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 static int __bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
693 if (min_ratio > 100 * BDI_RATIO_SCALE)
696 spin_lock_bh(&bdi_lock);
697 if (min_ratio > bdi->max_ratio) {
700 if (min_ratio < bdi->min_ratio) {
701 delta = bdi->min_ratio - min_ratio;
702 bdi_min_ratio -= delta;
703 bdi->min_ratio = min_ratio;
705 delta = min_ratio - bdi->min_ratio;
706 if (bdi_min_ratio + delta < 100 * BDI_RATIO_SCALE) {
707 bdi_min_ratio += delta;
708 bdi->min_ratio = min_ratio;
714 spin_unlock_bh(&bdi_lock);
719 static int __bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio)
723 if (max_ratio > 100 * BDI_RATIO_SCALE)
726 spin_lock_bh(&bdi_lock);
727 if (bdi->min_ratio > max_ratio) {
730 bdi->max_ratio = max_ratio;
731 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) /
732 (100 * BDI_RATIO_SCALE);
734 spin_unlock_bh(&bdi_lock);
739 int bdi_set_min_ratio_no_scale(struct backing_dev_info *bdi, unsigned int min_ratio)
741 return __bdi_set_min_ratio(bdi, min_ratio);
744 int bdi_set_max_ratio_no_scale(struct backing_dev_info *bdi, unsigned int max_ratio)
746 return __bdi_set_max_ratio(bdi, max_ratio);
749 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
751 return __bdi_set_min_ratio(bdi, min_ratio * BDI_RATIO_SCALE);
754 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio)
756 return __bdi_set_max_ratio(bdi, max_ratio * BDI_RATIO_SCALE);
758 EXPORT_SYMBOL(bdi_set_max_ratio);
760 u64 bdi_get_min_bytes(struct backing_dev_info *bdi)
762 return bdi_get_bytes(bdi->min_ratio);
765 int bdi_set_min_bytes(struct backing_dev_info *bdi, u64 min_bytes)
768 unsigned long pages = min_bytes >> PAGE_SHIFT;
769 unsigned long min_ratio;
771 ret = bdi_check_pages_limit(pages);
775 min_ratio = bdi_ratio_from_pages(pages);
776 return __bdi_set_min_ratio(bdi, min_ratio);
779 u64 bdi_get_max_bytes(struct backing_dev_info *bdi)
781 return bdi_get_bytes(bdi->max_ratio);
784 int bdi_set_max_bytes(struct backing_dev_info *bdi, u64 max_bytes)
787 unsigned long pages = max_bytes >> PAGE_SHIFT;
788 unsigned long max_ratio;
790 ret = bdi_check_pages_limit(pages);
794 max_ratio = bdi_ratio_from_pages(pages);
795 return __bdi_set_max_ratio(bdi, max_ratio);
798 int bdi_set_strict_limit(struct backing_dev_info *bdi, unsigned int strict_limit)
800 if (strict_limit > 1)
803 spin_lock_bh(&bdi_lock);
805 bdi->capabilities |= BDI_CAP_STRICTLIMIT;
807 bdi->capabilities &= ~BDI_CAP_STRICTLIMIT;
808 spin_unlock_bh(&bdi_lock);
813 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
814 unsigned long bg_thresh)
816 return (thresh + bg_thresh) / 2;
819 static unsigned long hard_dirty_limit(struct wb_domain *dom,
820 unsigned long thresh)
822 return max(thresh, dom->dirty_limit);
826 * Memory which can be further allocated to a memcg domain is capped by
827 * system-wide clean memory excluding the amount being used in the domain.
829 static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
830 unsigned long filepages, unsigned long headroom)
832 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
833 unsigned long clean = filepages - min(filepages, mdtc->dirty);
834 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
835 unsigned long other_clean = global_clean - min(global_clean, clean);
837 mdtc->avail = filepages + min(headroom, other_clean);
841 * __wb_calc_thresh - @wb's share of dirty threshold
842 * @dtc: dirty_throttle_context of interest
843 * @thresh: dirty throttling or dirty background threshold of wb_domain in @dtc
845 * Note that balance_dirty_pages() will only seriously take dirty throttling
846 * threshold as a hard limit when sleeping max_pause per page is not enough
847 * to keep the dirty pages under control. For example, when the device is
848 * completely stalled due to some error conditions, or when there are 1000
849 * dd tasks writing to a slow 10MB/s USB key.
850 * In the other normal situations, it acts more gently by throttling the tasks
851 * more (rather than completely block them) when the wb dirty pages go high.
853 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
854 * - starving fast devices
855 * - piling up dirty pages (that will take long time to sync) on slow devices
857 * The wb's share of dirty limit will be adapting to its throughput and
858 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
860 * Return: @wb's dirty limit in pages. For dirty throttling limit, the term
861 * "dirty" in the context of dirty balancing includes all PG_dirty and
862 * PG_writeback pages.
864 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc,
865 unsigned long thresh)
867 struct wb_domain *dom = dtc_dom(dtc);
869 unsigned long numerator, denominator;
870 unsigned long wb_min_ratio, wb_max_ratio;
873 * Calculate this wb's share of the thresh ratio.
875 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
876 &numerator, &denominator);
878 wb_thresh = (thresh * (100 * BDI_RATIO_SCALE - bdi_min_ratio)) / (100 * BDI_RATIO_SCALE);
879 wb_thresh *= numerator;
880 wb_thresh = div64_ul(wb_thresh, denominator);
882 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
884 wb_thresh += (thresh * wb_min_ratio) / (100 * BDI_RATIO_SCALE);
885 if (wb_thresh > (thresh * wb_max_ratio) / (100 * BDI_RATIO_SCALE))
886 wb_thresh = thresh * wb_max_ratio / (100 * BDI_RATIO_SCALE);
891 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
893 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
895 return __wb_calc_thresh(&gdtc, thresh);
898 unsigned long cgwb_calc_thresh(struct bdi_writeback *wb)
900 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
901 struct dirty_throttle_control mdtc = { MDTC_INIT(wb, &gdtc) };
902 unsigned long filepages = 0, headroom = 0, writeback = 0;
904 gdtc.avail = global_dirtyable_memory();
905 gdtc.dirty = global_node_page_state(NR_FILE_DIRTY) +
906 global_node_page_state(NR_WRITEBACK);
908 mem_cgroup_wb_stats(wb, &filepages, &headroom,
909 &mdtc.dirty, &writeback);
910 mdtc.dirty += writeback;
911 mdtc_calc_avail(&mdtc, filepages, headroom);
912 domain_dirty_limits(&mdtc);
914 return __wb_calc_thresh(&mdtc, mdtc.thresh);
919 * f(dirty) := 1.0 + (----------------)
922 * it's a 3rd order polynomial that subjects to
924 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
925 * (2) f(setpoint) = 1.0 => the balance point
926 * (3) f(limit) = 0 => the hard limit
927 * (4) df/dx <= 0 => negative feedback control
928 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
929 * => fast response on large errors; small oscillation near setpoint
931 static long long pos_ratio_polynom(unsigned long setpoint,
938 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
939 (limit - setpoint) | 1);
941 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
942 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
943 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
945 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
949 * Dirty position control.
951 * (o) global/bdi setpoints
953 * We want the dirty pages be balanced around the global/wb setpoints.
954 * When the number of dirty pages is higher/lower than the setpoint, the
955 * dirty position control ratio (and hence task dirty ratelimit) will be
956 * decreased/increased to bring the dirty pages back to the setpoint.
958 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
960 * if (dirty < setpoint) scale up pos_ratio
961 * if (dirty > setpoint) scale down pos_ratio
963 * if (wb_dirty < wb_setpoint) scale up pos_ratio
964 * if (wb_dirty > wb_setpoint) scale down pos_ratio
966 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
968 * (o) global control line
972 * | |<===== global dirty control scope ======>|
980 * 1.0 ................................*
986 * 0 +------------.------------------.----------------------*------------->
987 * freerun^ setpoint^ limit^ dirty pages
989 * (o) wb control line
997 * | * |<=========== span ============>|
998 * 1.0 .......................*
1010 * 1/4 ...............................................* * * * * * * * * * * *
1014 * 0 +----------------------.-------------------------------.------------->
1015 * wb_setpoint^ x_intercept^
1017 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
1018 * be smoothly throttled down to normal if it starts high in situations like
1019 * - start writing to a slow SD card and a fast disk at the same time. The SD
1020 * card's wb_dirty may rush to many times higher than wb_setpoint.
1021 * - the wb dirty thresh drops quickly due to change of JBOD workload
1023 static void wb_position_ratio(struct dirty_throttle_control *dtc)
1025 struct bdi_writeback *wb = dtc->wb;
1026 unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth);
1027 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1028 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1029 unsigned long wb_thresh = dtc->wb_thresh;
1030 unsigned long x_intercept;
1031 unsigned long setpoint; /* dirty pages' target balance point */
1032 unsigned long wb_setpoint;
1034 long long pos_ratio; /* for scaling up/down the rate limit */
1039 if (unlikely(dtc->dirty >= limit))
1045 * See comment for pos_ratio_polynom().
