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/export.h>
17 #include <linux/spinlock.h>
20 #include <linux/swap.h>
21 #include <linux/slab.h>
22 #include <linux/pagemap.h>
23 #include <linux/writeback.h>
24 #include <linux/init.h>
25 #include <linux/backing-dev.h>
26 #include <linux/task_io_accounting_ops.h>
27 #include <linux/blkdev.h>
28 #include <linux/mpage.h>
29 #include <linux/rmap.h>
30 #include <linux/percpu.h>
31 #include <linux/smp.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/syscalls.h>
35 #include <linux/pagevec.h>
36 #include <linux/timer.h>
37 #include <linux/sched/rt.h>
38 #include <linux/sched/signal.h>
39 #include <linux/mm_inline.h>
40 #include <trace/events/writeback.h>
45 * Sleep at most 200ms at a time in balance_dirty_pages().
47 #define MAX_PAUSE max(HZ/5, 1)
50 * Try to keep balance_dirty_pages() call intervals higher than this many pages
51 * by raising pause time to max_pause when falls below it.
53 #define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
56 * Estimate write bandwidth at 200ms intervals.
58 #define BANDWIDTH_INTERVAL max(HZ/5, 1)
60 #define RATELIMIT_CALC_SHIFT 10
63 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
64 * will look to see if it needs to force writeback or throttling.
66 static long ratelimit_pages = 32;
68 /* The following parameters are exported via /proc/sys/vm */
71 * Start background writeback (via writeback threads) at this percentage
73 static int dirty_background_ratio = 10;
76 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
77 * dirty_background_ratio * the amount of dirtyable memory
79 static unsigned long dirty_background_bytes;
82 * free highmem will not be subtracted from the total free memory
83 * for calculating free ratios if vm_highmem_is_dirtyable is true
85 static int vm_highmem_is_dirtyable;
88 * The generator of dirty data starts writeback at this percentage
90 static int vm_dirty_ratio = 20;
93 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
94 * vm_dirty_ratio * the amount of dirtyable memory
96 static unsigned long vm_dirty_bytes;
99 * The interval between `kupdate'-style writebacks
101 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
103 EXPORT_SYMBOL_GPL(dirty_writeback_interval);
106 * The longest time for which data is allowed to remain dirty
108 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
111 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
112 * a full sync is triggered after this time elapses without any disk activity.
116 EXPORT_SYMBOL(laptop_mode);
118 /* End of sysctl-exported parameters */
120 struct wb_domain global_wb_domain;
122 /* consolidated parameters for balance_dirty_pages() and its subroutines */
123 struct dirty_throttle_control {
124 #ifdef CONFIG_CGROUP_WRITEBACK
125 struct wb_domain *dom;
126 struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */
128 struct bdi_writeback *wb;
129 struct fprop_local_percpu *wb_completions;
131 unsigned long avail; /* dirtyable */
132 unsigned long dirty; /* file_dirty + write + nfs */
133 unsigned long thresh; /* dirty threshold */
134 unsigned long bg_thresh; /* dirty background threshold */
136 unsigned long wb_dirty; /* per-wb counterparts */
137 unsigned long wb_thresh;
138 unsigned long wb_bg_thresh;
140 unsigned long pos_ratio;
144 * Length of period for aging writeout fractions of bdis. This is an
145 * arbitrarily chosen number. The longer the period, the slower fractions will
146 * reflect changes in current writeout rate.
148 #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
150 #ifdef CONFIG_CGROUP_WRITEBACK
152 #define GDTC_INIT(__wb) .wb = (__wb), \
153 .dom = &global_wb_domain, \
154 .wb_completions = &(__wb)->completions
156 #define GDTC_INIT_NO_WB .dom = &global_wb_domain
158 #define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
159 .dom = mem_cgroup_wb_domain(__wb), \
160 .wb_completions = &(__wb)->memcg_completions, \
163 static bool mdtc_valid(struct dirty_throttle_control *dtc)
168 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
173 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
178 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
180 return &wb->memcg_completions;
183 static void wb_min_max_ratio(struct bdi_writeback *wb,
184 unsigned long *minp, unsigned long *maxp)
186 unsigned long this_bw = READ_ONCE(wb->avg_write_bandwidth);
187 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
188 unsigned long long min = wb->bdi->min_ratio;
189 unsigned long long max = wb->bdi->max_ratio;
192 * @wb may already be clean by the time control reaches here and
193 * the total may not include its bw.
195 if (this_bw < tot_bw) {
198 min = div64_ul(min, tot_bw);
200 if (max < 100 * BDI_RATIO_SCALE) {
202 max = div64_ul(max, tot_bw);
210 #else /* CONFIG_CGROUP_WRITEBACK */
212 #define GDTC_INIT(__wb) .wb = (__wb), \
213 .wb_completions = &(__wb)->completions
214 #define GDTC_INIT_NO_WB
215 #define MDTC_INIT(__wb, __gdtc)
217 static bool mdtc_valid(struct dirty_throttle_control *dtc)
222 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
224 return &global_wb_domain;
227 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
232 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
237 static void wb_min_max_ratio(struct bdi_writeback *wb,
238 unsigned long *minp, unsigned long *maxp)
240 *minp = wb->bdi->min_ratio;
241 *maxp = wb->bdi->max_ratio;
244 #endif /* CONFIG_CGROUP_WRITEBACK */
247 * In a memory zone, there is a certain amount of pages we consider
248 * available for the page cache, which is essentially the number of
249 * free and reclaimable pages, minus some zone reserves to protect
250 * lowmem and the ability to uphold the zone's watermarks without
251 * requiring writeback.
253 * This number of dirtyable pages is the base value of which the
254 * user-configurable dirty ratio is the effective number of pages that
255 * are allowed to be actually dirtied. Per individual zone, or
256 * globally by using the sum of dirtyable pages over all zones.
258 * Because the user is allowed to specify the dirty limit globally as
259 * absolute number of bytes, calculating the per-zone dirty limit can
260 * require translating the configured limit into a percentage of
261 * global dirtyable memory first.
265 * node_dirtyable_memory - number of dirtyable pages in a node
268 * Return: the node's number of pages potentially available for dirty
269 * page cache. This is the base value for the per-node dirty limits.
271 static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
273 unsigned long nr_pages = 0;
276 for (z = 0; z < MAX_NR_ZONES; z++) {
277 struct zone *zone = pgdat->node_zones + z;
279 if (!populated_zone(zone))
282 nr_pages += zone_page_state(zone, NR_FREE_PAGES);
286 * Pages reserved for the kernel should not be considered
287 * dirtyable, to prevent a situation where reclaim has to
288 * clean pages in order to balance the zones.
