Loading...
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/notifier.h>
31#include <linux/smp.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/syscalls.h>
35#include <linux/buffer_head.h> /* __set_page_dirty_buffers */
36#include <linux/pagevec.h>
37#include <linux/timer.h>
38#include <linux/sched/rt.h>
39#include <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 */
66static 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 */
73int 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 */
79unsigned 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 */
85int vm_highmem_is_dirtyable;
86
87/*
88 * The generator of dirty data starts writeback at this percentage
89 */
90int 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 */
96unsigned long vm_dirty_bytes;
97
98/*
99 * The interval between `kupdate'-style writebacks
100 */
101unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
102
103EXPORT_SYMBOL_GPL(dirty_writeback_interval);
104
105/*
106 * The longest time for which data is allowed to remain dirty
107 */
108unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
109
110/*
111 * Flag that makes the machine dump writes/reads and block dirtyings.
112 */
113int 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 */
119int laptop_mode;
120
121EXPORT_SYMBOL(laptop_mode);
122
123/* End of sysctl-exported parameters */
124
125struct wb_domain global_wb_domain;
126
127/* consolidated parameters for balance_dirty_pages() and its subroutines */
128struct 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
168static bool mdtc_valid(struct dirty_throttle_control *dtc)
169{
170 return dtc->dom;
171}
172
173static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
174{
175 return dtc->dom;
176}
177
178static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
179{
180 return mdtc->gdtc;
181}
182
183static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
184{
185 return &wb->memcg_completions;
186}
187
188static 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
222static bool mdtc_valid(struct dirty_throttle_control *dtc)
223{
224 return false;
225}
226
227static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
228{
229 return &global_wb_domain;
230}
231
232static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
233{
234 return NULL;
235}
236
237static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
238{
239 return NULL;
240}
241
242static 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 * zone_dirtyable_memory - number of dirtyable pages in a zone
271 * @zone: the zone
272 *
273 * Returns the zone's number of pages potentially available for dirty
274 * page cache. This is the base value for the per-zone dirty limits.
275 */
276static unsigned long zone_dirtyable_memory(struct zone *zone)
277{
278 unsigned long nr_pages;
279
280 nr_pages = zone_page_state(zone, NR_FREE_PAGES);
281 /*
282 * Pages reserved for the kernel should not be considered
283 * dirtyable, to prevent a situation where reclaim has to
284 * clean pages in order to balance the zones.
285 */
286 nr_pages -= min(nr_pages, zone->totalreserve_pages);
287
288 nr_pages += zone_page_state(zone, NR_INACTIVE_FILE);
289 nr_pages += zone_page_state(zone, NR_ACTIVE_FILE);
290
291 return nr_pages;
292}
293
294static unsigned long highmem_dirtyable_memory(unsigned long total)
295{
296#ifdef CONFIG_HIGHMEM
297 int node;
298 unsigned long x = 0;
299
300 for_each_node_state(node, N_HIGH_MEMORY) {
301 struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
302
303 x += zone_dirtyable_memory(z);
304 }
305 /*
306 * Unreclaimable memory (kernel memory or anonymous memory
307 * without swap) can bring down the dirtyable pages below
308 * the zone's dirty balance reserve and the above calculation
309 * will underflow. However we still want to add in nodes
310 * which are below threshold (negative values) to get a more
311 * accurate calculation but make sure that the total never
312 * underflows.
313 */
314 if ((long)x < 0)
315 x = 0;
316
317 /*
318 * Make sure that the number of highmem pages is never larger
319 * than the number of the total dirtyable memory. This can only
320 * occur in very strange VM situations but we want to make sure
321 * that this does not occur.
322 */
323 return min(x, total);
324#else
325 return 0;
326#endif
327}
328
329/**
330 * global_dirtyable_memory - number of globally dirtyable pages
331 *
332 * Returns the global number of pages potentially available for dirty
333 * page cache. This is the base value for the global dirty limits.
334 */
335static unsigned long global_dirtyable_memory(void)
336{
337 unsigned long x;
338
339 x = global_page_state(NR_FREE_PAGES);
340 /*
341 * Pages reserved for the kernel should not be considered
342 * dirtyable, to prevent a situation where reclaim has to
343 * clean pages in order to balance the zones.
344 */
345 x -= min(x, totalreserve_pages);
346
347 x += global_page_state(NR_INACTIVE_FILE);
348 x += global_page_state(NR_ACTIVE_FILE);
349
350 if (!vm_highmem_is_dirtyable)
351 x -= highmem_dirtyable_memory(x);
352
353 return x + 1; /* Ensure that we never return 0 */
354}
355
356/**
357 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
358 * @dtc: dirty_throttle_control of interest
359 *
360 * Calculate @dtc->thresh and ->bg_thresh considering
361 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
362 * must ensure that @dtc->avail is set before calling this function. The
363 * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
364 * real-time tasks.
365 */
366static void domain_dirty_limits(struct dirty_throttle_control *dtc)
367{
368 const unsigned long available_memory = dtc->avail;
369 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
370 unsigned long bytes = vm_dirty_bytes;
371 unsigned long bg_bytes = dirty_background_bytes;
372 unsigned long ratio = vm_dirty_ratio;
373 unsigned long bg_ratio = dirty_background_ratio;
374 unsigned long thresh;
375 unsigned long bg_thresh;
376 struct task_struct *tsk;
377
378 /* gdtc is !NULL iff @dtc is for memcg domain */
379 if (gdtc) {
380 unsigned long global_avail = gdtc->avail;
381
382 /*
383 * The byte settings can't be applied directly to memcg
384 * domains. Convert them to ratios by scaling against
385 * globally available memory.
386 */
387 if (bytes)
388 ratio = min(DIV_ROUND_UP(bytes, PAGE_SIZE) * 100 /
389 global_avail, 100UL);
390 if (bg_bytes)
391 bg_ratio = min(DIV_ROUND_UP(bg_bytes, PAGE_SIZE) * 100 /
392 global_avail, 100UL);
393 bytes = bg_bytes = 0;
394 }
395
396 if (bytes)
397 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
398 else
399 thresh = (ratio * available_memory) / 100;
400
401 if (bg_bytes)
402 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
403 else
404 bg_thresh = (bg_ratio * available_memory) / 100;
405
406 if (bg_thresh >= thresh)
407 bg_thresh = thresh / 2;
408 tsk = current;
409 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
410 bg_thresh += bg_thresh / 4;
411 thresh += thresh / 4;
412 }
413 dtc->thresh = thresh;
414 dtc->bg_thresh = bg_thresh;
415
416 /* we should eventually report the domain in the TP */
417 if (!gdtc)
418 trace_global_dirty_state(bg_thresh, thresh);
419}
420
421/**
422 * global_dirty_limits - background-writeback and dirty-throttling thresholds
423 * @pbackground: out parameter for bg_thresh
424 * @pdirty: out parameter for thresh
425 *
426 * Calculate bg_thresh and thresh for global_wb_domain. See
427 * domain_dirty_limits() for details.
428 */
429void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
430{
431 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
432
433 gdtc.avail = global_dirtyable_memory();
434 domain_dirty_limits(&gdtc);
435
436 *pbackground = gdtc.bg_thresh;
437 *pdirty = gdtc.thresh;
438}
439
440/**
441 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
442 * @zone: the zone
443 *
444 * Returns the maximum number of dirty pages allowed in a zone, based
445 * on the zone's dirtyable memory.
446 */
447static unsigned long zone_dirty_limit(struct zone *zone)
448{
449 unsigned long zone_memory = zone_dirtyable_memory(zone);
450 struct task_struct *tsk = current;
451 unsigned long dirty;
452
453 if (vm_dirty_bytes)
454 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
455 zone_memory / global_dirtyable_memory();
456 else
457 dirty = vm_dirty_ratio * zone_memory / 100;
458
459 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
460 dirty += dirty / 4;
461
462 return dirty;
463}
464
465/**
466 * zone_dirty_ok - tells whether a zone is within its dirty limits
467 * @zone: the zone to check
468 *
469 * Returns %true when the dirty pages in @zone are within the zone's
470 * dirty limit, %false if the limit is exceeded.
471 */
472bool zone_dirty_ok(struct zone *zone)
473{
474 unsigned long limit = zone_dirty_limit(zone);
475
476 return zone_page_state(zone, NR_FILE_DIRTY) +
477 zone_page_state(zone, NR_UNSTABLE_NFS) +
478 zone_page_state(zone, NR_WRITEBACK) <= limit;
479}
480
481int dirty_background_ratio_handler(struct ctl_table *table, int write,
482 void __user *buffer, size_t *lenp,
483 loff_t *ppos)
484{
485 int ret;
486
487 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
488 if (ret == 0 && write)
489 dirty_background_bytes = 0;
490 return ret;
491}
492
493int dirty_background_bytes_handler(struct ctl_table *table, int write,
494 void __user *buffer, size_t *lenp,
495 loff_t *ppos)
496{
497 int ret;
498
499 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
500 if (ret == 0 && write)
501 dirty_background_ratio = 0;
502 return ret;
503}
504
505int dirty_ratio_handler(struct ctl_table *table, int write,
506 void __user *buffer, size_t *lenp,
507 loff_t *ppos)
508{
509 int old_ratio = vm_dirty_ratio;
510 int ret;
511
512 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
513 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
514 writeback_set_ratelimit();
515 vm_dirty_bytes = 0;
516 }
517 return ret;
518}
519
520int dirty_bytes_handler(struct ctl_table *table, int write,
521 void __user *buffer, size_t *lenp,
522 loff_t *ppos)
523{
524 unsigned long old_bytes = vm_dirty_bytes;
525 int ret;
526
527 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
528 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
529 writeback_set_ratelimit();
530 vm_dirty_ratio = 0;
531 }
532 return ret;
533}
534
535static unsigned long wp_next_time(unsigned long cur_time)
536{
537 cur_time += VM_COMPLETIONS_PERIOD_LEN;
538 /* 0 has a special meaning... */
539 if (!cur_time)
540 return 1;
541 return cur_time;
542}
543
544static void wb_domain_writeout_inc(struct wb_domain *dom,
545 struct fprop_local_percpu *completions,
546 unsigned int max_prop_frac)
547{
548 __fprop_inc_percpu_max(&dom->completions, completions,
549 max_prop_frac);
550 /* First event after period switching was turned off? */
551 if (!unlikely(dom->period_time)) {
552 /*
553 * We can race with other __bdi_writeout_inc calls here but
554 * it does not cause any harm since the resulting time when
555 * timer will fire and what is in writeout_period_time will be
556 * roughly the same.
557 */
558 dom->period_time = wp_next_time(jiffies);
559 mod_timer(&dom->period_timer, dom->period_time);
560 }
561}
562
563/*
564 * Increment @wb's writeout completion count and the global writeout
565 * completion count. Called from test_clear_page_writeback().
566 */
567static inline void __wb_writeout_inc(struct bdi_writeback *wb)
568{
569 struct wb_domain *cgdom;
570
571 __inc_wb_stat(wb, WB_WRITTEN);
572 wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
573 wb->bdi->max_prop_frac);
574
575 cgdom = mem_cgroup_wb_domain(wb);
576 if (cgdom)
577 wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
578 wb->bdi->max_prop_frac);
579}
580
581void wb_writeout_inc(struct bdi_writeback *wb)
582{
583 unsigned long flags;
584
585 local_irq_save(flags);
586 __wb_writeout_inc(wb);
587 local_irq_restore(flags);
588}
589EXPORT_SYMBOL_GPL(wb_writeout_inc);
590
591/*
592 * On idle system, we can be called long after we scheduled because we use
593 * deferred timers so count with missed periods.
594 */
595static void writeout_period(unsigned long t)
596{
597 struct wb_domain *dom = (void *)t;
598 int miss_periods = (jiffies - dom->period_time) /
599 VM_COMPLETIONS_PERIOD_LEN;
600
601 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
602 dom->period_time = wp_next_time(dom->period_time +
603 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
604 mod_timer(&dom->period_timer, dom->period_time);
605 } else {
606 /*
607 * Aging has zeroed all fractions. Stop wasting CPU on period
608 * updates.
609 */
610 dom->period_time = 0;
611 }
612}
613
614int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
615{
616 memset(dom, 0, sizeof(*dom));
617
618 spin_lock_init(&dom->lock);
619
620 init_timer_deferrable(&dom->period_timer);
621 dom->period_timer.function = writeout_period;
622 dom->period_timer.data = (unsigned long)dom;
623
624 dom->dirty_limit_tstamp = jiffies;
625
626 return fprop_global_init(&dom->completions, gfp);
627}
628
629#ifdef CONFIG_CGROUP_WRITEBACK
630void wb_domain_exit(struct wb_domain *dom)
631{
632 del_timer_sync(&dom->period_timer);
633 fprop_global_destroy(&dom->completions);
634}
635#endif
636
637/*
638 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
639 * registered backing devices, which, for obvious reasons, can not
640 * exceed 100%.
641 */
642static unsigned int bdi_min_ratio;
643
644int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
645{
646 int ret = 0;
647
648 spin_lock_bh(&bdi_lock);
649 if (min_ratio > bdi->max_ratio) {
650 ret = -EINVAL;
651 } else {
652 min_ratio -= bdi->min_ratio;
653 if (bdi_min_ratio + min_ratio < 100) {
654 bdi_min_ratio += min_ratio;
655 bdi->min_ratio += min_ratio;
656 } else {
657 ret = -EINVAL;
658 }
659 }
660 spin_unlock_bh(&bdi_lock);
661
662 return ret;
663}
664
665int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
666{
667 int ret = 0;
668
669 if (max_ratio > 100)
670 return -EINVAL;
671
672 spin_lock_bh(&bdi_lock);
673 if (bdi->min_ratio > max_ratio) {
674 ret = -EINVAL;
675 } else {
676 bdi->max_ratio = max_ratio;
677 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
678 }
679 spin_unlock_bh(&bdi_lock);
680
681 return ret;
682}
683EXPORT_SYMBOL(bdi_set_max_ratio);
684
685static unsigned long dirty_freerun_ceiling(unsigned long thresh,
686 unsigned long bg_thresh)
687{
688 return (thresh + bg_thresh) / 2;
689}
690
691static unsigned long hard_dirty_limit(struct wb_domain *dom,
692 unsigned long thresh)
693{
694 return max(thresh, dom->dirty_limit);
695}
696
697/*
698 * Memory which can be further allocated to a memcg domain is capped by
699 * system-wide clean memory excluding the amount being used in the domain.
700 */
701static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
702 unsigned long filepages, unsigned long headroom)
703{
704 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
705 unsigned long clean = filepages - min(filepages, mdtc->dirty);
706 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
707 unsigned long other_clean = global_clean - min(global_clean, clean);
708
709 mdtc->avail = filepages + min(headroom, other_clean);
710}
711
712/**
713 * __wb_calc_thresh - @wb's share of dirty throttling threshold
714 * @dtc: dirty_throttle_context of interest
715 *
716 * Returns @wb's dirty limit in pages. The term "dirty" in the context of
717 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
718 *
719 * Note that balance_dirty_pages() will only seriously take it as a hard limit
720 * when sleeping max_pause per page is not enough to keep the dirty pages under
721 * control. For example, when the device is completely stalled due to some error
722 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
723 * In the other normal situations, it acts more gently by throttling the tasks
724 * more (rather than completely block them) when the wb dirty pages go high.
725 *
726 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
727 * - starving fast devices
728 * - piling up dirty pages (that will take long time to sync) on slow devices
729 *
730 * The wb's share of dirty limit will be adapting to its throughput and
731 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
732 */
733static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
734{
735 struct wb_domain *dom = dtc_dom(dtc);
736 unsigned long thresh = dtc->thresh;
737 u64 wb_thresh;
738 long numerator, denominator;
739 unsigned long wb_min_ratio, wb_max_ratio;
740
741 /*
742 * Calculate this BDI's share of the thresh ratio.
743 */
744 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
745 &numerator, &denominator);
746
747 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
748 wb_thresh *= numerator;
749 do_div(wb_thresh, denominator);
750
751 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
752
753 wb_thresh += (thresh * wb_min_ratio) / 100;
754 if (wb_thresh > (thresh * wb_max_ratio) / 100)
755 wb_thresh = thresh * wb_max_ratio / 100;
756
757 return wb_thresh;
758}
759
760unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
761{
762 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
763 .thresh = thresh };
764 return __wb_calc_thresh(&gdtc);
765}
766
767/*
768 * setpoint - dirty 3
769 * f(dirty) := 1.0 + (----------------)
770 * limit - setpoint
771 *
772 * it's a 3rd order polynomial that subjects to
773 *
774 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
775 * (2) f(setpoint) = 1.0 => the balance point
776 * (3) f(limit) = 0 => the hard limit
777 * (4) df/dx <= 0 => negative feedback control
778 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
779 * => fast response on large errors; small oscillation near setpoint
780 */
781static long long pos_ratio_polynom(unsigned long setpoint,
782 unsigned long dirty,
783 unsigned long limit)
784{
785 long long pos_ratio;
786 long x;
787
788 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
789 (limit - setpoint) | 1);
790 pos_ratio = x;
791 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
792 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
793 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
794
795 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
796}
797
798/*
799 * Dirty position control.
800 *
801 * (o) global/bdi setpoints
802 *
803 * We want the dirty pages be balanced around the global/wb setpoints.
804 * When the number of dirty pages is higher/lower than the setpoint, the
805 * dirty position control ratio (and hence task dirty ratelimit) will be
806 * decreased/increased to bring the dirty pages back to the setpoint.
807 *
808 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
809 *
810 * if (dirty < setpoint) scale up pos_ratio
811 * if (dirty > setpoint) scale down pos_ratio
812 *
813 * if (wb_dirty < wb_setpoint) scale up pos_ratio
814 * if (wb_dirty > wb_setpoint) scale down pos_ratio
815 *
816 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
817 *
818 * (o) global control line
819 *
820 * ^ pos_ratio
821 * |
822 * | |<===== global dirty control scope ======>|
823 * 2.0 .............*
824 * | .*
825 * | . *
826 * | . *
827 * | . *
828 * | . *
829 * | . *
830 * 1.0 ................................*
831 * | . . *
832 * | . . *
833 * | . . *
834 * | . . *
835 * | . . *
836 * 0 +------------.------------------.----------------------*------------->
837 * freerun^ setpoint^ limit^ dirty pages
838 *
839 * (o) wb control line
840 *
841 * ^ pos_ratio
842 * |
843 * | *
844 * | *
845 * | *
846 * | *
847 * | * |<=========== span ============>|
848 * 1.0 .......................*
849 * | . *
850 * | . *
851 * | . *
852 * | . *
853 * | . *
854 * | . *
855 * | . *
856 * | . *
857 * | . *
858 * | . *
859 * | . *
860 * 1/4 ...............................................* * * * * * * * * * * *
861 * | . .
862 * | . .
863 * | . .
864 * 0 +----------------------.-------------------------------.------------->
865 * wb_setpoint^ x_intercept^
866 *
867 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
868 * be smoothly throttled down to normal if it starts high in situations like
869 * - start writing to a slow SD card and a fast disk at the same time. The SD
870 * card's wb_dirty may rush to many times higher than wb_setpoint.
871 * - the wb dirty thresh drops quickly due to change of JBOD workload
872 */
873static void wb_position_ratio(struct dirty_throttle_control *dtc)
874{
875 struct bdi_writeback *wb = dtc->wb;
876 unsigned long write_bw = wb->avg_write_bandwidth;
877 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
878 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
879 unsigned long wb_thresh = dtc->wb_thresh;
880 unsigned long x_intercept;
881 unsigned long setpoint; /* dirty pages' target balance point */
882 unsigned long wb_setpoint;
883 unsigned long span;
884 long long pos_ratio; /* for scaling up/down the rate limit */
885 long x;
886
887 dtc->pos_ratio = 0;
888
889 if (unlikely(dtc->dirty >= limit))
890 return;
891
892 /*
893 * global setpoint
894 *
895 * See comment for pos_ratio_polynom().
896 */
897 setpoint = (freerun + limit) / 2;
898 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
899
900 /*
901 * The strictlimit feature is a tool preventing mistrusted filesystems
902 * from growing a large number of dirty pages before throttling. For
903 * such filesystems balance_dirty_pages always checks wb counters
904 * against wb limits. Even if global "nr_dirty" is under "freerun".
905 * This is especially important for fuse which sets bdi->max_ratio to
906 * 1% by default. Without strictlimit feature, fuse writeback may
907 * consume arbitrary amount of RAM because it is accounted in
908 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
909 *
910 * Here, in wb_position_ratio(), we calculate pos_ratio based on
911 * two values: wb_dirty and wb_thresh. Let's consider an example:
912 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
913 * limits are set by default to 10% and 20% (background and throttle).
914 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
915 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
916 * about ~6K pages (as the average of background and throttle wb
917 * limits). The 3rd order polynomial will provide positive feedback if
918 * wb_dirty is under wb_setpoint and vice versa.
919 *
920 * Note, that we cannot use global counters in these calculations
921 * because we want to throttle process writing to a strictlimit wb
922 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
923 * in the example above).
