1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 2002, Linus Torvalds.
6 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
8 * Contains functions related to writing back dirty pages at the
11 * 10Apr2002 Andrew Morton
15 #include <linux/kernel.h>
16 #include <linux/math64.h>
17 #include <linux/export.h>
18 #include <linux/spinlock.h>
21 #include <linux/swap.h>
22 #include <linux/slab.h>
23 #include <linux/pagemap.h>
24 #include <linux/writeback.h>
25 #include <linux/init.h>
26 #include <linux/backing-dev.h>
27 #include <linux/task_io_accounting_ops.h>
28 #include <linux/blkdev.h>
29 #include <linux/mpage.h>
30 #include <linux/rmap.h>
31 #include <linux/percpu.h>
32 #include <linux/smp.h>
33 #include <linux/sysctl.h>
34 #include <linux/cpu.h>
35 #include <linux/syscalls.h>
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>
46 * Sleep at most 200ms at a time in balance_dirty_pages().
48 #define MAX_PAUSE max(HZ/5, 1)
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.
54 #define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
57 * Estimate write bandwidth at 200ms intervals.
59 #define BANDWIDTH_INTERVAL max(HZ/5, 1)
61 #define RATELIMIT_CALC_SHIFT 10
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.
67 static long ratelimit_pages = 32;
69 /* The following parameters are exported via /proc/sys/vm */
72 * Start background writeback (via writeback threads) at this percentage
74 static int dirty_background_ratio = 10;
77 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
78 * dirty_background_ratio * the amount of dirtyable memory
80 static unsigned long dirty_background_bytes;
83 * free highmem will not be subtracted from the total free memory
84 * for calculating free ratios if vm_highmem_is_dirtyable is true
86 static int vm_highmem_is_dirtyable;
89 * The generator of dirty data starts writeback at this percentage
91 static int vm_dirty_ratio = 20;
94 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
95 * vm_dirty_ratio * the amount of dirtyable memory
97 static unsigned long vm_dirty_bytes;
100 * The interval between `kupdate'-style writebacks
102 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
104 EXPORT_SYMBOL_GPL(dirty_writeback_interval);
107 * The longest time for which data is allowed to remain dirty
109 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
112 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
113 * a full sync is triggered after this time elapses without any disk activity.
117 EXPORT_SYMBOL(laptop_mode);
119 /* End of sysctl-exported parameters */
121 struct wb_domain global_wb_domain;
123 /* consolidated parameters for balance_dirty_pages() and its subroutines */
124 struct dirty_throttle_control {
125 #ifdef CONFIG_CGROUP_WRITEBACK
126 struct wb_domain *dom;
127 struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */
129 struct bdi_writeback *wb;
130 struct fprop_local_percpu *wb_completions;
132 unsigned long avail; /* dirtyable */
133 unsigned long dirty; /* file_dirty + write + nfs */
134 unsigned long thresh; /* dirty threshold */
135 unsigned long bg_thresh; /* dirty background threshold */
137 unsigned long wb_dirty; /* per-wb counterparts */
138 unsigned long wb_thresh;
139 unsigned long wb_bg_thresh;
141 unsigned long pos_ratio;
145 * Length of period for aging writeout fractions of bdis. This is an
146 * arbitrarily chosen number. The longer the period, the slower fractions will
147 * reflect changes in current writeout rate.
149 #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
151 #ifdef CONFIG_CGROUP_WRITEBACK
153 #define GDTC_INIT(__wb) .wb = (__wb), \
154 .dom = &global_wb_domain, \
155 .wb_completions = &(__wb)->completions
157 #define GDTC_INIT_NO_WB .dom = &global_wb_domain
159 #define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
160 .dom = mem_cgroup_wb_domain(__wb), \
161 .wb_completions = &(__wb)->memcg_completions, \
164 static bool mdtc_valid(struct dirty_throttle_control *dtc)
169 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
174 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
179 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
181 return &wb->memcg_completions;
184 static void wb_min_max_ratio(struct bdi_writeback *wb,
185 unsigned long *minp, unsigned long *maxp)
187 unsigned long this_bw = READ_ONCE(wb->avg_write_bandwidth);
188 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
189 unsigned long long min = wb->bdi->min_ratio;
190 unsigned long long max = wb->bdi->max_ratio;
193 * @wb may already be clean by the time control reaches here and
194 * the total may not include its bw.
196 if (this_bw < tot_bw) {
199 min = div64_ul(min, tot_bw);
201 if (max < 100 * BDI_RATIO_SCALE) {
203 max = div64_ul(max, tot_bw);
211 #else /* CONFIG_CGROUP_WRITEBACK */
213 #define GDTC_INIT(__wb) .wb = (__wb), \
214 .wb_completions = &(__wb)->completions
215 #define GDTC_INIT_NO_WB
216 #define MDTC_INIT(__wb, __gdtc)
218 static bool mdtc_valid(struct dirty_throttle_control *dtc)
223 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
225 return &global_wb_domain;
228 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
233 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
238 static void wb_min_max_ratio(struct bdi_writeback *wb,
239 unsigned long *minp, unsigned long *maxp)
241 *minp = wb->bdi->min_ratio;
242 *maxp = wb->bdi->max_ratio;
245 #endif /* CONFIG_CGROUP_WRITEBACK */
248 * In a memory zone, there is a certain amount of pages we consider
249 * available for the page cache, which is essentially the number of
250 * free and reclaimable pages, minus some zone reserves to protect
251 * lowmem and the ability to uphold the zone's watermarks without
252 * requiring writeback.
254 * This number of dirtyable pages is the base value of which the
255 * user-configurable dirty ratio is the effective number of pages that
256 * are allowed to be actually dirtied. Per individual zone, or
257 * globally by using the sum of dirtyable pages over all zones.
259 * Because the user is allowed to specify the dirty limit globally as
260 * absolute number of bytes, calculating the per-zone dirty limit can
261 * require translating the configured limit into a percentage of
262 * global dirtyable memory first.
266 * node_dirtyable_memory - number of dirtyable pages in a node
269 * Return: the node's number of pages potentially available for dirty
270 * page cache. This is the base value for the per-node dirty limits.
272 static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
274 unsigned long nr_pages = 0;
277 for (z = 0; z < MAX_NR_ZONES; z++) {
278 struct zone *zone = pgdat->node_zones + z;
280 if (!populated_zone(zone))
283 nr_pages += zone_page_state(zone, NR_FREE_PAGES);
287 * Pages reserved for the kernel should not be considered
288 * dirtyable, to prevent a situation where reclaim has to
289 * clean pages in order to balance the zones.
291 nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
293 nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
294 nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
299 static unsigned long highmem_dirtyable_memory(unsigned long total)
301 #ifdef CONFIG_HIGHMEM
306 for_each_node_state(node, N_HIGH_MEMORY) {
307 for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
309 unsigned long nr_pages;
311 if (!is_highmem_idx(i))
314 z = &NODE_DATA(node)->node_zones[i];
315 if (!populated_zone(z))
318 nr_pages = zone_page_state(z, NR_FREE_PAGES);
319 /* watch for underflows */
320 nr_pages -= min(nr_pages, high_wmark_pages(z));
321 nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
322 nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
328 * Make sure that the number of highmem pages is never larger
329 * than the number of the total dirtyable memory. This can only
330 * occur in very strange VM situations but we want to make sure
331 * that this does not occur.
333 return min(x, total);
340 * global_dirtyable_memory - number of globally dirtyable pages
342 * Return: the global number of pages potentially available for dirty
343 * page cache. This is the base value for the global dirty limits.
345 static unsigned long global_dirtyable_memory(void)
349 x = global_zone_page_state(NR_FREE_PAGES);
351 * Pages reserved for the kernel should not be considered
352 * dirtyable, to prevent a situation where reclaim has to
353 * clean pages in order to balance the zones.
355 x -= min(x, totalreserve_pages);
357 x += global_node_page_state(NR_INACTIVE_FILE);
358 x += global_node_page_state(NR_ACTIVE_FILE);
360 if (!vm_highmem_is_dirtyable)
361 x -= highmem_dirtyable_memory(x);
363 return x + 1; /* Ensure that we never return 0 */
367 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
368 * @dtc: dirty_throttle_control of interest
370 * Calculate @dtc->thresh and ->bg_thresh considering
371 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
372 * must ensure that @dtc->avail is set before calling this function. The
373 * dirty limits will be lifted by 1/4 for real-time tasks.
375 static void domain_dirty_limits(struct dirty_throttle_control *dtc)
377 const unsigned long available_memory = dtc->avail;
378 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
379 unsigned long bytes = vm_dirty_bytes;
380 unsigned long bg_bytes = dirty_background_bytes;
381 /* convert ratios to per-PAGE_SIZE for higher precision */
382 unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
383 unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
384 unsigned long thresh;
385 unsigned long bg_thresh;
386 struct task_struct *tsk;
388 /* gdtc is !NULL iff @dtc is for memcg domain */
390 unsigned long global_avail = gdtc->avail;
393 * The byte settings can't be applied directly to memcg
394 * domains. Convert them to ratios by scaling against
395 * globally available memory. As the ratios are in
396 * per-PAGE_SIZE, they can be obtained by dividing bytes by
400 ratio = min(DIV_ROUND_UP(bytes, global_avail),
403 bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
405 bytes = bg_bytes = 0;
409 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
411 thresh = (ratio * available_memory) / PAGE_SIZE;
414 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
416 bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
418 if (bg_thresh >= thresh)
419 bg_thresh = thresh / 2;
422 bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
423 thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
425 dtc->thresh = thresh;
426 dtc->bg_thresh = bg_thresh;
428 /* we should eventually report the domain in the TP */
430 trace_global_dirty_state(bg_thresh, thresh);
434 * global_dirty_limits - background-writeback and dirty-throttling thresholds
435 * @pbackground: out parameter for bg_thresh
436 * @pdirty: out parameter for thresh
438 * Calculate bg_thresh and thresh for global_wb_domain. See
439 * domain_dirty_limits() for details.
441 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
443 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
445 gdtc.avail = global_dirtyable_memory();
446 domain_dirty_limits(&gdtc);
448 *pbackground = gdtc.bg_thresh;
449 *pdirty = gdtc.thresh;
453 * node_dirty_limit - maximum number of dirty pages allowed in a node
456 * Return: the maximum number of dirty pages allowed in a node, based
457 * on the node's dirtyable memory.
459 static unsigned long node_dirty_limit(struct pglist_data *pgdat)
461 unsigned long node_memory = node_dirtyable_memory(pgdat);
462 struct task_struct *tsk = current;
466 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
467 node_memory / global_dirtyable_memory();
469 dirty = vm_dirty_ratio * node_memory / 100;
478 * node_dirty_ok - tells whether a node is within its dirty limits
479 * @pgdat: the node to check
481 * Return: %true when the dirty pages in @pgdat are within the node's
482 * dirty limit, %false if the limit is exceeded.