290 nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
292 nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
293 nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
298 static unsigned long highmem_dirtyable_memory(unsigned long total)
300 #ifdef CONFIG_HIGHMEM
305 for_each_node_state(node, N_HIGH_MEMORY) {
306 for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
308 unsigned long nr_pages;
310 if (!is_highmem_idx(i))
313 z = &NODE_DATA(node)->node_zones[i];
314 if (!populated_zone(z))
317 nr_pages = zone_page_state(z, NR_FREE_PAGES);
318 /* watch for underflows */
319 nr_pages -= min(nr_pages, high_wmark_pages(z));
320 nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
321 nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
327 * Make sure that the number of highmem pages is never larger
328 * than the number of the total dirtyable memory. This can only
329 * occur in very strange VM situations but we want to make sure
330 * that this does not occur.
332 return min(x, total);
339 * global_dirtyable_memory - number of globally dirtyable pages
341 * Return: the global number of pages potentially available for dirty
342 * page cache. This is the base value for the global dirty limits.
344 static unsigned long global_dirtyable_memory(void)
348 x = global_zone_page_state(NR_FREE_PAGES);
350 * Pages reserved for the kernel should not be considered
351 * dirtyable, to prevent a situation where reclaim has to
352 * clean pages in order to balance the zones.
354 x -= min(x, totalreserve_pages);
356 x += global_node_page_state(NR_INACTIVE_FILE);
357 x += global_node_page_state(NR_ACTIVE_FILE);
359 if (!vm_highmem_is_dirtyable)
360 x -= highmem_dirtyable_memory(x);
362 return x + 1; /* Ensure that we never return 0 */
366 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
367 * @dtc: dirty_throttle_control of interest
369 * Calculate @dtc->thresh and ->bg_thresh considering
370 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
371 * must ensure that @dtc->avail is set before calling this function. The
372 * dirty limits will be lifted by 1/4 for real-time tasks.
374 static void domain_dirty_limits(struct dirty_throttle_control *dtc)
376 const unsigned long available_memory = dtc->avail;
377 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
378 unsigned long bytes = vm_dirty_bytes;
379 unsigned long bg_bytes = dirty_background_bytes;
380 /* convert ratios to per-PAGE_SIZE for higher precision */
381 unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
382 unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
383 unsigned long thresh;
384 unsigned long bg_thresh;
385 struct task_struct *tsk;
387 /* gdtc is !NULL iff @dtc is for memcg domain */
389 unsigned long global_avail = gdtc->avail;
392 * The byte settings can't be applied directly to memcg
393 * domains. Convert them to ratios by scaling against
394 * globally available memory. As the ratios are in
395 * per-PAGE_SIZE, they can be obtained by dividing bytes by
399 ratio = min(DIV_ROUND_UP(bytes, global_avail),
402 bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
404 bytes = bg_bytes = 0;
408 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
410 thresh = (ratio * available_memory) / PAGE_SIZE;
413 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
415 bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
417 if (bg_thresh >= thresh)
418 bg_thresh = thresh / 2;
421 bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
422 thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
424 dtc->thresh = thresh;
425 dtc->bg_thresh = bg_thresh;
427 /* we should eventually report the domain in the TP */
429 trace_global_dirty_state(bg_thresh, thresh);
433 * global_dirty_limits - background-writeback and dirty-throttling thresholds
434 * @pbackground: out parameter for bg_thresh
435 * @pdirty: out parameter for thresh
437 * Calculate bg_thresh and thresh for global_wb_domain. See
438 * domain_dirty_limits() for details.
440 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
442 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
444 gdtc.avail = global_dirtyable_memory();
445 domain_dirty_limits(&gdtc);
447 *pbackground = gdtc.bg_thresh;
448 *pdirty = gdtc.thresh;
452 * node_dirty_limit - maximum number of dirty pages allowed in a node
455 * Return: the maximum number of dirty pages allowed in a node, based
456 * on the node's dirtyable memory.
458 static unsigned long node_dirty_limit(struct pglist_data *pgdat)
460 unsigned long node_memory = node_dirtyable_memory(pgdat);
461 struct task_struct *tsk = current;
465 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
466 node_memory / global_dirtyable_memory();
468 dirty = vm_dirty_ratio * node_memory / 100;
477 * node_dirty_ok - tells whether a node is within its dirty limits
478 * @pgdat: the node to check
480 * Return: %true when the dirty pages in @pgdat are within the node's
481 * dirty limit, %false if the limit is exceeded.
483 bool node_dirty_ok(struct pglist_data *pgdat)
485 unsigned long limit = node_dirty_limit(pgdat);
486 unsigned long nr_pages = 0;
488 nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
489 nr_pages += node_page_state(pgdat, NR_WRITEBACK);
491 return nr_pages <= limit;
495 static int dirty_background_ratio_handler(struct ctl_table *table, int write,
496 void *buffer, size_t *lenp, loff_t *ppos)
500 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
501 if (ret == 0 && write)
502 dirty_background_bytes = 0;
506 static int dirty_background_bytes_handler(struct ctl_table *table, int write,
507 void *buffer, size_t *lenp, loff_t *ppos)
511 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
512 if (ret == 0 && write)
513 dirty_background_ratio = 0;
517 static int dirty_ratio_handler(struct ctl_table *table, int write, void *buffer,
518 size_t *lenp, loff_t *ppos)
520 int old_ratio = vm_dirty_ratio;
523 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
524 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
525 writeback_set_ratelimit();
531 static int dirty_bytes_handler(struct ctl_table *table, int write,
532 void *buffer, size_t *lenp, loff_t *ppos)
534 unsigned long old_bytes = vm_dirty_bytes;
537 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
538 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
539 writeback_set_ratelimit();
546 static unsigned long wp_next_time(unsigned long cur_time)
548 cur_time += VM_COMPLETIONS_PERIOD_LEN;
549 /* 0 has a special meaning... */
555 static void wb_domain_writeout_add(struct wb_domain *dom,
556 struct fprop_local_percpu *completions,
557 unsigned int max_prop_frac, long nr)
559 __fprop_add_percpu_max(&dom->completions, completions,
561 /* First event after period switching was turned off? */
562 if (unlikely(!dom->period_time)) {
564 * We can race with other __bdi_writeout_inc calls here but
565 * it does not cause any harm since the resulting time when
566 * timer will fire and what is in writeout_period_time will be
569 dom->period_time = wp_next_time(jiffies);
570 mod_timer(&dom->period_timer, dom->period_time);
575 * Increment @wb's writeout completion count and the global writeout
576 * completion count. Called from __folio_end_writeback().