924 */
925 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
926 long long wb_pos_ratio;
927
928 if (dtc->wb_dirty < 8) {
929 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
930 2 << RATELIMIT_CALC_SHIFT);
931 return;
932 }
933
934 if (dtc->wb_dirty >= wb_thresh)
935 return;
936
937 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
938 dtc->wb_bg_thresh);
939
940 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
941 return;
942
943 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
944 wb_thresh);
945
946 /*
947 * Typically, for strictlimit case, wb_setpoint << setpoint
948 * and pos_ratio >> wb_pos_ratio. In the other words global
949 * state ("dirty") is not limiting factor and we have to
950 * make decision based on wb counters. But there is an
951 * important case when global pos_ratio should get precedence:
952 * global limits are exceeded (e.g. due to activities on other
953 * wb's) while given strictlimit wb is below limit.
954 *
955 * "pos_ratio * wb_pos_ratio" would work for the case above,
956 * but it would look too non-natural for the case of all
957 * activity in the system coming from a single strictlimit wb
958 * with bdi->max_ratio == 100%.
959 *
960 * Note that min() below somewhat changes the dynamics of the
961 * control system. Normally, pos_ratio value can be well over 3
962 * (when globally we are at freerun and wb is well below wb
963 * setpoint). Now the maximum pos_ratio in the same situation
964 * is 2. We might want to tweak this if we observe the control
965 * system is too slow to adapt.
966 */
967 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
968 return;
969 }
970
971 /*
972 * We have computed basic pos_ratio above based on global situation. If
973 * the wb is over/under its share of dirty pages, we want to scale
974 * pos_ratio further down/up. That is done by the following mechanism.
975 */
976
977 /*
978 * wb setpoint
979 *
980 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
981 *
982 * x_intercept - wb_dirty
983 * := --------------------------
984 * x_intercept - wb_setpoint
985 *
986 * The main wb control line is a linear function that subjects to
987 *
988 * (1) f(wb_setpoint) = 1.0
989 * (2) k = - 1 / (8 * write_bw) (in single wb case)
990 * or equally: x_intercept = wb_setpoint + 8 * write_bw
991 *
992 * For single wb case, the dirty pages are observed to fluctuate
993 * regularly within range
994 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
995 * for various filesystems, where (2) can yield in a reasonable 12.5%
996 * fluctuation range for pos_ratio.
997 *
998 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
999 * own size, so move the slope over accordingly and choose a slope that
1000 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1001 */
1002 if (unlikely(wb_thresh > dtc->thresh))
1003 wb_thresh = dtc->thresh;
1004 /*
1005 * It's very possible that wb_thresh is close to 0 not because the
1006 * device is slow, but that it has remained inactive for long time.
1007 * Honour such devices a reasonable good (hopefully IO efficient)
1008 * threshold, so that the occasional writes won't be blocked and active
1009 * writes can rampup the threshold quickly.
1010 */
1011 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1012 /*
1013 * scale global setpoint to wb's:
1014 * wb_setpoint = setpoint * wb_thresh / thresh
1015 */
1016 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1017 wb_setpoint = setpoint * (u64)x >> 16;
1018 /*
1019 * Use span=(8*write_bw) in single wb case as indicated by
1020 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1021 *
1022 * wb_thresh thresh - wb_thresh
1023 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1024 * thresh thresh
1025 */
1026 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1027 x_intercept = wb_setpoint + span;
1028
1029 if (dtc->wb_dirty < x_intercept - span / 4) {
1030 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1031 (x_intercept - wb_setpoint) | 1);
1032 } else
1033 pos_ratio /= 4;
1034
1035 /*
1036 * wb reserve area, safeguard against dirty pool underrun and disk idle
1037 * It may push the desired control point of global dirty pages higher
1038 * than setpoint.
1039 */
1040 x_intercept = wb_thresh / 2;
1041 if (dtc->wb_dirty < x_intercept) {
1042 if (dtc->wb_dirty > x_intercept / 8)
1043 pos_ratio = div_u64(pos_ratio * x_intercept,
1044 dtc->wb_dirty);
1045 else
1046 pos_ratio *= 8;
1047 }
1048
1049 dtc->pos_ratio = pos_ratio;
1050}
1051
1052static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1053 unsigned long elapsed,
1054 unsigned long written)
1055{
1056 const unsigned long period = roundup_pow_of_two(3 * HZ);
1057 unsigned long avg = wb->avg_write_bandwidth;
1058 unsigned long old = wb->write_bandwidth;
1059 u64 bw;
1060
1061 /*
1062 * bw = written * HZ / elapsed
1063 *
1064 * bw * elapsed + write_bandwidth * (period - elapsed)
1065 * write_bandwidth = ---------------------------------------------------
1066 * period
1067 *
1068 * @written may have decreased due to account_page_redirty().
1069 * Avoid underflowing @bw calculation.
1070 */
1071 bw = written - min(written, wb->written_stamp);
1072 bw *= HZ;
1073 if (unlikely(elapsed > period)) {
1074 do_div(bw, elapsed);
1075 avg = bw;
1076 goto out;
1077 }
1078 bw += (u64)wb->write_bandwidth * (period - elapsed);
1079 bw >>= ilog2(period);
1080
1081 /*
1082 * one more level of smoothing, for filtering out sudden spikes
1083 */
1084 if (avg > old && old >= (unsigned long)bw)
1085 avg -= (avg - old) >> 3;
1086
1087 if (avg < old && old <= (unsigned long)bw)
1088 avg += (old - avg) >> 3;
1089
1090out:
1091 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1092 avg = max(avg, 1LU);
1093 if (wb_has_dirty_io(wb)) {
1094 long delta = avg - wb->avg_write_bandwidth;
1095 WARN_ON_ONCE(atomic_long_add_return(delta,
1096 &wb->bdi->tot_write_bandwidth) <= 0);
1097 }
1098 wb->write_bandwidth = bw;
1099 wb->avg_write_bandwidth = avg;
1100}
1101
1102static void update_dirty_limit(struct dirty_throttle_control *dtc)
1103{
1104 struct wb_domain *dom = dtc_dom(dtc);
1105 unsigned long thresh = dtc->thresh;
1106 unsigned long limit = dom->dirty_limit;
1107
1108 /*
1109 * Follow up in one step.
1110 */
1111 if (limit < thresh) {
1112 limit = thresh;
1113 goto update;
1114 }
1115
1116 /*
1117 * Follow down slowly. Use the higher one as the target, because thresh
1118 * may drop below dirty. This is exactly the reason to introduce
1119 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1120 */
1121 thresh = max(thresh, dtc->dirty);
1122 if (limit > thresh) {
1123 limit -= (limit - thresh) >> 5;
1124 goto update;
1125 }
1126 return;
1127update:
1128 dom->dirty_limit = limit;
1129}
1130
1131static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
1132 unsigned long now)
1133{
1134 struct wb_domain *dom = dtc_dom(dtc);
1135
1136 /*
1137 * check locklessly first to optimize away locking for the most time
1138 */
1139 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1140 return;
1141
1142 spin_lock(&dom->lock);
1143 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1144 update_dirty_limit(dtc);
1145 dom->dirty_limit_tstamp = now;
1146 }
1147 spin_unlock(&dom->lock);
1148}
1149
1150/*
1151 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1152 *
1153 * Normal wb tasks will be curbed at or below it in long term.
1154 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1155 */
1156static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1157 unsigned long dirtied,
1158 unsigned long elapsed)
1159{
1160 struct bdi_writeback *wb = dtc->wb;
1161 unsigned long dirty = dtc->dirty;
1162 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1163 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1164 unsigned long setpoint = (freerun + limit) / 2;
1165 unsigned long write_bw = wb->avg_write_bandwidth;
1166 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1167 unsigned long dirty_rate;
1168 unsigned long task_ratelimit;
1169 unsigned long balanced_dirty_ratelimit;
1170 unsigned long step;
1171 unsigned long x;
1172 unsigned long shift;
1173
1174 /*
1175 * The dirty rate will match the writeout rate in long term, except
1176 * when dirty pages are truncated by userspace or re-dirtied by FS.
1177 */
1178 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1179
1180 /*
1181 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1182 */
1183 task_ratelimit = (u64)dirty_ratelimit *
1184 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1185 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1186
1187 /*
1188 * A linear estimation of the "balanced" throttle rate. The theory is,
1189 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1190 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1191 * formula will yield the balanced rate limit (write_bw / N).
1192 *
1193 * Note that the expanded form is not a pure rate feedback:
1194 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1195 * but also takes pos_ratio into account:
1196 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1197 *
1198 * (1) is not realistic because pos_ratio also takes part in balancing
1199 * the dirty rate. Consider the state
1200 * pos_ratio = 0.5 (3)
1201 * rate = 2 * (write_bw / N) (4)
1202 * If (1) is used, it will stuck in that state! Because each dd will
1203 * be throttled at
1204 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1205 * yielding
1206 * dirty_rate = N * task_ratelimit = write_bw (6)
1207 * put (6) into (1) we get
1208 * rate_(i+1) = rate_(i) (7)
1209 *
1210 * So we end up using (2) to always keep
1211 * rate_(i+1) ~= (write_bw / N) (8)
1212 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1213 * pos_ratio is able to drive itself to 1.0, which is not only where
1214 * the dirty count meet the setpoint, but also where the slope of
1215 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1216 */
1217 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1218 dirty_rate | 1);
1219 /*
1220 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1221 */
1222 if (unlikely(balanced_dirty_ratelimit > write_bw))
1223 balanced_dirty_ratelimit = write_bw;
1224
1225 /*
1226 * We could safely do this and return immediately:
1227 *
1228 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1229 *
1230 * However to get a more stable dirty_ratelimit, the below elaborated
1231 * code makes use of task_ratelimit to filter out singular points and
1232 * limit the step size.
1233 *
1234 * The below code essentially only uses the relative value of
1235 *
1236 * task_ratelimit - dirty_ratelimit
1237 * = (pos_ratio - 1) * dirty_ratelimit
1238 *
1239 * which reflects the direction and size of dirty position error.
1240 */
1241
1242 /*
1243 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1244 * task_ratelimit is on the same side of dirty_ratelimit, too.
1245 * For example, when
1246 * - dirty_ratelimit > balanced_dirty_ratelimit
1247 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1248 * lowering dirty_ratelimit will help meet both the position and rate
1249 * control targets. Otherwise, don't update dirty_ratelimit if it will
1250 * only help meet the rate target. After all, what the users ultimately
1251 * feel and care are stable dirty rate and small position error.
1252 *
1253 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1254 * and filter out the singular points of balanced_dirty_ratelimit. Which
1255 * keeps jumping around randomly and can even leap far away at times
1256 * due to the small 200ms estimation period of dirty_rate (we want to
1257 * keep that period small to reduce time lags).
1258 */
1259 step = 0;
1260
1261 /*
1262 * For strictlimit case, calculations above were based on wb counters
1263 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1264 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1265 * Hence, to calculate "step" properly, we have to use wb_dirty as
1266 * "dirty" and wb_setpoint as "setpoint".
1267 *
1268 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1269 * it's possible that wb_thresh is close to zero due to inactivity
1270 * of backing device.
1271 */
1272 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1273 dirty = dtc->wb_dirty;
1274 if (dtc->wb_dirty < 8)
1275 setpoint = dtc->wb_dirty + 1;
1276 else
1277 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1278 }
1279
1280 if (dirty < setpoint) {
1281 x = min3(wb->balanced_dirty_ratelimit,
1282 balanced_dirty_ratelimit, task_ratelimit);
1283 if (dirty_ratelimit < x)
1284 step = x - dirty_ratelimit;
1285 } else {
1286 x = max3(wb->balanced_dirty_ratelimit,
1287 balanced_dirty_ratelimit, task_ratelimit);
1288 if (dirty_ratelimit > x)
1289 step = dirty_ratelimit - x;
1290 }
1291
1292 /*
1293 * Don't pursue 100% rate matching. It's impossible since the balanced
1294 * rate itself is constantly fluctuating. So decrease the track speed
1295 * when it gets close to the target. Helps eliminate pointless tremors.
1296 */
1297 shift = dirty_ratelimit / (2 * step + 1);
1298 if (shift < BITS_PER_LONG)
1299 step = DIV_ROUND_UP(step >> shift, 8);
1300 else
1301 step = 0;
1302
1303 if (dirty_ratelimit < balanced_dirty_ratelimit)
1304 dirty_ratelimit += step;
1305 else
1306 dirty_ratelimit -= step;
1307
1308 wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1309 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1310
1311 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1312}
1313
1314static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1315 struct dirty_throttle_control *mdtc,
1316 unsigned long start_time,
1317 bool update_ratelimit)
1318{
1319 struct bdi_writeback *wb = gdtc->wb;
1320 unsigned long now = jiffies;
1321 unsigned long elapsed = now - wb->bw_time_stamp;
1322 unsigned long dirtied;
1323 unsigned long written;
1324
1325 lockdep_assert_held(&wb->list_lock);
1326
1327 /*
1328 * rate-limit, only update once every 200ms.
1329 */
1330 if (elapsed < BANDWIDTH_INTERVAL)
1331 return;
1332
1333 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1334 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1335
1336 /*
1337 * Skip quiet periods when disk bandwidth is under-utilized.
1338 * (at least 1s idle time between two flusher runs)
1339 */
1340 if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1341 goto snapshot;
1342
1343 if (update_ratelimit) {
1344 domain_update_bandwidth(gdtc, now);
1345 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1346
1347 /*
1348 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1349 * compiler has no way to figure that out. Help it.
1350 */
1351 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1352 domain_update_bandwidth(mdtc, now);
1353 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1354 }
1355 }
1356 wb_update_write_bandwidth(wb, elapsed, written);
1357
1358snapshot:
1359 wb->dirtied_stamp = dirtied;
1360 wb->written_stamp = written;
1361 wb->bw_time_stamp = now;
1362}
1363
1364void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1365{
1366 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1367
1368 __wb_update_bandwidth(&gdtc, NULL, start_time, false);
1369}
1370
1371/*
1372 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1373 * will look to see if it needs to start dirty throttling.
1374 *
1375 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1376 * global_page_state() too often. So scale it near-sqrt to the safety margin
1377 * (the number of pages we may dirty without exceeding the dirty limits).
1378 */
1379static unsigned long dirty_poll_interval(unsigned long dirty,
1380 unsigned long thresh)
1381{
1382 if (thresh > dirty)
1383 return 1UL << (ilog2(thresh - dirty) >> 1);
1384
1385 return 1;
1386}
1387
1388static unsigned long wb_max_pause(struct bdi_writeback *wb,
1389 unsigned long wb_dirty)
1390{
1391 unsigned long bw = wb->avg_write_bandwidth;
1392 unsigned long t;
1393
1394 /*
1395 * Limit pause time for small memory systems. If sleeping for too long
1396 * time, a small pool of dirty/writeback pages may go empty and disk go
1397 * idle.
1398 *
1399 * 8 serves as the safety ratio.
1400 */
1401 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1402 t++;
1403
1404 return min_t(unsigned long, t, MAX_PAUSE);
1405}
1406
1407static long wb_min_pause(struct bdi_writeback *wb,
1408 long max_pause,
1409 unsigned long task_ratelimit,
1410 unsigned long dirty_ratelimit,
1411 int *nr_dirtied_pause)
1412{
1413 long hi = ilog2(wb->avg_write_bandwidth);
1414 long lo = ilog2(wb->dirty_ratelimit);
1415 long t; /* target pause */
1416 long pause; /* estimated next pause */
1417 int pages; /* target nr_dirtied_pause */
1418
1419 /* target for 10ms pause on 1-dd case */
1420 t = max(1, HZ / 100);
1421
1422 /*
1423 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1424 * overheads.
1425 *
1426 * (N * 10ms) on 2^N concurrent tasks.
1427 */
1428 if (hi > lo)
1429 t += (hi - lo) * (10 * HZ) / 1024;
1430
1431 /*
1432 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1433 * on the much more stable dirty_ratelimit. However the next pause time
1434 * will be computed based on task_ratelimit and the two rate limits may
1435 * depart considerably at some time. Especially if task_ratelimit goes
1436 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1437 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1438 * result task_ratelimit won't be executed faithfully, which could
1439 * eventually bring down dirty_ratelimit.
1440 *
1441 * We apply two rules to fix it up:
1442 * 1) try to estimate the next pause time and if necessary, use a lower
1443 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1444 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1445 * 2) limit the target pause time to max_pause/2, so that the normal
1446 * small fluctuations of task_ratelimit won't trigger rule (1) and
1447 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1448 */
1449 t = min(t, 1 + max_pause / 2);
1450 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1451
1452 /*
1453 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1454 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1455 * When the 16 consecutive reads are often interrupted by some dirty
1456 * throttling pause during the async writes, cfq will go into idles
1457 * (deadline is fine). So push nr_dirtied_pause as high as possible
1458 * until reaches DIRTY_POLL_THRESH=32 pages.
1459 */
1460 if (pages < DIRTY_POLL_THRESH) {
1461 t = max_pause;
1462 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1463 if (pages > DIRTY_POLL_THRESH) {
1464 pages = DIRTY_POLL_THRESH;
1465 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1466 }
1467 }
1468
1469 pause = HZ * pages / (task_ratelimit + 1);
1470 if (pause > max_pause) {
1471 t = max_pause;
1472 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1473 }
1474
1475 *nr_dirtied_pause = pages;
1476 /*
1477 * The minimal pause time will normally be half the target pause time.
1478 */
1479 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1480}
1481
1482static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1483{
1484 struct bdi_writeback *wb = dtc->wb;
1485 unsigned long wb_reclaimable;
1486
1487 /*
1488 * wb_thresh is not treated as some limiting factor as
1489 * dirty_thresh, due to reasons
1490 * - in JBOD setup, wb_thresh can fluctuate a lot
1491 * - in a system with HDD and USB key, the USB key may somehow
1492 * go into state (wb_dirty >> wb_thresh) either because
1493 * wb_dirty starts high, or because wb_thresh drops low.
1494 * In this case we don't want to hard throttle the USB key
1495 * dirtiers for 100 seconds until wb_dirty drops under
1496 * wb_thresh. Instead the auxiliary wb control line in
1497 * wb_position_ratio() will let the dirtier task progress
1498 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1499 */
1500 dtc->wb_thresh = __wb_calc_thresh(dtc);
1501 dtc->wb_bg_thresh = dtc->thresh ?
1502 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1503
1504 /*
1505 * In order to avoid the stacked BDI deadlock we need
1506 * to ensure we accurately count the 'dirty' pages when
1507 * the threshold is low.
1508 *
1509 * Otherwise it would be possible to get thresh+n pages
1510 * reported dirty, even though there are thresh-m pages
1511 * actually dirty; with m+n sitting in the percpu
1512 * deltas.
1513 */
1514 if (dtc->wb_thresh < 2 * wb_stat_error(wb)) {
1515 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1516 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1517 } else {
1518 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1519 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1520 }
1521}
1522
1523/*
1524 * balance_dirty_pages() must be called by processes which are generating dirty
1525 * data. It looks at the number of dirty pages in the machine and will force
1526 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1527 * If we're over `background_thresh' then the writeback threads are woken to
1528 * perform some writeout.
1529 */
1530static void balance_dirty_pages(struct address_space *mapping,
1531 struct bdi_writeback *wb,
1532 unsigned long pages_dirtied)
1533{
1534 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1535 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1536 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1537 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1538 &mdtc_stor : NULL;
1539 struct dirty_throttle_control *sdtc;
1540 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1541 long period;
1542 long pause;
1543 long max_pause;
1544 long min_pause;
1545 int nr_dirtied_pause;
1546 bool dirty_exceeded = false;
1547 unsigned long task_ratelimit;
1548 unsigned long dirty_ratelimit;
1549 struct backing_dev_info *bdi = wb->bdi;
1550 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1551 unsigned long start_time = jiffies;
1552
1553 for (;;) {
1554 unsigned long now = jiffies;
1555 unsigned long dirty, thresh, bg_thresh;
1556 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1557 unsigned long m_thresh = 0;
1558 unsigned long m_bg_thresh = 0;
1559
1560 /*
1561 * Unstable writes are a feature of certain networked
1562 * filesystems (i.e. NFS) in which data may have been
1563 * written to the server's write cache, but has not yet
1564 * been flushed to permanent storage.
1565 */
1566 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1567 global_page_state(NR_UNSTABLE_NFS);
1568 gdtc->avail = global_dirtyable_memory();
1569 gdtc->dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1570
1571 domain_dirty_limits(gdtc);
1572
1573 if (unlikely(strictlimit)) {
1574 wb_dirty_limits(gdtc);
1575
1576 dirty = gdtc->wb_dirty;
1577 thresh = gdtc->wb_thresh;
1578 bg_thresh = gdtc->wb_bg_thresh;
1579 } else {
1580 dirty = gdtc->dirty;
1581 thresh = gdtc->thresh;
1582 bg_thresh = gdtc->bg_thresh;
1583 }
1584
1585 if (mdtc) {
1586 unsigned long filepages, headroom, writeback;
1587
1588 /*
1589 * If @wb belongs to !root memcg, repeat the same
1590 * basic calculations for the memcg domain.