484 bool node_dirty_ok(struct pglist_data *pgdat)
486 unsigned long limit = node_dirty_limit(pgdat);
487 unsigned long nr_pages = 0;
489 nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
490 nr_pages += node_page_state(pgdat, NR_WRITEBACK);
492 return nr_pages <= limit;
496 static int dirty_background_ratio_handler(struct ctl_table *table, int write,
497 void *buffer, size_t *lenp, loff_t *ppos)
501 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
502 if (ret == 0 && write)
503 dirty_background_bytes = 0;
507 static int dirty_background_bytes_handler(struct ctl_table *table, int write,
508 void *buffer, size_t *lenp, loff_t *ppos)
512 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
513 if (ret == 0 && write)
514 dirty_background_ratio = 0;
518 static int dirty_ratio_handler(struct ctl_table *table, int write, void *buffer,
519 size_t *lenp, loff_t *ppos)
521 int old_ratio = vm_dirty_ratio;
524 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
525 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
526 writeback_set_ratelimit();
532 static int dirty_bytes_handler(struct ctl_table *table, int write,
533 void *buffer, size_t *lenp, loff_t *ppos)
535 unsigned long old_bytes = vm_dirty_bytes;
538 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
539 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
540 writeback_set_ratelimit();
547 static unsigned long wp_next_time(unsigned long cur_time)
549 cur_time += VM_COMPLETIONS_PERIOD_LEN;
550 /* 0 has a special meaning... */
556 static void wb_domain_writeout_add(struct wb_domain *dom,
557 struct fprop_local_percpu *completions,
558 unsigned int max_prop_frac, long nr)
560 __fprop_add_percpu_max(&dom->completions, completions,
562 /* First event after period switching was turned off? */
563 if (unlikely(!dom->period_time)) {
565 * We can race with other __bdi_writeout_inc calls here but
566 * it does not cause any harm since the resulting time when
567 * timer will fire and what is in writeout_period_time will be
570 dom->period_time = wp_next_time(jiffies);
571 mod_timer(&dom->period_timer, dom->period_time);
576 * Increment @wb's writeout completion count and the global writeout
577 * completion count. Called from __folio_end_writeback().
579 static inline void __wb_writeout_add(struct bdi_writeback *wb, long nr)
581 struct wb_domain *cgdom;
583 wb_stat_mod(wb, WB_WRITTEN, nr);
584 wb_domain_writeout_add(&global_wb_domain, &wb->completions,
585 wb->bdi->max_prop_frac, nr);
587 cgdom = mem_cgroup_wb_domain(wb);
589 wb_domain_writeout_add(cgdom, wb_memcg_completions(wb),
590 wb->bdi->max_prop_frac, nr);
593 void wb_writeout_inc(struct bdi_writeback *wb)
597 local_irq_save(flags);
598 __wb_writeout_add(wb, 1);
599 local_irq_restore(flags);
601 EXPORT_SYMBOL_GPL(wb_writeout_inc);
604 * On idle system, we can be called long after we scheduled because we use
605 * deferred timers so count with missed periods.
607 static void writeout_period(struct timer_list *t)
609 struct wb_domain *dom = from_timer(dom, t, period_timer);
610 int miss_periods = (jiffies - dom->period_time) /
611 VM_COMPLETIONS_PERIOD_LEN;
613 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
614 dom->period_time = wp_next_time(dom->period_time +
615 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
616 mod_timer(&dom->period_timer, dom->period_time);
619 * Aging has zeroed all fractions. Stop wasting CPU on period
622 dom->period_time = 0;
626 int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
628 memset(dom, 0, sizeof(*dom));
630 spin_lock_init(&dom->lock);
632 timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
634 dom->dirty_limit_tstamp = jiffies;
636 return fprop_global_init(&dom->completions, gfp);
639 #ifdef CONFIG_CGROUP_WRITEBACK
640 void wb_domain_exit(struct wb_domain *dom)
642 del_timer_sync(&dom->period_timer);
643 fprop_global_destroy(&dom->completions);
648 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
649 * registered backing devices, which, for obvious reasons, can not
652 static unsigned int bdi_min_ratio;
654 static int bdi_check_pages_limit(unsigned long pages)
656 unsigned long max_dirty_pages = global_dirtyable_memory();
658 if (pages > max_dirty_pages)
664 static unsigned long bdi_ratio_from_pages(unsigned long pages)
666 unsigned long background_thresh;
667 unsigned long dirty_thresh;
670 global_dirty_limits(&background_thresh, &dirty_thresh);
671 ratio = div64_u64(pages * 100ULL * BDI_RATIO_SCALE, dirty_thresh);
676 static u64 bdi_get_bytes(unsigned int ratio)
678 unsigned long background_thresh;
679 unsigned long dirty_thresh;
682 global_dirty_limits(&background_thresh, &dirty_thresh);
683 bytes = (dirty_thresh * PAGE_SIZE * ratio) / BDI_RATIO_SCALE / 100;
688 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
693 min_ratio *= BDI_RATIO_SCALE;
695 spin_lock_bh(&bdi_lock);
696 if (min_ratio > bdi->max_ratio) {
699 if (min_ratio < bdi->min_ratio) {
700 delta = bdi->min_ratio - min_ratio;
701 bdi_min_ratio -= delta;
702 bdi->min_ratio = min_ratio;
704 delta = min_ratio - bdi->min_ratio;
705 if (bdi_min_ratio + delta < 100 * BDI_RATIO_SCALE) {
706 bdi_min_ratio += delta;
707 bdi->min_ratio = min_ratio;
713 spin_unlock_bh(&bdi_lock);
718 static int __bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio)
722 spin_lock_bh(&bdi_lock);
723 if (bdi->min_ratio > max_ratio) {
726 bdi->max_ratio = max_ratio;
727 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
729 spin_unlock_bh(&bdi_lock);
734 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio)
739 return __bdi_set_max_ratio(bdi, max_ratio * BDI_RATIO_SCALE);
741 EXPORT_SYMBOL(bdi_set_max_ratio);
743 u64 bdi_get_min_bytes(struct backing_dev_info *bdi)
745 return bdi_get_bytes(bdi->min_ratio);
748 u64 bdi_get_max_bytes(struct backing_dev_info *bdi)
750 return bdi_get_bytes(bdi->max_ratio);
753 int bdi_set_max_bytes(struct backing_dev_info *bdi, u64 max_bytes)
756 unsigned long pages = max_bytes >> PAGE_SHIFT;
757 unsigned long max_ratio;
759 ret = bdi_check_pages_limit(pages);
763 max_ratio = bdi_ratio_from_pages(pages);
764 return __bdi_set_max_ratio(bdi, max_ratio);
767 int bdi_set_strict_limit(struct backing_dev_info *bdi, unsigned int strict_limit)
769 if (strict_limit > 1)
772 spin_lock_bh(&bdi_lock);
774 bdi->capabilities |= BDI_CAP_STRICTLIMIT;
776 bdi->capabilities &= ~BDI_CAP_STRICTLIMIT;
777 spin_unlock_bh(&bdi_lock);
782 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
783 unsigned long bg_thresh)
785 return (thresh + bg_thresh) / 2;
788 static unsigned long hard_dirty_limit(struct wb_domain *dom,
789 unsigned long thresh)
791 return max(thresh, dom->dirty_limit);
795 * Memory which can be further allocated to a memcg domain is capped by
796 * system-wide clean memory excluding the amount being used in the domain.
798 static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
799 unsigned long filepages, unsigned long headroom)
801 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
802 unsigned long clean = filepages - min(filepages, mdtc->dirty);
803 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
804 unsigned long other_clean = global_clean - min(global_clean, clean);
806 mdtc->avail = filepages + min(headroom, other_clean);
810 * __wb_calc_thresh - @wb's share of dirty throttling threshold
811 * @dtc: dirty_throttle_context of interest
813 * Note that balance_dirty_pages() will only seriously take it as a hard limit
814 * when sleeping max_pause per page is not enough to keep the dirty pages under
815 * control. For example, when the device is completely stalled due to some error
816 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
817 * In the other normal situations, it acts more gently by throttling the tasks
818 * more (rather than completely block them) when the wb dirty pages go high.
820 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
821 * - starving fast devices
822 * - piling up dirty pages (that will take long time to sync) on slow devices
824 * The wb's share of dirty limit will be adapting to its throughput and
825 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
827 * Return: @wb's dirty limit in pages. The term "dirty" in the context of
828 * dirty balancing includes all PG_dirty and PG_writeback pages.
830 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
832 struct wb_domain *dom = dtc_dom(dtc);
833 unsigned long thresh = dtc->thresh;
835 unsigned long numerator, denominator;
836 unsigned long wb_min_ratio, wb_max_ratio;
839 * Calculate this BDI's share of the thresh ratio.
841 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
842 &numerator, &denominator);
844 wb_thresh = (thresh * (100 * BDI_RATIO_SCALE - bdi_min_ratio)) / (100 * BDI_RATIO_SCALE);
845 wb_thresh *= numerator;
846 wb_thresh = div64_ul(wb_thresh, denominator);
848 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
850 wb_thresh += (thresh * wb_min_ratio) / (100 * BDI_RATIO_SCALE);
851 if (wb_thresh > (thresh * wb_max_ratio) / (100 * BDI_RATIO_SCALE))
852 wb_thresh = thresh * wb_max_ratio / (100 * BDI_RATIO_SCALE);
857 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
859 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
861 return __wb_calc_thresh(&gdtc);
866 * f(dirty) := 1.0 + (----------------)
869 * it's a 3rd order polynomial that subjects to
871 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
872 * (2) f(setpoint) = 1.0 => the balance point
873 * (3) f(limit) = 0 => the hard limit
874 * (4) df/dx <= 0 => negative feedback control
875 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
876 * => fast response on large errors; small oscillation near setpoint
878 static long long pos_ratio_polynom(unsigned long setpoint,
885 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
886 (limit - setpoint) | 1);
888 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
889 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
890 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
892 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
896 * Dirty position control.
898 * (o) global/bdi setpoints
900 * We want the dirty pages be balanced around the global/wb setpoints.
901 * When the number of dirty pages is higher/lower than the setpoint, the
902 * dirty position control ratio (and hence task dirty ratelimit) will be
903 * decreased/increased to bring the dirty pages back to the setpoint.