1047 setpoint = (freerun + limit) / 2;
1048 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
1051 * The strictlimit feature is a tool preventing mistrusted filesystems
1052 * from growing a large number of dirty pages before throttling. For
1053 * such filesystems balance_dirty_pages always checks wb counters
1054 * against wb limits. Even if global "nr_dirty" is under "freerun".
1055 * This is especially important for fuse which sets bdi->max_ratio to
1056 * 1% by default. Without strictlimit feature, fuse writeback may
1057 * consume arbitrary amount of RAM because it is accounted in
1058 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
1060 * Here, in wb_position_ratio(), we calculate pos_ratio based on
1061 * two values: wb_dirty and wb_thresh. Let's consider an example:
1062 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
1063 * limits are set by default to 10% and 20% (background and throttle).
1064 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
1065 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
1066 * about ~6K pages (as the average of background and throttle wb
1067 * limits). The 3rd order polynomial will provide positive feedback if
1068 * wb_dirty is under wb_setpoint and vice versa.
1070 * Note, that we cannot use global counters in these calculations
1071 * because we want to throttle process writing to a strictlimit wb
1072 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
1073 * in the example above).
1075 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1076 long long wb_pos_ratio;
1078 if (dtc->wb_dirty < 8) {
1079 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
1080 2 << RATELIMIT_CALC_SHIFT);
1084 if (dtc->wb_dirty >= wb_thresh)
1087 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
1090 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
1093 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
1097 * Typically, for strictlimit case, wb_setpoint << setpoint
1098 * and pos_ratio >> wb_pos_ratio. In the other words global
1099 * state ("dirty") is not limiting factor and we have to
1100 * make decision based on wb counters. But there is an
1101 * important case when global pos_ratio should get precedence:
1102 * global limits are exceeded (e.g. due to activities on other
1103 * wb's) while given strictlimit wb is below limit.
1105 * "pos_ratio * wb_pos_ratio" would work for the case above,
1106 * but it would look too non-natural for the case of all
1107 * activity in the system coming from a single strictlimit wb
1108 * with bdi->max_ratio == 100%.
1110 * Note that min() below somewhat changes the dynamics of the
1111 * control system. Normally, pos_ratio value can be well over 3
1112 * (when globally we are at freerun and wb is well below wb
1113 * setpoint). Now the maximum pos_ratio in the same situation
1114 * is 2. We might want to tweak this if we observe the control
1115 * system is too slow to adapt.
1117 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
1122 * We have computed basic pos_ratio above based on global situation. If
1123 * the wb is over/under its share of dirty pages, we want to scale
1124 * pos_ratio further down/up. That is done by the following mechanism.
1130 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1132 * x_intercept - wb_dirty
1133 * := --------------------------
1134 * x_intercept - wb_setpoint
1136 * The main wb control line is a linear function that subjects to
1138 * (1) f(wb_setpoint) = 1.0
1139 * (2) k = - 1 / (8 * write_bw) (in single wb case)
1140 * or equally: x_intercept = wb_setpoint + 8 * write_bw
1142 * For single wb case, the dirty pages are observed to fluctuate
1143 * regularly within range
1144 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1145 * for various filesystems, where (2) can yield in a reasonable 12.5%
1146 * fluctuation range for pos_ratio.
1148 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1149 * own size, so move the slope over accordingly and choose a slope that
1150 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1152 if (unlikely(wb_thresh > dtc->thresh))
1153 wb_thresh = dtc->thresh;
1155 * It's very possible that wb_thresh is close to 0 not because the
1156 * device is slow, but that it has remained inactive for long time.
1157 * Honour such devices a reasonable good (hopefully IO efficient)
1158 * threshold, so that the occasional writes won't be blocked and active
1159 * writes can rampup the threshold quickly.
1161 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1163 * scale global setpoint to wb's:
1164 * wb_setpoint = setpoint * wb_thresh / thresh
1166 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1167 wb_setpoint = setpoint * (u64)x >> 16;
1169 * Use span=(8*write_bw) in single wb case as indicated by
1170 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1172 * wb_thresh thresh - wb_thresh
1173 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1176 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1177 x_intercept = wb_setpoint + span;
1179 if (dtc->wb_dirty < x_intercept - span / 4) {
1180 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1181 (x_intercept - wb_setpoint) | 1);
1186 * wb reserve area, safeguard against dirty pool underrun and disk idle
1187 * It may push the desired control point of global dirty pages higher
1190 x_intercept = wb_thresh / 2;
1191 if (dtc->wb_dirty < x_intercept) {
1192 if (dtc->wb_dirty > x_intercept / 8)
1193 pos_ratio = div_u64(pos_ratio * x_intercept,
1199 dtc->pos_ratio = pos_ratio;
1202 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1203 unsigned long elapsed,
1204 unsigned long written)
1206 const unsigned long period = roundup_pow_of_two(3 * HZ);
1207 unsigned long avg = wb->avg_write_bandwidth;
1208 unsigned long old = wb->write_bandwidth;
1212 * bw = written * HZ / elapsed
1214 * bw * elapsed + write_bandwidth * (period - elapsed)
1215 * write_bandwidth = ---------------------------------------------------
1218 * @written may have decreased due to folio_redirty_for_writepage().
1219 * Avoid underflowing @bw calculation.
1221 bw = written - min(written, wb->written_stamp);
1223 if (unlikely(elapsed > period)) {
1224 bw = div64_ul(bw, elapsed);
1228 bw += (u64)wb->write_bandwidth * (period - elapsed);
1229 bw >>= ilog2(period);
1232 * one more level of smoothing, for filtering out sudden spikes
1234 if (avg > old && old >= (unsigned long)bw)
1235 avg -= (avg - old) >> 3;
1237 if (avg < old && old <= (unsigned long)bw)
1238 avg += (old - avg) >> 3;
1241 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1242 avg = max(avg, 1LU);
1243 if (wb_has_dirty_io(wb)) {
1244 long delta = avg - wb->avg_write_bandwidth;
1245 WARN_ON_ONCE(atomic_long_add_return(delta,
1246 &wb->bdi->tot_write_bandwidth) <= 0);
1248 wb->write_bandwidth = bw;
1249 WRITE_ONCE(wb->avg_write_bandwidth, avg);
1252 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1254 struct wb_domain *dom = dtc_dom(dtc);
1255 unsigned long thresh = dtc->thresh;
1256 unsigned long limit = dom->dirty_limit;
1259 * Follow up in one step.
1261 if (limit < thresh) {
1267 * Follow down slowly. Use the higher one as the target, because thresh
1268 * may drop below dirty. This is exactly the reason to introduce
1269 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1271 thresh = max(thresh, dtc->dirty);
1272 if (limit > thresh) {
1273 limit -= (limit - thresh) >> 5;
1278 dom->dirty_limit = limit;
1281 static void domain_update_dirty_limit(struct dirty_throttle_control *dtc,
1284 struct wb_domain *dom = dtc_dom(dtc);
1287 * check locklessly first to optimize away locking for the most time
1289 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1292 spin_lock(&dom->lock);
1293 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1294 update_dirty_limit(dtc);
1295 dom->dirty_limit_tstamp = now;
1297 spin_unlock(&dom->lock);
1301 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1303 * Normal wb tasks will be curbed at or below it in long term.
1304 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1306 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1307 unsigned long dirtied,
1308 unsigned long elapsed)
1310 struct bdi_writeback *wb = dtc->wb;
1311 unsigned long dirty = dtc->dirty;
1312 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1313 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1314 unsigned long setpoint = (freerun + limit) / 2;
1315 unsigned long write_bw = wb->avg_write_bandwidth;
1316 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1317 unsigned long dirty_rate;
1318 unsigned long task_ratelimit;
1319 unsigned long balanced_dirty_ratelimit;
1322 unsigned long shift;
1325 * The dirty rate will match the writeout rate in long term, except
1326 * when dirty pages are truncated by userspace or re-dirtied by FS.
1328 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1331 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1333 task_ratelimit = (u64)dirty_ratelimit *
1334 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1335 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1338 * A linear estimation of the "balanced" throttle rate. The theory is,
1339 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1340 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1341 * formula will yield the balanced rate limit (write_bw / N).
1343 * Note that the expanded form is not a pure rate feedback:
1344 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1345 * but also takes pos_ratio into account:
1346 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1348 * (1) is not realistic because pos_ratio also takes part in balancing
1349 * the dirty rate. Consider the state
1350 * pos_ratio = 0.5 (3)
1351 * rate = 2 * (write_bw / N) (4)
1352 * If (1) is used, it will stuck in that state! Because each dd will
1354 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1356 * dirty_rate = N * task_ratelimit = write_bw (6)
1357 * put (6) into (1) we get
1358 * rate_(i+1) = rate_(i) (7)
1360 * So we end up using (2) to always keep
1361 * rate_(i+1) ~= (write_bw / N) (8)
1362 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1363 * pos_ratio is able to drive itself to 1.0, which is not only where
1364 * the dirty count meet the setpoint, but also where the slope of
1365 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1367 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1370 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1372 if (unlikely(balanced_dirty_ratelimit > write_bw))
1373 balanced_dirty_ratelimit = write_bw;
1376 * We could safely do this and return immediately:
1378 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1380 * However to get a more stable dirty_ratelimit, the below elaborated
1381 * code makes use of task_ratelimit to filter out singular points and
1382 * limit the step size.