578 static inline void __wb_writeout_add(struct bdi_writeback *wb, long nr)
580 struct wb_domain *cgdom;
582 wb_stat_mod(wb, WB_WRITTEN, nr);
583 wb_domain_writeout_add(&global_wb_domain, &wb->completions,
584 wb->bdi->max_prop_frac, nr);
586 cgdom = mem_cgroup_wb_domain(wb);
588 wb_domain_writeout_add(cgdom, wb_memcg_completions(wb),
589 wb->bdi->max_prop_frac, nr);
592 void wb_writeout_inc(struct bdi_writeback *wb)
596 local_irq_save(flags);
597 __wb_writeout_add(wb, 1);
598 local_irq_restore(flags);
600 EXPORT_SYMBOL_GPL(wb_writeout_inc);
603 * On idle system, we can be called long after we scheduled because we use
604 * deferred timers so count with missed periods.
606 static void writeout_period(struct timer_list *t)
608 struct wb_domain *dom = from_timer(dom, t, period_timer);
609 int miss_periods = (jiffies - dom->period_time) /
610 VM_COMPLETIONS_PERIOD_LEN;
612 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
613 dom->period_time = wp_next_time(dom->period_time +
614 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
615 mod_timer(&dom->period_timer, dom->period_time);
618 * Aging has zeroed all fractions. Stop wasting CPU on period
621 dom->period_time = 0;
625 int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
627 memset(dom, 0, sizeof(*dom));
629 spin_lock_init(&dom->lock);
631 timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
633 dom->dirty_limit_tstamp = jiffies;
635 return fprop_global_init(&dom->completions, gfp);
638 #ifdef CONFIG_CGROUP_WRITEBACK
639 void wb_domain_exit(struct wb_domain *dom)
641 del_timer_sync(&dom->period_timer);
642 fprop_global_destroy(&dom->completions);
647 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
648 * registered backing devices, which, for obvious reasons, can not
651 static unsigned int bdi_min_ratio;
653 static u64 bdi_get_bytes(unsigned int ratio)
655 unsigned long background_thresh;
656 unsigned long dirty_thresh;
659 global_dirty_limits(&background_thresh, &dirty_thresh);
660 bytes = (dirty_thresh * PAGE_SIZE * ratio) / BDI_RATIO_SCALE / 100;
665 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
670 min_ratio *= BDI_RATIO_SCALE;
672 spin_lock_bh(&bdi_lock);
673 if (min_ratio > bdi->max_ratio) {
676 if (min_ratio < bdi->min_ratio) {
677 delta = bdi->min_ratio - min_ratio;
678 bdi_min_ratio -= delta;
679 bdi->min_ratio = min_ratio;
681 delta = min_ratio - bdi->min_ratio;
682 if (bdi_min_ratio + delta < 100 * BDI_RATIO_SCALE) {
683 bdi_min_ratio += delta;
684 bdi->min_ratio = min_ratio;
690 spin_unlock_bh(&bdi_lock);
695 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
701 max_ratio *= BDI_RATIO_SCALE;
703 spin_lock_bh(&bdi_lock);
704 if (bdi->min_ratio > max_ratio) {
707 bdi->max_ratio = max_ratio;
708 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
710 spin_unlock_bh(&bdi_lock);
714 EXPORT_SYMBOL(bdi_set_max_ratio);
716 u64 bdi_get_max_bytes(struct backing_dev_info *bdi)
718 return bdi_get_bytes(bdi->max_ratio);
721 int bdi_set_strict_limit(struct backing_dev_info *bdi, unsigned int strict_limit)
723 if (strict_limit > 1)
726 spin_lock_bh(&bdi_lock);
728 bdi->capabilities |= BDI_CAP_STRICTLIMIT;
730 bdi->capabilities &= ~BDI_CAP_STRICTLIMIT;
731 spin_unlock_bh(&bdi_lock);
736 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
737 unsigned long bg_thresh)
739 return (thresh + bg_thresh) / 2;
742 static unsigned long hard_dirty_limit(struct wb_domain *dom,
743 unsigned long thresh)
745 return max(thresh, dom->dirty_limit);
749 * Memory which can be further allocated to a memcg domain is capped by
750 * system-wide clean memory excluding the amount being used in the domain.
752 static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
753 unsigned long filepages, unsigned long headroom)
755 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
756 unsigned long clean = filepages - min(filepages, mdtc->dirty);
757 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
758 unsigned long other_clean = global_clean - min(global_clean, clean);
760 mdtc->avail = filepages + min(headroom, other_clean);
764 * __wb_calc_thresh - @wb's share of dirty throttling threshold
765 * @dtc: dirty_throttle_context of interest
767 * Note that balance_dirty_pages() will only seriously take it as a hard limit
768 * when sleeping max_pause per page is not enough to keep the dirty pages under
769 * control. For example, when the device is completely stalled due to some error
770 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
771 * In the other normal situations, it acts more gently by throttling the tasks
772 * more (rather than completely block them) when the wb dirty pages go high.
774 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
775 * - starving fast devices
776 * - piling up dirty pages (that will take long time to sync) on slow devices
778 * The wb's share of dirty limit will be adapting to its throughput and
779 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
781 * Return: @wb's dirty limit in pages. The term "dirty" in the context of
782 * dirty balancing includes all PG_dirty and PG_writeback pages.
784 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
786 struct wb_domain *dom = dtc_dom(dtc);
787 unsigned long thresh = dtc->thresh;
789 unsigned long numerator, denominator;
790 unsigned long wb_min_ratio, wb_max_ratio;
793 * Calculate this BDI's share of the thresh ratio.
795 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
796 &numerator, &denominator);
798 wb_thresh = (thresh * (100 * BDI_RATIO_SCALE - bdi_min_ratio)) / (100 * BDI_RATIO_SCALE);
799 wb_thresh *= numerator;
800 wb_thresh = div64_ul(wb_thresh, denominator);
802 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
804 wb_thresh += (thresh * wb_min_ratio) / (100 * BDI_RATIO_SCALE);
805 if (wb_thresh > (thresh * wb_max_ratio) / (100 * BDI_RATIO_SCALE))
806 wb_thresh = thresh * wb_max_ratio / (100 * BDI_RATIO_SCALE);
811 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
813 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
815 return __wb_calc_thresh(&gdtc);
820 * f(dirty) := 1.0 + (----------------)
823 * it's a 3rd order polynomial that subjects to
825 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
826 * (2) f(setpoint) = 1.0 => the balance point
827 * (3) f(limit) = 0 => the hard limit
828 * (4) df/dx <= 0 => negative feedback control
829 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
830 * => fast response on large errors; small oscillation near setpoint
832 static long long pos_ratio_polynom(unsigned long setpoint,
839 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
840 (limit - setpoint) | 1);
842 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
843 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
844 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
846 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
850 * Dirty position control.