905 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
907 * if (dirty < setpoint) scale up pos_ratio
908 * if (dirty > setpoint) scale down pos_ratio
910 * if (wb_dirty < wb_setpoint) scale up pos_ratio
911 * if (wb_dirty > wb_setpoint) scale down pos_ratio
913 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
915 * (o) global control line
919 * | |<===== global dirty control scope ======>|
927 * 1.0 ................................*
933 * 0 +------------.------------------.----------------------*------------->
934 * freerun^ setpoint^ limit^ dirty pages
936 * (o) wb control line
944 * | * |<=========== span ============>|
945 * 1.0 .......................*
957 * 1/4 ...............................................* * * * * * * * * * * *
961 * 0 +----------------------.-------------------------------.------------->
962 * wb_setpoint^ x_intercept^
964 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
965 * be smoothly throttled down to normal if it starts high in situations like
966 * - start writing to a slow SD card and a fast disk at the same time. The SD
967 * card's wb_dirty may rush to many times higher than wb_setpoint.
968 * - the wb dirty thresh drops quickly due to change of JBOD workload
970 static void wb_position_ratio(struct dirty_throttle_control *dtc)
972 struct bdi_writeback *wb = dtc->wb;
973 unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth);
974 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
975 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
976 unsigned long wb_thresh = dtc->wb_thresh;
977 unsigned long x_intercept;
978 unsigned long setpoint; /* dirty pages' target balance point */
979 unsigned long wb_setpoint;
981 long long pos_ratio; /* for scaling up/down the rate limit */
986 if (unlikely(dtc->dirty >= limit))
992 * See comment for pos_ratio_polynom().
994 setpoint = (freerun + limit) / 2;
995 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
998 * The strictlimit feature is a tool preventing mistrusted filesystems
999 * from growing a large number of dirty pages before throttling. For
1000 * such filesystems balance_dirty_pages always checks wb counters
1001 * against wb limits. Even if global "nr_dirty" is under "freerun".
1002 * This is especially important for fuse which sets bdi->max_ratio to
1003 * 1% by default. Without strictlimit feature, fuse writeback may
1004 * consume arbitrary amount of RAM because it is accounted in
1005 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
1007 * Here, in wb_position_ratio(), we calculate pos_ratio based on
1008 * two values: wb_dirty and wb_thresh. Let's consider an example:
1009 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
1010 * limits are set by default to 10% and 20% (background and throttle).
1011 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
1012 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
1013 * about ~6K pages (as the average of background and throttle wb
1014 * limits). The 3rd order polynomial will provide positive feedback if
1015 * wb_dirty is under wb_setpoint and vice versa.
1017 * Note, that we cannot use global counters in these calculations
1018 * because we want to throttle process writing to a strictlimit wb
1019 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
1020 * in the example above).
1022 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1023 long long wb_pos_ratio;
1025 if (dtc->wb_dirty < 8) {
1026 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
1027 2 << RATELIMIT_CALC_SHIFT);
1031 if (dtc->wb_dirty >= wb_thresh)
1034 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
1037 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
1040 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
1044 * Typically, for strictlimit case, wb_setpoint << setpoint
1045 * and pos_ratio >> wb_pos_ratio. In the other words global
1046 * state ("dirty") is not limiting factor and we have to
1047 * make decision based on wb counters. But there is an
1048 * important case when global pos_ratio should get precedence:
1049 * global limits are exceeded (e.g. due to activities on other
1050 * wb's) while given strictlimit wb is below limit.
1052 * "pos_ratio * wb_pos_ratio" would work for the case above,
1053 * but it would look too non-natural for the case of all
1054 * activity in the system coming from a single strictlimit wb
1055 * with bdi->max_ratio == 100%.
1057 * Note that min() below somewhat changes the dynamics of the
1058 * control system. Normally, pos_ratio value can be well over 3
1059 * (when globally we are at freerun and wb is well below wb
1060 * setpoint). Now the maximum pos_ratio in the same situation
1061 * is 2. We might want to tweak this if we observe the control
1062 * system is too slow to adapt.
1064 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
1069 * We have computed basic pos_ratio above based on global situation. If
1070 * the wb is over/under its share of dirty pages, we want to scale
1071 * pos_ratio further down/up. That is done by the following mechanism.
1077 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1079 * x_intercept - wb_dirty
1080 * := --------------------------
1081 * x_intercept - wb_setpoint
1083 * The main wb control line is a linear function that subjects to
1085 * (1) f(wb_setpoint) = 1.0
1086 * (2) k = - 1 / (8 * write_bw) (in single wb case)
1087 * or equally: x_intercept = wb_setpoint + 8 * write_bw
1089 * For single wb case, the dirty pages are observed to fluctuate
1090 * regularly within range
1091 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1092 * for various filesystems, where (2) can yield in a reasonable 12.5%
1093 * fluctuation range for pos_ratio.
1095 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1096 * own size, so move the slope over accordingly and choose a slope that
1097 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1099 if (unlikely(wb_thresh > dtc->thresh))
1100 wb_thresh = dtc->thresh;
1102 * It's very possible that wb_thresh is close to 0 not because the
1103 * device is slow, but that it has remained inactive for long time.
1104 * Honour such devices a reasonable good (hopefully IO efficient)
1105 * threshold, so that the occasional writes won't be blocked and active
1106 * writes can rampup the threshold quickly.
1108 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1110 * scale global setpoint to wb's:
1111 * wb_setpoint = setpoint * wb_thresh / thresh
1113 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1114 wb_setpoint = setpoint * (u64)x >> 16;
1116 * Use span=(8*write_bw) in single wb case as indicated by
1117 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1119 * wb_thresh thresh - wb_thresh
1120 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1123 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1124 x_intercept = wb_setpoint + span;
1126 if (dtc->wb_dirty < x_intercept - span / 4) {
1127 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1128 (x_intercept - wb_setpoint) | 1);
1133 * wb reserve area, safeguard against dirty pool underrun and disk idle
1134 * It may push the desired control point of global dirty pages higher
1137 x_intercept = wb_thresh / 2;
1138 if (dtc->wb_dirty < x_intercept) {
1139 if (dtc->wb_dirty > x_intercept / 8)
1140 pos_ratio = div_u64(pos_ratio * x_intercept,
1146 dtc->pos_ratio = pos_ratio;
1149 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1150 unsigned long elapsed,
1151 unsigned long written)
1153 const unsigned long period = roundup_pow_of_two(3 * HZ);
1154 unsigned long avg = wb->avg_write_bandwidth;
1155 unsigned long old = wb->write_bandwidth;
1159 * bw = written * HZ / elapsed
1161 * bw * elapsed + write_bandwidth * (period - elapsed)
1162 * write_bandwidth = ---------------------------------------------------
1165 * @written may have decreased due to folio_account_redirty().
1166 * Avoid underflowing @bw calculation.
1168 bw = written - min(written, wb->written_stamp);
1170 if (unlikely(elapsed > period)) {
1171 bw = div64_ul(bw, elapsed);
1175 bw += (u64)wb->write_bandwidth * (period - elapsed);
1176 bw >>= ilog2(period);
1179 * one more level of smoothing, for filtering out sudden spikes
1181 if (avg > old && old >= (unsigned long)bw)
1182 avg -= (avg - old) >> 3;
1184 if (avg < old && old <= (unsigned long)bw)
1185 avg += (old - avg) >> 3;
1188 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1189 avg = max(avg, 1LU);
1190 if (wb_has_dirty_io(wb)) {
1191 long delta = avg - wb->avg_write_bandwidth;
1192 WARN_ON_ONCE(atomic_long_add_return(delta,
1193 &wb->bdi->tot_write_bandwidth) <= 0);
1195 wb->write_bandwidth = bw;
1196 WRITE_ONCE(wb->avg_write_bandwidth, avg);
1199 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1201 struct wb_domain *dom = dtc_dom(dtc);
1202 unsigned long thresh = dtc->thresh;
1203 unsigned long limit = dom->dirty_limit;
1206 * Follow up in one step.
1208 if (limit < thresh) {
1214 * Follow down slowly. Use the higher one as the target, because thresh
1215 * may drop below dirty. This is exactly the reason to introduce
1216 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1218 thresh = max(thresh, dtc->dirty);
1219 if (limit > thresh) {
1220 limit -= (limit - thresh) >> 5;
1225 dom->dirty_limit = limit;
1228 static void domain_update_dirty_limit(struct dirty_throttle_control *dtc,
1231 struct wb_domain *dom = dtc_dom(dtc);
1234 * check locklessly first to optimize away locking for the most time
1236 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1239 spin_lock(&dom->lock);
1240 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1241 update_dirty_limit(dtc);
1242 dom->dirty_limit_tstamp = now;
1244 spin_unlock(&dom->lock);
1248 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1250 * Normal wb tasks will be curbed at or below it in long term.
1251 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1253 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1254 unsigned long dirtied,
1255 unsigned long elapsed)
1257 struct bdi_writeback *wb = dtc->wb;
1258 unsigned long dirty = dtc->dirty;
1259 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1260 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1261 unsigned long setpoint = (freerun + limit) / 2;
1262 unsigned long write_bw = wb->avg_write_bandwidth;
1263 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1264 unsigned long dirty_rate;
1265 unsigned long task_ratelimit;
1266 unsigned long balanced_dirty_ratelimit;
1269 unsigned long shift;
1272 * The dirty rate will match the writeout rate in long term, except
1273 * when dirty pages are truncated by userspace or re-dirtied by FS.
1275 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1278 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1280 task_ratelimit = (u64)dirty_ratelimit *
1281 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1282 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1285 * A linear estimation of the "balanced" throttle rate. The theory is,
1286 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1287 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1288 * formula will yield the balanced rate limit (write_bw / N).
1290 * Note that the expanded form is not a pure rate feedback:
1291 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1292 * but also takes pos_ratio into account:
1293 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1295 * (1) is not realistic because pos_ratio also takes part in balancing
1296 * the dirty rate. Consider the state
1297 * pos_ratio = 0.5 (3)
1298 * rate = 2 * (write_bw / N) (4)
1299 * If (1) is used, it will stuck in that state! Because each dd will
1301 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1303 * dirty_rate = N * task_ratelimit = write_bw (6)
1304 * put (6) into (1) we get
1305 * rate_(i+1) = rate_(i) (7)
1307 * So we end up using (2) to always keep
1308 * rate_(i+1) ~= (write_bw / N) (8)
1309 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1310 * pos_ratio is able to drive itself to 1.0, which is not only where
1311 * the dirty count meet the setpoint, but also where the slope of
1312 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1314 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1317 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1319 if (unlikely(balanced_dirty_ratelimit > write_bw))
1320 balanced_dirty_ratelimit = write_bw;
1323 * We could safely do this and return immediately:
1325 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1327 * However to get a more stable dirty_ratelimit, the below elaborated
1328 * code makes use of task_ratelimit to filter out singular points and
1329 * limit the step size.