1384 * The below code essentially only uses the relative value of
1386 * task_ratelimit - dirty_ratelimit
1387 * = (pos_ratio - 1) * dirty_ratelimit
1389 * which reflects the direction and size of dirty position error.
1393 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1394 * task_ratelimit is on the same side of dirty_ratelimit, too.
1396 * - dirty_ratelimit > balanced_dirty_ratelimit
1397 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1398 * lowering dirty_ratelimit will help meet both the position and rate
1399 * control targets. Otherwise, don't update dirty_ratelimit if it will
1400 * only help meet the rate target. After all, what the users ultimately
1401 * feel and care are stable dirty rate and small position error.
1403 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1404 * and filter out the singular points of balanced_dirty_ratelimit. Which
1405 * keeps jumping around randomly and can even leap far away at times
1406 * due to the small 200ms estimation period of dirty_rate (we want to
1407 * keep that period small to reduce time lags).
1412 * For strictlimit case, calculations above were based on wb counters
1413 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1414 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1415 * Hence, to calculate "step" properly, we have to use wb_dirty as
1416 * "dirty" and wb_setpoint as "setpoint".
1418 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1419 * it's possible that wb_thresh is close to zero due to inactivity
1420 * of backing device.
1422 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1423 dirty = dtc->wb_dirty;
1424 if (dtc->wb_dirty < 8)
1425 setpoint = dtc->wb_dirty + 1;
1427 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1430 if (dirty < setpoint) {
1431 x = min3(wb->balanced_dirty_ratelimit,
1432 balanced_dirty_ratelimit, task_ratelimit);
1433 if (dirty_ratelimit < x)
1434 step = x - dirty_ratelimit;
1436 x = max3(wb->balanced_dirty_ratelimit,
1437 balanced_dirty_ratelimit, task_ratelimit);
1438 if (dirty_ratelimit > x)
1439 step = dirty_ratelimit - x;
1443 * Don't pursue 100% rate matching. It's impossible since the balanced
1444 * rate itself is constantly fluctuating. So decrease the track speed
1445 * when it gets close to the target. Helps eliminate pointless tremors.
1447 shift = dirty_ratelimit / (2 * step + 1);
1448 if (shift < BITS_PER_LONG)
1449 step = DIV_ROUND_UP(step >> shift, 8);
1453 if (dirty_ratelimit < balanced_dirty_ratelimit)
1454 dirty_ratelimit += step;
1456 dirty_ratelimit -= step;
1458 WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL));
1459 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1461 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1464 static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1465 struct dirty_throttle_control *mdtc,
1466 bool update_ratelimit)
1468 struct bdi_writeback *wb = gdtc->wb;
1469 unsigned long now = jiffies;
1470 unsigned long elapsed;
1471 unsigned long dirtied;
1472 unsigned long written;
1474 spin_lock(&wb->list_lock);
1477 * Lockless checks for elapsed time are racy and delayed update after
1478 * IO completion doesn't do it at all (to make sure written pages are
1479 * accounted reasonably quickly). Make sure elapsed >= 1 to avoid
1482 elapsed = max(now - wb->bw_time_stamp, 1UL);
1483 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1484 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1486 if (update_ratelimit) {
1487 domain_update_dirty_limit(gdtc, now);
1488 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1491 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1492 * compiler has no way to figure that out. Help it.
1494 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1495 domain_update_dirty_limit(mdtc, now);
1496 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1499 wb_update_write_bandwidth(wb, elapsed, written);
1501 wb->dirtied_stamp = dirtied;
1502 wb->written_stamp = written;
1503 WRITE_ONCE(wb->bw_time_stamp, now);
1504 spin_unlock(&wb->list_lock);
1507 void wb_update_bandwidth(struct bdi_writeback *wb)
1509 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1511 __wb_update_bandwidth(&gdtc, NULL, false);
1514 /* Interval after which we consider wb idle and don't estimate bandwidth */
1515 #define WB_BANDWIDTH_IDLE_JIF (HZ)
1517 static void wb_bandwidth_estimate_start(struct bdi_writeback *wb)
1519 unsigned long now = jiffies;
1520 unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp);
1522 if (elapsed > WB_BANDWIDTH_IDLE_JIF &&
1523 !atomic_read(&wb->writeback_inodes)) {
1524 spin_lock(&wb->list_lock);
1525 wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED);
1526 wb->written_stamp = wb_stat(wb, WB_WRITTEN);
1527 WRITE_ONCE(wb->bw_time_stamp, now);
1528 spin_unlock(&wb->list_lock);
1533 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1534 * will look to see if it needs to start dirty throttling.
1536 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1537 * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1538 * (the number of pages we may dirty without exceeding the dirty limits).
1540 static unsigned long dirty_poll_interval(unsigned long dirty,
1541 unsigned long thresh)
1544 return 1UL << (ilog2(thresh - dirty) >> 1);
1549 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1550 unsigned long wb_dirty)
1552 unsigned long bw = READ_ONCE(wb->avg_write_bandwidth);
1556 * Limit pause time for small memory systems. If sleeping for too long
1557 * time, a small pool of dirty/writeback pages may go empty and disk go
1560 * 8 serves as the safety ratio.
1562 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1565 return min_t(unsigned long, t, MAX_PAUSE);
1568 static long wb_min_pause(struct bdi_writeback *wb,
1570 unsigned long task_ratelimit,
1571 unsigned long dirty_ratelimit,
1572 int *nr_dirtied_pause)
1574 long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth));
1575 long lo = ilog2(READ_ONCE(wb->dirty_ratelimit));
1576 long t; /* target pause */
1577 long pause; /* estimated next pause */
1578 int pages; /* target nr_dirtied_pause */
1580 /* target for 10ms pause on 1-dd case */
1581 t = max(1, HZ / 100);
1584 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1587 * (N * 10ms) on 2^N concurrent tasks.
1590 t += (hi - lo) * (10 * HZ) / 1024;
1593 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1594 * on the much more stable dirty_ratelimit. However the next pause time
1595 * will be computed based on task_ratelimit and the two rate limits may
1596 * depart considerably at some time. Especially if task_ratelimit goes
1597 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1598 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1599 * result task_ratelimit won't be executed faithfully, which could
1600 * eventually bring down dirty_ratelimit.
1602 * We apply two rules to fix it up:
1603 * 1) try to estimate the next pause time and if necessary, use a lower
1604 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1605 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1606 * 2) limit the target pause time to max_pause/2, so that the normal
1607 * small fluctuations of task_ratelimit won't trigger rule (1) and
1608 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1610 t = min(t, 1 + max_pause / 2);
1611 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1614 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1615 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1616 * When the 16 consecutive reads are often interrupted by some dirty
1617 * throttling pause during the async writes, cfq will go into idles
1618 * (deadline is fine). So push nr_dirtied_pause as high as possible
1619 * until reaches DIRTY_POLL_THRESH=32 pages.
1621 if (pages < DIRTY_POLL_THRESH) {
1623 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1624 if (pages > DIRTY_POLL_THRESH) {
1625 pages = DIRTY_POLL_THRESH;
1626 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1630 pause = HZ * pages / (task_ratelimit + 1);
1631 if (pause > max_pause) {
1633 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1636 *nr_dirtied_pause = pages;
1638 * The minimal pause time will normally be half the target pause time.
1640 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1643 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1645 struct bdi_writeback *wb = dtc->wb;
1646 unsigned long wb_reclaimable;
1649 * wb_thresh is not treated as some limiting factor as
1650 * dirty_thresh, due to reasons
1651 * - in JBOD setup, wb_thresh can fluctuate a lot
1652 * - in a system with HDD and USB key, the USB key may somehow
1653 * go into state (wb_dirty >> wb_thresh) either because
1654 * wb_dirty starts high, or because wb_thresh drops low.
1655 * In this case we don't want to hard throttle the USB key
1656 * dirtiers for 100 seconds until wb_dirty drops under
1657 * wb_thresh. Instead the auxiliary wb control line in
1658 * wb_position_ratio() will let the dirtier task progress
1659 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1661 dtc->wb_thresh = __wb_calc_thresh(dtc, dtc->thresh);
1662 dtc->wb_bg_thresh = dtc->thresh ?
1663 div64_u64(dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1666 * In order to avoid the stacked BDI deadlock we need
1667 * to ensure we accurately count the 'dirty' pages when
1668 * the threshold is low.
1670 * Otherwise it would be possible to get thresh+n pages
1671 * reported dirty, even though there are thresh-m pages
1672 * actually dirty; with m+n sitting in the percpu
1675 if (dtc->wb_thresh < 2 * wb_stat_error()) {
1676 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1677 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1679 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1680 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1685 * balance_dirty_pages() must be called by processes which are generating dirty
1686 * data. It looks at the number of dirty pages in the machine and will force
1687 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1688 * If we're over `background_thresh' then the writeback threads are woken to
1689 * perform some writeout.