852 * (o) global/bdi setpoints
854 * We want the dirty pages be balanced around the global/wb setpoints.
855 * When the number of dirty pages is higher/lower than the setpoint, the
856 * dirty position control ratio (and hence task dirty ratelimit) will be
857 * decreased/increased to bring the dirty pages back to the setpoint.
859 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
861 * if (dirty < setpoint) scale up pos_ratio
862 * if (dirty > setpoint) scale down pos_ratio
864 * if (wb_dirty < wb_setpoint) scale up pos_ratio
865 * if (wb_dirty > wb_setpoint) scale down pos_ratio
867 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
869 * (o) global control line
873 * | |<===== global dirty control scope ======>|
881 * 1.0 ................................*
887 * 0 +------------.------------------.----------------------*------------->
888 * freerun^ setpoint^ limit^ dirty pages
890 * (o) wb control line
898 * | * |<=========== span ============>|
899 * 1.0 .......................*
911 * 1/4 ...............................................* * * * * * * * * * * *
915 * 0 +----------------------.-------------------------------.------------->
916 * wb_setpoint^ x_intercept^
918 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
919 * be smoothly throttled down to normal if it starts high in situations like
920 * - start writing to a slow SD card and a fast disk at the same time. The SD
921 * card's wb_dirty may rush to many times higher than wb_setpoint.
922 * - the wb dirty thresh drops quickly due to change of JBOD workload
924 static void wb_position_ratio(struct dirty_throttle_control *dtc)
926 struct bdi_writeback *wb = dtc->wb;
927 unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth);
928 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
929 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
930 unsigned long wb_thresh = dtc->wb_thresh;
931 unsigned long x_intercept;
932 unsigned long setpoint; /* dirty pages' target balance point */
933 unsigned long wb_setpoint;
935 long long pos_ratio; /* for scaling up/down the rate limit */
940 if (unlikely(dtc->dirty >= limit))
946 * See comment for pos_ratio_polynom().
948 setpoint = (freerun + limit) / 2;
949 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
952 * The strictlimit feature is a tool preventing mistrusted filesystems
953 * from growing a large number of dirty pages before throttling. For
954 * such filesystems balance_dirty_pages always checks wb counters
955 * against wb limits. Even if global "nr_dirty" is under "freerun".
956 * This is especially important for fuse which sets bdi->max_ratio to
957 * 1% by default. Without strictlimit feature, fuse writeback may
958 * consume arbitrary amount of RAM because it is accounted in
959 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
961 * Here, in wb_position_ratio(), we calculate pos_ratio based on
962 * two values: wb_dirty and wb_thresh. Let's consider an example:
963 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
964 * limits are set by default to 10% and 20% (background and throttle).
965 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
966 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
967 * about ~6K pages (as the average of background and throttle wb
968 * limits). The 3rd order polynomial will provide positive feedback if
969 * wb_dirty is under wb_setpoint and vice versa.
971 * Note, that we cannot use global counters in these calculations
972 * because we want to throttle process writing to a strictlimit wb
973 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
974 * in the example above).
976 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
977 long long wb_pos_ratio;
979 if (dtc->wb_dirty < 8) {
980 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
981 2 << RATELIMIT_CALC_SHIFT);
985 if (dtc->wb_dirty >= wb_thresh)
988 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
991 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
994 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
998 * Typically, for strictlimit case, wb_setpoint << setpoint
999 * and pos_ratio >> wb_pos_ratio. In the other words global
1000 * state ("dirty") is not limiting factor and we have to
1001 * make decision based on wb counters. But there is an
1002 * important case when global pos_ratio should get precedence:
1003 * global limits are exceeded (e.g. due to activities on other
1004 * wb's) while given strictlimit wb is below limit.
1006 * "pos_ratio * wb_pos_ratio" would work for the case above,
1007 * but it would look too non-natural for the case of all
1008 * activity in the system coming from a single strictlimit wb
1009 * with bdi->max_ratio == 100%.
1011 * Note that min() below somewhat changes the dynamics of the
1012 * control system. Normally, pos_ratio value can be well over 3
1013 * (when globally we are at freerun and wb is well below wb
1014 * setpoint). Now the maximum pos_ratio in the same situation
1015 * is 2. We might want to tweak this if we observe the control
1016 * system is too slow to adapt.
1018 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
1023 * We have computed basic pos_ratio above based on global situation. If
1024 * the wb is over/under its share of dirty pages, we want to scale
1025 * pos_ratio further down/up. That is done by the following mechanism.
1031 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1033 * x_intercept - wb_dirty
1034 * := --------------------------
1035 * x_intercept - wb_setpoint
1037 * The main wb control line is a linear function that subjects to
1039 * (1) f(wb_setpoint) = 1.0
1040 * (2) k = - 1 / (8 * write_bw) (in single wb case)
1041 * or equally: x_intercept = wb_setpoint + 8 * write_bw
1043 * For single wb case, the dirty pages are observed to fluctuate
1044 * regularly within range
1045 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1046 * for various filesystems, where (2) can yield in a reasonable 12.5%
1047 * fluctuation range for pos_ratio.
1049 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1050 * own size, so move the slope over accordingly and choose a slope that
1051 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1053 if (unlikely(wb_thresh > dtc->thresh))
1054 wb_thresh = dtc->thresh;
1056 * It's very possible that wb_thresh is close to 0 not because the
1057 * device is slow, but that it has remained inactive for long time.
1058 * Honour such devices a reasonable good (hopefully IO efficient)
1059 * threshold, so that the occasional writes won't be blocked and active
1060 * writes can rampup the threshold quickly.