1591 */
1592 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1593 &mdtc->dirty, &writeback);
1594 mdtc->dirty += writeback;
1595 mdtc_calc_avail(mdtc, filepages, headroom);
1596
1597 domain_dirty_limits(mdtc);
1598
1599 if (unlikely(strictlimit)) {
1600 wb_dirty_limits(mdtc);
1601 m_dirty = mdtc->wb_dirty;
1602 m_thresh = mdtc->wb_thresh;
1603 m_bg_thresh = mdtc->wb_bg_thresh;
1604 } else {
1605 m_dirty = mdtc->dirty;
1606 m_thresh = mdtc->thresh;
1607 m_bg_thresh = mdtc->bg_thresh;
1608 }
1609 }
1610
1611 /*
1612 * Throttle it only when the background writeback cannot
1613 * catch-up. This avoids (excessively) small writeouts
1614 * when the wb limits are ramping up in case of !strictlimit.
1615 *
1616 * In strictlimit case make decision based on the wb counters
1617 * and limits. Small writeouts when the wb limits are ramping
1618 * up are the price we consciously pay for strictlimit-ing.
1619 *
1620 * If memcg domain is in effect, @dirty should be under
1621 * both global and memcg freerun ceilings.
1622 */
1623 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1624 (!mdtc ||
1625 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1626 unsigned long intv = dirty_poll_interval(dirty, thresh);
1627 unsigned long m_intv = ULONG_MAX;
1628
1629 current->dirty_paused_when = now;
1630 current->nr_dirtied = 0;
1631 if (mdtc)
1632 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1633 current->nr_dirtied_pause = min(intv, m_intv);
1634 break;
1635 }
1636
1637 if (unlikely(!writeback_in_progress(wb)))
1638 wb_start_background_writeback(wb);
1639
1640 /*
1641 * Calculate global domain's pos_ratio and select the
1642 * global dtc by default.
1643 */
1644 if (!strictlimit)
1645 wb_dirty_limits(gdtc);
1646
1647 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1648 ((gdtc->dirty > gdtc->thresh) || strictlimit);
1649
1650 wb_position_ratio(gdtc);
1651 sdtc = gdtc;
1652
1653 if (mdtc) {
1654 /*
1655 * If memcg domain is in effect, calculate its
1656 * pos_ratio. @wb should satisfy constraints from
1657 * both global and memcg domains. Choose the one
1658 * w/ lower pos_ratio.
1659 */
1660 if (!strictlimit)
1661 wb_dirty_limits(mdtc);
1662
1663 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1664 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1665
1666 wb_position_ratio(mdtc);
1667 if (mdtc->pos_ratio < gdtc->pos_ratio)
1668 sdtc = mdtc;
1669 }
1670
1671 if (dirty_exceeded && !wb->dirty_exceeded)
1672 wb->dirty_exceeded = 1;
1673
1674 if (time_is_before_jiffies(wb->bw_time_stamp +
1675 BANDWIDTH_INTERVAL)) {
1676 spin_lock(&wb->list_lock);
1677 __wb_update_bandwidth(gdtc, mdtc, start_time, true);
1678 spin_unlock(&wb->list_lock);
1679 }
1680
1681 /* throttle according to the chosen dtc */
1682 dirty_ratelimit = wb->dirty_ratelimit;
1683 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1684 RATELIMIT_CALC_SHIFT;
1685 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1686 min_pause = wb_min_pause(wb, max_pause,
1687 task_ratelimit, dirty_ratelimit,
1688 &nr_dirtied_pause);
1689
1690 if (unlikely(task_ratelimit == 0)) {
1691 period = max_pause;
1692 pause = max_pause;
1693 goto pause;
1694 }
1695 period = HZ * pages_dirtied / task_ratelimit;
1696 pause = period;
1697 if (current->dirty_paused_when)
1698 pause -= now - current->dirty_paused_when;
1699 /*
1700 * For less than 1s think time (ext3/4 may block the dirtier
1701 * for up to 800ms from time to time on 1-HDD; so does xfs,
1702 * however at much less frequency), try to compensate it in
1703 * future periods by updating the virtual time; otherwise just
1704 * do a reset, as it may be a light dirtier.
1705 */
1706 if (pause < min_pause) {
1707 trace_balance_dirty_pages(wb,
1708 sdtc->thresh,
1709 sdtc->bg_thresh,
1710 sdtc->dirty,
1711 sdtc->wb_thresh,
1712 sdtc->wb_dirty,
1713 dirty_ratelimit,
1714 task_ratelimit,
1715 pages_dirtied,
1716 period,
1717 min(pause, 0L),
1718 start_time);
1719 if (pause < -HZ) {
1720 current->dirty_paused_when = now;
1721 current->nr_dirtied = 0;
1722 } else if (period) {
1723 current->dirty_paused_when += period;
1724 current->nr_dirtied = 0;
1725 } else if (current->nr_dirtied_pause <= pages_dirtied)
1726 current->nr_dirtied_pause += pages_dirtied;
1727 break;
1728 }
1729 if (unlikely(pause > max_pause)) {
1730 /* for occasional dropped task_ratelimit */
1731 now += min(pause - max_pause, max_pause);
1732 pause = max_pause;
1733 }
1734
1735pause:
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 pause,
1747 start_time);
1748 __set_current_state(TASK_KILLABLE);
1749 io_schedule_timeout(pause);
1750
1751 current->dirty_paused_when = now + pause;
1752 current->nr_dirtied = 0;
1753 current->nr_dirtied_pause = nr_dirtied_pause;
1754
1755 /*
1756 * This is typically equal to (dirty < thresh) and can also
1757 * keep "1000+ dd on a slow USB stick" under control.
1758 */
1759 if (task_ratelimit)
1760 break;
1761
1762 /*
1763 * In the case of an unresponding NFS server and the NFS dirty
1764 * pages exceeds dirty_thresh, give the other good wb's a pipe
1765 * to go through, so that tasks on them still remain responsive.
1766 *
1767 * In theory 1 page is enough to keep the comsumer-producer
1768 * pipe going: the flusher cleans 1 page => the task dirties 1
1769 * more page. However wb_dirty has accounting errors. So use
1770 * the larger and more IO friendly wb_stat_error.
1771 */
1772 if (sdtc->wb_dirty <= wb_stat_error(wb))
1773 break;
1774
1775 if (fatal_signal_pending(current))
1776 break;
1777 }
1778
1779 if (!dirty_exceeded && wb->dirty_exceeded)
1780 wb->dirty_exceeded = 0;
1781
1782 if (writeback_in_progress(wb))
1783 return;
1784
1785 /*
1786 * In laptop mode, we wait until hitting the higher threshold before
1787 * starting background writeout, and then write out all the way down
1788 * to the lower threshold. So slow writers cause minimal disk activity.
1789 *
1790 * In normal mode, we start background writeout at the lower
1791 * background_thresh, to keep the amount of dirty memory low.
1792 */
1793 if (laptop_mode)
1794 return;
1795
1796 if (nr_reclaimable > gdtc->bg_thresh)
1797 wb_start_background_writeback(wb);
1798}
1799
1800static DEFINE_PER_CPU(int, bdp_ratelimits);
1801
1802/*
1803 * Normal tasks are throttled by
1804 * loop {
1805 * dirty tsk->nr_dirtied_pause pages;
1806 * take a snap in balance_dirty_pages();
1807 * }
1808 * However there is a worst case. If every task exit immediately when dirtied
1809 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1810 * called to throttle the page dirties. The solution is to save the not yet
1811 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1812 * randomly into the running tasks. This works well for the above worst case,
1813 * as the new task will pick up and accumulate the old task's leaked dirty
1814 * count and eventually get throttled.
1815 */
1816DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1817
1818/**
1819 * balance_dirty_pages_ratelimited - balance dirty memory state
1820 * @mapping: address_space which was dirtied
1821 *
1822 * Processes which are dirtying memory should call in here once for each page
1823 * which was newly dirtied. The function will periodically check the system's
1824 * dirty state and will initiate writeback if needed.
1825 *
1826 * On really big machines, get_writeback_state is expensive, so try to avoid
1827 * calling it too often (ratelimiting). But once we're over the dirty memory
1828 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1829 * from overshooting the limit by (ratelimit_pages) each.
1830 */
1831void balance_dirty_pages_ratelimited(struct address_space *mapping)
1832{
1833 struct inode *inode = mapping->host;
1834 struct backing_dev_info *bdi = inode_to_bdi(inode);
1835 struct bdi_writeback *wb = NULL;
1836 int ratelimit;
1837 int *p;
1838
1839 if (!bdi_cap_account_dirty(bdi))
1840 return;
1841
1842 if (inode_cgwb_enabled(inode))
1843 wb = wb_get_create_current(bdi, GFP_KERNEL);
1844 if (!wb)
1845 wb = &bdi->wb;
1846
1847 ratelimit = current->nr_dirtied_pause;
1848 if (wb->dirty_exceeded)
1849 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1850
1851 preempt_disable();
1852 /*
1853 * This prevents one CPU to accumulate too many dirtied pages without
1854 * calling into balance_dirty_pages(), which can happen when there are
1855 * 1000+ tasks, all of them start dirtying pages at exactly the same
1856 * time, hence all honoured too large initial task->nr_dirtied_pause.
1857 */
1858 p = this_cpu_ptr(&bdp_ratelimits);
1859 if (unlikely(current->nr_dirtied >= ratelimit))
1860 *p = 0;
1861 else if (unlikely(*p >= ratelimit_pages)) {
1862 *p = 0;
1863 ratelimit = 0;
1864 }
1865 /*
1866 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1867 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1868 * the dirty throttling and livelock other long-run dirtiers.
1869 */
1870 p = this_cpu_ptr(&dirty_throttle_leaks);
1871 if (*p > 0 && current->nr_dirtied < ratelimit) {
1872 unsigned long nr_pages_dirtied;
1873 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1874 *p -= nr_pages_dirtied;
1875 current->nr_dirtied += nr_pages_dirtied;
1876 }
1877 preempt_enable();
1878
1879 if (unlikely(current->nr_dirtied >= ratelimit))
1880 balance_dirty_pages(mapping, wb, current->nr_dirtied);
1881
1882 wb_put(wb);
1883}
1884EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1885
1886/**
1887 * wb_over_bg_thresh - does @wb need to be written back?
1888 * @wb: bdi_writeback of interest
1889 *
1890 * Determines whether background writeback should keep writing @wb or it's
1891 * clean enough. Returns %true if writeback should continue.
1892 */
1893bool wb_over_bg_thresh(struct bdi_writeback *wb)
1894{
1895 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1896 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1897 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1898 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1899 &mdtc_stor : NULL;
1900
1901 /*
1902 * Similar to balance_dirty_pages() but ignores pages being written
1903 * as we're trying to decide whether to put more under writeback.
1904 */
1905 gdtc->avail = global_dirtyable_memory();
1906 gdtc->dirty = global_page_state(NR_FILE_DIRTY) +
1907 global_page_state(NR_UNSTABLE_NFS);
1908 domain_dirty_limits(gdtc);
1909
1910 if (gdtc->dirty > gdtc->bg_thresh)
1911 return true;
1912
1913 if (wb_stat(wb, WB_RECLAIMABLE) >
1914 wb_calc_thresh(gdtc->wb, gdtc->bg_thresh))
1915 return true;
1916
1917 if (mdtc) {
1918 unsigned long filepages, headroom, writeback;
1919
1920 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1921 &writeback);
1922 mdtc_calc_avail(mdtc, filepages, headroom);
1923 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1924
1925 if (mdtc->dirty > mdtc->bg_thresh)
1926 return true;
1927
1928 if (wb_stat(wb, WB_RECLAIMABLE) >
1929 wb_calc_thresh(mdtc->wb, mdtc->bg_thresh))
1930 return true;
1931 }
1932
1933 return false;
1934}
1935
1936void throttle_vm_writeout(gfp_t gfp_mask)
1937{
1938 unsigned long background_thresh;
1939 unsigned long dirty_thresh;
1940
1941 for ( ; ; ) {
1942 global_dirty_limits(&background_thresh, &dirty_thresh);
1943 dirty_thresh = hard_dirty_limit(&global_wb_domain, dirty_thresh);
1944
1945 /*
1946 * Boost the allowable dirty threshold a bit for page
1947 * allocators so they don't get DoS'ed by heavy writers
1948 */
1949 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1950
1951 if (global_page_state(NR_UNSTABLE_NFS) +
1952 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1953 break;
1954 congestion_wait(BLK_RW_ASYNC, HZ/10);
1955
1956 /*
1957 * The caller might hold locks which can prevent IO completion
1958 * or progress in the filesystem. So we cannot just sit here
1959 * waiting for IO to complete.
1960 */
1961 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1962 break;
1963 }
1964}
1965
1966/*
1967 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1968 */
1969int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1970 void __user *buffer, size_t *length, loff_t *ppos)
1971{
1972 proc_dointvec(table, write, buffer, length, ppos);
1973 return 0;
1974}
1975
1976#ifdef CONFIG_BLOCK
1977void laptop_mode_timer_fn(unsigned long data)
1978{
1979 struct request_queue *q = (struct request_queue *)data;
1980 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1981 global_page_state(NR_UNSTABLE_NFS);
1982 struct bdi_writeback *wb;
1983
1984 /*
1985 * We want to write everything out, not just down to the dirty
1986 * threshold
1987 */
1988 if (!bdi_has_dirty_io(&q->backing_dev_info))
1989 return;
1990
1991 rcu_read_lock();
1992 list_for_each_entry_rcu(wb, &q->backing_dev_info.wb_list, bdi_node)
1993 if (wb_has_dirty_io(wb))
1994 wb_start_writeback(wb, nr_pages, true,
1995 WB_REASON_LAPTOP_TIMER);
1996 rcu_read_unlock();
1997}
1998
1999/*
2000 * We've spun up the disk and we're in laptop mode: schedule writeback
2001 * of all dirty data a few seconds from now. If the flush is already scheduled
2002 * then push it back - the user is still using the disk.
2003 */
2004void laptop_io_completion(struct backing_dev_info *info)
2005{
2006 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2007}
2008
2009/*
2010 * We're in laptop mode and we've just synced. The sync's writes will have
2011 * caused another writeback to be scheduled by laptop_io_completion.
2012 * Nothing needs to be written back anymore, so we unschedule the writeback.
2013 */
2014void laptop_sync_completion(void)
2015{
2016 struct backing_dev_info *bdi;
2017
2018 rcu_read_lock();
2019
2020 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2021 del_timer(&bdi->laptop_mode_wb_timer);
2022
2023 rcu_read_unlock();
2024}
2025#endif
2026
2027/*
2028 * If ratelimit_pages is too high then we can get into dirty-data overload
2029 * if a large number of processes all perform writes at the same time.
2030 * If it is too low then SMP machines will call the (expensive)
2031 * get_writeback_state too often.
2032 *
2033 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2034 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2035 * thresholds.
2036 */
2037
2038void writeback_set_ratelimit(void)
2039{
2040 struct wb_domain *dom = &global_wb_domain;
2041 unsigned long background_thresh;
2042 unsigned long dirty_thresh;
2043
2044 global_dirty_limits(&background_thresh, &dirty_thresh);
2045 dom->dirty_limit = dirty_thresh;
2046 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2047 if (ratelimit_pages < 16)
2048 ratelimit_pages = 16;
2049}
2050
2051static int
2052ratelimit_handler(struct notifier_block *self, unsigned long action,
2053 void *hcpu)
2054{
2055
2056 switch (action & ~CPU_TASKS_FROZEN) {
2057 case CPU_ONLINE:
2058 case CPU_DEAD:
2059 writeback_set_ratelimit();
2060 return NOTIFY_OK;
2061 default:
2062 return NOTIFY_DONE;
2063 }
2064}
2065
2066static struct notifier_block ratelimit_nb = {
2067 .notifier_call = ratelimit_handler,
2068 .next = NULL,
2069};
2070
2071/*
2072 * Called early on to tune the page writeback dirty limits.
2073 *
2074 * We used to scale dirty pages according to how total memory
2075 * related to pages that could be allocated for buffers (by
2076 * comparing nr_free_buffer_pages() to vm_total_pages.
2077 *
2078 * However, that was when we used "dirty_ratio" to scale with
2079 * all memory, and we don't do that any more. "dirty_ratio"
2080 * is now applied to total non-HIGHPAGE memory (by subtracting
2081 * totalhigh_pages from vm_total_pages), and as such we can't
2082 * get into the old insane situation any more where we had
2083 * large amounts of dirty pages compared to a small amount of
2084 * non-HIGHMEM memory.
2085 *
2086 * But we might still want to scale the dirty_ratio by how
2087 * much memory the box has..
2088 */
2089void __init page_writeback_init(void)
2090{
2091 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2092
2093 writeback_set_ratelimit();
2094 register_cpu_notifier(&ratelimit_nb);
2095}
2096
2097/**
2098 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2099 * @mapping: address space structure to write
2100 * @start: starting page index
2101 * @end: ending page index (inclusive)
2102 *
2103 * This function scans the page range from @start to @end (inclusive) and tags
2104 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2105 * that write_cache_pages (or whoever calls this function) will then use
2106 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2107 * used to avoid livelocking of writeback by a process steadily creating new
2108 * dirty pages in the file (thus it is important for this function to be quick
2109 * so that it can tag pages faster than a dirtying process can create them).
2110 */
2111/*
2112 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
2113 */
2114void tag_pages_for_writeback(struct address_space *mapping,
2115 pgoff_t start, pgoff_t end)
2116{
2117#define WRITEBACK_TAG_BATCH 4096
2118 unsigned long tagged;
2119
2120 do {
2121 spin_lock_irq(&mapping->tree_lock);
2122 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
2123 &start, end, WRITEBACK_TAG_BATCH,
2124 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
2125 spin_unlock_irq(&mapping->tree_lock);
2126 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
2127 cond_resched();
2128 /* We check 'start' to handle wrapping when end == ~0UL */
2129 } while (tagged >= WRITEBACK_TAG_BATCH && start);
2130}
2131EXPORT_SYMBOL(tag_pages_for_writeback);
2132
2133/**
2134 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2135 * @mapping: address space structure to write
2136 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2137 * @writepage: function called for each page
2138 * @data: data passed to writepage function
2139 *
2140 * If a page is already under I/O, write_cache_pages() skips it, even
2141 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2142 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2143 * and msync() need to guarantee that all the data which was dirty at the time
2144 * the call was made get new I/O started against them. If wbc->sync_mode is
2145 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2146 * existing IO to complete.
2147 *
2148 * To avoid livelocks (when other process dirties new pages), we first tag
2149 * pages which should be written back with TOWRITE tag and only then start
2150 * writing them. For data-integrity sync we have to be careful so that we do
2151 * not miss some pages (e.g., because some other process has cleared TOWRITE
2152 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2153 * by the process clearing the DIRTY tag (and submitting the page for IO).
2154 */
2155int write_cache_pages(struct address_space *mapping,
2156 struct writeback_control *wbc, writepage_t writepage,
2157 void *data)
2158{
2159 int ret = 0;
2160 int done = 0;
2161 struct pagevec pvec;
2162 int nr_pages;
2163 pgoff_t uninitialized_var(writeback_index);
2164 pgoff_t index;
2165 pgoff_t end; /* Inclusive */
2166 pgoff_t done_index;
2167 int cycled;
2168 int range_whole = 0;
2169 int tag;
2170
2171 pagevec_init(&pvec, 0);
2172 if (wbc->range_cyclic) {
2173 writeback_index = mapping->writeback_index; /* prev offset */
2174 index = writeback_index;
2175 if (index == 0)
2176 cycled = 1;
2177 else
2178 cycled = 0;
2179 end = -1;
2180 } else {
2181 index = wbc->range_start >> PAGE_SHIFT;
2182 end = wbc->range_end >> PAGE_SHIFT;
2183 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2184 range_whole = 1;
2185 cycled = 1; /* ignore range_cyclic tests */
2186 }
2187 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2188 tag = PAGECACHE_TAG_TOWRITE;
2189 else
2190 tag = PAGECACHE_TAG_DIRTY;
2191retry:
2192 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2193 tag_pages_for_writeback(mapping, index, end);
2194 done_index = index;
2195 while (!done && (index <= end)) {
2196 int i;
2197
2198 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2199 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2200 if (nr_pages == 0)
2201 break;
2202
2203 for (i = 0; i < nr_pages; i++) {
2204 struct page *page = pvec.pages[i];
2205
2206 /*
2207 * At this point, the page may be truncated or
2208 * invalidated (changing page->mapping to NULL), or
2209 * even swizzled back from swapper_space to tmpfs file
2210 * mapping. However, page->index will not change
2211 * because we have a reference on the page.
2212 */
2213 if (page->index > end) {
2214 /*
2215 * can't be range_cyclic (1st pass) because
2216 * end == -1 in that case.
2217 */
2218 done = 1;
2219 break;
2220 }
2221
2222 done_index = page->index;
2223
2224 lock_page(page);
2225
2226 /*
2227 * Page truncated or invalidated. We can freely skip it
2228 * then, even for data integrity operations: the page
2229 * has disappeared concurrently, so there could be no
2230 * real expectation of this data interity operation
2231 * even if there is now a new, dirty page at the same
2232 * pagecache address.