1331 * The below code essentially only uses the relative value of
1333 * task_ratelimit - dirty_ratelimit
1334 * = (pos_ratio - 1) * dirty_ratelimit
1336 * which reflects the direction and size of dirty position error.
1340 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1341 * task_ratelimit is on the same side of dirty_ratelimit, too.
1343 * - dirty_ratelimit > balanced_dirty_ratelimit
1344 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1345 * lowering dirty_ratelimit will help meet both the position and rate
1346 * control targets. Otherwise, don't update dirty_ratelimit if it will
1347 * only help meet the rate target. After all, what the users ultimately
1348 * feel and care are stable dirty rate and small position error.
1350 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1351 * and filter out the singular points of balanced_dirty_ratelimit. Which
1352 * keeps jumping around randomly and can even leap far away at times
1353 * due to the small 200ms estimation period of dirty_rate (we want to
1354 * keep that period small to reduce time lags).
1359 * For strictlimit case, calculations above were based on wb counters
1360 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1361 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1362 * Hence, to calculate "step" properly, we have to use wb_dirty as
1363 * "dirty" and wb_setpoint as "setpoint".
1365 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1366 * it's possible that wb_thresh is close to zero due to inactivity
1367 * of backing device.
1369 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1370 dirty = dtc->wb_dirty;
1371 if (dtc->wb_dirty < 8)
1372 setpoint = dtc->wb_dirty + 1;
1374 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1377 if (dirty < setpoint) {
1378 x = min3(wb->balanced_dirty_ratelimit,
1379 balanced_dirty_ratelimit, task_ratelimit);
1380 if (dirty_ratelimit < x)
1381 step = x - dirty_ratelimit;
1383 x = max3(wb->balanced_dirty_ratelimit,
1384 balanced_dirty_ratelimit, task_ratelimit);
1385 if (dirty_ratelimit > x)
1386 step = dirty_ratelimit - x;
1390 * Don't pursue 100% rate matching. It's impossible since the balanced
1391 * rate itself is constantly fluctuating. So decrease the track speed
1392 * when it gets close to the target. Helps eliminate pointless tremors.
1394 shift = dirty_ratelimit / (2 * step + 1);
1395 if (shift < BITS_PER_LONG)
1396 step = DIV_ROUND_UP(step >> shift, 8);
1400 if (dirty_ratelimit < balanced_dirty_ratelimit)
1401 dirty_ratelimit += step;
1403 dirty_ratelimit -= step;
1405 WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL));
1406 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1408 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1411 static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1412 struct dirty_throttle_control *mdtc,
1413 bool update_ratelimit)
1415 struct bdi_writeback *wb = gdtc->wb;
1416 unsigned long now = jiffies;
1417 unsigned long elapsed;
1418 unsigned long dirtied;
1419 unsigned long written;
1421 spin_lock(&wb->list_lock);
1424 * Lockless checks for elapsed time are racy and delayed update after
1425 * IO completion doesn't do it at all (to make sure written pages are
1426 * accounted reasonably quickly). Make sure elapsed >= 1 to avoid
1429 elapsed = max(now - wb->bw_time_stamp, 1UL);
1430 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1431 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1433 if (update_ratelimit) {
1434 domain_update_dirty_limit(gdtc, now);
1435 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1438 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1439 * compiler has no way to figure that out. Help it.
1441 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1442 domain_update_dirty_limit(mdtc, now);
1443 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1446 wb_update_write_bandwidth(wb, elapsed, written);
1448 wb->dirtied_stamp = dirtied;
1449 wb->written_stamp = written;
1450 WRITE_ONCE(wb->bw_time_stamp, now);
1451 spin_unlock(&wb->list_lock);
1454 void wb_update_bandwidth(struct bdi_writeback *wb)
1456 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1458 __wb_update_bandwidth(&gdtc, NULL, false);
1461 /* Interval after which we consider wb idle and don't estimate bandwidth */
1462 #define WB_BANDWIDTH_IDLE_JIF (HZ)
1464 static void wb_bandwidth_estimate_start(struct bdi_writeback *wb)
1466 unsigned long now = jiffies;
1467 unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp);
1469 if (elapsed > WB_BANDWIDTH_IDLE_JIF &&
1470 !atomic_read(&wb->writeback_inodes)) {
1471 spin_lock(&wb->list_lock);
1472 wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED);
1473 wb->written_stamp = wb_stat(wb, WB_WRITTEN);
1474 WRITE_ONCE(wb->bw_time_stamp, now);
1475 spin_unlock(&wb->list_lock);
1480 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1481 * will look to see if it needs to start dirty throttling.
1483 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1484 * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1485 * (the number of pages we may dirty without exceeding the dirty limits).
1487 static unsigned long dirty_poll_interval(unsigned long dirty,
1488 unsigned long thresh)
1491 return 1UL << (ilog2(thresh - dirty) >> 1);
1496 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1497 unsigned long wb_dirty)
1499 unsigned long bw = READ_ONCE(wb->avg_write_bandwidth);
1503 * Limit pause time for small memory systems. If sleeping for too long
1504 * time, a small pool of dirty/writeback pages may go empty and disk go
1507 * 8 serves as the safety ratio.
1509 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1512 return min_t(unsigned long, t, MAX_PAUSE);
1515 static long wb_min_pause(struct bdi_writeback *wb,
1517 unsigned long task_ratelimit,
1518 unsigned long dirty_ratelimit,
1519 int *nr_dirtied_pause)
1521 long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth));
1522 long lo = ilog2(READ_ONCE(wb->dirty_ratelimit));
1523 long t; /* target pause */
1524 long pause; /* estimated next pause */
1525 int pages; /* target nr_dirtied_pause */
1527 /* target for 10ms pause on 1-dd case */
1528 t = max(1, HZ / 100);
1531 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1534 * (N * 10ms) on 2^N concurrent tasks.
1537 t += (hi - lo) * (10 * HZ) / 1024;
1540 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1541 * on the much more stable dirty_ratelimit. However the next pause time
1542 * will be computed based on task_ratelimit and the two rate limits may
1543 * depart considerably at some time. Especially if task_ratelimit goes
1544 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1545 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1546 * result task_ratelimit won't be executed faithfully, which could
1547 * eventually bring down dirty_ratelimit.
1549 * We apply two rules to fix it up:
1550 * 1) try to estimate the next pause time and if necessary, use a lower
1551 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1552 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1553 * 2) limit the target pause time to max_pause/2, so that the normal
1554 * small fluctuations of task_ratelimit won't trigger rule (1) and
1555 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1557 t = min(t, 1 + max_pause / 2);
1558 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1561 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1562 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1563 * When the 16 consecutive reads are often interrupted by some dirty
1564 * throttling pause during the async writes, cfq will go into idles
1565 * (deadline is fine). So push nr_dirtied_pause as high as possible
1566 * until reaches DIRTY_POLL_THRESH=32 pages.
1568 if (pages < DIRTY_POLL_THRESH) {
1570 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1571 if (pages > DIRTY_POLL_THRESH) {
1572 pages = DIRTY_POLL_THRESH;
1573 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1577 pause = HZ * pages / (task_ratelimit + 1);
1578 if (pause > max_pause) {
1580 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1583 *nr_dirtied_pause = pages;
1585 * The minimal pause time will normally be half the target pause time.
1587 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1590 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1592 struct bdi_writeback *wb = dtc->wb;
1593 unsigned long wb_reclaimable;
1596 * wb_thresh is not treated as some limiting factor as
1597 * dirty_thresh, due to reasons
1598 * - in JBOD setup, wb_thresh can fluctuate a lot
1599 * - in a system with HDD and USB key, the USB key may somehow
1600 * go into state (wb_dirty >> wb_thresh) either because
1601 * wb_dirty starts high, or because wb_thresh drops low.
1602 * In this case we don't want to hard throttle the USB key
1603 * dirtiers for 100 seconds until wb_dirty drops under
1604 * wb_thresh. Instead the auxiliary wb control line in
1605 * wb_position_ratio() will let the dirtier task progress
1606 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1608 dtc->wb_thresh = __wb_calc_thresh(dtc);
1609 dtc->wb_bg_thresh = dtc->thresh ?
1610 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1613 * In order to avoid the stacked BDI deadlock we need
1614 * to ensure we accurately count the 'dirty' pages when
1615 * the threshold is low.
1617 * Otherwise it would be possible to get thresh+n pages
1618 * reported dirty, even though there are thresh-m pages
1619 * actually dirty; with m+n sitting in the percpu
1622 if (dtc->wb_thresh < 2 * wb_stat_error()) {
1623 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1624 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1626 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1627 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1632 * balance_dirty_pages() must be called by processes which are generating dirty
1633 * data. It looks at the number of dirty pages in the machine and will force
1634 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1635 * If we're over `background_thresh' then the writeback threads are woken to
1636 * perform some writeout.
1638 static int balance_dirty_pages(struct bdi_writeback *wb,
1639 unsigned long pages_dirtied, unsigned int flags)
1641 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1642 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1643 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1644 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1646 struct dirty_throttle_control *sdtc;
1647 unsigned long nr_reclaimable; /* = file_dirty */
1652 int nr_dirtied_pause;
1653 bool dirty_exceeded = false;
1654 unsigned long task_ratelimit;
1655 unsigned long dirty_ratelimit;
1656 struct backing_dev_info *bdi = wb->bdi;
1657 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1658 unsigned long start_time = jiffies;
1662 unsigned long now = jiffies;
1663 unsigned long dirty, thresh, bg_thresh;
1664 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1665 unsigned long m_thresh = 0;
1666 unsigned long m_bg_thresh = 0;
1668 nr_reclaimable = global_node_page_state(NR_FILE_DIRTY);
1669 gdtc->avail = global_dirtyable_memory();
1670 gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1672 domain_dirty_limits(gdtc);
1674 if (unlikely(strictlimit)) {
1675 wb_dirty_limits(gdtc);
1677 dirty = gdtc->wb_dirty;
1678 thresh = gdtc->wb_thresh;
1679 bg_thresh = gdtc->wb_bg_thresh;
1681 dirty = gdtc->dirty;
1682 thresh = gdtc->thresh;
1683 bg_thresh = gdtc->bg_thresh;
1687 unsigned long filepages, headroom, writeback;
1690 * If @wb belongs to !root memcg, repeat the same
1691 * basic calculations for the memcg domain.