1062 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1064 * scale global setpoint to wb's:
1065 * wb_setpoint = setpoint * wb_thresh / thresh
1067 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1068 wb_setpoint = setpoint * (u64)x >> 16;
1070 * Use span=(8*write_bw) in single wb case as indicated by
1071 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1073 * wb_thresh thresh - wb_thresh
1074 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1077 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1078 x_intercept = wb_setpoint + span;
1080 if (dtc->wb_dirty < x_intercept - span / 4) {
1081 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1082 (x_intercept - wb_setpoint) | 1);
1087 * wb reserve area, safeguard against dirty pool underrun and disk idle
1088 * It may push the desired control point of global dirty pages higher
1091 x_intercept = wb_thresh / 2;
1092 if (dtc->wb_dirty < x_intercept) {
1093 if (dtc->wb_dirty > x_intercept / 8)
1094 pos_ratio = div_u64(pos_ratio * x_intercept,
1100 dtc->pos_ratio = pos_ratio;
1103 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1104 unsigned long elapsed,
1105 unsigned long written)
1107 const unsigned long period = roundup_pow_of_two(3 * HZ);
1108 unsigned long avg = wb->avg_write_bandwidth;
1109 unsigned long old = wb->write_bandwidth;
1113 * bw = written * HZ / elapsed
1115 * bw * elapsed + write_bandwidth * (period - elapsed)
1116 * write_bandwidth = ---------------------------------------------------
1119 * @written may have decreased due to folio_account_redirty().
1120 * Avoid underflowing @bw calculation.
1122 bw = written - min(written, wb->written_stamp);
1124 if (unlikely(elapsed > period)) {
1125 bw = div64_ul(bw, elapsed);
1129 bw += (u64)wb->write_bandwidth * (period - elapsed);
1130 bw >>= ilog2(period);
1133 * one more level of smoothing, for filtering out sudden spikes
1135 if (avg > old && old >= (unsigned long)bw)
1136 avg -= (avg - old) >> 3;
1138 if (avg < old && old <= (unsigned long)bw)
1139 avg += (old - avg) >> 3;
1142 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1143 avg = max(avg, 1LU);
1144 if (wb_has_dirty_io(wb)) {
1145 long delta = avg - wb->avg_write_bandwidth;
1146 WARN_ON_ONCE(atomic_long_add_return(delta,
1147 &wb->bdi->tot_write_bandwidth) <= 0);
1149 wb->write_bandwidth = bw;
1150 WRITE_ONCE(wb->avg_write_bandwidth, avg);
1153 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1155 struct wb_domain *dom = dtc_dom(dtc);
1156 unsigned long thresh = dtc->thresh;
1157 unsigned long limit = dom->dirty_limit;
1160 * Follow up in one step.
1162 if (limit < thresh) {
1168 * Follow down slowly. Use the higher one as the target, because thresh
1169 * may drop below dirty. This is exactly the reason to introduce
1170 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1172 thresh = max(thresh, dtc->dirty);
1173 if (limit > thresh) {
1174 limit -= (limit - thresh) >> 5;
1179 dom->dirty_limit = limit;
1182 static void domain_update_dirty_limit(struct dirty_throttle_control *dtc,
1185 struct wb_domain *dom = dtc_dom(dtc);
1188 * check locklessly first to optimize away locking for the most time
1190 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1193 spin_lock(&dom->lock);
1194 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1195 update_dirty_limit(dtc);
1196 dom->dirty_limit_tstamp = now;
1198 spin_unlock(&dom->lock);
1202 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1204 * Normal wb tasks will be curbed at or below it in long term.
1205 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1207 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1208 unsigned long dirtied,
1209 unsigned long elapsed)
1211 struct bdi_writeback *wb = dtc->wb;
1212 unsigned long dirty = dtc->dirty;
1213 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1214 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1215 unsigned long setpoint = (freerun + limit) / 2;
1216 unsigned long write_bw = wb->avg_write_bandwidth;
1217 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1218 unsigned long dirty_rate;
1219 unsigned long task_ratelimit;
1220 unsigned long balanced_dirty_ratelimit;
1223 unsigned long shift;
1226 * The dirty rate will match the writeout rate in long term, except
1227 * when dirty pages are truncated by userspace or re-dirtied by FS.
1229 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1232 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1234 task_ratelimit = (u64)dirty_ratelimit *
1235 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1236 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1239 * A linear estimation of the "balanced" throttle rate. The theory is,
1240 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1241 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1242 * formula will yield the balanced rate limit (write_bw / N).
1244 * Note that the expanded form is not a pure rate feedback:
1245 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1246 * but also takes pos_ratio into account:
1247 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1249 * (1) is not realistic because pos_ratio also takes part in balancing
1250 * the dirty rate. Consider the state
1251 * pos_ratio = 0.5 (3)
1252 * rate = 2 * (write_bw / N) (4)
1253 * If (1) is used, it will stuck in that state! Because each dd will
1255 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1257 * dirty_rate = N * task_ratelimit = write_bw (6)
1258 * put (6) into (1) we get
1259 * rate_(i+1) = rate_(i) (7)
1261 * So we end up using (2) to always keep
1262 * rate_(i+1) ~= (write_bw / N) (8)
1263 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1264 * pos_ratio is able to drive itself to 1.0, which is not only where
1265 * the dirty count meet the setpoint, but also where the slope of
1266 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1268 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1271 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1273 if (unlikely(balanced_dirty_ratelimit > write_bw))
1274 balanced_dirty_ratelimit = write_bw;
1277 * We could safely do this and return immediately:
1279 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1281 * However to get a more stable dirty_ratelimit, the below elaborated
1282 * code makes use of task_ratelimit to filter out singular points and
1283 * limit the step size.
1285 * The below code essentially only uses the relative value of
1287 * task_ratelimit - dirty_ratelimit
1288 * = (pos_ratio - 1) * dirty_ratelimit
1290 * which reflects the direction and size of dirty position error.
1294 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1295 * task_ratelimit is on the same side of dirty_ratelimit, too.
1297 * - dirty_ratelimit > balanced_dirty_ratelimit
1298 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1299 * lowering dirty_ratelimit will help meet both the position and rate
1300 * control targets. Otherwise, don't update dirty_ratelimit if it will
1301 * only help meet the rate target. After all, what the users ultimately
1302 * feel and care are stable dirty rate and small position error.
1304 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1305 * and filter out the singular points of balanced_dirty_ratelimit. Which
1306 * keeps jumping around randomly and can even leap far away at times
1307 * due to the small 200ms estimation period of dirty_rate (we want to
1308 * keep that period small to reduce time lags).
1313 * For strictlimit case, calculations above were based on wb counters
1314 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1315 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1316 * Hence, to calculate "step" properly, we have to use wb_dirty as
1317 * "dirty" and wb_setpoint as "setpoint".
1319 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1320 * it's possible that wb_thresh is close to zero due to inactivity
1321 * of backing device.