2233 */
2234 if (unlikely(page->mapping != mapping)) {
2235continue_unlock:
2236 unlock_page(page);
2237 continue;
2238 }
2239
2240 if (!PageDirty(page)) {
2241 /* someone wrote it for us */
2242 goto continue_unlock;
2243 }
2244
2245 if (PageWriteback(page)) {
2246 if (wbc->sync_mode != WB_SYNC_NONE)
2247 wait_on_page_writeback(page);
2248 else
2249 goto continue_unlock;
2250 }
2251
2252 BUG_ON(PageWriteback(page));
2253 if (!clear_page_dirty_for_io(page))
2254 goto continue_unlock;
2255
2256 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2257 ret = (*writepage)(page, wbc, data);
2258 if (unlikely(ret)) {
2259 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2260 unlock_page(page);
2261 ret = 0;
2262 } else {
2263 /*
2264 * done_index is set past this page,
2265 * so media errors will not choke
2266 * background writeout for the entire
2267 * file. This has consequences for
2268 * range_cyclic semantics (ie. it may
2269 * not be suitable for data integrity
2270 * writeout).
2271 */
2272 done_index = page->index + 1;
2273 done = 1;
2274 break;
2275 }
2276 }
2277
2278 /*
2279 * We stop writing back only if we are not doing
2280 * integrity sync. In case of integrity sync we have to
2281 * keep going until we have written all the pages
2282 * we tagged for writeback prior to entering this loop.
2283 */
2284 if (--wbc->nr_to_write <= 0 &&
2285 wbc->sync_mode == WB_SYNC_NONE) {
2286 done = 1;
2287 break;
2288 }
2289 }
2290 pagevec_release(&pvec);
2291 cond_resched();
2292 }
2293 if (!cycled && !done) {
2294 /*
2295 * range_cyclic:
2296 * We hit the last page and there is more work to be done: wrap
2297 * back to the start of the file
2298 */
2299 cycled = 1;
2300 index = 0;
2301 end = writeback_index - 1;
2302 goto retry;
2303 }
2304 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2305 mapping->writeback_index = done_index;
2306
2307 return ret;
2308}
2309EXPORT_SYMBOL(write_cache_pages);
2310
2311/*
2312 * Function used by generic_writepages to call the real writepage
2313 * function and set the mapping flags on error
2314 */
2315static int __writepage(struct page *page, struct writeback_control *wbc,
2316 void *data)
2317{
2318 struct address_space *mapping = data;
2319 int ret = mapping->a_ops->writepage(page, wbc);
2320 mapping_set_error(mapping, ret);
2321 return ret;
2322}
2323
2324/**
2325 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2326 * @mapping: address space structure to write
2327 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2328 *
2329 * This is a library function, which implements the writepages()
2330 * address_space_operation.
2331 */
2332int generic_writepages(struct address_space *mapping,
2333 struct writeback_control *wbc)
2334{
2335 struct blk_plug plug;
2336 int ret;
2337
2338 /* deal with chardevs and other special file */
2339 if (!mapping->a_ops->writepage)
2340 return 0;
2341
2342 blk_start_plug(&plug);
2343 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2344 blk_finish_plug(&plug);
2345 return ret;
2346}
2347
2348EXPORT_SYMBOL(generic_writepages);
2349
2350int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2351{
2352 int ret;
2353
2354 if (wbc->nr_to_write <= 0)
2355 return 0;
2356 if (mapping->a_ops->writepages)
2357 ret = mapping->a_ops->writepages(mapping, wbc);
2358 else
2359 ret = generic_writepages(mapping, wbc);
2360 return ret;
2361}
2362
2363/**
2364 * write_one_page - write out a single page and optionally wait on I/O
2365 * @page: the page to write
2366 * @wait: if true, wait on writeout
2367 *
2368 * The page must be locked by the caller and will be unlocked upon return.
2369 *
2370 * write_one_page() returns a negative error code if I/O failed.
2371 */
2372int write_one_page(struct page *page, int wait)
2373{
2374 struct address_space *mapping = page->mapping;
2375 int ret = 0;
2376 struct writeback_control wbc = {
2377 .sync_mode = WB_SYNC_ALL,
2378 .nr_to_write = 1,
2379 };
2380
2381 BUG_ON(!PageLocked(page));
2382
2383 if (wait)
2384 wait_on_page_writeback(page);
2385
2386 if (clear_page_dirty_for_io(page)) {
2387 get_page(page);
2388 ret = mapping->a_ops->writepage(page, &wbc);
2389 if (ret == 0 && wait) {
2390 wait_on_page_writeback(page);
2391 if (PageError(page))
2392 ret = -EIO;
2393 }
2394 put_page(page);
2395 } else {
2396 unlock_page(page);
2397 }
2398 return ret;
2399}
2400EXPORT_SYMBOL(write_one_page);
2401
2402/*
2403 * For address_spaces which do not use buffers nor write back.
2404 */
2405int __set_page_dirty_no_writeback(struct page *page)
2406{
2407 if (!PageDirty(page))
2408 return !TestSetPageDirty(page);
2409 return 0;
2410}
2411
2412/*
2413 * Helper function for set_page_dirty family.
2414 *
2415 * Caller must hold lock_page_memcg().
2416 *
2417 * NOTE: This relies on being atomic wrt interrupts.
2418 */
2419void account_page_dirtied(struct page *page, struct address_space *mapping)
2420{
2421 struct inode *inode = mapping->host;
2422
2423 trace_writeback_dirty_page(page, mapping);
2424
2425 if (mapping_cap_account_dirty(mapping)) {
2426 struct bdi_writeback *wb;
2427
2428 inode_attach_wb(inode, page);
2429 wb = inode_to_wb(inode);
2430
2431 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_DIRTY);
2432 __inc_zone_page_state(page, NR_FILE_DIRTY);
2433 __inc_zone_page_state(page, NR_DIRTIED);
2434 __inc_wb_stat(wb, WB_RECLAIMABLE);
2435 __inc_wb_stat(wb, WB_DIRTIED);
2436 task_io_account_write(PAGE_SIZE);
2437 current->nr_dirtied++;
2438 this_cpu_inc(bdp_ratelimits);
2439 }
2440}
2441EXPORT_SYMBOL(account_page_dirtied);
2442
2443/*
2444 * Helper function for deaccounting dirty page without writeback.
2445 *
2446 * Caller must hold lock_page_memcg().
2447 */
2448void account_page_cleaned(struct page *page, struct address_space *mapping,
2449 struct bdi_writeback *wb)
2450{
2451 if (mapping_cap_account_dirty(mapping)) {
2452 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_DIRTY);
2453 dec_zone_page_state(page, NR_FILE_DIRTY);
2454 dec_wb_stat(wb, WB_RECLAIMABLE);
2455 task_io_account_cancelled_write(PAGE_SIZE);
2456 }
2457}
2458
2459/*
2460 * For address_spaces which do not use buffers. Just tag the page as dirty in
2461 * its radix tree.
2462 *
2463 * This is also used when a single buffer is being dirtied: we want to set the
2464 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2465 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2466 *
2467 * The caller must ensure this doesn't race with truncation. Most will simply
2468 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2469 * the pte lock held, which also locks out truncation.
2470 */
2471int __set_page_dirty_nobuffers(struct page *page)
2472{
2473 lock_page_memcg(page);
2474 if (!TestSetPageDirty(page)) {
2475 struct address_space *mapping = page_mapping(page);
2476 unsigned long flags;
2477
2478 if (!mapping) {
2479 unlock_page_memcg(page);
2480 return 1;
2481 }
2482
2483 spin_lock_irqsave(&mapping->tree_lock, flags);
2484 BUG_ON(page_mapping(page) != mapping);
2485 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2486 account_page_dirtied(page, mapping);
2487 radix_tree_tag_set(&mapping->page_tree, page_index(page),
2488 PAGECACHE_TAG_DIRTY);
2489 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2490 unlock_page_memcg(page);
2491
2492 if (mapping->host) {
2493 /* !PageAnon && !swapper_space */
2494 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2495 }
2496 return 1;
2497 }
2498 unlock_page_memcg(page);
2499 return 0;
2500}
2501EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2502
2503/*
2504 * Call this whenever redirtying a page, to de-account the dirty counters
2505 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2506 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2507 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2508 * control.
2509 */
2510void account_page_redirty(struct page *page)
2511{
2512 struct address_space *mapping = page->mapping;
2513
2514 if (mapping && mapping_cap_account_dirty(mapping)) {
2515 struct inode *inode = mapping->host;
2516 struct bdi_writeback *wb;
2517 bool locked;
2518
2519 wb = unlocked_inode_to_wb_begin(inode, &locked);
2520 current->nr_dirtied--;
2521 dec_zone_page_state(page, NR_DIRTIED);
2522 dec_wb_stat(wb, WB_DIRTIED);
2523 unlocked_inode_to_wb_end(inode, locked);
2524 }
2525}
2526EXPORT_SYMBOL(account_page_redirty);
2527
2528/*
2529 * When a writepage implementation decides that it doesn't want to write this
2530 * page for some reason, it should redirty the locked page via
2531 * redirty_page_for_writepage() and it should then unlock the page and return 0
2532 */
2533int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2534{
2535 int ret;
2536
2537 wbc->pages_skipped++;
2538 ret = __set_page_dirty_nobuffers(page);
2539 account_page_redirty(page);
2540 return ret;
2541}
2542EXPORT_SYMBOL(redirty_page_for_writepage);
2543
2544/*
2545 * Dirty a page.
2546 *
2547 * For pages with a mapping this should be done under the page lock
2548 * for the benefit of asynchronous memory errors who prefer a consistent
2549 * dirty state. This rule can be broken in some special cases,
2550 * but should be better not to.
2551 *
2552 * If the mapping doesn't provide a set_page_dirty a_op, then
2553 * just fall through and assume that it wants buffer_heads.
2554 */
2555int set_page_dirty(struct page *page)
2556{
2557 struct address_space *mapping = page_mapping(page);
2558
2559 if (likely(mapping)) {
2560 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2561 /*
2562 * readahead/lru_deactivate_page could remain
2563 * PG_readahead/PG_reclaim due to race with end_page_writeback
2564 * About readahead, if the page is written, the flags would be
2565 * reset. So no problem.
2566 * About lru_deactivate_page, if the page is redirty, the flag
2567 * will be reset. So no problem. but if the page is used by readahead
2568 * it will confuse readahead and make it restart the size rampup
2569 * process. But it's a trivial problem.
2570 */
2571 if (PageReclaim(page))
2572 ClearPageReclaim(page);
2573#ifdef CONFIG_BLOCK
2574 if (!spd)
2575 spd = __set_page_dirty_buffers;
2576#endif
2577 return (*spd)(page);
2578 }
2579 if (!PageDirty(page)) {
2580 if (!TestSetPageDirty(page))
2581 return 1;
2582 }
2583 return 0;
2584}
2585EXPORT_SYMBOL(set_page_dirty);
2586
2587/*
2588 * set_page_dirty() is racy if the caller has no reference against
2589 * page->mapping->host, and if the page is unlocked. This is because another
2590 * CPU could truncate the page off the mapping and then free the mapping.
2591 *
2592 * Usually, the page _is_ locked, or the caller is a user-space process which
2593 * holds a reference on the inode by having an open file.
2594 *
2595 * In other cases, the page should be locked before running set_page_dirty().
2596 */
2597int set_page_dirty_lock(struct page *page)
2598{
2599 int ret;
2600
2601 lock_page(page);
2602 ret = set_page_dirty(page);
2603 unlock_page(page);
2604 return ret;
2605}
2606EXPORT_SYMBOL(set_page_dirty_lock);
2607
2608/*
2609 * This cancels just the dirty bit on the kernel page itself, it does NOT
2610 * actually remove dirty bits on any mmap's that may be around. It also
2611 * leaves the page tagged dirty, so any sync activity will still find it on
2612 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2613 * look at the dirty bits in the VM.
2614 *
2615 * Doing this should *normally* only ever be done when a page is truncated,
2616 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2617 * this when it notices that somebody has cleaned out all the buffers on a
2618 * page without actually doing it through the VM. Can you say "ext3 is
2619 * horribly ugly"? Thought you could.
2620 */
2621void cancel_dirty_page(struct page *page)
2622{
2623 struct address_space *mapping = page_mapping(page);
2624
2625 if (mapping_cap_account_dirty(mapping)) {
2626 struct inode *inode = mapping->host;
2627 struct bdi_writeback *wb;
2628 bool locked;
2629
2630 lock_page_memcg(page);
2631 wb = unlocked_inode_to_wb_begin(inode, &locked);
2632
2633 if (TestClearPageDirty(page))
2634 account_page_cleaned(page, mapping, wb);
2635
2636 unlocked_inode_to_wb_end(inode, locked);
2637 unlock_page_memcg(page);
2638 } else {
2639 ClearPageDirty(page);
2640 }
2641}
2642EXPORT_SYMBOL(cancel_dirty_page);
2643
2644/*
2645 * Clear a page's dirty flag, while caring for dirty memory accounting.
2646 * Returns true if the page was previously dirty.
2647 *
2648 * This is for preparing to put the page under writeout. We leave the page
2649 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2650 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2651 * implementation will run either set_page_writeback() or set_page_dirty(),
2652 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2653 * back into sync.
2654 *
2655 * This incoherency between the page's dirty flag and radix-tree tag is
2656 * unfortunate, but it only exists while the page is locked.
2657 */
2658int clear_page_dirty_for_io(struct page *page)
2659{
2660 struct address_space *mapping = page_mapping(page);
2661 int ret = 0;
2662
2663 BUG_ON(!PageLocked(page));
2664
2665 if (mapping && mapping_cap_account_dirty(mapping)) {
2666 struct inode *inode = mapping->host;
2667 struct bdi_writeback *wb;
2668 bool locked;
2669
2670 /*
2671 * Yes, Virginia, this is indeed insane.
2672 *
2673 * We use this sequence to make sure that
2674 * (a) we account for dirty stats properly
2675 * (b) we tell the low-level filesystem to
2676 * mark the whole page dirty if it was
2677 * dirty in a pagetable. Only to then
2678 * (c) clean the page again and return 1 to
2679 * cause the writeback.
2680 *
2681 * This way we avoid all nasty races with the
2682 * dirty bit in multiple places and clearing
2683 * them concurrently from different threads.
2684 *
2685 * Note! Normally the "set_page_dirty(page)"
2686 * has no effect on the actual dirty bit - since
2687 * that will already usually be set. But we
2688 * need the side effects, and it can help us
2689 * avoid races.
2690 *
2691 * We basically use the page "master dirty bit"
2692 * as a serialization point for all the different
2693 * threads doing their things.
2694 */
2695 if (page_mkclean(page))
2696 set_page_dirty(page);
2697 /*
2698 * We carefully synchronise fault handlers against
2699 * installing a dirty pte and marking the page dirty
2700 * at this point. We do this by having them hold the
2701 * page lock while dirtying the page, and pages are
2702 * always locked coming in here, so we get the desired
2703 * exclusion.
2704 */
2705 wb = unlocked_inode_to_wb_begin(inode, &locked);
2706 if (TestClearPageDirty(page)) {
2707 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_DIRTY);
2708 dec_zone_page_state(page, NR_FILE_DIRTY);
2709 dec_wb_stat(wb, WB_RECLAIMABLE);
2710 ret = 1;
2711 }
2712 unlocked_inode_to_wb_end(inode, locked);
2713 return ret;
2714 }
2715 return TestClearPageDirty(page);
2716}
2717EXPORT_SYMBOL(clear_page_dirty_for_io);
2718
2719int test_clear_page_writeback(struct page *page)
2720{
2721 struct address_space *mapping = page_mapping(page);
2722 int ret;
2723
2724 lock_page_memcg(page);
2725 if (mapping) {
2726 struct inode *inode = mapping->host;
2727 struct backing_dev_info *bdi = inode_to_bdi(inode);
2728 unsigned long flags;
2729
2730 spin_lock_irqsave(&mapping->tree_lock, flags);
2731 ret = TestClearPageWriteback(page);
2732 if (ret) {
2733 radix_tree_tag_clear(&mapping->page_tree,
2734 page_index(page),
2735 PAGECACHE_TAG_WRITEBACK);
2736 if (bdi_cap_account_writeback(bdi)) {
2737 struct bdi_writeback *wb = inode_to_wb(inode);
2738
2739 __dec_wb_stat(wb, WB_WRITEBACK);
2740 __wb_writeout_inc(wb);
2741 }
2742 }
2743 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2744 } else {
2745 ret = TestClearPageWriteback(page);
2746 }
2747 if (ret) {
2748 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
2749 dec_zone_page_state(page, NR_WRITEBACK);
2750 inc_zone_page_state(page, NR_WRITTEN);
2751 }
2752 unlock_page_memcg(page);
2753 return ret;
2754}
2755
2756int __test_set_page_writeback(struct page *page, bool keep_write)
2757{
2758 struct address_space *mapping = page_mapping(page);
2759 int ret;
2760
2761 lock_page_memcg(page);
2762 if (mapping) {
2763 struct inode *inode = mapping->host;
2764 struct backing_dev_info *bdi = inode_to_bdi(inode);
2765 unsigned long flags;
2766
2767 spin_lock_irqsave(&mapping->tree_lock, flags);
2768 ret = TestSetPageWriteback(page);
2769 if (!ret) {
2770 radix_tree_tag_set(&mapping->page_tree,
2771 page_index(page),
2772 PAGECACHE_TAG_WRITEBACK);
2773 if (bdi_cap_account_writeback(bdi))
2774 __inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2775 }
2776 if (!PageDirty(page))
2777 radix_tree_tag_clear(&mapping->page_tree,
2778 page_index(page),
2779 PAGECACHE_TAG_DIRTY);
2780 if (!keep_write)
2781 radix_tree_tag_clear(&mapping->page_tree,
2782 page_index(page),
2783 PAGECACHE_TAG_TOWRITE);
2784 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2785 } else {
2786 ret = TestSetPageWriteback(page);
2787 }
2788 if (!ret) {
2789 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
2790 inc_zone_page_state(page, NR_WRITEBACK);
2791 }
2792 unlock_page_memcg(page);
2793 return ret;
2794
2795}
2796EXPORT_SYMBOL(__test_set_page_writeback);
2797
2798/*
2799 * Return true if any of the pages in the mapping are marked with the
2800 * passed tag.
2801 */
2802int mapping_tagged(struct address_space *mapping, int tag)
2803{
2804 return radix_tree_tagged(&mapping->page_tree, tag);
2805}
2806EXPORT_SYMBOL(mapping_tagged);
2807
2808/**
2809 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2810 * @page: The page to wait on.
2811 *
2812 * This function determines if the given page is related to a backing device
2813 * that requires page contents to be held stable during writeback. If so, then
2814 * it will wait for any pending writeback to complete.
2815 */
2816void wait_for_stable_page(struct page *page)
2817{
2818 if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2819 wait_on_page_writeback(page);
2820}
2821EXPORT_SYMBOL_GPL(wait_for_stable_page);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * mm/page-writeback.c
4 *
5 * Copyright (C) 2002, Linus Torvalds.
6 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
7 *
8 * Contains functions related to writing back dirty pages at the
9 * address_space level.
10 *
11 * 10Apr2002 Andrew Morton
12 * Initial version
13 */
14
15#include <linux/kernel.h>
16#include <linux/export.h>
17#include <linux/spinlock.h>
18#include <linux/fs.h>
19#include <linux/mm.h>
20#include <linux/swap.h>
21#include <linux/slab.h>
22#include <linux/pagemap.h>
23#include <linux/writeback.h>
24#include <linux/init.h>
25#include <linux/backing-dev.h>
26#include <linux/task_io_accounting_ops.h>
27#include <linux/blkdev.h>
28#include <linux/mpage.h>
29#include <linux/rmap.h>
30#include <linux/percpu.h>
31#include <linux/smp.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/syscalls.h>
35#include <linux/buffer_head.h> /* __set_page_dirty_buffers */
36#include <linux/pagevec.h>
37#include <linux/timer.h>
38#include <linux/sched/rt.h>
39#include <linux/sched/signal.h>
40#include <linux/mm_inline.h>
41#include <trace/events/writeback.h>
42
43#include "internal.h"
44
45/*
46 * Sleep at most 200ms at a time in balance_dirty_pages().
47 */
48#define MAX_PAUSE max(HZ/5, 1)
49
50/*
51 * Try to keep balance_dirty_pages() call intervals higher than this many pages
52 * by raising pause time to max_pause when falls below it.
53 */
54#define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
55
56/*
57 * Estimate write bandwidth at 200ms intervals.
58 */
59#define BANDWIDTH_INTERVAL max(HZ/5, 1)
60
61#define RATELIMIT_CALC_SHIFT 10
62
63/*
64 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
65 * will look to see if it needs to force writeback or throttling.