1691 static int balance_dirty_pages(struct bdi_writeback *wb,
1692 unsigned long pages_dirtied, unsigned int flags)
1694 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1695 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1696 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1697 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1699 struct dirty_throttle_control *sdtc;
1700 unsigned long nr_dirty;
1705 int nr_dirtied_pause;
1706 bool dirty_exceeded = false;
1707 unsigned long task_ratelimit;
1708 unsigned long dirty_ratelimit;
1709 struct backing_dev_info *bdi = wb->bdi;
1710 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1711 unsigned long start_time = jiffies;
1715 unsigned long now = jiffies;
1716 unsigned long dirty, thresh, bg_thresh;
1717 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1718 unsigned long m_thresh = 0;
1719 unsigned long m_bg_thresh = 0;
1721 nr_dirty = global_node_page_state(NR_FILE_DIRTY);
1722 gdtc->avail = global_dirtyable_memory();
1723 gdtc->dirty = nr_dirty + global_node_page_state(NR_WRITEBACK);
1725 domain_dirty_limits(gdtc);
1727 if (unlikely(strictlimit)) {
1728 wb_dirty_limits(gdtc);
1730 dirty = gdtc->wb_dirty;
1731 thresh = gdtc->wb_thresh;
1732 bg_thresh = gdtc->wb_bg_thresh;
1734 dirty = gdtc->dirty;
1735 thresh = gdtc->thresh;
1736 bg_thresh = gdtc->bg_thresh;
1740 unsigned long filepages, headroom, writeback;
1743 * If @wb belongs to !root memcg, repeat the same
1744 * basic calculations for the memcg domain.
1746 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1747 &mdtc->dirty, &writeback);
1748 mdtc->dirty += writeback;
1749 mdtc_calc_avail(mdtc, filepages, headroom);
1751 domain_dirty_limits(mdtc);
1753 if (unlikely(strictlimit)) {
1754 wb_dirty_limits(mdtc);
1755 m_dirty = mdtc->wb_dirty;
1756 m_thresh = mdtc->wb_thresh;
1757 m_bg_thresh = mdtc->wb_bg_thresh;
1759 m_dirty = mdtc->dirty;
1760 m_thresh = mdtc->thresh;
1761 m_bg_thresh = mdtc->bg_thresh;
1766 * In laptop mode, we wait until hitting the higher threshold
1767 * before starting background writeout, and then write out all
1768 * the way down to the lower threshold. So slow writers cause
1769 * minimal disk activity.
1771 * In normal mode, we start background writeout at the lower
1772 * background_thresh, to keep the amount of dirty memory low.
1774 if (!laptop_mode && nr_dirty > gdtc->bg_thresh &&
1775 !writeback_in_progress(wb))
1776 wb_start_background_writeback(wb);
1779 * Throttle it only when the background writeback cannot
1780 * catch-up. This avoids (excessively) small writeouts
1781 * when the wb limits are ramping up in case of !strictlimit.
1783 * In strictlimit case make decision based on the wb counters
1784 * and limits. Small writeouts when the wb limits are ramping
1785 * up are the price we consciously pay for strictlimit-ing.
1787 * If memcg domain is in effect, @dirty should be under
1788 * both global and memcg freerun ceilings.
1790 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1792 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1794 unsigned long m_intv;
1797 intv = dirty_poll_interval(dirty, thresh);
1800 current->dirty_paused_when = now;
1801 current->nr_dirtied = 0;
1803 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1804 current->nr_dirtied_pause = min(intv, m_intv);
1808 /* Start writeback even when in laptop mode */
1809 if (unlikely(!writeback_in_progress(wb)))
1810 wb_start_background_writeback(wb);
1812 mem_cgroup_flush_foreign(wb);
1815 * Calculate global domain's pos_ratio and select the
1816 * global dtc by default.
1819 wb_dirty_limits(gdtc);
1821 if ((current->flags & PF_LOCAL_THROTTLE) &&
1823 dirty_freerun_ceiling(gdtc->wb_thresh,
1824 gdtc->wb_bg_thresh))
1826 * LOCAL_THROTTLE tasks must not be throttled
1827 * when below the per-wb freerun ceiling.
1832 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1833 ((gdtc->dirty > gdtc->thresh) || strictlimit);
1835 wb_position_ratio(gdtc);
1840 * If memcg domain is in effect, calculate its
1841 * pos_ratio. @wb should satisfy constraints from
1842 * both global and memcg domains. Choose the one
1843 * w/ lower pos_ratio.
1846 wb_dirty_limits(mdtc);
1848 if ((current->flags & PF_LOCAL_THROTTLE) &&
1850 dirty_freerun_ceiling(mdtc->wb_thresh,
1851 mdtc->wb_bg_thresh))
1853 * LOCAL_THROTTLE tasks must not be
1854 * throttled when below the per-wb
1859 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1860 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1862 wb_position_ratio(mdtc);
1863 if (mdtc->pos_ratio < gdtc->pos_ratio)
1867 if (dirty_exceeded != wb->dirty_exceeded)
1868 wb->dirty_exceeded = dirty_exceeded;
1870 if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
1871 BANDWIDTH_INTERVAL))
1872 __wb_update_bandwidth(gdtc, mdtc, true);
1874 /* throttle according to the chosen dtc */
1875 dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit);
1876 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1877 RATELIMIT_CALC_SHIFT;
1878 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1879 min_pause = wb_min_pause(wb, max_pause,
1880 task_ratelimit, dirty_ratelimit,
1883 if (unlikely(task_ratelimit == 0)) {
1888 period = HZ * pages_dirtied / task_ratelimit;
1890 if (current->dirty_paused_when)
1891 pause -= now - current->dirty_paused_when;
1893 * For less than 1s think time (ext3/4 may block the dirtier
1894 * for up to 800ms from time to time on 1-HDD; so does xfs,
1895 * however at much less frequency), try to compensate it in
1896 * future periods by updating the virtual time; otherwise just
1897 * do a reset, as it may be a light dirtier.
1899 if (pause < min_pause) {
1900 trace_balance_dirty_pages(wb,
1913 current->dirty_paused_when = now;
1914 current->nr_dirtied = 0;
1915 } else if (period) {
1916 current->dirty_paused_when += period;
1917 current->nr_dirtied = 0;
1918 } else if (current->nr_dirtied_pause <= pages_dirtied)
1919 current->nr_dirtied_pause += pages_dirtied;
1922 if (unlikely(pause > max_pause)) {
1923 /* for occasional dropped task_ratelimit */
1924 now += min(pause - max_pause, max_pause);
1929 trace_balance_dirty_pages(wb,
1941 if (flags & BDP_ASYNC) {
1945 __set_current_state(TASK_KILLABLE);
1946 bdi->last_bdp_sleep = jiffies;
1947 io_schedule_timeout(pause);
1949 current->dirty_paused_when = now + pause;
1950 current->nr_dirtied = 0;
1951 current->nr_dirtied_pause = nr_dirtied_pause;
1954 * This is typically equal to (dirty < thresh) and can also
1955 * keep "1000+ dd on a slow USB stick" under control.
1961 * In the case of an unresponsive NFS server and the NFS dirty
1962 * pages exceeds dirty_thresh, give the other good wb's a pipe
1963 * to go through, so that tasks on them still remain responsive.
1965 * In theory 1 page is enough to keep the consumer-producer
1966 * pipe going: the flusher cleans 1 page => the task dirties 1
1967 * more page. However wb_dirty has accounting errors. So use
1968 * the larger and more IO friendly wb_stat_error.
1970 if (sdtc->wb_dirty <= wb_stat_error())
1973 if (fatal_signal_pending(current))
1979 static DEFINE_PER_CPU(int, bdp_ratelimits);
1982 * Normal tasks are throttled by
1984 * dirty tsk->nr_dirtied_pause pages;
1985 * take a snap in balance_dirty_pages();
1987 * However there is a worst case. If every task exit immediately when dirtied
1988 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1989 * called to throttle the page dirties. The solution is to save the not yet
1990 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1991 * randomly into the running tasks. This works well for the above worst case,
1992 * as the new task will pick up and accumulate the old task's leaked dirty
1993 * count and eventually get throttled.
1995 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1998 * balance_dirty_pages_ratelimited_flags - Balance dirty memory state.
1999 * @mapping: address_space which was dirtied.
2000 * @flags: BDP flags.
2002 * Processes which are dirtying memory should call in here once for each page
2003 * which was newly dirtied. The function will periodically check the system's
2004 * dirty state and will initiate writeback if needed.
2006 * See balance_dirty_pages_ratelimited() for details.
2008 * Return: If @flags contains BDP_ASYNC, it may return -EAGAIN to
2009 * indicate that memory is out of balance and the caller must wait
2010 * for I/O to complete. Otherwise, it will return 0 to indicate
2011 * that either memory was already in balance, or it was able to sleep
2012 * until the amount of dirty memory returned to balance.