1323 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1324 dirty = dtc->wb_dirty;
1325 if (dtc->wb_dirty < 8)
1326 setpoint = dtc->wb_dirty + 1;
1328 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1331 if (dirty < setpoint) {
1332 x = min3(wb->balanced_dirty_ratelimit,
1333 balanced_dirty_ratelimit, task_ratelimit);
1334 if (dirty_ratelimit < x)
1335 step = x - dirty_ratelimit;
1337 x = max3(wb->balanced_dirty_ratelimit,
1338 balanced_dirty_ratelimit, task_ratelimit);
1339 if (dirty_ratelimit > x)
1340 step = dirty_ratelimit - x;
1344 * Don't pursue 100% rate matching. It's impossible since the balanced
1345 * rate itself is constantly fluctuating. So decrease the track speed
1346 * when it gets close to the target. Helps eliminate pointless tremors.
1348 shift = dirty_ratelimit / (2 * step + 1);
1349 if (shift < BITS_PER_LONG)
1350 step = DIV_ROUND_UP(step >> shift, 8);
1354 if (dirty_ratelimit < balanced_dirty_ratelimit)
1355 dirty_ratelimit += step;
1357 dirty_ratelimit -= step;
1359 WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL));
1360 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1362 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1365 static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1366 struct dirty_throttle_control *mdtc,
1367 bool update_ratelimit)
1369 struct bdi_writeback *wb = gdtc->wb;
1370 unsigned long now = jiffies;
1371 unsigned long elapsed;
1372 unsigned long dirtied;
1373 unsigned long written;
1375 spin_lock(&wb->list_lock);
1378 * Lockless checks for elapsed time are racy and delayed update after
1379 * IO completion doesn't do it at all (to make sure written pages are
1380 * accounted reasonably quickly). Make sure elapsed >= 1 to avoid
1383 elapsed = max(now - wb->bw_time_stamp, 1UL);
1384 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1385 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1387 if (update_ratelimit) {
1388 domain_update_dirty_limit(gdtc, now);
1389 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1392 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1393 * compiler has no way to figure that out. Help it.
1395 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1396 domain_update_dirty_limit(mdtc, now);
1397 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1400 wb_update_write_bandwidth(wb, elapsed, written);
1402 wb->dirtied_stamp = dirtied;
1403 wb->written_stamp = written;
1404 WRITE_ONCE(wb->bw_time_stamp, now);
1405 spin_unlock(&wb->list_lock);
1408 void wb_update_bandwidth(struct bdi_writeback *wb)
1410 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1412 __wb_update_bandwidth(&gdtc, NULL, false);
1415 /* Interval after which we consider wb idle and don't estimate bandwidth */
1416 #define WB_BANDWIDTH_IDLE_JIF (HZ)
1418 static void wb_bandwidth_estimate_start(struct bdi_writeback *wb)
1420 unsigned long now = jiffies;
1421 unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp);
1423 if (elapsed > WB_BANDWIDTH_IDLE_JIF &&
1424 !atomic_read(&wb->writeback_inodes)) {
1425 spin_lock(&wb->list_lock);
1426 wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED);
1427 wb->written_stamp = wb_stat(wb, WB_WRITTEN);
1428 WRITE_ONCE(wb->bw_time_stamp, now);
1429 spin_unlock(&wb->list_lock);
1434 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1435 * will look to see if it needs to start dirty throttling.
1437 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1438 * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1439 * (the number of pages we may dirty without exceeding the dirty limits).
1441 static unsigned long dirty_poll_interval(unsigned long dirty,
1442 unsigned long thresh)
1445 return 1UL << (ilog2(thresh - dirty) >> 1);
1450 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1451 unsigned long wb_dirty)
1453 unsigned long bw = READ_ONCE(wb->avg_write_bandwidth);
1457 * Limit pause time for small memory systems. If sleeping for too long
1458 * time, a small pool of dirty/writeback pages may go empty and disk go
1461 * 8 serves as the safety ratio.
1463 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1466 return min_t(unsigned long, t, MAX_PAUSE);
1469 static long wb_min_pause(struct bdi_writeback *wb,
1471 unsigned long task_ratelimit,
1472 unsigned long dirty_ratelimit,
1473 int *nr_dirtied_pause)
1475 long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth));
1476 long lo = ilog2(READ_ONCE(wb->dirty_ratelimit));
1477 long t; /* target pause */
1478 long pause; /* estimated next pause */
1479 int pages; /* target nr_dirtied_pause */
1481 /* target for 10ms pause on 1-dd case */
1482 t = max(1, HZ / 100);
1485 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1488 * (N * 10ms) on 2^N concurrent tasks.
1491 t += (hi - lo) * (10 * HZ) / 1024;
1494 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1495 * on the much more stable dirty_ratelimit. However the next pause time
1496 * will be computed based on task_ratelimit and the two rate limits may
1497 * depart considerably at some time. Especially if task_ratelimit goes
1498 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1499 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1500 * result task_ratelimit won't be executed faithfully, which could
1501 * eventually bring down dirty_ratelimit.
1503 * We apply two rules to fix it up:
1504 * 1) try to estimate the next pause time and if necessary, use a lower
1505 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1506 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1507 * 2) limit the target pause time to max_pause/2, so that the normal
1508 * small fluctuations of task_ratelimit won't trigger rule (1) and
1509 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1511 t = min(t, 1 + max_pause / 2);
1512 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1515 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1516 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1517 * When the 16 consecutive reads are often interrupted by some dirty
1518 * throttling pause during the async writes, cfq will go into idles
1519 * (deadline is fine). So push nr_dirtied_pause as high as possible
1520 * until reaches DIRTY_POLL_THRESH=32 pages.
1522 if (pages < DIRTY_POLL_THRESH) {
1524 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1525 if (pages > DIRTY_POLL_THRESH) {
1526 pages = DIRTY_POLL_THRESH;
1527 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1531 pause = HZ * pages / (task_ratelimit + 1);
1532 if (pause > max_pause) {
1534 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1537 *nr_dirtied_pause = pages;
1539 * The minimal pause time will normally be half the target pause time.
1541 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1544 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1546 struct bdi_writeback *wb = dtc->wb;
1547 unsigned long wb_reclaimable;
1550 * wb_thresh is not treated as some limiting factor as
1551 * dirty_thresh, due to reasons
1552 * - in JBOD setup, wb_thresh can fluctuate a lot
1553 * - in a system with HDD and USB key, the USB key may somehow
1554 * go into state (wb_dirty >> wb_thresh) either because
1555 * wb_dirty starts high, or because wb_thresh drops low.