66 */
67static long ratelimit_pages = 32;
68
69/* The following parameters are exported via /proc/sys/vm */
70
71/*
72 * Start background writeback (via writeback threads) at this percentage
73 */
74int dirty_background_ratio = 10;
75
76/*
77 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
78 * dirty_background_ratio * the amount of dirtyable memory
79 */
80unsigned long dirty_background_bytes;
81
82/*
83 * free highmem will not be subtracted from the total free memory
84 * for calculating free ratios if vm_highmem_is_dirtyable is true
85 */
86int vm_highmem_is_dirtyable;
87
88/*
89 * The generator of dirty data starts writeback at this percentage
90 */
91int vm_dirty_ratio = 20;
92
93/*
94 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
95 * vm_dirty_ratio * the amount of dirtyable memory
96 */
97unsigned long vm_dirty_bytes;
98
99/*
100 * The interval between `kupdate'-style writebacks
101 */
102unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
103
104EXPORT_SYMBOL_GPL(dirty_writeback_interval);
105
106/*
107 * The longest time for which data is allowed to remain dirty
108 */
109unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
110
111/*
112 * Flag that makes the machine dump writes/reads and block dirtyings.
113 */
114int block_dump;
115
116/*
117 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
118 * a full sync is triggered after this time elapses without any disk activity.
119 */
120int laptop_mode;
121
122EXPORT_SYMBOL(laptop_mode);
123
124/* End of sysctl-exported parameters */
125
126struct wb_domain global_wb_domain;
127
128/* consolidated parameters for balance_dirty_pages() and its subroutines */
129struct dirty_throttle_control {
130#ifdef CONFIG_CGROUP_WRITEBACK
131 struct wb_domain *dom;
132 struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */
133#endif
134 struct bdi_writeback *wb;
135 struct fprop_local_percpu *wb_completions;
136
137 unsigned long avail; /* dirtyable */
138 unsigned long dirty; /* file_dirty + write + nfs */
139 unsigned long thresh; /* dirty threshold */
140 unsigned long bg_thresh; /* dirty background threshold */
141
142 unsigned long wb_dirty; /* per-wb counterparts */
143 unsigned long wb_thresh;
144 unsigned long wb_bg_thresh;
145
146 unsigned long pos_ratio;
147};
148
149/*
150 * Length of period for aging writeout fractions of bdis. This is an
151 * arbitrarily chosen number. The longer the period, the slower fractions will
152 * reflect changes in current writeout rate.
153 */
154#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
155
156#ifdef CONFIG_CGROUP_WRITEBACK
157
158#define GDTC_INIT(__wb) .wb = (__wb), \
159 .dom = &global_wb_domain, \
160 .wb_completions = &(__wb)->completions
161
162#define GDTC_INIT_NO_WB .dom = &global_wb_domain
163
164#define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
165 .dom = mem_cgroup_wb_domain(__wb), \
166 .wb_completions = &(__wb)->memcg_completions, \
167 .gdtc = __gdtc
168
169static bool mdtc_valid(struct dirty_throttle_control *dtc)
170{
171 return dtc->dom;
172}
173
174static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
175{
176 return dtc->dom;
177}
178
179static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
180{
181 return mdtc->gdtc;
182}
183
184static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
185{
186 return &wb->memcg_completions;
187}
188
189static void wb_min_max_ratio(struct bdi_writeback *wb,
190 unsigned long *minp, unsigned long *maxp)
191{
192 unsigned long this_bw = wb->avg_write_bandwidth;
193 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
194 unsigned long long min = wb->bdi->min_ratio;
195 unsigned long long max = wb->bdi->max_ratio;
196
197 /*
198 * @wb may already be clean by the time control reaches here and
199 * the total may not include its bw.
200 */
201 if (this_bw < tot_bw) {
202 if (min) {
203 min *= this_bw;
204 min = div64_ul(min, tot_bw);
205 }
206 if (max < 100) {
207 max *= this_bw;
208 max = div64_ul(max, tot_bw);
209 }
210 }
211
212 *minp = min;
213 *maxp = max;
214}
215
216#else /* CONFIG_CGROUP_WRITEBACK */
217
218#define GDTC_INIT(__wb) .wb = (__wb), \
219 .wb_completions = &(__wb)->completions
220#define GDTC_INIT_NO_WB
221#define MDTC_INIT(__wb, __gdtc)
222
223static bool mdtc_valid(struct dirty_throttle_control *dtc)
224{
225 return false;
226}
227
228static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
229{
230 return &global_wb_domain;
231}
232
233static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
234{
235 return NULL;
236}
237
238static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
239{
240 return NULL;
241}
242
243static void wb_min_max_ratio(struct bdi_writeback *wb,
244 unsigned long *minp, unsigned long *maxp)
245{
246 *minp = wb->bdi->min_ratio;
247 *maxp = wb->bdi->max_ratio;
248}
249
250#endif /* CONFIG_CGROUP_WRITEBACK */
251
252/*
253 * In a memory zone, there is a certain amount of pages we consider
254 * available for the page cache, which is essentially the number of
255 * free and reclaimable pages, minus some zone reserves to protect
256 * lowmem and the ability to uphold the zone's watermarks without
257 * requiring writeback.
258 *
259 * This number of dirtyable pages is the base value of which the
260 * user-configurable dirty ratio is the effective number of pages that
261 * are allowed to be actually dirtied. Per individual zone, or
262 * globally by using the sum of dirtyable pages over all zones.
263 *
264 * Because the user is allowed to specify the dirty limit globally as
265 * absolute number of bytes, calculating the per-zone dirty limit can
266 * require translating the configured limit into a percentage of
267 * global dirtyable memory first.
268 */
269
270/**
271 * node_dirtyable_memory - number of dirtyable pages in a node
272 * @pgdat: the node
273 *
274 * Return: the node's number of pages potentially available for dirty
275 * page cache. This is the base value for the per-node dirty limits.
276 */
277static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
278{
279 unsigned long nr_pages = 0;
280 int z;
281
282 for (z = 0; z < MAX_NR_ZONES; z++) {
283 struct zone *zone = pgdat->node_zones + z;
284
285 if (!populated_zone(zone))
286 continue;
287
288 nr_pages += zone_page_state(zone, NR_FREE_PAGES);
289 }
290
291 /*
292 * Pages reserved for the kernel should not be considered
293 * dirtyable, to prevent a situation where reclaim has to
294 * clean pages in order to balance the zones.
295 */
296 nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
297
298 nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
299 nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
300
301 return nr_pages;
302}
303
304static unsigned long highmem_dirtyable_memory(unsigned long total)
305{
306#ifdef CONFIG_HIGHMEM
307 int node;
308 unsigned long x = 0;
309 int i;
310
311 for_each_node_state(node, N_HIGH_MEMORY) {
312 for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
313 struct zone *z;
314 unsigned long nr_pages;
315
316 if (!is_highmem_idx(i))
317 continue;
318
319 z = &NODE_DATA(node)->node_zones[i];
320 if (!populated_zone(z))
321 continue;
322
323 nr_pages = zone_page_state(z, NR_FREE_PAGES);
324 /* watch for underflows */
325 nr_pages -= min(nr_pages, high_wmark_pages(z));
326 nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
327 nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
328 x += nr_pages;
329 }
330 }
331
332 /*
333 * Unreclaimable memory (kernel memory or anonymous memory
334 * without swap) can bring down the dirtyable pages below
335 * the zone's dirty balance reserve and the above calculation
336 * will underflow. However we still want to add in nodes
337 * which are below threshold (negative values) to get a more
338 * accurate calculation but make sure that the total never
339 * underflows.
340 */
341 if ((long)x < 0)
342 x = 0;
343
344 /*
345 * Make sure that the number of highmem pages is never larger
346 * than the number of the total dirtyable memory. This can only
347 * occur in very strange VM situations but we want to make sure
348 * that this does not occur.
349 */
350 return min(x, total);
351#else
352 return 0;
353#endif
354}
355
356/**
357 * global_dirtyable_memory - number of globally dirtyable pages
358 *
359 * Return: the global number of pages potentially available for dirty
360 * page cache. This is the base value for the global dirty limits.
361 */
362static unsigned long global_dirtyable_memory(void)
363{
364 unsigned long x;
365
366 x = global_zone_page_state(NR_FREE_PAGES);
367 /*
368 * Pages reserved for the kernel should not be considered
369 * dirtyable, to prevent a situation where reclaim has to
370 * clean pages in order to balance the zones.
371 */
372 x -= min(x, totalreserve_pages);
373
374 x += global_node_page_state(NR_INACTIVE_FILE);
375 x += global_node_page_state(NR_ACTIVE_FILE);
376
377 if (!vm_highmem_is_dirtyable)
378 x -= highmem_dirtyable_memory(x);
379
380 return x + 1; /* Ensure that we never return 0 */
381}
382
383/**
384 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
385 * @dtc: dirty_throttle_control of interest
386 *
387 * Calculate @dtc->thresh and ->bg_thresh considering
388 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
389 * must ensure that @dtc->avail is set before calling this function. The
390 * dirty limits will be lifted by 1/4 for real-time tasks.
391 */
392static 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 (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 */
458void 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 * Return: the maximum number of dirty pages allowed in a node, based
474 * on the node's dirtyable memory.
475 */
476static 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 (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 * Return: %true when the dirty pages in @pgdat are within the node's
499 * dirty limit, %false if the limit is exceeded.
500 */
501bool 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_WRITEBACK);
508
509 return nr_pages <= limit;
510}
511
512int dirty_background_ratio_handler(struct ctl_table *table, int write,
513 void *buffer, size_t *lenp, loff_t *ppos)
514{
515 int ret;
516
517 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
518 if (ret == 0 && write)
519 dirty_background_bytes = 0;
520 return ret;
521}
522
523int dirty_background_bytes_handler(struct ctl_table *table, int write,
524 void *buffer, size_t *lenp, loff_t *ppos)
525{
526 int ret;
527
528 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
529 if (ret == 0 && write)
530 dirty_background_ratio = 0;
531 return ret;
532}
533
534int dirty_ratio_handler(struct ctl_table *table, int write, void *buffer,
535 size_t *lenp, loff_t *ppos)
536{
537 int old_ratio = vm_dirty_ratio;
538 int ret;
539
540 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
541 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
542 writeback_set_ratelimit();
543 vm_dirty_bytes = 0;
544 }
545 return ret;
546}
547
548int dirty_bytes_handler(struct ctl_table *table, int write,
549 void *buffer, size_t *lenp, loff_t *ppos)
550{
551 unsigned long old_bytes = vm_dirty_bytes;
552 int ret;
553
554 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
555 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
556 writeback_set_ratelimit();
557 vm_dirty_ratio = 0;
558 }
559 return ret;
560}
561
562static unsigned long wp_next_time(unsigned long cur_time)
563{
564 cur_time += VM_COMPLETIONS_PERIOD_LEN;
565 /* 0 has a special meaning... */
566 if (!cur_time)
567 return 1;
568 return cur_time;
569}
570
571static void wb_domain_writeout_inc(struct wb_domain *dom,
572 struct fprop_local_percpu *completions,
573 unsigned int max_prop_frac)
574{
575 __fprop_inc_percpu_max(&dom->completions, completions,
576 max_prop_frac);
577 /* First event after period switching was turned off? */
578 if (unlikely(!dom->period_time)) {
579 /*
580 * We can race with other __bdi_writeout_inc calls here but
581 * it does not cause any harm since the resulting time when
582 * timer will fire and what is in writeout_period_time will be
583 * roughly the same.
584 */
585 dom->period_time = wp_next_time(jiffies);
586 mod_timer(&dom->period_timer, dom->period_time);
587 }
588}
589
590/*
591 * Increment @wb's writeout completion count and the global writeout
592 * completion count. Called from test_clear_page_writeback().
593 */
594static inline void __wb_writeout_inc(struct bdi_writeback *wb)
595{
596 struct wb_domain *cgdom;
597
598 inc_wb_stat(wb, WB_WRITTEN);
599 wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
600 wb->bdi->max_prop_frac);
601
602 cgdom = mem_cgroup_wb_domain(wb);
603 if (cgdom)
604 wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
605 wb->bdi->max_prop_frac);
606}
607
608void wb_writeout_inc(struct bdi_writeback *wb)
609{
610 unsigned long flags;
611
612 local_irq_save(flags);
613 __wb_writeout_inc(wb);
614 local_irq_restore(flags);
615}
616EXPORT_SYMBOL_GPL(wb_writeout_inc);
617
618/*
619 * On idle system, we can be called long after we scheduled because we use
620 * deferred timers so count with missed periods.
621 */
622static void writeout_period(struct timer_list *t)
623{
624 struct wb_domain *dom = from_timer(dom, t, period_timer);
625 int miss_periods = (jiffies - dom->period_time) /
626 VM_COMPLETIONS_PERIOD_LEN;
627
628 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
629 dom->period_time = wp_next_time(dom->period_time +
630 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
631 mod_timer(&dom->period_timer, dom->period_time);
632 } else {
633 /*
634 * Aging has zeroed all fractions. Stop wasting CPU on period
635 * updates.
636 */
637 dom->period_time = 0;
638 }
639}
640
641int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
642{
643 memset(dom, 0, sizeof(*dom));
644
645 spin_lock_init(&dom->lock);
646
647 timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
648
649 dom->dirty_limit_tstamp = jiffies;
650
651 return fprop_global_init(&dom->completions, gfp);
652}
653
654#ifdef CONFIG_CGROUP_WRITEBACK
655void wb_domain_exit(struct wb_domain *dom)
656{
657 del_timer_sync(&dom->period_timer);
658 fprop_global_destroy(&dom->completions);
659}
660#endif
661
662/*
663 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
664 * registered backing devices, which, for obvious reasons, can not
665 * exceed 100%.
666 */
667static unsigned int bdi_min_ratio;
668
669int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
670{
671 int ret = 0;
672
673 spin_lock_bh(&bdi_lock);
674 if (min_ratio > bdi->max_ratio) {
675 ret = -EINVAL;
676 } else {
677 min_ratio -= bdi->min_ratio;
678 if (bdi_min_ratio + min_ratio < 100) {
679 bdi_min_ratio += min_ratio;
680 bdi->min_ratio += min_ratio;
681 } else {
682 ret = -EINVAL;
683 }
684 }
685 spin_unlock_bh(&bdi_lock);
686
687 return ret;
688}
689
690int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
691{
692 int ret = 0;
693
694 if (max_ratio > 100)
695 return -EINVAL;
696
697 spin_lock_bh(&bdi_lock);
698 if (bdi->min_ratio > max_ratio) {
699 ret = -EINVAL;
700 } else {
701 bdi->max_ratio = max_ratio;
702 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
703 }
704 spin_unlock_bh(&bdi_lock);
705
706 return ret;
707}
708EXPORT_SYMBOL(bdi_set_max_ratio);
709
710static unsigned long dirty_freerun_ceiling(unsigned long thresh,
711 unsigned long bg_thresh)
712{
713 return (thresh + bg_thresh) / 2;
714}
715
716static unsigned long hard_dirty_limit(struct wb_domain *dom,
717 unsigned long thresh)
718{
719 return max(thresh, dom->dirty_limit);
720}
721
722/*
723 * Memory which can be further allocated to a memcg domain is capped by
724 * system-wide clean memory excluding the amount being used in the domain.
725 */
726static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
727 unsigned long filepages, unsigned long headroom)
728{
729 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
730 unsigned long clean = filepages - min(filepages, mdtc->dirty);
731 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
732 unsigned long other_clean = global_clean - min(global_clean, clean);
733
734 mdtc->avail = filepages + min(headroom, other_clean);
735}
736
737/**
738 * __wb_calc_thresh - @wb's share of dirty throttling threshold
739 * @dtc: dirty_throttle_context of interest
740 *
741 * Note that balance_dirty_pages() will only seriously take it as a hard limit
742 * when sleeping max_pause per page is not enough to keep the dirty pages under
743 * control. For example, when the device is completely stalled due to some error
744 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
745 * In the other normal situations, it acts more gently by throttling the tasks
746 * more (rather than completely block them) when the wb dirty pages go high.
747 *
748 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
749 * - starving fast devices
750 * - piling up dirty pages (that will take long time to sync) on slow devices
751 *
752 * The wb's share of dirty limit will be adapting to its throughput and
753 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
754 *
755 * Return: @wb's dirty limit in pages. The term "dirty" in the context of
756 * dirty balancing includes all PG_dirty and PG_writeback pages.
757 */
758static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
759{
760 struct wb_domain *dom = dtc_dom(dtc);
761 unsigned long thresh = dtc->thresh;
762 u64 wb_thresh;
763 unsigned long numerator, denominator;
764 unsigned long wb_min_ratio, wb_max_ratio;
765
766 /*
767 * Calculate this BDI's share of the thresh ratio.
768 */
769 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
770 &numerator, &denominator);
771
772 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
773 wb_thresh *= numerator;
774 wb_thresh = div64_ul(wb_thresh, denominator);
775
776 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
777
778 wb_thresh += (thresh * wb_min_ratio) / 100;
779 if (wb_thresh > (thresh * wb_max_ratio) / 100)
780 wb_thresh = thresh * wb_max_ratio / 100;
781
782 return wb_thresh;
783}
784
785unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
786{
787 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
788 .thresh = thresh };
789 return __wb_calc_thresh(&gdtc);
790}
791
792/*
793 * setpoint - dirty 3
794 * f(dirty) := 1.0 + (----------------)
795 * limit - setpoint
796 *
797 * it's a 3rd order polynomial that subjects to
798 *
799 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
800 * (2) f(setpoint) = 1.0 => the balance point
801 * (3) f(limit) = 0 => the hard limit
802 * (4) df/dx <= 0 => negative feedback control
803 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
804 * => fast response on large errors; small oscillation near setpoint
805 */
806static long long pos_ratio_polynom(unsigned long setpoint,
807 unsigned long dirty,
808 unsigned long limit)
809{
810 long long pos_ratio;
811 long x;
812
813 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
814 (limit - setpoint) | 1);
815 pos_ratio = x;
816 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
817 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
818 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
819
820 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
821}
822
823/*
824 * Dirty position control.
825 *
826 * (o) global/bdi setpoints
827 *
828 * We want the dirty pages be balanced around the global/wb setpoints.
829 * When the number of dirty pages is higher/lower than the setpoint, the
830 * dirty position control ratio (and hence task dirty ratelimit) will be
831 * decreased/increased to bring the dirty pages back to the setpoint.
832 *
833 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
834 *
835 * if (dirty < setpoint) scale up pos_ratio
836 * if (dirty > setpoint) scale down pos_ratio
837 *
838 * if (wb_dirty < wb_setpoint) scale up pos_ratio
839 * if (wb_dirty > wb_setpoint) scale down pos_ratio
840 *
841 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
842 *
843 * (o) global control line
844 *
845 * ^ pos_ratio
846 * |
847 * | |<===== global dirty control scope ======>|
848 * 2.0 .............*
849 * | .*
850 * | . *
851 * | . *
852 * | . *
853 * | . *
854 * | . *
855 * 1.0 ................................*
856 * | . . *
857 * | . . *
858 * | . . *
859 * | . . *
860 * | . . *
861 * 0 +------------.------------------.----------------------*------------->
862 * freerun^ setpoint^ limit^ dirty pages
863 *
864 * (o) wb control line
865 *
866 * ^ pos_ratio
867 * |
868 * | *
869 * | *
870 * | *
871 * | *
872 * | * |<=========== span ============>|
873 * 1.0 .......................*
874 * | . *
875 * | . *
876 * | . *
877 * | . *
878 * | . *
879 * | . *
880 * | . *
881 * | . *
882 * | . *
883 * | . *
884 * | . *
885 * 1/4 ...............................................* * * * * * * * * * * *
886 * | . .
887 * | . .
888 * | . .
889 * 0 +----------------------.-------------------------------.------------->
890 * wb_setpoint^ x_intercept^
891 *
892 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
893 * be smoothly throttled down to normal if it starts high in situations like
894 * - start writing to a slow SD card and a fast disk at the same time. The SD
895 * card's wb_dirty may rush to many times higher than wb_setpoint.
896 * - the wb dirty thresh drops quickly due to change of JBOD workload
897 */
898static void wb_position_ratio(struct dirty_throttle_control *dtc)
899{
900 struct bdi_writeback *wb = dtc->wb;
901 unsigned long write_bw = wb->avg_write_bandwidth;
902 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
903 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
904 unsigned long wb_thresh = dtc->wb_thresh;
905 unsigned long x_intercept;
906 unsigned long setpoint; /* dirty pages' target balance point */
907 unsigned long wb_setpoint;
908 unsigned long span;
909 long long pos_ratio; /* for scaling up/down the rate limit */
910 long x;
911
912 dtc->pos_ratio = 0;
913
914 if (unlikely(dtc->dirty >= limit))
915 return;
916
917 /*
918 * global setpoint
919 *
920 * See comment for pos_ratio_polynom().
921 */
922 setpoint = (freerun + limit) / 2;
923 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
924
925 /*
926 * The strictlimit feature is a tool preventing mistrusted filesystems
927 * from growing a large number of dirty pages before throttling. For
928 * such filesystems balance_dirty_pages always checks wb counters
929 * against wb limits. Even if global "nr_dirty" is under "freerun".