1693 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1694 &mdtc->dirty, &writeback);
1695 mdtc->dirty += writeback;
1696 mdtc_calc_avail(mdtc, filepages, headroom);
1698 domain_dirty_limits(mdtc);
1700 if (unlikely(strictlimit)) {
1701 wb_dirty_limits(mdtc);
1702 m_dirty = mdtc->wb_dirty;
1703 m_thresh = mdtc->wb_thresh;
1704 m_bg_thresh = mdtc->wb_bg_thresh;
1706 m_dirty = mdtc->dirty;
1707 m_thresh = mdtc->thresh;
1708 m_bg_thresh = mdtc->bg_thresh;
1713 * In laptop mode, we wait until hitting the higher threshold
1714 * before starting background writeout, and then write out all
1715 * the way down to the lower threshold. So slow writers cause
1716 * minimal disk activity.
1718 * In normal mode, we start background writeout at the lower
1719 * background_thresh, to keep the amount of dirty memory low.
1721 if (!laptop_mode && nr_reclaimable > gdtc->bg_thresh &&
1722 !writeback_in_progress(wb))
1723 wb_start_background_writeback(wb);
1726 * Throttle it only when the background writeback cannot
1727 * catch-up. This avoids (excessively) small writeouts
1728 * when the wb limits are ramping up in case of !strictlimit.
1730 * In strictlimit case make decision based on the wb counters
1731 * and limits. Small writeouts when the wb limits are ramping
1732 * up are the price we consciously pay for strictlimit-ing.
1734 * If memcg domain is in effect, @dirty should be under
1735 * both global and memcg freerun ceilings.
1737 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1739 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1741 unsigned long m_intv;
1744 intv = dirty_poll_interval(dirty, thresh);
1747 current->dirty_paused_when = now;
1748 current->nr_dirtied = 0;
1750 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1751 current->nr_dirtied_pause = min(intv, m_intv);
1755 /* Start writeback even when in laptop mode */
1756 if (unlikely(!writeback_in_progress(wb)))
1757 wb_start_background_writeback(wb);
1759 mem_cgroup_flush_foreign(wb);
1762 * Calculate global domain's pos_ratio and select the
1763 * global dtc by default.
1766 wb_dirty_limits(gdtc);
1768 if ((current->flags & PF_LOCAL_THROTTLE) &&
1770 dirty_freerun_ceiling(gdtc->wb_thresh,
1771 gdtc->wb_bg_thresh))
1773 * LOCAL_THROTTLE tasks must not be throttled
1774 * when below the per-wb freerun ceiling.
1779 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1780 ((gdtc->dirty > gdtc->thresh) || strictlimit);
1782 wb_position_ratio(gdtc);
1787 * If memcg domain is in effect, calculate its
1788 * pos_ratio. @wb should satisfy constraints from
1789 * both global and memcg domains. Choose the one
1790 * w/ lower pos_ratio.
1793 wb_dirty_limits(mdtc);
1795 if ((current->flags & PF_LOCAL_THROTTLE) &&
1797 dirty_freerun_ceiling(mdtc->wb_thresh,
1798 mdtc->wb_bg_thresh))
1800 * LOCAL_THROTTLE tasks must not be
1801 * throttled when below the per-wb
1806 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1807 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1809 wb_position_ratio(mdtc);
1810 if (mdtc->pos_ratio < gdtc->pos_ratio)
1814 if (dirty_exceeded != wb->dirty_exceeded)
1815 wb->dirty_exceeded = dirty_exceeded;
1817 if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
1818 BANDWIDTH_INTERVAL))
1819 __wb_update_bandwidth(gdtc, mdtc, true);
1821 /* throttle according to the chosen dtc */
1822 dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit);
1823 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1824 RATELIMIT_CALC_SHIFT;
1825 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1826 min_pause = wb_min_pause(wb, max_pause,
1827 task_ratelimit, dirty_ratelimit,
1830 if (unlikely(task_ratelimit == 0)) {
1835 period = HZ * pages_dirtied / task_ratelimit;
1837 if (current->dirty_paused_when)
1838 pause -= now - current->dirty_paused_when;
1840 * For less than 1s think time (ext3/4 may block the dirtier
1841 * for up to 800ms from time to time on 1-HDD; so does xfs,
1842 * however at much less frequency), try to compensate it in
1843 * future periods by updating the virtual time; otherwise just
1844 * do a reset, as it may be a light dirtier.
1846 if (pause < min_pause) {
1847 trace_balance_dirty_pages(wb,
1860 current->dirty_paused_when = now;
1861 current->nr_dirtied = 0;
1862 } else if (period) {
1863 current->dirty_paused_when += period;
1864 current->nr_dirtied = 0;
1865 } else if (current->nr_dirtied_pause <= pages_dirtied)
1866 current->nr_dirtied_pause += pages_dirtied;
1869 if (unlikely(pause > max_pause)) {
1870 /* for occasional dropped task_ratelimit */
1871 now += min(pause - max_pause, max_pause);
1876 trace_balance_dirty_pages(wb,
1888 if (flags & BDP_ASYNC) {
1892 __set_current_state(TASK_KILLABLE);
1893 wb->dirty_sleep = now;
1894 io_schedule_timeout(pause);
1896 current->dirty_paused_when = now + pause;
1897 current->nr_dirtied = 0;
1898 current->nr_dirtied_pause = nr_dirtied_pause;
1901 * This is typically equal to (dirty < thresh) and can also
1902 * keep "1000+ dd on a slow USB stick" under control.
1908 * In the case of an unresponsive NFS server and the NFS dirty
1909 * pages exceeds dirty_thresh, give the other good wb's a pipe
1910 * to go through, so that tasks on them still remain responsive.
1912 * In theory 1 page is enough to keep the consumer-producer
1913 * pipe going: the flusher cleans 1 page => the task dirties 1
1914 * more page. However wb_dirty has accounting errors. So use
1915 * the larger and more IO friendly wb_stat_error.
1917 if (sdtc->wb_dirty <= wb_stat_error())
1920 if (fatal_signal_pending(current))
1926 static DEFINE_PER_CPU(int, bdp_ratelimits);
1929 * Normal tasks are throttled by
1931 * dirty tsk->nr_dirtied_pause pages;
1932 * take a snap in balance_dirty_pages();
1934 * However there is a worst case. If every task exit immediately when dirtied
1935 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1936 * called to throttle the page dirties. The solution is to save the not yet
1937 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1938 * randomly into the running tasks. This works well for the above worst case,
1939 * as the new task will pick up and accumulate the old task's leaked dirty
1940 * count and eventually get throttled.
1942 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1945 * balance_dirty_pages_ratelimited_flags - Balance dirty memory state.
1946 * @mapping: address_space which was dirtied.
1947 * @flags: BDP flags.
1949 * Processes which are dirtying memory should call in here once for each page
1950 * which was newly dirtied. The function will periodically check the system's
1951 * dirty state and will initiate writeback if needed.
1953 * See balance_dirty_pages_ratelimited() for details.
1955 * Return: If @flags contains BDP_ASYNC, it may return -EAGAIN to
1956 * indicate that memory is out of balance and the caller must wait
1957 * for I/O to complete. Otherwise, it will return 0 to indicate
1958 * that either memory was already in balance, or it was able to sleep
1959 * until the amount of dirty memory returned to balance.
1961 int balance_dirty_pages_ratelimited_flags(struct address_space *mapping,
1964 struct inode *inode = mapping->host;
1965 struct backing_dev_info *bdi = inode_to_bdi(inode);
1966 struct bdi_writeback *wb = NULL;
1971 if (!(bdi->capabilities & BDI_CAP_WRITEBACK))
1974 if (inode_cgwb_enabled(inode))
1975 wb = wb_get_create_current(bdi, GFP_KERNEL);
1979 ratelimit = current->nr_dirtied_pause;
1980 if (wb->dirty_exceeded)
1981 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1985 * This prevents one CPU to accumulate too many dirtied pages without
1986 * calling into balance_dirty_pages(), which can happen when there are
1987 * 1000+ tasks, all of them start dirtying pages at exactly the same
1988 * time, hence all honoured too large initial task->nr_dirtied_pause.
1990 p = this_cpu_ptr(&bdp_ratelimits);
1991 if (unlikely(current->nr_dirtied >= ratelimit))
1993 else if (unlikely(*p >= ratelimit_pages)) {
1998 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1999 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
2000 * the dirty throttling and livelock other long-run dirtiers.
2002 p = this_cpu_ptr(&dirty_throttle_leaks);
2003 if (*p > 0 && current->nr_dirtied < ratelimit) {
2004 unsigned long nr_pages_dirtied;
2005 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
2006 *p -= nr_pages_dirtied;
2007 current->nr_dirtied += nr_pages_dirtied;
2011 if (unlikely(current->nr_dirtied >= ratelimit))
2012 ret = balance_dirty_pages(wb, current->nr_dirtied, flags);
2017 EXPORT_SYMBOL_GPL(balance_dirty_pages_ratelimited_flags);
2020 * balance_dirty_pages_ratelimited - balance dirty memory state.
2021 * @mapping: address_space which was dirtied.
2023 * Processes which are dirtying memory should call in here once for each page
2024 * which was newly dirtied. The function will periodically check the system's
2025 * dirty state and will initiate writeback if needed.
2027 * Once we're over the dirty memory limit we decrease the ratelimiting
2028 * by a lot, to prevent individual processes from overshooting the limit
2029 * by (ratelimit_pages) each.
2031 void balance_dirty_pages_ratelimited(struct address_space *mapping)
2033 balance_dirty_pages_ratelimited_flags(mapping, 0);
2035 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
2038 * wb_over_bg_thresh - does @wb need to be written back?
2039 * @wb: bdi_writeback of interest
2041 * Determines whether background writeback should keep writing @wb or it's
2044 * Return: %true if writeback should continue.
2046 bool wb_over_bg_thresh(struct bdi_writeback *wb)
2048 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
2049 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
2050 struct dirty_throttle_control * const gdtc = &gdtc_stor;
2051 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
2053 unsigned long reclaimable;
2054 unsigned long thresh;
2057 * Similar to balance_dirty_pages() but ignores pages being written
2058 * as we're trying to decide whether to put more under writeback.
