty a_op, then
2540  * just fall through and assume that it wants buffer_heads.
2541  */
2542 int set_page_dirty(struct page *page)
2543 {
2544         struct address_space *mapping = page_mapping(page);
2545
2546         page = compound_head(page);
2547         if (likely(mapping)) {
2548                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2549                 /*
2550                  * readahead/lru_deactivate_page could remain
2551                  * PG_readahead/PG_reclaim due to race with end_page_writeback
2552                  * About readahead, if the page is written, the flags would be
2553                  * reset. So no problem.
2554                  * About lru_deactivate_page, if the page is redirty, the flag
2555                  * will be reset. So no problem. but if the page is used by readahead
2556                  * it will confuse readahead and make it restart the size rampup
2557                  * process. But it's a trivial problem.
2558                  */
2559                 if (PageReclaim(page))
2560                         ClearPageReclaim(page);
2561 #ifdef CONFIG_BLOCK
2562                 if (!spd)
2563                         spd = __set_page_dirty_buffers;
2564 #endif
2565                 return (*spd)(page);
2566         }
2567         if (!PageDirty(page)) {
2568                 if (!TestSetPageDirty(page))
2569                         return 1;
2570         }
2571         return 0;
2572 }
2573 EXPORT_SYMBOL(set_page_dirty);
2574
2575 /*
2576  * set_page_dirty() is racy if the caller has no reference against
2577  * page->mapping->host, and if the page is unlocked.  This is because another
2578  * CPU could truncate the page off the mapping and then free the mapping.
2579  *
2580  * Usually, the page _is_ locked, or the caller is a user-space process which
2581  * holds a reference on the inode by having an open file.
2582  *
2583  * In other cases, the page should be locked before running set_page_dirty().
2584  */
2585 int set_page_dirty_lock(struct page *page)
2586 {
2587         int ret;
2588
2589         lock_page(page);
2590         ret = set_page_dirty(page);
2591         unlock_page(page);
2592         return ret;
2593 }
2594 EXPORT_SYMBOL(set_page_dirty_lock);
2595
2596 /*
2597  * This cancels just the dirty bit on the kernel page itself, it does NOT
2598  * actually remove dirty bits on any mmap's that may be around. It also
2599  * leaves the page tagged dirty, so any sync activity will still find it on
2600  * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2601  * look at the dirty bits in the VM.
2602  *
2603  * Doing this should *normally* only ever be done when a page is truncated,
2604  * and is not actually mapped anywhere at all. However, fs/buffer.c does
2605  * this when it notices that somebody has cleaned out all the buffers on a
2606  * page without actually doing it through the VM. Can you say "ext3 is
2607  * horribly ugly"? Thought you could.
2608  */
2609 void __cancel_dirty_page(struct page *page)
2610 {
2611         struct address_space *mapping = page_mapping(page);
2612
2613         if (mapping_cap_account_dirty(mapping)) {
2614                 struct inode *inode = mapping->host;
2615                 struct bdi_writeback *wb;
2616                 struct wb_lock_cookie cookie = {};
2617
2618                 lock_page_memcg(page);
2619                 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2620
2621                 if (TestClearPageDirty(page))
2622                         account_page_cleaned(page, mapping, wb);
2623
2624                 unlocked_inode_to_wb_end(inode, &cookie);
2625                 unlock_page_memcg(page);
2626         } else {
2627                 ClearPageDirty(page);
2628         }
2629 }
2630 EXPORT_SYMBOL(__cancel_dirty_page);
2631
2632 /*
2633  * Clear a page's dirty flag, while caring for dirty memory accounting.
2634  * Returns true if the page was previously dirty.
2635  *
2636  * This is for preparing to put the page under writeout.  We leave the page
2637  * tagged as dirty in the radix tree so that a concurrent write-for-sync
2638  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
2639  * implementation will run either set_page_writeback() or set_page_dirty(),
2640  * at which stage we bring the page's dirty flag and radix-tree dirty tag
2641  * back into sync.
2642  *
2643  * This incoherency between the page's dirty flag and radix-tree tag is
2644  * unfortunate, but it only exists while the page is locked.
2645  */
2646 int clear_page_dirty_for_io(struct page *page)
2647 {
2648         struct address_space *mapping = page_mapping(page);
2649         int ret = 0;
2650
2651         BUG_ON(!PageLocked(page));
2652
2653         if (mapping && mapping_cap_account_dirty(mapping)) {
2654                 struct inode *inode = mapping->host;
2655                 struct bdi_writeback *wb;
2656                 struct wb_lock_cookie cookie = {};
2657
2658                 /*
2659                  * Yes, Virginia, this is indeed insane.
2660                  *
2661                  * We use this sequence to make sure that
2662                  *  (a) we account for dirty stats properly
2663                  *  (b) we tell the low-level filesystem to
2664                  *      mark the whole page dirty if it was
2665                  *      dirty in a pagetable. Only to then
2666                  *  (c) clean the page again and return 1 to
2667                  *      cause the writeback.
2668                  *
2669                  * This way we avoid all nasty races with the
2670                  * dirty bit in multiple places and clearing
2671                  * them concurrently from different threads.
2672                  *
2673                  * Note! Normally the "set_page_dirty(page)"
2674                  * has no effect on the actual dirty bit - since
2675                  * that will already usually be set. But we
2676                  * need the side effects, and it can help us
2677                  * avoid races.
2678                  *
2679                  * We basically use the page "master dirty bit"
2680                  * as a serialization point for all the different
2681                  * threads doing their things.
2682                  */
2683                 if (page_mkclean(page))
2684                         set_page_dirty(page);
2685                 /*
2686                  * We carefully synchronise fault handlers against
2687                  * installing a dirty pte and marking the page dirty
2688                  * at this point.  We do this by having them hold the
2689                  * page lock while dirtying the page, and pages are
2690                  * always locked coming in here, so we get the desired
2691                  * exclusion.