930 * This is especially important for fuse which sets bdi->max_ratio to
931 * 1% by default. Without strictlimit feature, fuse writeback may
932 * consume arbitrary amount of RAM because it is accounted in
933 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
934 *
935 * Here, in wb_position_ratio(), we calculate pos_ratio based on
936 * two values: wb_dirty and wb_thresh. Let's consider an example:
937 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
938 * limits are set by default to 10% and 20% (background and throttle).
939 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
940 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
941 * about ~6K pages (as the average of background and throttle wb
942 * limits). The 3rd order polynomial will provide positive feedback if
943 * wb_dirty is under wb_setpoint and vice versa.
944 *
945 * Note, that we cannot use global counters in these calculations
946 * because we want to throttle process writing to a strictlimit wb
947 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
948 * in the example above).
949 */
950 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
951 long long wb_pos_ratio;
952
953 if (dtc->wb_dirty < 8) {
954 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
955 2 << RATELIMIT_CALC_SHIFT);
956 return;
957 }
958
959 if (dtc->wb_dirty >= wb_thresh)
960 return;
961
962 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
963 dtc->wb_bg_thresh);
964
965 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
966 return;
967
968 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
969 wb_thresh);
970
971 /*
972 * Typically, for strictlimit case, wb_setpoint << setpoint
973 * and pos_ratio >> wb_pos_ratio. In the other words global
974 * state ("dirty") is not limiting factor and we have to
975 * make decision based on wb counters. But there is an
976 * important case when global pos_ratio should get precedence:
977 * global limits are exceeded (e.g. due to activities on other
978 * wb's) while given strictlimit wb is below limit.
979 *
980 * "pos_ratio * wb_pos_ratio" would work for the case above,
981 * but it would look too non-natural for the case of all
982 * activity in the system coming from a single strictlimit wb
983 * with bdi->max_ratio == 100%.
984 *
985 * Note that min() below somewhat changes the dynamics of the
986 * control system. Normally, pos_ratio value can be well over 3
987 * (when globally we are at freerun and wb is well below wb
988 * setpoint). Now the maximum pos_ratio in the same situation
989 * is 2. We might want to tweak this if we observe the control
990 * system is too slow to adapt.
991 */
992 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
993 return;
994 }
995
996 /*
997 * We have computed basic pos_ratio above based on global situation. If
998 * the wb is over/under its share of dirty pages, we want to scale
999 * pos_ratio further down/up. That is done by the following mechanism.
1000 */
1001
1002 /*
1003 * wb setpoint
1004 *
1005 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1006 *
1007 * x_intercept - wb_dirty
1008 * := --------------------------
1009 * x_intercept - wb_setpoint
1010 *
1011 * The main wb control line is a linear function that subjects to
1012 *
1013 * (1) f(wb_setpoint) = 1.0
1014 * (2) k = - 1 / (8 * write_bw) (in single wb case)
1015 * or equally: x_intercept = wb_setpoint + 8 * write_bw
1016 *
1017 * For single wb case, the dirty pages are observed to fluctuate
1018 * regularly within range
1019 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1020 * for various filesystems, where (2) can yield in a reasonable 12.5%
1021 * fluctuation range for pos_ratio.
1022 *
1023 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1024 * own size, so move the slope over accordingly and choose a slope that
1025 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1026 */
1027 if (unlikely(wb_thresh > dtc->thresh))
1028 wb_thresh = dtc->thresh;
1029 /*
1030 * It's very possible that wb_thresh is close to 0 not because the
1031 * device is slow, but that it has remained inactive for long time.
1032 * Honour such devices a reasonable good (hopefully IO efficient)
1033 * threshold, so that the occasional writes won't be blocked and active
1034 * writes can rampup the threshold quickly.
1035 */
1036 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1037 /*
1038 * scale global setpoint to wb's:
1039 * wb_setpoint = setpoint * wb_thresh / thresh
1040 */
1041 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1042 wb_setpoint = setpoint * (u64)x >> 16;
1043 /*
1044 * Use span=(8*write_bw) in single wb case as indicated by
1045 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1046 *
1047 * wb_thresh thresh - wb_thresh
1048 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1049 * thresh thresh
1050 */
1051 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1052 x_intercept = wb_setpoint + span;
1053
1054 if (dtc->wb_dirty < x_intercept - span / 4) {
1055 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1056 (x_intercept - wb_setpoint) | 1);
1057 } else
1058 pos_ratio /= 4;
1059
1060 /*
1061 * wb reserve area, safeguard against dirty pool underrun and disk idle
1062 * It may push the desired control point of global dirty pages higher
1063 * than setpoint.
1064 */
1065 x_intercept = wb_thresh / 2;
1066 if (dtc->wb_dirty < x_intercept) {
1067 if (dtc->wb_dirty > x_intercept / 8)
1068 pos_ratio = div_u64(pos_ratio * x_intercept,
1069 dtc->wb_dirty);
1070 else
1071 pos_ratio *= 8;
1072 }
1073
1074 dtc->pos_ratio = pos_ratio;
1075}
1076
1077static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1078 unsigned long elapsed,
1079 unsigned long written)
1080{
1081 const unsigned long period = roundup_pow_of_two(3 * HZ);
1082 unsigned long avg = wb->avg_write_bandwidth;
1083 unsigned long old = wb->write_bandwidth;
1084 u64 bw;
1085
1086 /*
1087 * bw = written * HZ / elapsed
1088 *
1089 * bw * elapsed + write_bandwidth * (period - elapsed)
1090 * write_bandwidth = ---------------------------------------------------
1091 * period
1092 *
1093 * @written may have decreased due to account_page_redirty().
1094 * Avoid underflowing @bw calculation.
1095 */
1096 bw = written - min(written, wb->written_stamp);
1097 bw *= HZ;
1098 if (unlikely(elapsed > period)) {
1099 bw = div64_ul(bw, elapsed);
1100 avg = bw;
1101 goto out;
1102 }
1103 bw += (u64)wb->write_bandwidth * (period - elapsed);
1104 bw >>= ilog2(period);
1105
1106 /*
1107 * one more level of smoothing, for filtering out sudden spikes
1108 */
1109 if (avg > old && old >= (unsigned long)bw)
1110 avg -= (avg - old) >> 3;
1111
1112 if (avg < old && old <= (unsigned long)bw)
1113 avg += (old - avg) >> 3;
1114
1115out:
1116 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1117 avg = max(avg, 1LU);
1118 if (wb_has_dirty_io(wb)) {
1119 long delta = avg - wb->avg_write_bandwidth;
1120 WARN_ON_ONCE(atomic_long_add_return(delta,
1121 &wb->bdi->tot_write_bandwidth) <= 0);
1122 }
1123 wb->write_bandwidth = bw;
1124 wb->avg_write_bandwidth = avg;
1125}
1126
1127static void update_dirty_limit(struct dirty_throttle_control *dtc)
1128{
1129 struct wb_domain *dom = dtc_dom(dtc);
1130 unsigned long thresh = dtc->thresh;
1131 unsigned long limit = dom->dirty_limit;
1132
1133 /*
1134 * Follow up in one step.
1135 */
1136 if (limit < thresh) {
1137 limit = thresh;
1138 goto update;
1139 }
1140
1141 /*
1142 * Follow down slowly. Use the higher one as the target, because thresh
1143 * may drop below dirty. This is exactly the reason to introduce
1144 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1145 */
1146 thresh = max(thresh, dtc->dirty);
1147 if (limit > thresh) {
1148 limit -= (limit - thresh) >> 5;
1149 goto update;
1150 }
1151 return;
1152update:
1153 dom->dirty_limit = limit;
1154}
1155
1156static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
1157 unsigned long now)
1158{
1159 struct wb_domain *dom = dtc_dom(dtc);
1160
1161 /*
1162 * check locklessly first to optimize away locking for the most time
1163 */
1164 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1165 return;
1166
1167 spin_lock(&dom->lock);
1168 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1169 update_dirty_limit(dtc);
1170 dom->dirty_limit_tstamp = now;
1171 }
1172 spin_unlock(&dom->lock);
1173}
1174
1175/*
1176 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1177 *
1178 * Normal wb tasks will be curbed at or below it in long term.
1179 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1180 */
1181static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1182 unsigned long dirtied,
1183 unsigned long elapsed)
1184{
1185 struct bdi_writeback *wb = dtc->wb;
1186 unsigned long dirty = dtc->dirty;
1187 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1188 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1189 unsigned long setpoint = (freerun + limit) / 2;
1190 unsigned long write_bw = wb->avg_write_bandwidth;
1191 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1192 unsigned long dirty_rate;
1193 unsigned long task_ratelimit;
1194 unsigned long balanced_dirty_ratelimit;
1195 unsigned long step;
1196 unsigned long x;
1197 unsigned long shift;
1198
1199 /*
1200 * The dirty rate will match the writeout rate in long term, except
1201 * when dirty pages are truncated by userspace or re-dirtied by FS.
1202 */
1203 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1204
1205 /*
1206 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1207 */
1208 task_ratelimit = (u64)dirty_ratelimit *
1209 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1210 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1211
1212 /*
1213 * A linear estimation of the "balanced" throttle rate. The theory is,
1214 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1215 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1216 * formula will yield the balanced rate limit (write_bw / N).
1217 *
1218 * Note that the expanded form is not a pure rate feedback:
1219 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1220 * but also takes pos_ratio into account:
1221 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1222 *
1223 * (1) is not realistic because pos_ratio also takes part in balancing
1224 * the dirty rate. Consider the state
1225 * pos_ratio = 0.5 (3)
1226 * rate = 2 * (write_bw / N) (4)
1227 * If (1) is used, it will stuck in that state! Because each dd will
1228 * be throttled at
1229 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1230 * yielding
1231 * dirty_rate = N * task_ratelimit = write_bw (6)
1232 * put (6) into (1) we get
1233 * rate_(i+1) = rate_(i) (7)
1234 *
1235 * So we end up using (2) to always keep
1236 * rate_(i+1) ~= (write_bw / N) (8)
1237 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1238 * pos_ratio is able to drive itself to 1.0, which is not only where
1239 * the dirty count meet the setpoint, but also where the slope of
1240 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1241 */
1242 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1243 dirty_rate | 1);
1244 /*
1245 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1246 */
1247 if (unlikely(balanced_dirty_ratelimit > write_bw))
1248 balanced_dirty_ratelimit = write_bw;
1249
1250 /*
1251 * We could safely do this and return immediately:
1252 *
1253 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1254 *
1255 * However to get a more stable dirty_ratelimit, the below elaborated
1256 * code makes use of task_ratelimit to filter out singular points and
1257 * limit the step size.
1258 *
1259 * The below code essentially only uses the relative value of
1260 *
1261 * task_ratelimit - dirty_ratelimit
1262 * = (pos_ratio - 1) * dirty_ratelimit
1263 *
1264 * which reflects the direction and size of dirty position error.
1265 */
1266
1267 /*
1268 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1269 * task_ratelimit is on the same side of dirty_ratelimit, too.
1270 * For example, when
1271 * - dirty_ratelimit > balanced_dirty_ratelimit
1272 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1273 * lowering dirty_ratelimit will help meet both the position and rate
1274 * control targets. Otherwise, don't update dirty_ratelimit if it will
1275 * only help meet the rate target. After all, what the users ultimately
1276 * feel and care are stable dirty rate and small position error.
1277 *
1278 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1279 * and filter out the singular points of balanced_dirty_ratelimit. Which
1280 * keeps jumping around randomly and can even leap far away at times
1281 * due to the small 200ms estimation period of dirty_rate (we want to
1282 * keep that period small to reduce time lags).
1283 */
1284 step = 0;
1285
1286 /*
1287 * For strictlimit case, calculations above were based on wb counters
1288 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1289 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1290 * Hence, to calculate "step" properly, we have to use wb_dirty as
1291 * "dirty" and wb_setpoint as "setpoint".
1292 *
1293 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1294 * it's possible that wb_thresh is close to zero due to inactivity
1295 * of backing device.
1296 */
1297 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1298 dirty = dtc->wb_dirty;
1299 if (dtc->wb_dirty < 8)
1300 setpoint = dtc->wb_dirty + 1;
1301 else
1302 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1303 }
1304
1305 if (dirty < setpoint) {
1306 x = min3(wb->balanced_dirty_ratelimit,
1307 balanced_dirty_ratelimit, task_ratelimit);
1308 if (dirty_ratelimit < x)
1309 step = x - dirty_ratelimit;
1310 } else {
1311 x = max3(wb->balanced_dirty_ratelimit,
1312 balanced_dirty_ratelimit, task_ratelimit);
1313 if (dirty_ratelimit > x)
1314 step = dirty_ratelimit - x;
1315 }
1316
1317 /*
1318 * Don't pursue 100% rate matching. It's impossible since the balanced
1319 * rate itself is constantly fluctuating. So decrease the track speed
1320 * when it gets close to the target. Helps eliminate pointless tremors.
1321 */
1322 shift = dirty_ratelimit / (2 * step + 1);
1323 if (shift < BITS_PER_LONG)
1324 step = DIV_ROUND_UP(step >> shift, 8);
1325 else
1326 step = 0;
1327
1328 if (dirty_ratelimit < balanced_dirty_ratelimit)
1329 dirty_ratelimit += step;
1330 else
1331 dirty_ratelimit -= step;
1332
1333 wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1334 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1335
1336 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1337}
1338
1339static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1340 struct dirty_throttle_control *mdtc,
1341 unsigned long start_time,
1342 bool update_ratelimit)
1343{
1344 struct bdi_writeback *wb = gdtc->wb;
1345 unsigned long now = jiffies;
1346 unsigned long elapsed = now - wb->bw_time_stamp;
1347 unsigned long dirtied;
1348 unsigned long written;
1349
1350 lockdep_assert_held(&wb->list_lock);
1351
1352 /*
1353 * rate-limit, only update once every 200ms.
1354 */
1355 if (elapsed < BANDWIDTH_INTERVAL)
1356 return;
1357
1358 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1359 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1360
1361 /*
1362 * Skip quiet periods when disk bandwidth is under-utilized.
1363 * (at least 1s idle time between two flusher runs)
1364 */
1365 if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1366 goto snapshot;
1367
1368 if (update_ratelimit) {
1369 domain_update_bandwidth(gdtc, now);
1370 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1371
1372 /*
1373 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1374 * compiler has no way to figure that out. Help it.
1375 */
1376 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1377 domain_update_bandwidth(mdtc, now);
1378 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1379 }
1380 }
1381 wb_update_write_bandwidth(wb, elapsed, written);
1382
1383snapshot:
1384 wb->dirtied_stamp = dirtied;
1385 wb->written_stamp = written;
1386 wb->bw_time_stamp = now;
1387}
1388
1389void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1390{
1391 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1392
1393 __wb_update_bandwidth(&gdtc, NULL, start_time, false);
1394}
1395
1396/*
1397 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1398 * will look to see if it needs to start dirty throttling.
1399 *
1400 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1401 * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1402 * (the number of pages we may dirty without exceeding the dirty limits).
1403 */
1404static unsigned long dirty_poll_interval(unsigned long dirty,
1405 unsigned long thresh)
1406{
1407 if (thresh > dirty)
1408 return 1UL << (ilog2(thresh - dirty) >> 1);
1409
1410 return 1;
1411}
1412
1413static unsigned long wb_max_pause(struct bdi_writeback *wb,
1414 unsigned long wb_dirty)
1415{
1416 unsigned long bw = wb->avg_write_bandwidth;
1417 unsigned long t;
1418
1419 /*
1420 * Limit pause time for small memory systems. If sleeping for too long
1421 * time, a small pool of dirty/writeback pages may go empty and disk go
1422 * idle.
1423 *
1424 * 8 serves as the safety ratio.
1425 */
1426 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1427 t++;
1428
1429 return min_t(unsigned long, t, MAX_PAUSE);
1430}
1431
1432static long wb_min_pause(struct bdi_writeback *wb,
1433 long max_pause,
1434 unsigned long task_ratelimit,
1435 unsigned long dirty_ratelimit,
1436 int *nr_dirtied_pause)
1437{
1438 long hi = ilog2(wb->avg_write_bandwidth);
1439 long lo = ilog2(wb->dirty_ratelimit);
1440 long t; /* target pause */
1441 long pause; /* estimated next pause */
1442 int pages; /* target nr_dirtied_pause */
1443
1444 /* target for 10ms pause on 1-dd case */
1445 t = max(1, HZ / 100);
1446
1447 /*
1448 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1449 * overheads.
1450 *
1451 * (N * 10ms) on 2^N concurrent tasks.
1452 */
1453 if (hi > lo)
1454 t += (hi - lo) * (10 * HZ) / 1024;
1455
1456 /*
1457 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1458 * on the much more stable dirty_ratelimit. However the next pause time
1459 * will be computed based on task_ratelimit and the two rate limits may
1460 * depart considerably at some time. Especially if task_ratelimit goes
1461 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1462 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1463 * result task_ratelimit won't be executed faithfully, which could
1464 * eventually bring down dirty_ratelimit.
1465 *
1466 * We apply two rules to fix it up:
1467 * 1) try to estimate the next pause time and if necessary, use a lower
1468 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1469 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1470 * 2) limit the target pause time to max_pause/2, so that the normal
1471 * small fluctuations of task_ratelimit won't trigger rule (1) and
1472 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1473 */
1474 t = min(t, 1 + max_pause / 2);
1475 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1476
1477 /*
1478 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1479 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1480 * When the 16 consecutive reads are often interrupted by some dirty
1481 * throttling pause during the async writes, cfq will go into idles
1482 * (deadline is fine). So push nr_dirtied_pause as high as possible
1483 * until reaches DIRTY_POLL_THRESH=32 pages.
1484 */
1485 if (pages < DIRTY_POLL_THRESH) {
1486 t = max_pause;
1487 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1488 if (pages > DIRTY_POLL_THRESH) {
1489 pages = DIRTY_POLL_THRESH;
1490 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1491 }
1492 }
1493
1494 pause = HZ * pages / (task_ratelimit + 1);
1495 if (pause > max_pause) {
1496 t = max_pause;
1497 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1498 }
1499
1500 *nr_dirtied_pause = pages;
1501 /*
1502 * The minimal pause time will normally be half the target pause time.
1503 */
1504 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1505}
1506
1507static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1508{
1509 struct bdi_writeback *wb = dtc->wb;
1510 unsigned long wb_reclaimable;
1511
1512 /*
1513 * wb_thresh is not treated as some limiting factor as
1514 * dirty_thresh, due to reasons
1515 * - in JBOD setup, wb_thresh can fluctuate a lot
1516 * - in a system with HDD and USB key, the USB key may somehow
1517 * go into state (wb_dirty >> wb_thresh) either because
1518 * wb_dirty starts high, or because wb_thresh drops low.
1519 * In this case we don't want to hard throttle the USB key
1520 * dirtiers for 100 seconds until wb_dirty drops under
1521 * wb_thresh. Instead the auxiliary wb control line in
1522 * wb_position_ratio() will let the dirtier task progress
1523 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1524 */
1525 dtc->wb_thresh = __wb_calc_thresh(dtc);
1526 dtc->wb_bg_thresh = dtc->thresh ?
1527 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1528
1529 /*
1530 * In order to avoid the stacked BDI deadlock we need
1531 * to ensure we accurately count the 'dirty' pages when
1532 * the threshold is low.
1533 *
1534 * Otherwise it would be possible to get thresh+n pages
1535 * reported dirty, even though there are thresh-m pages
1536 * actually dirty; with m+n sitting in the percpu
1537 * deltas.
1538 */
1539 if (dtc->wb_thresh < 2 * wb_stat_error()) {
1540 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1541 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1542 } else {
1543 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1544 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1545 }
1546}
1547
1548/*
1549 * balance_dirty_pages() must be called by processes which are generating dirty
1550 * data. It looks at the number of dirty pages in the machine and will force
1551 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1552 * If we're over `background_thresh' then the writeback threads are woken to
1553 * perform some writeout.
1554 */
1555static void balance_dirty_pages(struct bdi_writeback *wb,
1556 unsigned long pages_dirtied)
1557{
1558 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1559 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1560 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1561 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1562 &mdtc_stor : NULL;
1563 struct dirty_throttle_control *sdtc;
1564 unsigned long nr_reclaimable; /* = file_dirty */
1565 long period;
1566 long pause;
1567 long max_pause;
1568 long min_pause;
1569 int nr_dirtied_pause;
1570 bool dirty_exceeded = false;
1571 unsigned long task_ratelimit;
1572 unsigned long dirty_ratelimit;
1573 struct backing_dev_info *bdi = wb->bdi;
1574 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1575 unsigned long start_time = jiffies;
1576
1577 for (;;) {
1578 unsigned long now = jiffies;
1579 unsigned long dirty, thresh, bg_thresh;
1580 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1581 unsigned long m_thresh = 0;
1582 unsigned long m_bg_thresh = 0;
1583
1584 nr_reclaimable = global_node_page_state(NR_FILE_DIRTY);
1585 gdtc->avail = global_dirtyable_memory();
1586 gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1587
1588 domain_dirty_limits(gdtc);
1589
1590 if (unlikely(strictlimit)) {
1591 wb_dirty_limits(gdtc);
1592
1593 dirty = gdtc->wb_dirty;
1594 thresh = gdtc->wb_thresh;
1595 bg_thresh = gdtc->wb_bg_thresh;
1596 } else {
1597 dirty = gdtc->dirty;
1598 thresh = gdtc->thresh;
1599 bg_thresh = gdtc->bg_thresh;
1600 }
1601
1602 if (mdtc) {
1603 unsigned long filepages, headroom, writeback;
1604
1605 /*
1606 * If @wb belongs to !root memcg, repeat the same
1607 * basic calculations for the memcg domain.