2014 int balance_dirty_pages_ratelimited_flags(struct address_space *mapping,
2017 struct inode *inode = mapping->host;
2018 struct backing_dev_info *bdi = inode_to_bdi(inode);
2019 struct bdi_writeback *wb = NULL;
2024 if (!(bdi->capabilities & BDI_CAP_WRITEBACK))
2027 if (inode_cgwb_enabled(inode))
2028 wb = wb_get_create_current(bdi, GFP_KERNEL);
2032 ratelimit = current->nr_dirtied_pause;
2033 if (wb->dirty_exceeded)
2034 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
2038 * This prevents one CPU to accumulate too many dirtied pages without
2039 * calling into balance_dirty_pages(), which can happen when there are
2040 * 1000+ tasks, all of them start dirtying pages at exactly the same
2041 * time, hence all honoured too large initial task->nr_dirtied_pause.
2043 p = this_cpu_ptr(&bdp_ratelimits);
2044 if (unlikely(current->nr_dirtied >= ratelimit))
2046 else if (unlikely(*p >= ratelimit_pages)) {
2051 * Pick up the dirtied pages by the exited tasks. This avoids lots of
2052 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
2053 * the dirty throttling and livelock other long-run dirtiers.
2055 p = this_cpu_ptr(&dirty_throttle_leaks);
2056 if (*p > 0 && current->nr_dirtied < ratelimit) {
2057 unsigned long nr_pages_dirtied;
2058 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
2059 *p -= nr_pages_dirtied;
2060 current->nr_dirtied += nr_pages_dirtied;
2064 if (unlikely(current->nr_dirtied >= ratelimit))
2065 ret = balance_dirty_pages(wb, current->nr_dirtied, flags);
2070 EXPORT_SYMBOL_GPL(balance_dirty_pages_ratelimited_flags);
2073 * balance_dirty_pages_ratelimited - balance dirty memory state.
2074 * @mapping: address_space which was dirtied.
2076 * Processes which are dirtying memory should call in here once for each page
2077 * which was newly dirtied. The function will periodically check the system's
2078 * dirty state and will initiate writeback if needed.
2080 * Once we're over the dirty memory limit we decrease the ratelimiting
2081 * by a lot, to prevent individual processes from overshooting the limit
2082 * by (ratelimit_pages) each.
2084 void balance_dirty_pages_ratelimited(struct address_space *mapping)
2086 balance_dirty_pages_ratelimited_flags(mapping, 0);
2088 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
2091 * wb_over_bg_thresh - does @wb need to be written back?
2092 * @wb: bdi_writeback of interest
2094 * Determines whether background writeback should keep writing @wb or it's
2097 * Return: %true if writeback should continue.
2099 bool wb_over_bg_thresh(struct bdi_writeback *wb)
2101 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
2102 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
2103 struct dirty_throttle_control * const gdtc = &gdtc_stor;
2104 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
2106 unsigned long reclaimable;
2107 unsigned long thresh;
2110 * Similar to balance_dirty_pages() but ignores pages being written
2111 * as we're trying to decide whether to put more under writeback.
2113 gdtc->avail = global_dirtyable_memory();
2114 gdtc->dirty = global_node_page_state(NR_FILE_DIRTY);
2115 domain_dirty_limits(gdtc);
2117 if (gdtc->dirty > gdtc->bg_thresh)
2120 thresh = __wb_calc_thresh(gdtc, gdtc->bg_thresh);
2121 if (thresh < 2 * wb_stat_error())
2122 reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
2124 reclaimable = wb_stat(wb, WB_RECLAIMABLE);
2126 if (reclaimable > thresh)
2130 unsigned long filepages, headroom, writeback;
2132 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
2134 mdtc_calc_avail(mdtc, filepages, headroom);
2135 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
2137 if (mdtc->dirty > mdtc->bg_thresh)
2140 thresh = __wb_calc_thresh(mdtc, mdtc->bg_thresh);
2141 if (thresh < 2 * wb_stat_error())
2142 reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
2144 reclaimable = wb_stat(wb, WB_RECLAIMABLE);
2146 if (reclaimable > thresh)
2153 #ifdef CONFIG_SYSCTL
2155 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
2157 static int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
2158 void *buffer, size_t *length, loff_t *ppos)
2160 unsigned int old_interval = dirty_writeback_interval;
2163 ret = proc_dointvec(table, write, buffer, length, ppos);
2166 * Writing 0 to dirty_writeback_interval will disable periodic writeback
2167 * and a different non-zero value will wakeup the writeback threads.
2168 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
2169 * iterate over all bdis and wbs.
2170 * The reason we do this is to make the change take effect immediately.
2172 if (!ret && write && dirty_writeback_interval &&
2173 dirty_writeback_interval != old_interval)
2174 wakeup_flusher_threads(WB_REASON_PERIODIC);
2180 void laptop_mode_timer_fn(struct timer_list *t)
2182 struct backing_dev_info *backing_dev_info =
2183 from_timer(backing_dev_info, t, laptop_mode_wb_timer);
2185 wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
2189 * We've spun up the disk and we're in laptop mode: schedule writeback
2190 * of all dirty data a few seconds from now. If the flush is already scheduled
2191 * then push it back - the user is still using the disk.
2193 void laptop_io_completion(struct backing_dev_info *info)
2195 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2199 * We're in laptop mode and we've just synced. The sync's writes will have
2200 * caused another writeback to be scheduled by laptop_io_completion.
2201 * Nothing needs to be written back anymore, so we unschedule the writeback.
2203 void laptop_sync_completion(void)
2205 struct backing_dev_info *bdi;
2209 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2210 del_timer(&bdi->laptop_mode_wb_timer);
2216 * If ratelimit_pages is too high then we can get into dirty-data overload
2217 * if a large number of processes all perform writes at the same time.
2219 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2220 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2224 void writeback_set_ratelimit(void)
2226 struct wb_domain *dom = &global_wb_domain;
2227 unsigned long background_thresh;
2228 unsigned long dirty_thresh;
2230 global_dirty_limits(&background_thresh, &dirty_thresh);
2231 dom->dirty_limit = dirty_thresh;
2232 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2233 if (ratelimit_pages < 16)
2234 ratelimit_pages = 16;
2237 static int page_writeback_cpu_online(unsigned int cpu)
2239 writeback_set_ratelimit();
2243 #ifdef CONFIG_SYSCTL
2245 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
2246 static const unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
2248 static struct ctl_table vm_page_writeback_sysctls[] = {
2250 .procname = "dirty_background_ratio",
2251 .data = &dirty_background_ratio,
2252 .maxlen = sizeof(dirty_background_ratio),
2254 .proc_handler = dirty_background_ratio_handler,
2255 .extra1 = SYSCTL_ZERO,
2256 .extra2 = SYSCTL_ONE_HUNDRED,
2259 .procname = "dirty_background_bytes",
2260 .data = &dirty_background_bytes,
2261 .maxlen = sizeof(dirty_background_bytes),
2263 .proc_handler = dirty_background_bytes_handler,
2264 .extra1 = SYSCTL_LONG_ONE,
2267 .procname = "dirty_ratio",
2268 .data = &vm_dirty_ratio,
2269 .maxlen = sizeof(vm_dirty_ratio),
2271 .proc_handler = dirty_ratio_handler,
2272 .extra1 = SYSCTL_ZERO,
2273 .extra2 = SYSCTL_ONE_HUNDRED,
2276 .procname = "dirty_bytes",
2277 .data = &vm_dirty_bytes,
2278 .maxlen = sizeof(vm_dirty_bytes),
2280 .proc_handler = dirty_bytes_handler,
2281 .extra1 = (void *)&dirty_bytes_min,
2284 .procname = "dirty_writeback_centisecs",
2285 .data = &dirty_writeback_interval,
2286 .maxlen = sizeof(dirty_writeback_interval),
2288 .proc_handler = dirty_writeback_centisecs_handler,
2291 .procname = "dirty_expire_centisecs",
2292 .data = &dirty_expire_interval,
2293 .maxlen = sizeof(dirty_expire_interval),
2295 .proc_handler = proc_dointvec_minmax,
2296 .extra1 = SYSCTL_ZERO,
2298 #ifdef CONFIG_HIGHMEM
2300 .procname = "highmem_is_dirtyable",
2301 .data = &vm_highmem_is_dirtyable,
2302 .maxlen = sizeof(vm_highmem_is_dirtyable),
2304 .proc_handler = proc_dointvec_minmax,
2305 .extra1 = SYSCTL_ZERO,
2306 .extra2 = SYSCTL_ONE,
2310 .procname = "laptop_mode",
2311 .data = &laptop_mode,
2312 .maxlen = sizeof(laptop_mode),
2314 .proc_handler = proc_dointvec_jiffies,
2320 * Called early on to tune the page writeback dirty limits.
2322 * We used to scale dirty pages according to how total memory
2323 * related to pages that could be allocated for buffers.