2060 gdtc->avail = global_dirtyable_memory();
2061 gdtc->dirty = global_node_page_state(NR_FILE_DIRTY);
2062 domain_dirty_limits(gdtc);
2064 if (gdtc->dirty > gdtc->bg_thresh)
2067 thresh = wb_calc_thresh(gdtc->wb, gdtc->bg_thresh);
2068 if (thresh < 2 * wb_stat_error())
2069 reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
2071 reclaimable = wb_stat(wb, WB_RECLAIMABLE);
2073 if (reclaimable > thresh)
2077 unsigned long filepages, headroom, writeback;
2079 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
2081 mdtc_calc_avail(mdtc, filepages, headroom);
2082 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
2084 if (mdtc->dirty > mdtc->bg_thresh)
2087 thresh = wb_calc_thresh(mdtc->wb, mdtc->bg_thresh);
2088 if (thresh < 2 * wb_stat_error())
2089 reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
2091 reclaimable = wb_stat(wb, WB_RECLAIMABLE);
2093 if (reclaimable > thresh)
2100 #ifdef CONFIG_SYSCTL
2102 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
2104 static int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
2105 void *buffer, size_t *length, loff_t *ppos)
2107 unsigned int old_interval = dirty_writeback_interval;
2110 ret = proc_dointvec(table, write, buffer, length, ppos);
2113 * Writing 0 to dirty_writeback_interval will disable periodic writeback
2114 * and a different non-zero value will wakeup the writeback threads.
2115 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
2116 * iterate over all bdis and wbs.
2117 * The reason we do this is to make the change take effect immediately.
2119 if (!ret && write && dirty_writeback_interval &&
2120 dirty_writeback_interval != old_interval)
2121 wakeup_flusher_threads(WB_REASON_PERIODIC);
2127 void laptop_mode_timer_fn(struct timer_list *t)
2129 struct backing_dev_info *backing_dev_info =
2130 from_timer(backing_dev_info, t, laptop_mode_wb_timer);
2132 wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
2136 * We've spun up the disk and we're in laptop mode: schedule writeback
2137 * of all dirty data a few seconds from now. If the flush is already scheduled
2138 * then push it back - the user is still using the disk.
2140 void laptop_io_completion(struct backing_dev_info *info)
2142 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2146 * We're in laptop mode and we've just synced. The sync's writes will have
2147 * caused another writeback to be scheduled by laptop_io_completion.
2148 * Nothing needs to be written back anymore, so we unschedule the writeback.
2150 void laptop_sync_completion(void)
2152 struct backing_dev_info *bdi;
2156 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2157 del_timer(&bdi->laptop_mode_wb_timer);
2163 * If ratelimit_pages is too high then we can get into dirty-data overload
2164 * if a large number of processes all perform writes at the same time.
2166 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2167 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2171 void writeback_set_ratelimit(void)
2173 struct wb_domain *dom = &global_wb_domain;
2174 unsigned long background_thresh;
2175 unsigned long dirty_thresh;
2177 global_dirty_limits(&background_thresh, &dirty_thresh);
2178 dom->dirty_limit = dirty_thresh;
2179 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2180 if (ratelimit_pages < 16)
2181 ratelimit_pages = 16;
2184 static int page_writeback_cpu_online(unsigned int cpu)
2186 writeback_set_ratelimit();
2190 #ifdef CONFIG_SYSCTL
2192 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
2193 static const unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
2195 static struct ctl_table vm_page_writeback_sysctls[] = {
2197 .procname = "dirty_background_ratio",
2198 .data = &dirty_background_ratio,
2199 .maxlen = sizeof(dirty_background_ratio),
2201 .proc_handler = dirty_background_ratio_handler,
2202 .extra1 = SYSCTL_ZERO,
2203 .extra2 = SYSCTL_ONE_HUNDRED,
2206 .procname = "dirty_background_bytes",
2207 .data = &dirty_background_bytes,
2208 .maxlen = sizeof(dirty_background_bytes),
2210 .proc_handler = dirty_background_bytes_handler,
2211 .extra1 = SYSCTL_LONG_ONE,
2214 .procname = "dirty_ratio",
2215 .data = &vm_dirty_ratio,
2216 .maxlen = sizeof(vm_dirty_ratio),
2218 .proc_handler = dirty_ratio_handler,
2219 .extra1 = SYSCTL_ZERO,
2220 .extra2 = SYSCTL_ONE_HUNDRED,
2223 .procname = "dirty_bytes",
2224 .data = &vm_dirty_bytes,
2225 .maxlen = sizeof(vm_dirty_bytes),
2227 .proc_handler = dirty_bytes_handler,
2228 .extra1 = (void *)&dirty_bytes_min,
2231 .procname = "dirty_writeback_centisecs",
2232 .data = &dirty_writeback_interval,
2233 .maxlen = sizeof(dirty_writeback_interval),
2235 .proc_handler = dirty_writeback_centisecs_handler,
2238 .procname = "dirty_expire_centisecs",
2239 .data = &dirty_expire_interval,
2240 .maxlen = sizeof(dirty_expire_interval),
2242 .proc_handler = proc_dointvec_minmax,
2243 .extra1 = SYSCTL_ZERO,
2245 #ifdef CONFIG_HIGHMEM
2247 .procname = "highmem_is_dirtyable",
2248 .data = &vm_highmem_is_dirtyable,
2249 .maxlen = sizeof(vm_highmem_is_dirtyable),
2251 .proc_handler = proc_dointvec_minmax,
2252 .extra1 = SYSCTL_ZERO,
2253 .extra2 = SYSCTL_ONE,
2257 .procname = "laptop_mode",
2258 .data = &laptop_mode,
2259 .maxlen = sizeof(laptop_mode),
2261 .proc_handler = proc_dointvec_jiffies,
2268 * Called early on to tune the page writeback dirty limits.
2270 * We used to scale dirty pages according to how total memory
2271 * related to pages that could be allocated for buffers.
2273 * However, that was when we used "dirty_ratio" to scale with
2274 * all memory, and we don't do that any more. "dirty_ratio"
2275 * is now applied to total non-HIGHPAGE memory, and as such we can't
2276 * get into the old insane situation any more where we had
2277 * large amounts of dirty pages compared to a small amount of
2278 * non-HIGHMEM memory.
2280 * But we might still want to scale the dirty_ratio by how
2281 * much memory the box has..
2283 void __init page_writeback_init(void)
2285 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2287 cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2288 page_writeback_cpu_online, NULL);
2289 cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2290 page_writeback_cpu_online);
2291 #ifdef CONFIG_SYSCTL
2292 register_sysctl_init("vm", vm_page_writeback_sysctls);
2297 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2298 * @mapping: address space structure to write
2299 * @start: starting page index
2300 * @end: ending page index (inclusive)
2302 * This function scans the page range from @start to @end (inclusive) and tags
2303 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2304 * that write_cache_pages (or whoever calls this function) will then use
2305 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2306 * used to avoid livelocking of writeback by a process steadily creating new
2307 * dirty pages in the file (thus it is important for this function to be quick
2308 * so that it can tag pages faster than a dirtying process can create them).
2310 void tag_pages_for_writeback(struct address_space *mapping,
2311 pgoff_t start, pgoff_t end)
2313 XA_STATE(xas, &mapping->i_pages, start);
2314 unsigned int tagged = 0;
2318 xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2319 xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2320 if (++tagged % XA_CHECK_SCHED)
2324 xas_unlock_irq(&xas);
2328 xas_unlock_irq(&xas);
2330 EXPORT_SYMBOL(tag_pages_for_writeback);
2333 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2334 * @mapping: address space structure to write
2335 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2336 * @writepage: function called for each page
2337 * @data: data passed to writepage function
2339 * If a page is already under I/O, write_cache_pages() skips it, even
2340 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2341 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2342 * and msync() need to guarantee that all the data which was dirty at the time
2343 * the call was made get new I/O started against them. If wbc->sync_mode is
2344 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2345 * existing IO to complete.
2347 * To avoid livelocks (when other process dirties new pages), we first tag
2348 * pages which should be written back with TOWRITE tag and only then start
2349 * writing them. For data-integrity sync we have to be careful so that we do
2350 * not miss some pages (e.g., because some other process has cleared TOWRITE
2351 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2352 * by the process clearing the DIRTY tag (and submitting the page for IO).
2354 * To avoid deadlocks between range_cyclic writeback and callers that hold
2355 * pages in PageWriteback to aggregate IO until write_cache_pages() returns,
2356 * we do not loop back to the start of the file. Doing so causes a page
2357 * lock/page writeback access order inversion - we should only ever lock
2358 * multiple pages in ascending page->index order, and looping back to the start
2359 * of the file violates that rule and causes deadlocks.
2361 * Return: %0 on success, negative error code otherwise
2363 int write_cache_pages(struct address_space *mapping,
2364 struct writeback_control *wbc, writepage_t writepage,
2370 struct pagevec pvec;
2373 pgoff_t end; /* Inclusive */
2375 int range_whole = 0;
2378 pagevec_init(&pvec);
2379 if (wbc->range_cyclic) {
2380 index = mapping->writeback_index; /* prev offset */
2383 index = wbc->range_start >> PAGE_SHIFT;
2384 end = wbc->range_end >> PAGE_SHIFT;
2385 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2388 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
2389 tag_pages_for_writeback(mapping, index, end);
2390 tag = PAGECACHE_TAG_TOWRITE;
2392 tag = PAGECACHE_TAG_DIRTY;
2395 while (!done && (index <= end)) {
2398 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2403 for (i = 0; i < nr_pages; i++) {
2404 struct page *page = pvec.pages[i];
2406 done_index = page->index;
2411 * Page truncated or invalidated. We can freely skip it
2412 * then, even for data integrity operations: the page
2413 * has disappeared concurrently, so there could be no
2414 * real expectation of this data integrity operation
2415 * even if there is now a new, dirty page at the same
2416 * pagecache address.
2418 if (unlikely(page->mapping != mapping)) {
2424 if (!PageDirty(page)) {
2425 /* someone wrote it for us */
2426 goto continue_unlock;
2429 if (PageWriteback(page)) {
2430 if (wbc->sync_mode != WB_SYNC_NONE)
2431 wait_on_page_writeback(page);
2433 goto continue_unlock;
2436 BUG_ON(PageWriteback(page));
2437 if (!clear_page_dirty_for_io(page))
2438 goto continue_unlock;
2440 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2441 error = (*writepage)(page, wbc, data);
2442 if (unlikely(error)) {
2444 * Handle errors according to the type of
2445 * writeback. There's no need to continue for
2446 * background writeback. Just push done_index
2447 * past this page so media errors won't choke
2448 * writeout for the entire file. For integrity
2449 * writeback, we must process the entire dirty
2450 * set regardless of errors because the fs may
2451 * still have state to clear for each page. In
2452 * that case we continue processing and return
2455 if (error == AOP_WRITEPAGE_ACTIVATE) {
2458 } else if (wbc->sync_mode != WB_SYNC_ALL) {
2460 done_index = page->index + 1;
2469 * We stop writing back only if we are not doing
2470 * integrity sync. In case of integrity sync we have to
2471 * keep going until we have written all the pages
2472 * we tagged for writeback prior to entering this loop.