1608 */
1609 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1610 &mdtc->dirty, &writeback);
1611 mdtc->dirty += writeback;
1612 mdtc_calc_avail(mdtc, filepages, headroom);
1613
1614 domain_dirty_limits(mdtc);
1615
1616 if (unlikely(strictlimit)) {
1617 wb_dirty_limits(mdtc);
1618 m_dirty = mdtc->wb_dirty;
1619 m_thresh = mdtc->wb_thresh;
1620 m_bg_thresh = mdtc->wb_bg_thresh;
1621 } else {
1622 m_dirty = mdtc->dirty;
1623 m_thresh = mdtc->thresh;
1624 m_bg_thresh = mdtc->bg_thresh;
1625 }
1626 }
1627
1628 /*
1629 * Throttle it only when the background writeback cannot
1630 * catch-up. This avoids (excessively) small writeouts
1631 * when the wb limits are ramping up in case of !strictlimit.
1632 *
1633 * In strictlimit case make decision based on the wb counters
1634 * and limits. Small writeouts when the wb limits are ramping
1635 * up are the price we consciously pay for strictlimit-ing.
1636 *
1637 * If memcg domain is in effect, @dirty should be under
1638 * both global and memcg freerun ceilings.
1639 */
1640 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1641 (!mdtc ||
1642 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1643 unsigned long intv;
1644 unsigned long m_intv;
1645
1646free_running:
1647 intv = dirty_poll_interval(dirty, thresh);
1648 m_intv = ULONG_MAX;
1649
1650 current->dirty_paused_when = now;
1651 current->nr_dirtied = 0;
1652 if (mdtc)
1653 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1654 current->nr_dirtied_pause = min(intv, m_intv);
1655 break;
1656 }
1657
1658 if (unlikely(!writeback_in_progress(wb)))
1659 wb_start_background_writeback(wb);
1660
1661 mem_cgroup_flush_foreign(wb);
1662
1663 /*
1664 * Calculate global domain's pos_ratio and select the
1665 * global dtc by default.
1666 */
1667 if (!strictlimit) {
1668 wb_dirty_limits(gdtc);
1669
1670 if ((current->flags & PF_LOCAL_THROTTLE) &&
1671 gdtc->wb_dirty <
1672 dirty_freerun_ceiling(gdtc->wb_thresh,
1673 gdtc->wb_bg_thresh))
1674 /*
1675 * LOCAL_THROTTLE tasks must not be throttled
1676 * when below the per-wb freerun ceiling.
1677 */
1678 goto free_running;
1679 }
1680
1681 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1682 ((gdtc->dirty > gdtc->thresh) || strictlimit);
1683
1684 wb_position_ratio(gdtc);
1685 sdtc = gdtc;
1686
1687 if (mdtc) {
1688 /*
1689 * If memcg domain is in effect, calculate its
1690 * pos_ratio. @wb should satisfy constraints from
1691 * both global and memcg domains. Choose the one
1692 * w/ lower pos_ratio.
1693 */
1694 if (!strictlimit) {
1695 wb_dirty_limits(mdtc);
1696
1697 if ((current->flags & PF_LOCAL_THROTTLE) &&
1698 mdtc->wb_dirty <
1699 dirty_freerun_ceiling(mdtc->wb_thresh,
1700 mdtc->wb_bg_thresh))
1701 /*
1702 * LOCAL_THROTTLE tasks must not be
1703 * throttled when below the per-wb
1704 * freerun ceiling.
1705 */
1706 goto free_running;
1707 }
1708 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1709 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1710
1711 wb_position_ratio(mdtc);
1712 if (mdtc->pos_ratio < gdtc->pos_ratio)
1713 sdtc = mdtc;
1714 }
1715
1716 if (dirty_exceeded && !wb->dirty_exceeded)
1717 wb->dirty_exceeded = 1;
1718
1719 if (time_is_before_jiffies(wb->bw_time_stamp +
1720 BANDWIDTH_INTERVAL)) {
1721 spin_lock(&wb->list_lock);
1722 __wb_update_bandwidth(gdtc, mdtc, start_time, true);
1723 spin_unlock(&wb->list_lock);
1724 }
1725
1726 /* throttle according to the chosen dtc */
1727 dirty_ratelimit = wb->dirty_ratelimit;
1728 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1729 RATELIMIT_CALC_SHIFT;
1730 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1731 min_pause = wb_min_pause(wb, max_pause,
1732 task_ratelimit, dirty_ratelimit,
1733 &nr_dirtied_pause);
1734
1735 if (unlikely(task_ratelimit == 0)) {
1736 period = max_pause;
1737 pause = max_pause;
1738 goto pause;
1739 }
1740 period = HZ * pages_dirtied / task_ratelimit;
1741 pause = period;
1742 if (current->dirty_paused_when)
1743 pause -= now - current->dirty_paused_when;
1744 /*
1745 * For less than 1s think time (ext3/4 may block the dirtier
1746 * for up to 800ms from time to time on 1-HDD; so does xfs,
1747 * however at much less frequency), try to compensate it in
1748 * future periods by updating the virtual time; otherwise just
1749 * do a reset, as it may be a light dirtier.
1750 */
1751 if (pause < min_pause) {
1752 trace_balance_dirty_pages(wb,
1753 sdtc->thresh,
1754 sdtc->bg_thresh,
1755 sdtc->dirty,
1756 sdtc->wb_thresh,
1757 sdtc->wb_dirty,
1758 dirty_ratelimit,
1759 task_ratelimit,
1760 pages_dirtied,
1761 period,
1762 min(pause, 0L),
1763 start_time);
1764 if (pause < -HZ) {
1765 current->dirty_paused_when = now;
1766 current->nr_dirtied = 0;
1767 } else if (period) {
1768 current->dirty_paused_when += period;
1769 current->nr_dirtied = 0;
1770 } else if (current->nr_dirtied_pause <= pages_dirtied)
1771 current->nr_dirtied_pause += pages_dirtied;
1772 break;
1773 }
1774 if (unlikely(pause > max_pause)) {
1775 /* for occasional dropped task_ratelimit */
1776 now += min(pause - max_pause, max_pause);
1777 pause = max_pause;
1778 }
1779
1780pause:
1781 trace_balance_dirty_pages(wb,
1782 sdtc->thresh,
1783 sdtc->bg_thresh,
1784 sdtc->dirty,
1785 sdtc->wb_thresh,
1786 sdtc->wb_dirty,
1787 dirty_ratelimit,
1788 task_ratelimit,
1789 pages_dirtied,
1790 period,
1791 pause,
1792 start_time);
1793 __set_current_state(TASK_KILLABLE);
1794 wb->dirty_sleep = now;
1795 io_schedule_timeout(pause);
1796
1797 current->dirty_paused_when = now + pause;
1798 current->nr_dirtied = 0;
1799 current->nr_dirtied_pause = nr_dirtied_pause;
1800
1801 /*
1802 * This is typically equal to (dirty < thresh) and can also
1803 * keep "1000+ dd on a slow USB stick" under control.
1804 */
1805 if (task_ratelimit)
1806 break;
1807
1808 /*
1809 * In the case of an unresponding NFS server and the NFS dirty
1810 * pages exceeds dirty_thresh, give the other good wb's a pipe
1811 * to go through, so that tasks on them still remain responsive.
1812 *
1813 * In theory 1 page is enough to keep the consumer-producer
1814 * pipe going: the flusher cleans 1 page => the task dirties 1
1815 * more page. However wb_dirty has accounting errors. So use
1816 * the larger and more IO friendly wb_stat_error.
1817 */
1818 if (sdtc->wb_dirty <= wb_stat_error())
1819 break;
1820
1821 if (fatal_signal_pending(current))
1822 break;
1823 }
1824
1825 if (!dirty_exceeded && wb->dirty_exceeded)
1826 wb->dirty_exceeded = 0;
1827
1828 if (writeback_in_progress(wb))
1829 return;
1830
1831 /*
1832 * In laptop mode, we wait until hitting the higher threshold before
1833 * starting background writeout, and then write out all the way down
1834 * to the lower threshold. So slow writers cause minimal disk activity.
1835 *
1836 * In normal mode, we start background writeout at the lower
1837 * background_thresh, to keep the amount of dirty memory low.
1838 */
1839 if (laptop_mode)
1840 return;
1841
1842 if (nr_reclaimable > gdtc->bg_thresh)
1843 wb_start_background_writeback(wb);
1844}
1845
1846static DEFINE_PER_CPU(int, bdp_ratelimits);
1847
1848/*
1849 * Normal tasks are throttled by
1850 * loop {
1851 * dirty tsk->nr_dirtied_pause pages;
1852 * take a snap in balance_dirty_pages();
1853 * }
1854 * However there is a worst case. If every task exit immediately when dirtied
1855 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1856 * called to throttle the page dirties. The solution is to save the not yet
1857 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1858 * randomly into the running tasks. This works well for the above worst case,
1859 * as the new task will pick up and accumulate the old task's leaked dirty
1860 * count and eventually get throttled.
1861 */
1862DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1863
1864/**
1865 * balance_dirty_pages_ratelimited - balance dirty memory state
1866 * @mapping: address_space which was dirtied
1867 *
1868 * Processes which are dirtying memory should call in here once for each page
1869 * which was newly dirtied. The function will periodically check the system's
1870 * dirty state and will initiate writeback if needed.
1871 *
1872 * On really big machines, get_writeback_state is expensive, so try to avoid
1873 * calling it too often (ratelimiting). But once we're over the dirty memory
1874 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1875 * from overshooting the limit by (ratelimit_pages) each.
1876 */
1877void balance_dirty_pages_ratelimited(struct address_space *mapping)
1878{
1879 struct inode *inode = mapping->host;
1880 struct backing_dev_info *bdi = inode_to_bdi(inode);
1881 struct bdi_writeback *wb = NULL;
1882 int ratelimit;
1883 int *p;
1884
1885 if (!bdi_cap_account_dirty(bdi))
1886 return;
1887
1888 if (inode_cgwb_enabled(inode))
1889 wb = wb_get_create_current(bdi, GFP_KERNEL);
1890 if (!wb)
1891 wb = &bdi->wb;
1892
1893 ratelimit = current->nr_dirtied_pause;
1894 if (wb->dirty_exceeded)
1895 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1896
1897 preempt_disable();
1898 /*
1899 * This prevents one CPU to accumulate too many dirtied pages without
1900 * calling into balance_dirty_pages(), which can happen when there are
1901 * 1000+ tasks, all of them start dirtying pages at exactly the same
1902 * time, hence all honoured too large initial task->nr_dirtied_pause.
1903 */
1904 p = this_cpu_ptr(&bdp_ratelimits);
1905 if (unlikely(current->nr_dirtied >= ratelimit))
1906 *p = 0;
1907 else if (unlikely(*p >= ratelimit_pages)) {
1908 *p = 0;
1909 ratelimit = 0;
1910 }
1911 /*
1912 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1913 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1914 * the dirty throttling and livelock other long-run dirtiers.
1915 */
1916 p = this_cpu_ptr(&dirty_throttle_leaks);
1917 if (*p > 0 && current->nr_dirtied < ratelimit) {
1918 unsigned long nr_pages_dirtied;
1919 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1920 *p -= nr_pages_dirtied;
1921 current->nr_dirtied += nr_pages_dirtied;
1922 }
1923 preempt_enable();
1924
1925 if (unlikely(current->nr_dirtied >= ratelimit))
1926 balance_dirty_pages(wb, current->nr_dirtied);
1927
1928 wb_put(wb);
1929}
1930EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1931
1932/**
1933 * wb_over_bg_thresh - does @wb need to be written back?
1934 * @wb: bdi_writeback of interest
1935 *
1936 * Determines whether background writeback should keep writing @wb or it's
1937 * clean enough.
1938 *
1939 * Return: %true if writeback should continue.
1940 */
1941bool wb_over_bg_thresh(struct bdi_writeback *wb)
1942{
1943 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1944 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1945 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1946 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1947 &mdtc_stor : NULL;
1948
1949 /*
1950 * Similar to balance_dirty_pages() but ignores pages being written
1951 * as we're trying to decide whether to put more under writeback.
1952 */
1953 gdtc->avail = global_dirtyable_memory();
1954 gdtc->dirty = global_node_page_state(NR_FILE_DIRTY);
1955 domain_dirty_limits(gdtc);
1956
1957 if (gdtc->dirty > gdtc->bg_thresh)
1958 return true;
1959
1960 if (wb_stat(wb, WB_RECLAIMABLE) >
1961 wb_calc_thresh(gdtc->wb, gdtc->bg_thresh))
1962 return true;
1963
1964 if (mdtc) {
1965 unsigned long filepages, headroom, writeback;
1966
1967 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1968 &writeback);
1969 mdtc_calc_avail(mdtc, filepages, headroom);
1970 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1971
1972 if (mdtc->dirty > mdtc->bg_thresh)
1973 return true;
1974
1975 if (wb_stat(wb, WB_RECLAIMABLE) >
1976 wb_calc_thresh(mdtc->wb, mdtc->bg_thresh))
1977 return true;
1978 }
1979
1980 return false;
1981}
1982
1983/*
1984 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1985 */
1986int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1987 void *buffer, size_t *length, loff_t *ppos)
1988{
1989 unsigned int old_interval = dirty_writeback_interval;
1990 int ret;
1991
1992 ret = proc_dointvec(table, write, buffer, length, ppos);
1993
1994 /*
1995 * Writing 0 to dirty_writeback_interval will disable periodic writeback
1996 * and a different non-zero value will wakeup the writeback threads.
1997 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
1998 * iterate over all bdis and wbs.
1999 * The reason we do this is to make the change take effect immediately.
2000 */
2001 if (!ret && write && dirty_writeback_interval &&
2002 dirty_writeback_interval != old_interval)
2003 wakeup_flusher_threads(WB_REASON_PERIODIC);
2004
2005 return ret;
2006}
2007
2008#ifdef CONFIG_BLOCK
2009void laptop_mode_timer_fn(struct timer_list *t)
2010{
2011 struct backing_dev_info *backing_dev_info =
2012 from_timer(backing_dev_info, t, laptop_mode_wb_timer);
2013
2014 wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
2015}
2016
2017/*
2018 * We've spun up the disk and we're in laptop mode: schedule writeback
2019 * of all dirty data a few seconds from now. If the flush is already scheduled
2020 * then push it back - the user is still using the disk.
2021 */
2022void laptop_io_completion(struct backing_dev_info *info)
2023{
2024 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2025}
2026
2027/*
2028 * We're in laptop mode and we've just synced. The sync's writes will have
2029 * caused another writeback to be scheduled by laptop_io_completion.
2030 * Nothing needs to be written back anymore, so we unschedule the writeback.
2031 */
2032void laptop_sync_completion(void)
2033{
2034 struct backing_dev_info *bdi;
2035
2036 rcu_read_lock();
2037
2038 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2039 del_timer(&bdi->laptop_mode_wb_timer);
2040
2041 rcu_read_unlock();
2042}
2043#endif
2044
2045/*
2046 * If ratelimit_pages is too high then we can get into dirty-data overload
2047 * if a large number of processes all perform writes at the same time.
2048 * If it is too low then SMP machines will call the (expensive)
2049 * get_writeback_state too often.
2050 *
2051 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2052 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2053 * thresholds.
2054 */
2055
2056void writeback_set_ratelimit(void)
2057{
2058 struct wb_domain *dom = &global_wb_domain;
2059 unsigned long background_thresh;
2060 unsigned long dirty_thresh;
2061
2062 global_dirty_limits(&background_thresh, &dirty_thresh);
2063 dom->dirty_limit = dirty_thresh;
2064 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2065 if (ratelimit_pages < 16)
2066 ratelimit_pages = 16;
2067}
2068
2069static int page_writeback_cpu_online(unsigned int cpu)
2070{
2071 writeback_set_ratelimit();
2072 return 0;
2073}
2074
2075/*
2076 * Called early on to tune the page writeback dirty limits.
2077 *
2078 * We used to scale dirty pages according to how total memory
2079 * related to pages that could be allocated for buffers.
2080 *
2081 * However, that was when we used "dirty_ratio" to scale with
2082 * all memory, and we don't do that any more. "dirty_ratio"
2083 * is now applied to total non-HIGHPAGE memory, and as such we can't
2084 * get into the old insane situation any more where we had
2085 * large amounts of dirty pages compared to a small amount of
2086 * non-HIGHMEM memory.
2087 *
2088 * But we might still want to scale the dirty_ratio by how
2089 * much memory the box has..
2090 */
2091void __init page_writeback_init(void)
2092{
2093 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2094
2095 cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2096 page_writeback_cpu_online, NULL);
2097 cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2098 page_writeback_cpu_online);
2099}
2100
2101/**
2102 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2103 * @mapping: address space structure to write
2104 * @start: starting page index
2105 * @end: ending page index (inclusive)
2106 *
2107 * This function scans the page range from @start to @end (inclusive) and tags
2108 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2109 * that write_cache_pages (or whoever calls this function) will then use
2110 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2111 * used to avoid livelocking of writeback by a process steadily creating new
2112 * dirty pages in the file (thus it is important for this function to be quick
2113 * so that it can tag pages faster than a dirtying process can create them).
2114 */
2115void tag_pages_for_writeback(struct address_space *mapping,
2116 pgoff_t start, pgoff_t end)
2117{
2118 XA_STATE(xas, &mapping->i_pages, start);
2119 unsigned int tagged = 0;
2120 void *page;
2121
2122 xas_lock_irq(&xas);
2123 xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2124 xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2125 if (++tagged % XA_CHECK_SCHED)
2126 continue;
2127
2128 xas_pause(&xas);
2129 xas_unlock_irq(&xas);
2130 cond_resched();
2131 xas_lock_irq(&xas);
2132 }
2133 xas_unlock_irq(&xas);
2134}
2135EXPORT_SYMBOL(tag_pages_for_writeback);
2136
2137/**
2138 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2139 * @mapping: address space structure to write
2140 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2141 * @writepage: function called for each page
2142 * @data: data passed to writepage function
2143 *
2144 * If a page is already under I/O, write_cache_pages() skips it, even
2145 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2146 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2147 * and msync() need to guarantee that all the data which was dirty at the time
2148 * the call was made get new I/O started against them. If wbc->sync_mode is
2149 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2150 * existing IO to complete.
2151 *
2152 * To avoid livelocks (when other process dirties new pages), we first tag
2153 * pages which should be written back with TOWRITE tag and only then start
2154 * writing them. For data-integrity sync we have to be careful so that we do
2155 * not miss some pages (e.g., because some other process has cleared TOWRITE
2156 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2157 * by the process clearing the DIRTY tag (and submitting the page for IO).
2158 *
2159 * To avoid deadlocks between range_cyclic writeback and callers that hold
2160 * pages in PageWriteback to aggregate IO until write_cache_pages() returns,
2161 * we do not loop back to the start of the file. Doing so causes a page
2162 * lock/page writeback access order inversion - we should only ever lock
2163 * multiple pages in ascending page->index order, and looping back to the start
2164 * of the file violates that rule and causes deadlocks.
2165 *
2166 * Return: %0 on success, negative error code otherwise
2167 */
2168int write_cache_pages(struct address_space *mapping,
2169 struct writeback_control *wbc, writepage_t writepage,
2170 void *data)
2171{
2172 int ret = 0;
2173 int done = 0;
2174 int error;
2175 struct pagevec pvec;
2176 int nr_pages;
2177 pgoff_t index;
2178 pgoff_t end; /* Inclusive */
2179 pgoff_t done_index;
2180 int range_whole = 0;
2181 xa_mark_t tag;
2182
2183 pagevec_init(&pvec);
2184 if (wbc->range_cyclic) {
2185 index = mapping->writeback_index; /* prev offset */
2186 end = -1;
2187 } else {
2188 index = wbc->range_start >> PAGE_SHIFT;
2189 end = wbc->range_end >> PAGE_SHIFT;
2190 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2191 range_whole = 1;
2192 }
2193 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
2194 tag_pages_for_writeback(mapping, index, end);
2195 tag = PAGECACHE_TAG_TOWRITE;
2196 } else {
2197 tag = PAGECACHE_TAG_DIRTY;
2198 }
2199 done_index = index;
2200 while (!done && (index <= end)) {
2201 int i;
2202
2203 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2204 tag);
2205 if (nr_pages == 0)
2206 break;
2207
2208 for (i = 0; i < nr_pages; i++) {
2209 struct page *page = pvec.pages[i];
2210
2211 done_index = page->index;
2212
2213 lock_page(page);
2214
2215 /*
2216 * Page truncated or invalidated. We can freely skip it
2217 * then, even for data integrity operations: the page
2218 * has disappeared concurrently, so there could be no
2219 * real expectation of this data interity operation
2220 * even if there is now a new, dirty page at the same
2221 * pagecache address.