2325 * However, that was when we used "dirty_ratio" to scale with
2326 * all memory, and we don't do that any more. "dirty_ratio"
2327 * is now applied to total non-HIGHPAGE memory, and as such we can't
2328 * get into the old insane situation any more where we had
2329 * large amounts of dirty pages compared to a small amount of
2330 * non-HIGHMEM memory.
2332 * But we might still want to scale the dirty_ratio by how
2333 * much memory the box has..
2335 void __init page_writeback_init(void)
2337 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2339 cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2340 page_writeback_cpu_online, NULL);
2341 cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2342 page_writeback_cpu_online);
2343 #ifdef CONFIG_SYSCTL
2344 register_sysctl_init("vm", vm_page_writeback_sysctls);
2349 * tag_pages_for_writeback - tag pages to be written by writeback
2350 * @mapping: address space structure to write
2351 * @start: starting page index
2352 * @end: ending page index (inclusive)
2354 * This function scans the page range from @start to @end (inclusive) and tags
2355 * all pages that have DIRTY tag set with a special TOWRITE tag. The caller
2356 * can then use the TOWRITE tag to identify pages eligible for writeback.
2357 * This mechanism is used to avoid livelocking of writeback by a process
2358 * steadily creating new dirty pages in the file (thus it is important for this
2359 * function to be quick so that it can tag pages faster than a dirtying process
2362 void tag_pages_for_writeback(struct address_space *mapping,
2363 pgoff_t start, pgoff_t end)
2365 XA_STATE(xas, &mapping->i_pages, start);
2366 unsigned int tagged = 0;
2370 xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2371 xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2372 if (++tagged % XA_CHECK_SCHED)
2376 xas_unlock_irq(&xas);
2380 xas_unlock_irq(&xas);
2382 EXPORT_SYMBOL(tag_pages_for_writeback);
2384 static bool folio_prepare_writeback(struct address_space *mapping,
2385 struct writeback_control *wbc, struct folio *folio)
2388 * Folio truncated or invalidated. We can freely skip it then,
2389 * even for data integrity operations: the folio has disappeared
2390 * concurrently, so there could be no real expectation of this
2391 * data integrity operation even if there is now a new, dirty
2392 * folio at the same pagecache index.
2394 if (unlikely(folio->mapping != mapping))
2398 * Did somebody else write it for us?
2400 if (!folio_test_dirty(folio))
2403 if (folio_test_writeback(folio)) {
2404 if (wbc->sync_mode == WB_SYNC_NONE)
2406 folio_wait_writeback(folio);
2408 BUG_ON(folio_test_writeback(folio));
2410 if (!folio_clear_dirty_for_io(folio))
2416 static xa_mark_t wbc_to_tag(struct writeback_control *wbc)
2418 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2419 return PAGECACHE_TAG_TOWRITE;
2420 return PAGECACHE_TAG_DIRTY;
2423 static pgoff_t wbc_end(struct writeback_control *wbc)
2425 if (wbc->range_cyclic)
2427 return wbc->range_end >> PAGE_SHIFT;
2430 static struct folio *writeback_get_folio(struct address_space *mapping,
2431 struct writeback_control *wbc)
2433 struct folio *folio;
2436 folio = folio_batch_next(&wbc->fbatch);
2438 folio_batch_release(&wbc->fbatch);
2440 filemap_get_folios_tag(mapping, &wbc->index, wbc_end(wbc),
2441 wbc_to_tag(wbc), &wbc->fbatch);
2442 folio = folio_batch_next(&wbc->fbatch);
2448 if (unlikely(!folio_prepare_writeback(mapping, wbc, folio))) {
2449 folio_unlock(folio);
2453 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2458 * writeback_iter - iterate folio of a mapping for writeback
2459 * @mapping: address space structure to write
2460 * @wbc: writeback context
2461 * @folio: previously iterated folio (%NULL to start)
2462 * @error: in-out pointer for writeback errors (see below)
2464 * This function returns the next folio for the writeback operation described by
2465 * @wbc on @mapping and should be called in a while loop in the ->writepages
2468 * To start the writeback operation, %NULL is passed in the @folio argument, and
2469 * for every subsequent iteration the folio returned previously should be passed
2472 * If there was an error in the per-folio writeback inside the writeback_iter()
2473 * loop, @error should be set to the error value.
2475 * Once the writeback described in @wbc has finished, this function will return
2476 * %NULL and if there was an error in any iteration restore it to @error.
2478 * Note: callers should not manually break out of the loop using break or goto
2479 * but must keep calling writeback_iter() until it returns %NULL.
2481 * Return: the folio to write or %NULL if the loop is done.
2483 struct folio *writeback_iter(struct address_space *mapping,
2484 struct writeback_control *wbc, struct folio *folio, int *error)
2487 folio_batch_init(&wbc->fbatch);
2488 wbc->saved_err = *error = 0;
2491 * For range cyclic writeback we remember where we stopped so
2492 * that we can continue where we stopped.
2494 * For non-cyclic writeback we always start at the beginning of
2495 * the passed in range.
2497 if (wbc->range_cyclic)
2498 wbc->index = mapping->writeback_index;
2500 wbc->index = wbc->range_start >> PAGE_SHIFT;
2503 * To avoid livelocks when other processes dirty new pages, we
2504 * first tag pages which should be written back and only then
2505 * start writing them.
2507 * For data-integrity writeback we have to be careful so that we
2508 * do not miss some pages (e.g., because some other process has
2509 * cleared the TOWRITE tag we set). The rule we follow is that
2510 * TOWRITE tag can be cleared only by the process clearing the
2511 * DIRTY tag (and submitting the page for I/O).
2513 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2514 tag_pages_for_writeback(mapping, wbc->index,
2517 wbc->nr_to_write -= folio_nr_pages(folio);
2519 WARN_ON_ONCE(*error > 0);
2522 * For integrity writeback we have to keep going until we have
2523 * written all the folios we tagged for writeback above, even if
2524 * we run past wbc->nr_to_write or encounter errors.
2525 * We stash away the first error we encounter in wbc->saved_err
2526 * so that it can be retrieved when we're done. This is because
2527 * the file system may still have state to clear for each folio.
2529 * For background writeback we exit as soon as we run past
2530 * wbc->nr_to_write or encounter the first error.
2532 if (wbc->sync_mode == WB_SYNC_ALL) {
2533 if (*error && !wbc->saved_err)
2534 wbc->saved_err = *error;
2536 if (*error || wbc->nr_to_write <= 0)
2541 folio = writeback_get_folio(mapping, wbc);
2544 * To avoid deadlocks between range_cyclic writeback and callers
2545 * that hold pages in PageWriteback to aggregate I/O until
2546 * the writeback iteration finishes, we do not loop back to the
2547 * start of the file. Doing so causes a page lock/page
2548 * writeback access order inversion - we should only ever lock
2549 * multiple pages in ascending page->index order, and looping
2550 * back to the start of the file violates that rule and causes
2553 if (wbc->range_cyclic)
2554 mapping->writeback_index = 0;
2557 * Return the first error we encountered (if there was any) to
2560 *error = wbc->saved_err;
2565 if (wbc->range_cyclic)
2566 mapping->writeback_index = folio->index + folio_nr_pages(folio);
2567 folio_batch_release(&wbc->fbatch);
2570 EXPORT_SYMBOL_GPL(writeback_iter);
2573 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2574 * @mapping: address space structure to write
2575 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2576 * @writepage: function called for each page
2577 * @data: data passed to writepage function
2579 * Return: %0 on success, negative error code otherwise
2581 * Note: please use writeback_iter() instead.
2583 int write_cache_pages(struct address_space *mapping,
2584 struct writeback_control *wbc, writepage_t writepage,
2587 struct folio *folio = NULL;
2590 while ((folio = writeback_iter(mapping, wbc, folio, &error))) {
2591 error = writepage(folio, wbc, data);
2592 if (error == AOP_WRITEPAGE_ACTIVATE) {
2593 folio_unlock(folio);
2600 EXPORT_SYMBOL(write_cache_pages);
2602 static int writeback_use_writepage(struct address_space *mapping,
2603 struct writeback_control *wbc)
2605 struct folio *folio = NULL;
2606 struct blk_plug plug;
2609 blk_start_plug(&plug);
2610 while ((folio = writeback_iter(mapping, wbc, folio, &err))) {
2611 err = mapping->a_ops->writepage(&folio->page, wbc);
2612 if (err == AOP_WRITEPAGE_ACTIVATE) {
2613 folio_unlock(folio);
2616 mapping_set_error(mapping, err);
2618 blk_finish_plug(&plug);
2623 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2626 struct bdi_writeback *wb;
2628 if (wbc->nr_to_write <= 0)
2630 wb = inode_to_wb_wbc(mapping->host, wbc);
2631 wb_bandwidth_estimate_start(wb);
2633 if (mapping->a_ops->writepages) {
2634 ret = mapping->a_ops->writepages(mapping, wbc);
2635 } else if (mapping->a_ops->writepage) {
2636 ret = writeback_use_writepage(mapping, wbc);
2638 /* deal with chardevs and other special files */
2641 if (ret != -ENOMEM || wbc->sync_mode != WB_SYNC_ALL)
2645 * Lacking an allocation context or the locality or writeback
2646 * state of any of the inode's pages, throttle based on
2647 * writeback activity on the local node. It's as good a
2650 reclaim_throttle(NODE_DATA(numa_node_id()),
2651 VMSCAN_THROTTLE_WRITEBACK);
2654 * Usually few pages are written by now from those we've just submitted
2655 * but if there's constant writeback being submitted, this makes sure
2656 * writeback bandwidth is updated once in a while.