2474 if (--wbc->nr_to_write <= 0 &&
2475 wbc->sync_mode == WB_SYNC_NONE) {
2480 pagevec_release(&pvec);
2485 * If we hit the last page and there is more work to be done: wrap
2486 * back the index back to the start of the file for the next
2487 * time we are called.
2489 if (wbc->range_cyclic && !done)
2491 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2492 mapping->writeback_index = done_index;
2496 EXPORT_SYMBOL(write_cache_pages);
2499 * Function used by generic_writepages to call the real writepage
2500 * function and set the mapping flags on error
2502 static int __writepage(struct page *page, struct writeback_control *wbc,
2505 struct address_space *mapping = data;
2506 int ret = mapping->a_ops->writepage(page, wbc);
2507 mapping_set_error(mapping, ret);
2512 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2513 * @mapping: address space structure to write
2514 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2516 * This is a library function, which implements the writepages()
2517 * address_space_operation.
2519 * Return: %0 on success, negative error code otherwise
2521 int generic_writepages(struct address_space *mapping,
2522 struct writeback_control *wbc)
2524 struct blk_plug plug;
2527 /* deal with chardevs and other special file */
2528 if (!mapping->a_ops->writepage)
2531 blk_start_plug(&plug);
2532 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2533 blk_finish_plug(&plug);
2537 EXPORT_SYMBOL(generic_writepages);
2539 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2542 struct bdi_writeback *wb;
2544 if (wbc->nr_to_write <= 0)
2546 wb = inode_to_wb_wbc(mapping->host, wbc);
2547 wb_bandwidth_estimate_start(wb);
2549 if (mapping->a_ops->writepages)
2550 ret = mapping->a_ops->writepages(mapping, wbc);
2552 ret = generic_writepages(mapping, wbc);
2553 if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2557 * Lacking an allocation context or the locality or writeback
2558 * state of any of the inode's pages, throttle based on
2559 * writeback activity on the local node. It's as good a
2562 reclaim_throttle(NODE_DATA(numa_node_id()),
2563 VMSCAN_THROTTLE_WRITEBACK);
2566 * Usually few pages are written by now from those we've just submitted
2567 * but if there's constant writeback being submitted, this makes sure
2568 * writeback bandwidth is updated once in a while.
2570 if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
2571 BANDWIDTH_INTERVAL))
2572 wb_update_bandwidth(wb);
2577 * folio_write_one - write out a single folio and wait on I/O.
2578 * @folio: The folio to write.
2580 * The folio must be locked by the caller and will be unlocked upon return.
2582 * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2585 * Return: %0 on success, negative error code otherwise
2587 int folio_write_one(struct folio *folio)
2589 struct address_space *mapping = folio->mapping;
2591 struct writeback_control wbc = {
2592 .sync_mode = WB_SYNC_ALL,
2593 .nr_to_write = folio_nr_pages(folio),
2596 BUG_ON(!folio_test_locked(folio));
2598 folio_wait_writeback(folio);
2600 if (folio_clear_dirty_for_io(folio)) {
2602 ret = mapping->a_ops->writepage(&folio->page, &wbc);
2604 folio_wait_writeback(folio);
2607 folio_unlock(folio);
2611 ret = filemap_check_errors(mapping);
2614 EXPORT_SYMBOL(folio_write_one);
2617 * For address_spaces which do not use buffers nor write back.
2619 bool noop_dirty_folio(struct address_space *mapping, struct folio *folio)
2621 if (!folio_test_dirty(folio))
2622 return !folio_test_set_dirty(folio);
2625 EXPORT_SYMBOL(noop_dirty_folio);
2628 * Helper function for set_page_dirty family.
2630 * Caller must hold lock_page_memcg().
2632 * NOTE: This relies on being atomic wrt interrupts.
2634 static void folio_account_dirtied(struct folio *folio,
2635 struct address_space *mapping)
2637 struct inode *inode = mapping->host;
2639 trace_writeback_dirty_folio(folio, mapping);
2641 if (mapping_can_writeback(mapping)) {
2642 struct bdi_writeback *wb;
2643 long nr = folio_nr_pages(folio);
2645 inode_attach_wb(inode, &folio->page);
2646 wb = inode_to_wb(inode);
2648 __lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, nr);
2649 __zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
2650 __node_stat_mod_folio(folio, NR_DIRTIED, nr);
2651 wb_stat_mod(wb, WB_RECLAIMABLE, nr);
2652 wb_stat_mod(wb, WB_DIRTIED, nr);
2653 task_io_account_write(nr * PAGE_SIZE);
2654 current->nr_dirtied += nr;
2655 __this_cpu_add(bdp_ratelimits, nr);
2657 mem_cgroup_track_foreign_dirty(folio, wb);
2662 * Helper function for deaccounting dirty page without writeback.
2664 * Caller must hold lock_page_memcg().
2666 void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb)
2668 long nr = folio_nr_pages(folio);
2670 lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2671 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2672 wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2673 task_io_account_cancelled_write(nr * PAGE_SIZE);
2677 * Mark the folio dirty, and set it dirty in the page cache, and mark
2680 * If warn is true, then emit a warning if the folio is not uptodate and has
2681 * not been truncated.
2683 * The caller must hold lock_page_memcg(). Most callers have the folio
2684 * locked. A few have the folio blocked from truncation through other
2685 * means (eg zap_page_range() has it mapped and is holding the page table
2686 * lock). This can also be called from mark_buffer_dirty(), which I
2687 * cannot prove is always protected against truncate.
2689 void __folio_mark_dirty(struct folio *folio, struct address_space *mapping,
2692 unsigned long flags;
2694 xa_lock_irqsave(&mapping->i_pages, flags);
2695 if (folio->mapping) { /* Race with truncate? */
2696 WARN_ON_ONCE(warn && !folio_test_uptodate(folio));
2697 folio_account_dirtied(folio, mapping);
2698 __xa_set_mark(&mapping->i_pages, folio_index(folio),
2699 PAGECACHE_TAG_DIRTY);
2701 xa_unlock_irqrestore(&mapping->i_pages, flags);
2705 * filemap_dirty_folio - Mark a folio dirty for filesystems which do not use buffer_heads.
2706 * @mapping: Address space this folio belongs to.
2707 * @folio: Folio to be marked as dirty.
2709 * Filesystems which do not use buffer heads should call this function
2710 * from their set_page_dirty address space operation. It ignores the
2711 * contents of folio_get_private(), so if the filesystem marks individual
2712 * blocks as dirty, the filesystem should handle that itself.
2714 * This is also sometimes used by filesystems which use buffer_heads when
2715 * a single buffer is being dirtied: we want to set the folio dirty in
2716 * that case, but not all the buffers. This is a "bottom-up" dirtying,
2717 * whereas block_dirty_folio() is a "top-down" dirtying.
2719 * The caller must ensure this doesn't race with truncation. Most will
2720 * simply hold the folio lock, but e.g. zap_pte_range() calls with the
2721 * folio mapped and the pte lock held, which also locks out truncation.
2723 bool filemap_dirty_folio(struct address_space *mapping, struct folio *folio)
2725 folio_memcg_lock(folio);
2726 if (folio_test_set_dirty(folio)) {
2727 folio_memcg_unlock(folio);
2731 __folio_mark_dirty(folio, mapping, !folio_test_private(folio));
2732 folio_memcg_unlock(folio);
2734 if (mapping->host) {
2735 /* !PageAnon && !swapper_space */
2736 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2740 EXPORT_SYMBOL(filemap_dirty_folio);
2743 * folio_account_redirty - Manually account for redirtying a page.
2744 * @folio: The folio which is being redirtied.
2746 * Most filesystems should call folio_redirty_for_writepage() instead
2747 * of this fuction. If your filesystem is doing writeback outside the
2748 * context of a writeback_control(), it can call this when redirtying
2749 * a folio, to de-account the dirty counters (NR_DIRTIED, WB_DIRTIED,
2750 * tsk->nr_dirtied), so that they match the written counters (NR_WRITTEN,
2751 * WB_WRITTEN) in long term. The mismatches will lead to systematic errors
2752 * in balanced_dirty_ratelimit and the dirty pages position control.
2754 void folio_account_redirty(struct folio *folio)
2756 struct address_space *mapping = folio->mapping;
2758 if (mapping && mapping_can_writeback(mapping)) {
2759 struct inode *inode = mapping->host;
2760 struct bdi_writeback *wb;
2761 struct wb_lock_cookie cookie = {};
2762 long nr = folio_nr_pages(folio);
2764 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2765 current->nr_dirtied -= nr;
2766 node_stat_mod_folio(folio, NR_DIRTIED, -nr);
2767 wb_stat_mod(wb, WB_DIRTIED, -nr);
2768 unlocked_inode_to_wb_end(inode, &cookie);
2771 EXPORT_SYMBOL(folio_account_redirty);
2774 * folio_redirty_for_writepage - Decline to write a dirty folio.
2775 * @wbc: The writeback control.
2776 * @folio: The folio.
2778 * When a writepage implementation decides that it doesn't want to write
2779 * @folio for some reason, it should call this function, unlock @folio and
2782 * Return: True if we redirtied the folio. False if someone else dirtied
2785 bool folio_redirty_for_writepage(struct writeback_control *wbc,
2786 struct folio *folio)
2789 long nr = folio_nr_pages(folio);
2791 wbc->pages_skipped += nr;
2792 ret = filemap_dirty_folio(folio->mapping, folio);
2793 folio_account_redirty(folio);
2797 EXPORT_SYMBOL(folio_redirty_for_writepage);
2800 * folio_mark_dirty - Mark a folio as being modified.
2801 * @folio: The folio.
2803 * The folio may not be truncated while this function is running.
2804 * Holding the folio lock is sufficient to prevent truncation, but some
2805 * callers cannot acquire a sleeping lock. These callers instead hold
2806 * the page table lock for a page table which contains at least one page
2807 * in this folio. Truncation will block on the page table lock as it
2808 * unmaps pages before removing the folio from its mapping.
2810 * Return: True if the folio was newly dirtied, false if it was already dirty.
2812 bool folio_mark_dirty(struct folio *folio)
2814 struct address_space *mapping = folio_mapping(folio);
2816 if (likely(mapping)) {
2818 * readahead/lru_deactivate_page could remain
2819 * PG_readahead/PG_reclaim due to race with folio_end_writeback
2820 * About readahead, if the folio is written, the flags would be
2821 * reset. So no problem.
2822 * About lru_deactivate_page, if the folio is redirtied,
2823 * the flag will be reset. So no problem. but if the
2824 * folio is used by readahead it will confuse readahead
2825 * and make it restart the size rampup process. But it's
2826 * a trivial problem.