2222 */
2223 if (unlikely(page->mapping != mapping)) {
2224continue_unlock:
2225 unlock_page(page);
2226 continue;
2227 }
2228
2229 if (!PageDirty(page)) {
2230 /* someone wrote it for us */
2231 goto continue_unlock;
2232 }
2233
2234 if (PageWriteback(page)) {
2235 if (wbc->sync_mode != WB_SYNC_NONE)
2236 wait_on_page_writeback(page);
2237 else
2238 goto continue_unlock;
2239 }
2240
2241 BUG_ON(PageWriteback(page));
2242 if (!clear_page_dirty_for_io(page))
2243 goto continue_unlock;
2244
2245 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2246 error = (*writepage)(page, wbc, data);
2247 if (unlikely(error)) {
2248 /*
2249 * Handle errors according to the type of
2250 * writeback. There's no need to continue for
2251 * background writeback. Just push done_index
2252 * past this page so media errors won't choke
2253 * writeout for the entire file. For integrity
2254 * writeback, we must process the entire dirty
2255 * set regardless of errors because the fs may
2256 * still have state to clear for each page. In
2257 * that case we continue processing and return
2258 * the first error.
2259 */
2260 if (error == AOP_WRITEPAGE_ACTIVATE) {
2261 unlock_page(page);
2262 error = 0;
2263 } else if (wbc->sync_mode != WB_SYNC_ALL) {
2264 ret = error;
2265 done_index = page->index + 1;
2266 done = 1;
2267 break;
2268 }
2269 if (!ret)
2270 ret = error;
2271 }
2272
2273 /*
2274 * We stop writing back only if we are not doing
2275 * integrity sync. In case of integrity sync we have to
2276 * keep going until we have written all the pages
2277 * we tagged for writeback prior to entering this loop.
2278 */
2279 if (--wbc->nr_to_write <= 0 &&
2280 wbc->sync_mode == WB_SYNC_NONE) {
2281 done = 1;
2282 break;
2283 }
2284 }
2285 pagevec_release(&pvec);
2286 cond_resched();
2287 }
2288
2289 /*
2290 * If we hit the last page and there is more work to be done: wrap
2291 * back the index back to the start of the file for the next
2292 * time we are called.
2293 */
2294 if (wbc->range_cyclic && !done)
2295 done_index = 0;
2296 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2297 mapping->writeback_index = done_index;
2298
2299 return ret;
2300}
2301EXPORT_SYMBOL(write_cache_pages);
2302
2303/*
2304 * Function used by generic_writepages to call the real writepage
2305 * function and set the mapping flags on error
2306 */
2307static int __writepage(struct page *page, struct writeback_control *wbc,
2308 void *data)
2309{
2310 struct address_space *mapping = data;
2311 int ret = mapping->a_ops->writepage(page, wbc);
2312 mapping_set_error(mapping, ret);
2313 return ret;
2314}
2315
2316/**
2317 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2318 * @mapping: address space structure to write
2319 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2320 *
2321 * This is a library function, which implements the writepages()
2322 * address_space_operation.
2323 *
2324 * Return: %0 on success, negative error code otherwise
2325 */
2326int generic_writepages(struct address_space *mapping,
2327 struct writeback_control *wbc)
2328{
2329 struct blk_plug plug;
2330 int ret;
2331
2332 /* deal with chardevs and other special file */
2333 if (!mapping->a_ops->writepage)
2334 return 0;
2335
2336 blk_start_plug(&plug);
2337 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2338 blk_finish_plug(&plug);
2339 return ret;
2340}
2341
2342EXPORT_SYMBOL(generic_writepages);
2343
2344int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2345{
2346 int ret;
2347
2348 if (wbc->nr_to_write <= 0)
2349 return 0;
2350 while (1) {
2351 if (mapping->a_ops->writepages)
2352 ret = mapping->a_ops->writepages(mapping, wbc);
2353 else
2354 ret = generic_writepages(mapping, wbc);
2355 if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2356 break;
2357 cond_resched();
2358 congestion_wait(BLK_RW_ASYNC, HZ/50);
2359 }
2360 return ret;
2361}
2362
2363/**
2364 * write_one_page - write out a single page and wait on I/O
2365 * @page: the page to write
2366 *
2367 * The page must be locked by the caller and will be unlocked upon return.
2368 *
2369 * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2370 * function returns.
2371 *
2372 * Return: %0 on success, negative error code otherwise
2373 */
2374int write_one_page(struct page *page)
2375{
2376 struct address_space *mapping = page->mapping;
2377 int ret = 0;
2378 struct writeback_control wbc = {
2379 .sync_mode = WB_SYNC_ALL,
2380 .nr_to_write = 1,
2381 };
2382
2383 BUG_ON(!PageLocked(page));
2384
2385 wait_on_page_writeback(page);
2386
2387 if (clear_page_dirty_for_io(page)) {
2388 get_page(page);
2389 ret = mapping->a_ops->writepage(page, &wbc);
2390 if (ret == 0)
2391 wait_on_page_writeback(page);
2392 put_page(page);
2393 } else {
2394 unlock_page(page);
2395 }
2396
2397 if (!ret)
2398 ret = filemap_check_errors(mapping);
2399 return ret;
2400}
2401EXPORT_SYMBOL(write_one_page);
2402
2403/*
2404 * For address_spaces which do not use buffers nor write back.
2405 */
2406int __set_page_dirty_no_writeback(struct page *page)
2407{
2408 if (!PageDirty(page))
2409 return !TestSetPageDirty(page);
2410 return 0;
2411}
2412
2413/*
2414 * Helper function for set_page_dirty family.
2415 *
2416 * Caller must hold lock_page_memcg().
2417 *
2418 * NOTE: This relies on being atomic wrt interrupts.
2419 */
2420void account_page_dirtied(struct page *page, struct address_space *mapping)
2421{
2422 struct inode *inode = mapping->host;
2423
2424 trace_writeback_dirty_page(page, mapping);
2425
2426 if (mapping_cap_account_dirty(mapping)) {
2427 struct bdi_writeback *wb;
2428
2429 inode_attach_wb(inode, page);
2430 wb = inode_to_wb(inode);
2431
2432 __inc_lruvec_page_state(page, NR_FILE_DIRTY);
2433 __inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2434 __inc_node_page_state(page, NR_DIRTIED);
2435 inc_wb_stat(wb, WB_RECLAIMABLE);
2436 inc_wb_stat(wb, WB_DIRTIED);
2437 task_io_account_write(PAGE_SIZE);
2438 current->nr_dirtied++;
2439 this_cpu_inc(bdp_ratelimits);
2440
2441 mem_cgroup_track_foreign_dirty(page, wb);
2442 }
2443}
2444
2445/*
2446 * Helper function for deaccounting dirty page without writeback.
2447 *
2448 * Caller must hold lock_page_memcg().
2449 */
2450void account_page_cleaned(struct page *page, struct address_space *mapping,
2451 struct bdi_writeback *wb)
2452{
2453 if (mapping_cap_account_dirty(mapping)) {
2454 dec_lruvec_page_state(page, NR_FILE_DIRTY);
2455 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2456 dec_wb_stat(wb, WB_RECLAIMABLE);
2457 task_io_account_cancelled_write(PAGE_SIZE);
2458 }
2459}
2460
2461/*
2462 * For address_spaces which do not use buffers. Just tag the page as dirty in
2463 * the xarray.
2464 *
2465 * This is also used when a single buffer is being dirtied: we want to set the
2466 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2467 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2468 *
2469 * The caller must ensure this doesn't race with truncation. Most will simply
2470 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2471 * the pte lock held, which also locks out truncation.
2472 */
2473int __set_page_dirty_nobuffers(struct page *page)
2474{
2475 lock_page_memcg(page);
2476 if (!TestSetPageDirty(page)) {
2477 struct address_space *mapping = page_mapping(page);
2478 unsigned long flags;
2479
2480 if (!mapping) {
2481 unlock_page_memcg(page);
2482 return 1;
2483 }
2484
2485 xa_lock_irqsave(&mapping->i_pages, flags);
2486 BUG_ON(page_mapping(page) != mapping);
2487 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2488 account_page_dirtied(page, mapping);
2489 __xa_set_mark(&mapping->i_pages, page_index(page),
2490 PAGECACHE_TAG_DIRTY);
2491 xa_unlock_irqrestore(&mapping->i_pages, flags);
2492 unlock_page_memcg(page);
2493
2494 if (mapping->host) {
2495 /* !PageAnon && !swapper_space */
2496 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2497 }
2498 return 1;
2499 }
2500 unlock_page_memcg(page);
2501 return 0;
2502}
2503EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2504
2505/*
2506 * Call this whenever redirtying a page, to de-account the dirty counters
2507 * (NR_DIRTIED, WB_DIRTIED, tsk->nr_dirtied), so that they match the written
2508 * counters (NR_WRITTEN, WB_WRITTEN) in long term. The mismatches will lead to
2509 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2510 * control.
2511 */
2512void account_page_redirty(struct page *page)
2513{
2514 struct address_space *mapping = page->mapping;
2515
2516 if (mapping && mapping_cap_account_dirty(mapping)) {
2517 struct inode *inode = mapping->host;
2518 struct bdi_writeback *wb;
2519 struct wb_lock_cookie cookie = {};
2520
2521 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2522 current->nr_dirtied--;
2523 dec_node_page_state(page, NR_DIRTIED);
2524 dec_wb_stat(wb, WB_DIRTIED);
2525 unlocked_inode_to_wb_end(inode, &cookie);
2526 }
2527}
2528EXPORT_SYMBOL(account_page_redirty);
2529
2530/*
2531 * When a writepage implementation decides that it doesn't want to write this
2532 * page for some reason, it should redirty the locked page via
2533 * redirty_page_for_writepage() and it should then unlock the page and return 0
2534 */
2535int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2536{
2537 int ret;
2538
2539 wbc->pages_skipped++;
2540 ret = __set_page_dirty_nobuffers(page);
2541 account_page_redirty(page);
2542 return ret;
2543}
2544EXPORT_SYMBOL(redirty_page_for_writepage);
2545
2546/*
2547 * Dirty a page.
2548 *
2549 * For pages with a mapping this should be done under the page lock
2550 * for the benefit of asynchronous memory errors who prefer a consistent
2551 * dirty state. This rule can be broken in some special cases,
2552 * but should be better not to.
2553 *
2554 * If the mapping doesn't provide a set_page_dirty a_op, then
2555 * just fall through and assume that it wants buffer_heads.
2556 */
2557int set_page_dirty(struct page *page)
2558{
2559 struct address_space *mapping = page_mapping(page);
2560
2561 page = compound_head(page);
2562 if (likely(mapping)) {
2563 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2564 /*
2565 * readahead/lru_deactivate_page could remain
2566 * PG_readahead/PG_reclaim due to race with end_page_writeback
2567 * About readahead, if the page is written, the flags would be
2568 * reset. So no problem.
2569 * About lru_deactivate_page, if the page is redirty, the flag
2570 * will be reset. So no problem. but if the page is used by readahead
2571 * it will confuse readahead and make it restart the size rampup
2572 * process. But it's a trivial problem.
2573 */
2574 if (PageReclaim(page))
2575 ClearPageReclaim(page);
2576#ifdef CONFIG_BLOCK
2577 if (!spd)
2578 spd = __set_page_dirty_buffers;
2579#endif
2580 return (*spd)(page);
2581 }
2582 if (!PageDirty(page)) {
2583 if (!TestSetPageDirty(page))
2584 return 1;
2585 }
2586 return 0;
2587}
2588EXPORT_SYMBOL(set_page_dirty);
2589
2590/*
2591 * set_page_dirty() is racy if the caller has no reference against
2592 * page->mapping->host, and if the page is unlocked. This is because another
2593 * CPU could truncate the page off the mapping and then free the mapping.
2594 *
2595 * Usually, the page _is_ locked, or the caller is a user-space process which
2596 * holds a reference on the inode by having an open file.
2597 *
2598 * In other cases, the page should be locked before running set_page_dirty().
2599 */
2600int set_page_dirty_lock(struct page *page)
2601{
2602 int ret;
2603
2604 lock_page(page);
2605 ret = set_page_dirty(page);
2606 unlock_page(page);
2607 return ret;
2608}
2609EXPORT_SYMBOL(set_page_dirty_lock);
2610
2611/*
2612 * This cancels just the dirty bit on the kernel page itself, it does NOT
2613 * actually remove dirty bits on any mmap's that may be around. It also
2614 * leaves the page tagged dirty, so any sync activity will still find it on
2615 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2616 * look at the dirty bits in the VM.
2617 *
2618 * Doing this should *normally* only ever be done when a page is truncated,
2619 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2620 * this when it notices that somebody has cleaned out all the buffers on a
2621 * page without actually doing it through the VM. Can you say "ext3 is
2622 * horribly ugly"? Thought you could.
2623 */
2624void __cancel_dirty_page(struct page *page)
2625{
2626 struct address_space *mapping = page_mapping(page);
2627
2628 if (mapping_cap_account_dirty(mapping)) {
2629 struct inode *inode = mapping->host;
2630 struct bdi_writeback *wb;
2631 struct wb_lock_cookie cookie = {};
2632
2633 lock_page_memcg(page);
2634 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2635
2636 if (TestClearPageDirty(page))
2637 account_page_cleaned(page, mapping, wb);
2638
2639 unlocked_inode_to_wb_end(inode, &cookie);
2640 unlock_page_memcg(page);
2641 } else {
2642 ClearPageDirty(page);
2643 }
2644}
2645EXPORT_SYMBOL(__cancel_dirty_page);
2646
2647/*
2648 * Clear a page's dirty flag, while caring for dirty memory accounting.
2649 * Returns true if the page was previously dirty.
2650 *
2651 * This is for preparing to put the page under writeout. We leave the page
2652 * tagged as dirty in the xarray so that a concurrent write-for-sync
2653 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2654 * implementation will run either set_page_writeback() or set_page_dirty(),
2655 * at which stage we bring the page's dirty flag and xarray dirty tag
2656 * back into sync.
2657 *
2658 * This incoherency between the page's dirty flag and xarray tag is
2659 * unfortunate, but it only exists while the page is locked.
2660 */
2661int clear_page_dirty_for_io(struct page *page)
2662{
2663 struct address_space *mapping = page_mapping(page);
2664 int ret = 0;
2665
2666 VM_BUG_ON_PAGE(!PageLocked(page), page);
2667
2668 if (mapping && mapping_cap_account_dirty(mapping)) {
2669 struct inode *inode = mapping->host;
2670 struct bdi_writeback *wb;
2671 struct wb_lock_cookie cookie = {};
2672
2673 /*
2674 * Yes, Virginia, this is indeed insane.
2675 *
2676 * We use this sequence to make sure that
2677 * (a) we account for dirty stats properly
2678 * (b) we tell the low-level filesystem to
2679 * mark the whole page dirty if it was
2680 * dirty in a pagetable. Only to then
2681 * (c) clean the page again and return 1 to
2682 * cause the writeback.
2683 *
2684 * This way we avoid all nasty races with the
2685 * dirty bit in multiple places and clearing
2686 * them concurrently from different threads.
2687 *
2688 * Note! Normally the "set_page_dirty(page)"
2689 * has no effect on the actual dirty bit - since
2690 * that will already usually be set. But we
2691 * need the side effects, and it can help us
2692 * avoid races.
2693 *
2694 * We basically use the page "master dirty bit"
2695 * as a serialization point for all the different
2696 * threads doing their things.
2697 */
2698 if (page_mkclean(page))
2699 set_page_dirty(page);
2700 /*
2701 * We carefully synchronise fault handlers against
2702 * installing a dirty pte and marking the page dirty
2703 * at this point. We do this by having them hold the
2704 * page lock while dirtying the page, and pages are
2705 * always locked coming in here, so we get the desired
2706 * exclusion.
2707 */
2708 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2709 if (TestClearPageDirty(page)) {
2710 dec_lruvec_page_state(page, NR_FILE_DIRTY);
2711 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2712 dec_wb_stat(wb, WB_RECLAIMABLE);
2713 ret = 1;
2714 }
2715 unlocked_inode_to_wb_end(inode, &cookie);
2716 return ret;
2717 }
2718 return TestClearPageDirty(page);
2719}
2720EXPORT_SYMBOL(clear_page_dirty_for_io);
2721
2722int test_clear_page_writeback(struct page *page)
2723{
2724 struct address_space *mapping = page_mapping(page);
2725 struct mem_cgroup *memcg;
2726 struct lruvec *lruvec;
2727 int ret;
2728
2729 memcg = lock_page_memcg(page);
2730 lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
2731 if (mapping && mapping_use_writeback_tags(mapping)) {
2732 struct inode *inode = mapping->host;
2733 struct backing_dev_info *bdi = inode_to_bdi(inode);
2734 unsigned long flags;
2735
2736 xa_lock_irqsave(&mapping->i_pages, flags);
2737 ret = TestClearPageWriteback(page);
2738 if (ret) {
2739 __xa_clear_mark(&mapping->i_pages, page_index(page),
2740 PAGECACHE_TAG_WRITEBACK);
2741 if (bdi_cap_account_writeback(bdi)) {
2742 struct bdi_writeback *wb = inode_to_wb(inode);
2743
2744 dec_wb_stat(wb, WB_WRITEBACK);
2745 __wb_writeout_inc(wb);
2746 }
2747 }
2748
2749 if (mapping->host && !mapping_tagged(mapping,
2750 PAGECACHE_TAG_WRITEBACK))
2751 sb_clear_inode_writeback(mapping->host);
2752
2753 xa_unlock_irqrestore(&mapping->i_pages, flags);
2754 } else {
2755 ret = TestClearPageWriteback(page);
2756 }
2757 /*
2758 * NOTE: Page might be free now! Writeback doesn't hold a page
2759 * reference on its own, it relies on truncation to wait for
2760 * the clearing of PG_writeback. The below can only access
2761 * page state that is static across allocation cycles.
2762 */
2763 if (ret) {
2764 dec_lruvec_state(lruvec, NR_WRITEBACK);
2765 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2766 inc_node_page_state(page, NR_WRITTEN);
2767 }
2768 __unlock_page_memcg(memcg);
2769 return ret;
2770}
2771
2772int __test_set_page_writeback(struct page *page, bool keep_write)
2773{
2774 struct address_space *mapping = page_mapping(page);
2775 int ret, access_ret;
2776
2777 lock_page_memcg(page);
2778 if (mapping && mapping_use_writeback_tags(mapping)) {
2779 XA_STATE(xas, &mapping->i_pages, page_index(page));
2780 struct inode *inode = mapping->host;
2781 struct backing_dev_info *bdi = inode_to_bdi(inode);
2782 unsigned long flags;
2783
2784 xas_lock_irqsave(&xas, flags);
2785 xas_load(&xas);
2786 ret = TestSetPageWriteback(page);
2787 if (!ret) {
2788 bool on_wblist;
2789
2790 on_wblist = mapping_tagged(mapping,
2791 PAGECACHE_TAG_WRITEBACK);
2792
2793 xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
2794 if (bdi_cap_account_writeback(bdi))
2795 inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2796
2797 /*
2798 * We can come through here when swapping anonymous
2799 * pages, so we don't necessarily have an inode to track
2800 * for sync.
2801 */
2802 if (mapping->host && !on_wblist)
2803 sb_mark_inode_writeback(mapping->host);
2804 }
2805 if (!PageDirty(page))
2806 xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
2807 if (!keep_write)
2808 xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
2809 xas_unlock_irqrestore(&xas, flags);
2810 } else {
2811 ret = TestSetPageWriteback(page);
2812 }
2813 if (!ret) {
2814 inc_lruvec_page_state(page, NR_WRITEBACK);
2815 inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2816 }
2817 unlock_page_memcg(page);
2818 access_ret = arch_make_page_accessible(page);
2819 /*
2820 * If writeback has been triggered on a page that cannot be made
2821 * accessible, it is too late to recover here.
2822 */
2823 VM_BUG_ON_PAGE(access_ret != 0, page);
2824
2825 return ret;
2826
2827}
2828EXPORT_SYMBOL(__test_set_page_writeback);
2829
2830/*
2831 * Wait for a page to complete writeback
2832 */
2833void wait_on_page_writeback(struct page *page)
2834{
2835 if (PageWriteback(page)) {
2836 trace_wait_on_page_writeback(page, page_mapping(page));
2837 wait_on_page_bit(page, PG_writeback);
2838 }
2839}
2840EXPORT_SYMBOL_GPL(wait_on_page_writeback);
2841
2842/**
2843 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2844 * @page: The page to wait on.
2845 *
2846 * This function determines if the given page is related to a backing device
2847 * that requires page contents to be held stable during writeback. If so, then
2848 * it will wait for any pending writeback to complete.
2849 */
2850void wait_for_stable_page(struct page *page)
2851{
2852 if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2853 wait_on_page_writeback(page);
2854}
2855EXPORT_SYMBOL_GPL(wait_for_stable_page);