2692                  */
2693                 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2694                 if (TestClearPageDirty(page)) {
2695                         dec_lruvec_page_state(page, NR_FILE_DIRTY);
2696                         dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2697                         dec_wb_stat(wb, WB_RECLAIMABLE);
2698                         ret = 1;
2699                 }
2700                 unlocked_inode_to_wb_end(inode, &cookie);
2701                 return ret;
2702         }
2703         return TestClearPageDirty(page);
2704 }
2705 EXPORT_SYMBOL(clear_page_dirty_for_io);
2706
2707 int test_clear_page_writeback(struct page *page)
2708 {
2709         struct address_space *mapping = page_mapping(page);
2710         struct mem_cgroup *memcg;
2711         struct lruvec *lruvec;
2712         int ret;
2713
2714         memcg = lock_page_memcg(page);
2715         lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
2716         if (mapping && mapping_use_writeback_tags(mapping)) {
2717                 struct inode *inode = mapping->host;
2718                 struct backing_dev_info *bdi = inode_to_bdi(inode);
2719                 unsigned long flags;
2720
2721                 xa_lock_irqsave(&mapping->i_pages, flags);
2722                 ret = TestClearPageWriteback(page);
2723                 if (ret) {
2724                         radix_tree_tag_clear(&mapping->i_pages, page_index(page),
2725                                                 PAGECACHE_TAG_WRITEBACK);
2726                         if (bdi_cap_account_writeback(bdi)) {
2727                                 struct bdi_writeback *wb = inode_to_wb(inode);
2728
2729                                 dec_wb_stat(wb, WB_WRITEBACK);
2730                                 __wb_writeout_inc(wb);
2731                         }
2732                 }
2733
2734                 if (mapping->host && !mapping_tagged(mapping,
2735                                                      PAGECACHE_TAG_WRITEBACK))
2736                         sb_clear_inode_writeback(mapping->host);
2737
2738                 xa_unlock_irqrestore(&mapping->i_pages, flags);
2739         } else {
2740                 ret = TestClearPageWriteback(page);
2741         }
2742         /*
2743          * NOTE: Page might be free now! Writeback doesn't hold a page
2744          * reference on its own, it relies on truncation to wait for
2745          * the clearing of PG_writeback. The below can only access
2746          * page state that is static across allocation cycles.
2747          */
2748         if (ret) {
2749                 dec_lruvec_state(lruvec, NR_WRITEBACK);
2750                 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2751                 inc_node_page_state(page, NR_WRITTEN);
2752         }
2753         __unlock_page_memcg(memcg);
2754         return ret;
2755 }
2756
2757 int __test_set_page_writeback(struct page *page, bool keep_write)
2758 {
2759         struct address_space *mapping = page_mapping(page);
2760         int ret;
2761
2762         lock_page_memcg(page);
2763         if (mapping && mapping_use_writeback_tags(mapping)) {
2764                 struct inode *inode = mapping->host;
2765                 struct backing_dev_info *bdi = inode_to_bdi(inode);
2766                 unsigned long flags;
2767
2768                 xa_lock_irqsave(&mapping->i_pages, flags);
2769                 ret = TestSetPageWriteback(page);
2770                 if (!ret) {
2771                         bool on_wblist;
2772
2773                         on_wblist = mapping_tagged(mapping,
2774                                                    PAGECACHE_TAG_WRITEBACK);
2775
2776                         radix_tree_tag_set(&mapping->i_pages, page_index(page),
2777                                                 PAGECACHE_TAG_WRITEBACK);
2778                         if (bdi_cap_account_writeback(bdi))
2779                                 inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2780
2781                         /*
2782                          * We can come through here when swapping anonymous
2783                          * pages, so we don't necessarily have an inode to track
2784                          * for sync.
2785                          */
2786                         if (mapping->host && !on_wblist)
2787                                 sb_mark_inode_writeback(mapping->host);
2788                 }
2789                 if (!PageDirty(page))
2790                         radix_tree_tag_clear(&mapping->i_pages, page_index(page),
2791                                                 PAGECACHE_TAG_DIRTY);
2792                 if (!keep_write)
2793                         radix_tree_tag_clear(&mapping->i_pages, page_index(page),
2794                                                 PAGECACHE_TAG_TOWRITE);
2795                 xa_unlock_irqrestore(&mapping->i_pages, flags);
2796         } else {
2797                 ret = TestSetPageWriteback(page);
2798         }
2799         if (!ret) {
2800                 inc_lruvec_page_state(page, NR_WRITEBACK);
2801                 inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2802         }
2803         unlock_page_memcg(page);
2804         return ret;
2805
2806 }
2807 EXPORT_SYMBOL(__test_set_page_writeback);
2808
2809 /*
2810  * Return true if any of the pages in the mapping are marked with the
2811  * passed tag.
2812  */
2813 int mapping_tagged(struct address_space *mapping, int tag)
2814 {
2815         return radix_tree_tagged(&mapping->i_pages, tag);
2816 }
2817 EXPORT_SYMBOL(mapping_tagged);
2818
2819 /**
2820  * wait_for_stable_page() - wait for writeback to finish, if necessary.
2821  * @page:       The page to wait on.
2822  *
2823  * This function determines if the given page is related to a backing device
2824  * that requires page contents to be held stable during writeback.  If so, then
2825  * it will wait for any pending writeback to complete.
2826  */
2827 void wait_for_stable_page(struct page *page)
2828 {
2829         if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2830                 wait_on_page_writeback(page);
2831 }
2832 EXPORT_SYMBOL_GPL(wait_for_stable_page);