2828 if (folio_test_reclaim(folio))
2829 folio_clear_reclaim(folio);
2830 return mapping->a_ops->dirty_folio(mapping, folio);
2833 return noop_dirty_folio(mapping, folio);
2835 EXPORT_SYMBOL(folio_mark_dirty);
2838 * set_page_dirty() is racy if the caller has no reference against
2839 * page->mapping->host, and if the page is unlocked. This is because another
2840 * CPU could truncate the page off the mapping and then free the mapping.
2842 * Usually, the page _is_ locked, or the caller is a user-space process which
2843 * holds a reference on the inode by having an open file.
2845 * In other cases, the page should be locked before running set_page_dirty().
2847 int set_page_dirty_lock(struct page *page)
2852 ret = set_page_dirty(page);
2856 EXPORT_SYMBOL(set_page_dirty_lock);
2859 * This cancels just the dirty bit on the kernel page itself, it does NOT
2860 * actually remove dirty bits on any mmap's that may be around. It also
2861 * leaves the page tagged dirty, so any sync activity will still find it on
2862 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2863 * look at the dirty bits in the VM.
2865 * Doing this should *normally* only ever be done when a page is truncated,
2866 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2867 * this when it notices that somebody has cleaned out all the buffers on a
2868 * page without actually doing it through the VM. Can you say "ext3 is
2869 * horribly ugly"? Thought you could.
2871 void __folio_cancel_dirty(struct folio *folio)
2873 struct address_space *mapping = folio_mapping(folio);
2875 if (mapping_can_writeback(mapping)) {
2876 struct inode *inode = mapping->host;
2877 struct bdi_writeback *wb;
2878 struct wb_lock_cookie cookie = {};
2880 folio_memcg_lock(folio);
2881 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2883 if (folio_test_clear_dirty(folio))
2884 folio_account_cleaned(folio, wb);
2886 unlocked_inode_to_wb_end(inode, &cookie);
2887 folio_memcg_unlock(folio);
2889 folio_clear_dirty(folio);
2892 EXPORT_SYMBOL(__folio_cancel_dirty);
2895 * Clear a folio's dirty flag, while caring for dirty memory accounting.
2896 * Returns true if the folio was previously dirty.
2898 * This is for preparing to put the folio under writeout. We leave
2899 * the folio tagged as dirty in the xarray so that a concurrent
2900 * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk.
2901 * The ->writepage implementation will run either folio_start_writeback()
2902 * or folio_mark_dirty(), at which stage we bring the folio's dirty flag
2903 * and xarray dirty tag back into sync.
2905 * This incoherency between the folio's dirty flag and xarray tag is
2906 * unfortunate, but it only exists while the folio is locked.
2908 bool folio_clear_dirty_for_io(struct folio *folio)
2910 struct address_space *mapping = folio_mapping(folio);
2913 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2915 if (mapping && mapping_can_writeback(mapping)) {
2916 struct inode *inode = mapping->host;
2917 struct bdi_writeback *wb;
2918 struct wb_lock_cookie cookie = {};
2921 * Yes, Virginia, this is indeed insane.
2923 * We use this sequence to make sure that
2924 * (a) we account for dirty stats properly
2925 * (b) we tell the low-level filesystem to
2926 * mark the whole folio dirty if it was
2927 * dirty in a pagetable. Only to then
2928 * (c) clean the folio again and return 1 to
2929 * cause the writeback.
2931 * This way we avoid all nasty races with the
2932 * dirty bit in multiple places and clearing
2933 * them concurrently from different threads.
2935 * Note! Normally the "folio_mark_dirty(folio)"
2936 * has no effect on the actual dirty bit - since
2937 * that will already usually be set. But we
2938 * need the side effects, and it can help us
2941 * We basically use the folio "master dirty bit"
2942 * as a serialization point for all the different
2943 * threads doing their things.
2945 if (folio_mkclean(folio))
2946 folio_mark_dirty(folio);
2948 * We carefully synchronise fault handlers against
2949 * installing a dirty pte and marking the folio dirty
2950 * at this point. We do this by having them hold the
2951 * page lock while dirtying the folio, and folios are
2952 * always locked coming in here, so we get the desired
2955 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2956 if (folio_test_clear_dirty(folio)) {
2957 long nr = folio_nr_pages(folio);
2958 lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2959 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2960 wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2963 unlocked_inode_to_wb_end(inode, &cookie);
2966 return folio_test_clear_dirty(folio);
2968 EXPORT_SYMBOL(folio_clear_dirty_for_io);
2970 static void wb_inode_writeback_start(struct bdi_writeback *wb)
2972 atomic_inc(&wb->writeback_inodes);
2975 static void wb_inode_writeback_end(struct bdi_writeback *wb)
2977 unsigned long flags;
2978 atomic_dec(&wb->writeback_inodes);
2980 * Make sure estimate of writeback throughput gets updated after
2981 * writeback completed. We delay the update by BANDWIDTH_INTERVAL
2982 * (which is the interval other bandwidth updates use for batching) so
2983 * that if multiple inodes end writeback at a similar time, they get
2984 * batched into one bandwidth update.
2986 spin_lock_irqsave(&wb->work_lock, flags);
2987 if (test_bit(WB_registered, &wb->state))
2988 queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL);
2989 spin_unlock_irqrestore(&wb->work_lock, flags);
2992 bool __folio_end_writeback(struct folio *folio)
2994 long nr = folio_nr_pages(folio);
2995 struct address_space *mapping = folio_mapping(folio);
2998 folio_memcg_lock(folio);
2999 if (mapping && mapping_use_writeback_tags(mapping)) {
3000 struct inode *inode = mapping->host;
3001 struct backing_dev_info *bdi = inode_to_bdi(inode);
3002 unsigned long flags;
3004 xa_lock_irqsave(&mapping->i_pages, flags);
3005 ret = folio_test_clear_writeback(folio);
3007 __xa_clear_mark(&mapping->i_pages, folio_index(folio),
3008 PAGECACHE_TAG_WRITEBACK);
3009 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
3010 struct bdi_writeback *wb = inode_to_wb(inode);
3012 wb_stat_mod(wb, WB_WRITEBACK, -nr);
3013 __wb_writeout_add(wb, nr);
3014 if (!mapping_tagged(mapping,
3015 PAGECACHE_TAG_WRITEBACK))
3016 wb_inode_writeback_end(wb);
3020 if (mapping->host && !mapping_tagged(mapping,
3021 PAGECACHE_TAG_WRITEBACK))
3022 sb_clear_inode_writeback(mapping->host);
3024 xa_unlock_irqrestore(&mapping->i_pages, flags);
3026 ret = folio_test_clear_writeback(folio);
3029 lruvec_stat_mod_folio(folio, NR_WRITEBACK, -nr);
3030 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
3031 node_stat_mod_folio(folio, NR_WRITTEN, nr);
3033 folio_memcg_unlock(folio);
3037 bool __folio_start_writeback(struct folio *folio, bool keep_write)
3039 long nr = folio_nr_pages(folio);
3040 struct address_space *mapping = folio_mapping(folio);
3044 folio_memcg_lock(folio);
3045 if (mapping && mapping_use_writeback_tags(mapping)) {
3046 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
3047 struct inode *inode = mapping->host;
3048 struct backing_dev_info *bdi = inode_to_bdi(inode);
3049 unsigned long flags;
3051 xas_lock_irqsave(&xas, flags);
3053 ret = folio_test_set_writeback(folio);
3057 on_wblist = mapping_tagged(mapping,
3058 PAGECACHE_TAG_WRITEBACK);
3060 xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
3061 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
3062 struct bdi_writeback *wb = inode_to_wb(inode);
3064 wb_stat_mod(wb, WB_WRITEBACK, nr);
3066 wb_inode_writeback_start(wb);
3070 * We can come through here when swapping
3071 * anonymous folios, so we don't necessarily
3072 * have an inode to track for sync.
3074 if (mapping->host && !on_wblist)
3075 sb_mark_inode_writeback(mapping->host);
3077 if (!folio_test_dirty(folio))
3078 xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
3080 xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
3081 xas_unlock_irqrestore(&xas, flags);
3083 ret = folio_test_set_writeback(folio);
3086 lruvec_stat_mod_folio(folio, NR_WRITEBACK, nr);
3087 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
3089 folio_memcg_unlock(folio);
3090 access_ret = arch_make_folio_accessible(folio);
3092 * If writeback has been triggered on a page that cannot be made
3093 * accessible, it is too late to recover here.
3095 VM_BUG_ON_FOLIO(access_ret != 0, folio);
3099 EXPORT_SYMBOL(__folio_start_writeback);
3102 * folio_wait_writeback - Wait for a folio to finish writeback.
3103 * @folio: The folio to wait for.
3105 * If the folio is currently being written back to storage, wait for the
3108 * Context: Sleeps. Must be called in process context and with
3109 * no spinlocks held. Caller should hold a reference on the folio.
3110 * If the folio is not locked, writeback may start again after writeback
3113 void folio_wait_writeback(struct folio *folio)
3115 while (folio_test_writeback(folio)) {
3116 trace_folio_wait_writeback(folio, folio_mapping(folio));
3117 folio_wait_bit(folio, PG_writeback);
3120 EXPORT_SYMBOL_GPL(folio_wait_writeback);
3123 * folio_wait_writeback_killable - Wait for a folio to finish writeback.
3124 * @folio: The folio to wait for.
3126 * If the folio is currently being written back to storage, wait for the
3127 * I/O to complete or a fatal signal to arrive.
3129 * Context: Sleeps. Must be called in process context and with
3130 * no spinlocks held. Caller should hold a reference on the folio.
3131 * If the folio is not locked, writeback may start again after writeback
3133 * Return: 0 on success, -EINTR if we get a fatal signal while waiting.
3135 int folio_wait_writeback_killable(struct folio *folio)
3137 while (folio_test_writeback(folio)) {
3138 trace_folio_wait_writeback(folio, folio_mapping(folio));
3139 if (folio_wait_bit_killable(folio, PG_writeback))
3145 EXPORT_SYMBOL_GPL(folio_wait_writeback_killable);
3148 * folio_wait_stable() - wait for writeback to finish, if necessary.
3149 * @folio: The folio to wait on.
3151 * This function determines if the given folio is related to a backing
3152 * device that requires folio contents to be held stable during writeback.
3153 * If so, then it will wait for any pending writeback to complete.
3155 * Context: Sleeps. Must be called in process context and with
3156 * no spinlocks held. Caller should hold a reference on the folio.
3157 * If the folio is not locked, writeback may start again after writeback
3160 void folio_wait_stable(struct folio *folio)
3162 if (folio_inode(folio)->i_sb->s_iflags & SB_I_STABLE_WRITES)
3163 folio_wait_writeback(folio);
3165 EXPORT_SYMBOL_GPL(folio_wait_stable);