2658 if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
2659 BANDWIDTH_INTERVAL))
2660 wb_update_bandwidth(wb);
2665 * For address_spaces which do not use buffers nor write back.
2667 bool noop_dirty_folio(struct address_space *mapping, struct folio *folio)
2669 if (!folio_test_dirty(folio))
2670 return !folio_test_set_dirty(folio);
2673 EXPORT_SYMBOL(noop_dirty_folio);
2676 * Helper function for set_page_dirty family.
2678 * Caller must hold folio_memcg_lock().
2680 * NOTE: This relies on being atomic wrt interrupts.
2682 static void folio_account_dirtied(struct folio *folio,
2683 struct address_space *mapping)
2685 struct inode *inode = mapping->host;
2687 trace_writeback_dirty_folio(folio, mapping);
2689 if (mapping_can_writeback(mapping)) {
2690 struct bdi_writeback *wb;
2691 long nr = folio_nr_pages(folio);
2693 inode_attach_wb(inode, folio);
2694 wb = inode_to_wb(inode);
2696 __lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, nr);
2697 __zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
2698 __node_stat_mod_folio(folio, NR_DIRTIED, nr);
2699 wb_stat_mod(wb, WB_RECLAIMABLE, nr);
2700 wb_stat_mod(wb, WB_DIRTIED, nr);
2701 task_io_account_write(nr * PAGE_SIZE);
2702 current->nr_dirtied += nr;
2703 __this_cpu_add(bdp_ratelimits, nr);
2705 mem_cgroup_track_foreign_dirty(folio, wb);
2710 * Helper function for deaccounting dirty page without writeback.
2712 * Caller must hold folio_memcg_lock().
2714 void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb)
2716 long nr = folio_nr_pages(folio);
2718 lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2719 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2720 wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2721 task_io_account_cancelled_write(nr * PAGE_SIZE);
2725 * Mark the folio dirty, and set it dirty in the page cache.
2727 * If warn is true, then emit a warning if the folio is not uptodate and has
2728 * not been truncated.
2730 * The caller must hold folio_memcg_lock(). It is the caller's
2731 * responsibility to prevent the folio from being truncated while
2732 * this function is in progress, although it may have been truncated
2733 * before this function is called. Most callers have the folio locked.
2734 * A few have the folio blocked from truncation through other means (e.g.
2735 * zap_vma_pages() has it mapped and is holding the page table lock).
2736 * When called from mark_buffer_dirty(), the filesystem should hold a
2737 * reference to the buffer_head that is being marked dirty, which causes
2738 * try_to_free_buffers() to fail.
2740 void __folio_mark_dirty(struct folio *folio, struct address_space *mapping,
2743 unsigned long flags;
2745 xa_lock_irqsave(&mapping->i_pages, flags);
2746 if (folio->mapping) { /* Race with truncate? */
2747 WARN_ON_ONCE(warn && !folio_test_uptodate(folio));
2748 folio_account_dirtied(folio, mapping);
2749 __xa_set_mark(&mapping->i_pages, folio_index(folio),
2750 PAGECACHE_TAG_DIRTY);
2752 xa_unlock_irqrestore(&mapping->i_pages, flags);
2756 * filemap_dirty_folio - Mark a folio dirty for filesystems which do not use buffer_heads.
2757 * @mapping: Address space this folio belongs to.
2758 * @folio: Folio to be marked as dirty.
2760 * Filesystems which do not use buffer heads should call this function
2761 * from their dirty_folio address space operation. It ignores the
2762 * contents of folio_get_private(), so if the filesystem marks individual
2763 * blocks as dirty, the filesystem should handle that itself.
2765 * This is also sometimes used by filesystems which use buffer_heads when
2766 * a single buffer is being dirtied: we want to set the folio dirty in
2767 * that case, but not all the buffers. This is a "bottom-up" dirtying,
2768 * whereas block_dirty_folio() is a "top-down" dirtying.
2770 * The caller must ensure this doesn't race with truncation. Most will
2771 * simply hold the folio lock, but e.g. zap_pte_range() calls with the
2772 * folio mapped and the pte lock held, which also locks out truncation.
2774 bool filemap_dirty_folio(struct address_space *mapping, struct folio *folio)
2776 folio_memcg_lock(folio);
2777 if (folio_test_set_dirty(folio)) {
2778 folio_memcg_unlock(folio);
2782 __folio_mark_dirty(folio, mapping, !folio_test_private(folio));
2783 folio_memcg_unlock(folio);
2785 if (mapping->host) {
2786 /* !PageAnon && !swapper_space */
2787 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2791 EXPORT_SYMBOL(filemap_dirty_folio);
2794 * folio_redirty_for_writepage - Decline to write a dirty folio.
2795 * @wbc: The writeback control.
2796 * @folio: The folio.
2798 * When a writepage implementation decides that it doesn't want to write
2799 * @folio for some reason, it should call this function, unlock @folio and
2802 * Return: True if we redirtied the folio. False if someone else dirtied
2805 bool folio_redirty_for_writepage(struct writeback_control *wbc,
2806 struct folio *folio)
2808 struct address_space *mapping = folio->mapping;
2809 long nr = folio_nr_pages(folio);
2812 wbc->pages_skipped += nr;
2813 ret = filemap_dirty_folio(mapping, folio);
2814 if (mapping && mapping_can_writeback(mapping)) {
2815 struct inode *inode = mapping->host;
2816 struct bdi_writeback *wb;
2817 struct wb_lock_cookie cookie = {};
2819 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2820 current->nr_dirtied -= nr;
2821 node_stat_mod_folio(folio, NR_DIRTIED, -nr);
2822 wb_stat_mod(wb, WB_DIRTIED, -nr);
2823 unlocked_inode_to_wb_end(inode, &cookie);
2827 EXPORT_SYMBOL(folio_redirty_for_writepage);
2830 * folio_mark_dirty - Mark a folio as being modified.
2831 * @folio: The folio.
2833 * The folio may not be truncated while this function is running.
2834 * Holding the folio lock is sufficient to prevent truncation, but some
2835 * callers cannot acquire a sleeping lock. These callers instead hold
2836 * the page table lock for a page table which contains at least one page
2837 * in this folio. Truncation will block on the page table lock as it
2838 * unmaps pages before removing the folio from its mapping.
2840 * Return: True if the folio was newly dirtied, false if it was already dirty.
2842 bool folio_mark_dirty(struct folio *folio)
2844 struct address_space *mapping = folio_mapping(folio);
2846 if (likely(mapping)) {
2848 * readahead/folio_deactivate could remain
2849 * PG_readahead/PG_reclaim due to race with folio_end_writeback
2850 * About readahead, if the folio is written, the flags would be
2851 * reset. So no problem.
2852 * About folio_deactivate, if the folio is redirtied,
2853 * the flag will be reset. So no problem. but if the
2854 * folio is used by readahead it will confuse readahead
2855 * and make it restart the size rampup process. But it's
2856 * a trivial problem.
2858 if (folio_test_reclaim(folio))
2859 folio_clear_reclaim(folio);
2860 return mapping->a_ops->dirty_folio(mapping, folio);
2863 return noop_dirty_folio(mapping, folio);
2865 EXPORT_SYMBOL(folio_mark_dirty);
2868 * set_page_dirty() is racy if the caller has no reference against
2869 * page->mapping->host, and if the page is unlocked. This is because another
2870 * CPU could truncate the page off the mapping and then free the mapping.
2872 * Usually, the page _is_ locked, or the caller is a user-space process which
2873 * holds a reference on the inode by having an open file.
2875 * In other cases, the page should be locked before running set_page_dirty().
2877 int set_page_dirty_lock(struct page *page)
2882 ret = set_page_dirty(page);
2886 EXPORT_SYMBOL(set_page_dirty_lock);
2889 * This cancels just the dirty bit on the kernel page itself, it does NOT
2890 * actually remove dirty bits on any mmap's that may be around. It also
2891 * leaves the page tagged dirty, so any sync activity will still find it on
2892 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2893 * look at the dirty bits in the VM.
2895 * Doing this should *normally* only ever be done when a page is truncated,
2896 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2897 * this when it notices that somebody has cleaned out all the buffers on a
2898 * page without actually doing it through the VM. Can you say "ext3 is
2899 * horribly ugly"? Thought you could.