1556 * In this case we don't want to hard throttle the USB key
1557 * dirtiers for 100 seconds until wb_dirty drops under
1558 * wb_thresh. Instead the auxiliary wb control line in
1559 * wb_position_ratio() will let the dirtier task progress
1560 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1562 dtc->wb_thresh = __wb_calc_thresh(dtc);
1563 dtc->wb_bg_thresh = dtc->thresh ?
1564 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1567 * In order to avoid the stacked BDI deadlock we need
1568 * to ensure we accurately count the 'dirty' pages when
1569 * the threshold is low.
1571 * Otherwise it would be possible to get thresh+n pages
1572 * reported dirty, even though there are thresh-m pages
1573 * actually dirty; with m+n sitting in the percpu
1576 if (dtc->wb_thresh < 2 * wb_stat_error()) {
1577 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1578 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1580 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1581 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1586 * balance_dirty_pages() must be called by processes which are generating dirty
1587 * data. It looks at the number of dirty pages in the machine and will force
1588 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1589 * If we're over `background_thresh' then the writeback threads are woken to
1590 * perform some writeout.
1592 static int balance_dirty_pages(struct bdi_writeback *wb,
1593 unsigned long pages_dirtied, unsigned int flags)
1595 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1596 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1597 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1598 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1600 struct dirty_throttle_control *sdtc;
1601 unsigned long nr_reclaimable; /* = file_dirty */
1606 int nr_dirtied_pause;
1607 bool dirty_exceeded = false;
1608 unsigned long task_ratelimit;
1609 unsigned long dirty_ratelimit;
1610 struct backing_dev_info *bdi = wb->bdi;
1611 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1612 unsigned long start_time = jiffies;
1616 unsigned long now = jiffies;
1617 unsigned long dirty, thresh, bg_thresh;
1618 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1619 unsigned long m_thresh = 0;
1620 unsigned long m_bg_thresh = 0;
1622 nr_reclaimable = global_node_page_state(NR_FILE_DIRTY);
1623 gdtc->avail = global_dirtyable_memory();
1624 gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1626 domain_dirty_limits(gdtc);
1628 if (unlikely(strictlimit)) {
1629 wb_dirty_limits(gdtc);
1631 dirty = gdtc->wb_dirty;
1632 thresh = gdtc->wb_thresh;
1633 bg_thresh = gdtc->wb_bg_thresh;
1635 dirty = gdtc->dirty;
1636 thresh = gdtc->thresh;
1637 bg_thresh = gdtc->bg_thresh;
1641 unsigned long filepages, headroom, writeback;
1644 * If @wb belongs to !root memcg, repeat the same
1645 * basic calculations for the memcg domain.
1647 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1648 &mdtc->dirty, &writeback);
1649 mdtc->dirty += writeback;
1650 mdtc_calc_avail(mdtc, filepages, headroom);
1652 domain_dirty_limits(mdtc);
1654 if (unlikely(strictlimit)) {
1655 wb_dirty_limits(mdtc);
1656 m_dirty = mdtc->wb_dirty;
1657 m_thresh = mdtc->wb_thresh;
1658 m_bg_thresh = mdtc->wb_bg_thresh;
1660 m_dirty = mdtc->dirty;
1661 m_thresh = mdtc->thresh;
1662 m_bg_thresh = mdtc->bg_thresh;
1667 * In laptop mode, we wait until hitting the higher threshold
1668 * before starting background writeout, and then write out all
1669 * the way down to the lower threshold. So slow writers cause
1670 * minimal disk activity.
1672 * In normal mode, we start background writeout at the lower
1673 * background_thresh, to keep the amount of dirty memory low.
1675 if (!laptop_mode && nr_reclaimable > gdtc->bg_thresh &&
1676 !writeback_in_progress(wb))
1677 wb_start_background_writeback(wb);
1680 * Throttle it only when the background writeback cannot
1681 * catch-up. This avoids (excessively) small writeouts
1682 * when the wb limits are ramping up in case of !strictlimit.
1684 * In strictlimit case make decision based on the wb counters
1685 * and limits. Small writeouts when the wb limits are ramping
1686 * up are the price we consciously pay for strictlimit-ing.
1688 * If memcg domain is in effect, @dirty should be under
1689 * both global and memcg freerun ceilings.
1691 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1693 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1695 unsigned long m_intv;
1698 intv = dirty_poll_interval(dirty, thresh);
1701 current->dirty_paused_when = now;
1702 current->nr_dirtied = 0;
1704 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1705 current->nr_dirtied_pause = min(intv, m_intv);
1709 /* Start writeback even when in laptop mode */
1710 if (unlikely(!writeback_in_progress(wb)))
1711 wb_start_background_writeback(wb);
1713 mem_cgroup_flush_foreign(wb);
1716 * Calculate global domain's pos_ratio and select the
1717 * global dtc by default.
1720 wb_dirty_limits(gdtc);
1722 if ((current->flags & PF_LOCAL_THROTTLE) &&
1724 dirty_freerun_ceiling(gdtc->wb_thresh,
1725 gdtc->wb_bg_thresh))
1727 * LOCAL_THROTTLE tasks must not be throttled
1728 * when below the per-wb freerun ceiling.
1733 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1734 ((gdtc->dirty > gdtc->thresh) || strictlimit);
1736 wb_position_ratio(gdtc);
1741 * If memcg domain is in effect, calculate its
1742 * pos_ratio. @wb should satisfy constraints from
1743 * both global and memcg domains. Choose the one
1744 * w/ lower pos_ratio.
1747 wb_dirty_limits(mdtc);
1749 if ((current->flags & PF_LOCAL_THROTTLE) &&
1751 dirty_freerun_ceiling(mdtc->wb_thresh,
1752 mdtc->wb_bg_thresh))
1754 * LOCAL_THROTTLE tasks must not be
1755 * throttled when below the per-wb
1760 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1761 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1763 wb_position_ratio(mdtc);
1764 if (mdtc->pos_ratio < gdtc->pos_ratio)
1768 if (dirty_exceeded != wb->dirty_exceeded)
1769 wb->dirty_exceeded = dirty_exceeded;
1771 if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
1772 BANDWIDTH_INTERVAL))
1773 __wb_update_bandwidth(gdtc, mdtc, true);
1775 /* throttle according to the chosen dtc */
1776 dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit);
1777 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1778 RATELIMIT_CALC_SHIFT;
1779 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1780 min_pause = wb_min_pause(wb, max_pause,
1781 task_ratelimit, dirty_ratelimit,
1784 if (unlikely(task_ratelimit == 0)) {
1789 period = HZ * pages_dirtied / task_ratelimit;
1791 if (current->dirty_paused_when)
1792 pause -= now - current->dirty_paused_when;
1794 * For less than 1s think time (ext3/4 may block the dirtier
1795 * for up to 800ms from time to time on 1-HDD; so does xfs,
1796 * however at much less frequency), try to compensate it in
1797 * future periods by updating the virtual time; otherwise just
1798 * do a reset, as it may be a light dirtier.