2901 void __folio_cancel_dirty(struct folio *folio)
2903 struct address_space *mapping = folio_mapping(folio);
2905 if (mapping_can_writeback(mapping)) {
2906 struct inode *inode = mapping->host;
2907 struct bdi_writeback *wb;
2908 struct wb_lock_cookie cookie = {};
2910 folio_memcg_lock(folio);
2911 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2913 if (folio_test_clear_dirty(folio))
2914 folio_account_cleaned(folio, wb);
2916 unlocked_inode_to_wb_end(inode, &cookie);
2917 folio_memcg_unlock(folio);
2919 folio_clear_dirty(folio);
2922 EXPORT_SYMBOL(__folio_cancel_dirty);
2925 * Clear a folio's dirty flag, while caring for dirty memory accounting.
2926 * Returns true if the folio was previously dirty.
2928 * This is for preparing to put the folio under writeout. We leave
2929 * the folio tagged as dirty in the xarray so that a concurrent
2930 * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk.
2931 * The ->writepage implementation will run either folio_start_writeback()
2932 * or folio_mark_dirty(), at which stage we bring the folio's dirty flag
2933 * and xarray dirty tag back into sync.
2935 * This incoherency between the folio's dirty flag and xarray tag is
2936 * unfortunate, but it only exists while the folio is locked.
2938 bool folio_clear_dirty_for_io(struct folio *folio)
2940 struct address_space *mapping = folio_mapping(folio);
2943 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2945 if (mapping && mapping_can_writeback(mapping)) {
2946 struct inode *inode = mapping->host;
2947 struct bdi_writeback *wb;
2948 struct wb_lock_cookie cookie = {};
2951 * Yes, Virginia, this is indeed insane.
2953 * We use this sequence to make sure that
2954 * (a) we account for dirty stats properly
2955 * (b) we tell the low-level filesystem to
2956 * mark the whole folio dirty if it was
2957 * dirty in a pagetable. Only to then
2958 * (c) clean the folio again and return 1 to
2959 * cause the writeback.
2961 * This way we avoid all nasty races with the
2962 * dirty bit in multiple places and clearing
2963 * them concurrently from different threads.
2965 * Note! Normally the "folio_mark_dirty(folio)"
2966 * has no effect on the actual dirty bit - since
2967 * that will already usually be set. But we
2968 * need the side effects, and it can help us
2971 * We basically use the folio "master dirty bit"
2972 * as a serialization point for all the different
2973 * threads doing their things.
2975 if (folio_mkclean(folio))
2976 folio_mark_dirty(folio);
2978 * We carefully synchronise fault handlers against
2979 * installing a dirty pte and marking the folio dirty
2980 * at this point. We do this by having them hold the
2981 * page lock while dirtying the folio, and folios are
2982 * always locked coming in here, so we get the desired
2985 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2986 if (folio_test_clear_dirty(folio)) {
2987 long nr = folio_nr_pages(folio);
2988 lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2989 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2990 wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2993 unlocked_inode_to_wb_end(inode, &cookie);
2996 return folio_test_clear_dirty(folio);
2998 EXPORT_SYMBOL(folio_clear_dirty_for_io);
3000 static void wb_inode_writeback_start(struct bdi_writeback *wb)
3002 atomic_inc(&wb->writeback_inodes);
3005 static void wb_inode_writeback_end(struct bdi_writeback *wb)
3007 unsigned long flags;
3008 atomic_dec(&wb->writeback_inodes);
3010 * Make sure estimate of writeback throughput gets updated after
3011 * writeback completed. We delay the update by BANDWIDTH_INTERVAL
3012 * (which is the interval other bandwidth updates use for batching) so
3013 * that if multiple inodes end writeback at a similar time, they get
3014 * batched into one bandwidth update.
3016 spin_lock_irqsave(&wb->work_lock, flags);
3017 if (test_bit(WB_registered, &wb->state))
3018 queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL);
3019 spin_unlock_irqrestore(&wb->work_lock, flags);
3022 bool __folio_end_writeback(struct folio *folio)
3024 long nr = folio_nr_pages(folio);
3025 struct address_space *mapping = folio_mapping(folio);
3028 folio_memcg_lock(folio);
3029 if (mapping && mapping_use_writeback_tags(mapping)) {
3030 struct inode *inode = mapping->host;
3031 struct backing_dev_info *bdi = inode_to_bdi(inode);
3032 unsigned long flags;
3034 xa_lock_irqsave(&mapping->i_pages, flags);
3035 ret = folio_xor_flags_has_waiters(folio, 1 << PG_writeback);
3036 __xa_clear_mark(&mapping->i_pages, folio_index(folio),
3037 PAGECACHE_TAG_WRITEBACK);
3038 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
3039 struct bdi_writeback *wb = inode_to_wb(inode);
3041 wb_stat_mod(wb, WB_WRITEBACK, -nr);
3042 __wb_writeout_add(wb, nr);
3043 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
3044 wb_inode_writeback_end(wb);
3047 if (mapping->host && !mapping_tagged(mapping,
3048 PAGECACHE_TAG_WRITEBACK))
3049 sb_clear_inode_writeback(mapping->host);
3051 xa_unlock_irqrestore(&mapping->i_pages, flags);
3053 ret = folio_xor_flags_has_waiters(folio, 1 << PG_writeback);
3056 lruvec_stat_mod_folio(folio, NR_WRITEBACK, -nr);
3057 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
3058 node_stat_mod_folio(folio, NR_WRITTEN, nr);
3059 folio_memcg_unlock(folio);
3064 void __folio_start_writeback(struct folio *folio, bool keep_write)
3066 long nr = folio_nr_pages(folio);
3067 struct address_space *mapping = folio_mapping(folio);
3070 VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio);
3072 folio_memcg_lock(folio);
3073 if (mapping && mapping_use_writeback_tags(mapping)) {
3074 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
3075 struct inode *inode = mapping->host;
3076 struct backing_dev_info *bdi = inode_to_bdi(inode);
3077 unsigned long flags;
3080 xas_lock_irqsave(&xas, flags);
3082 folio_test_set_writeback(folio);
3084 on_wblist = mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK);
3086 xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
3087 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
3088 struct bdi_writeback *wb = inode_to_wb(inode);
3090 wb_stat_mod(wb, WB_WRITEBACK, nr);
3092 wb_inode_writeback_start(wb);
3096 * We can come through here when swapping anonymous
3097 * folios, so we don't necessarily have an inode to
3100 if (mapping->host && !on_wblist)
3101 sb_mark_inode_writeback(mapping->host);
3102 if (!folio_test_dirty(folio))
3103 xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
3105 xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
3106 xas_unlock_irqrestore(&xas, flags);
3108 folio_test_set_writeback(folio);
3111 lruvec_stat_mod_folio(folio, NR_WRITEBACK, nr);
3112 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
3113 folio_memcg_unlock(folio);
3115 access_ret = arch_make_folio_accessible(folio);
3117 * If writeback has been triggered on a page that cannot be made
3118 * accessible, it is too late to recover here.
3120 VM_BUG_ON_FOLIO(access_ret != 0, folio);
3122 EXPORT_SYMBOL(__folio_start_writeback);
3125 * folio_wait_writeback - Wait for a folio to finish writeback.
3126 * @folio: The folio to wait for.
3128 * If the folio is currently being written back to storage, wait for the
3131 * Context: Sleeps. Must be called in process context and with
3132 * no spinlocks held. Caller should hold a reference on the folio.
3133 * If the folio is not locked, writeback may start again after writeback
3136 void folio_wait_writeback(struct folio *folio)
3138 while (folio_test_writeback(folio)) {
3139 trace_folio_wait_writeback(folio, folio_mapping(folio));
3140 folio_wait_bit(folio, PG_writeback);
3143 EXPORT_SYMBOL_GPL(folio_wait_writeback);
3146 * folio_wait_writeback_killable - Wait for a folio to finish writeback.
3147 * @folio: The folio to wait for.
3149 * If the folio is currently being written back to storage, wait for the
3150 * I/O to complete or a fatal signal to arrive.
3152 * Context: Sleeps. Must be called in process context and with
3153 * no spinlocks held. Caller should hold a reference on the folio.
3154 * If the folio is not locked, writeback may start again after writeback
3156 * Return: 0 on success, -EINTR if we get a fatal signal while waiting.
3158 int folio_wait_writeback_killable(struct folio *folio)
3160 while (folio_test_writeback(folio)) {
3161 trace_folio_wait_writeback(folio, folio_mapping(folio));
3162 if (folio_wait_bit_killable(folio, PG_writeback))
3168 EXPORT_SYMBOL_GPL(folio_wait_writeback_killable);
3171 * folio_wait_stable() - wait for writeback to finish, if necessary.
3172 * @folio: The folio to wait on.
3174 * This function determines if the given folio is related to a backing
3175 * device that requires folio contents to be held stable during writeback.
3176 * If so, then it will wait for any pending writeback to complete.
3178 * Context: Sleeps. Must be called in process context and with
3179 * no spinlocks held. Caller should hold a reference on the folio.
3180 * If the folio is not locked, writeback may start again after writeback
3183 void folio_wait_stable(struct folio *folio)
3185 if (mapping_stable_writes(folio_mapping(folio)))
3186 folio_wait_writeback(folio);
3188 EXPORT_SYMBOL_GPL(folio_wait_stable);