1800 if (pause < min_pause) {
1801 trace_balance_dirty_pages(wb,
1814 current->dirty_paused_when = now;
1815 current->nr_dirtied = 0;
1816 } else if (period) {
1817 current->dirty_paused_when += period;
1818 current->nr_dirtied = 0;
1819 } else if (current->nr_dirtied_pause <= pages_dirtied)
1820 current->nr_dirtied_pause += pages_dirtied;
1823 if (unlikely(pause > max_pause)) {
1824 /* for occasional dropped task_ratelimit */
1825 now += min(pause - max_pause, max_pause);
1830 trace_balance_dirty_pages(wb,
1842 if (flags & BDP_ASYNC) {
1846 __set_current_state(TASK_KILLABLE);
1847 wb->dirty_sleep = now;
1848 io_schedule_timeout(pause);
1850 current->dirty_paused_when = now + pause;
1851 current->nr_dirtied = 0;
1852 current->nr_dirtied_pause = nr_dirtied_pause;
1855 * This is typically equal to (dirty < thresh) and can also
1856 * keep "1000+ dd on a slow USB stick" under control.
1862 * In the case of an unresponsive NFS server and the NFS dirty
1863 * pages exceeds dirty_thresh, give the other good wb's a pipe
1864 * to go through, so that tasks on them still remain responsive.
1866 * In theory 1 page is enough to keep the consumer-producer
1867 * pipe going: the flusher cleans 1 page => the task dirties 1
1868 * more page. However wb_dirty has accounting errors. So use
1869 * the larger and more IO friendly wb_stat_error.
1871 if (sdtc->wb_dirty <= wb_stat_error())
1874 if (fatal_signal_pending(current))
1880 static DEFINE_PER_CPU(int, bdp_ratelimits);
1883 * Normal tasks are throttled by
1885 * dirty tsk->nr_dirtied_pause pages;
1886 * take a snap in balance_dirty_pages();
1888 * However there is a worst case. If every task exit immediately when dirtied
1889 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1890 * called to throttle the page dirties. The solution is to save the not yet
1891 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1892 * randomly into the running tasks. This works well for the above worst case,
1893 * as the new task will pick up and accumulate the old task's leaked dirty
1894 * count and eventually get throttled.
1896 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1899 * balance_dirty_pages_ratelimited_flags - Balance dirty memory state.
1900 * @mapping: address_space which was dirtied.
1901 * @flags: BDP flags.
1903 * Processes which are dirtying memory should call in here once for each page
1904 * which was newly dirtied. The function will periodically check the system's
1905 * dirty state and will initiate writeback if needed.
1907 * See balance_dirty_pages_ratelimited() for details.
1909 * Return: If @flags contains BDP_ASYNC, it may return -EAGAIN to
1910 * indicate that memory is out of balance and the caller must wait
1911 * for I/O to complete. Otherwise, it will return 0 to indicate
1912 * that either memory was already in balance, or it was able to sleep
1913 * until the amount of dirty memory returned to balance.
1915 int balance_dirty_pages_ratelimited_flags(struct address_space *mapping,
1918 struct inode *inode = mapping->host;
1919 struct backing_dev_info *bdi = inode_to_bdi(inode);
1920 struct bdi_writeback *wb = NULL;
1925 if (!(bdi->capabilities & BDI_CAP_WRITEBACK))
1928 if (inode_cgwb_enabled(inode))
1929 wb = wb_get_create_current(bdi, GFP_KERNEL);
1933 ratelimit = current->nr_dirtied_pause;
1934 if (wb->dirty_exceeded)
1935 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1939 * This prevents one CPU to accumulate too many dirtied pages without
1940 * calling into balance_dirty_pages(), which can happen when there are
1941 * 1000+ tasks, all of them start dirtying pages at exactly the same
1942 * time, hence all honoured too large initial task->nr_dirtied_pause.
1944 p = this_cpu_ptr(&bdp_ratelimits);
1945 if (unlikely(current->nr_dirtied >= ratelimit))
1947 else if (unlikely(*p >= ratelimit_pages)) {
1952 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1953 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1954 * the dirty throttling and livelock other long-run dirtiers.
1956 p = this_cpu_ptr(&dirty_throttle_leaks);
1957 if (*p > 0 && current->nr_dirtied < ratelimit) {
1958 unsigned long nr_pages_dirtied;
1959 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1960 *p -= nr_pages_dirtied;
1961 current->nr_dirtied += nr_pages_dirtied;
1965 if (unlikely(current->nr_dirtied >= ratelimit))
1966 ret = balance_dirty_pages(wb, current->nr_dirtied, flags);
1971 EXPORT_SYMBOL_GPL(balance_dirty_pages_ratelimited_flags);
1974 * balance_dirty_pages_ratelimited - balance dirty memory state.
1975 * @mapping: address_space which was dirtied.
1977 * Processes which are dirtying memory should call in here once for each page
1978 * which was newly dirtied. The function will periodically check the system's
1979 * dirty state and will initiate writeback if needed.
1981 * Once we're over the dirty memory limit we decrease the ratelimiting
1982 * by a lot, to prevent individual processes from overshooting the limit
1983 * by (ratelimit_pages) each.
1985 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1987 balance_dirty_pages_ratelimited_flags(mapping, 0);
1989 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1992 * wb_over_bg_thresh - does @wb need to be written back?
1993 * @wb: bdi_writeback of interest
1995 * Determines whether background writeback should keep writing @wb or it's
1998 * Return: %true if writeback should continue.
2000 bool wb_over_bg_thresh(struct bdi_writeback *wb)
2002 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
2003 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
2004 struct dirty_throttle_control * const gdtc = &gdtc_stor;
2005 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
2007 unsigned long reclaimable;
2008 unsigned long thresh;
2011 * Similar to balance_dirty_pages() but ignores pages being written
2012 * as we're trying to decide whether to put more under writeback.
2014 gdtc->avail = global_dirtyable_memory();
2015 gdtc->dirty = global_node_page_state(NR_FILE_DIRTY);
2016 domain_dirty_limits(gdtc);
2018 if (gdtc->dirty > gdtc->bg_thresh)
2021 thresh = wb_calc_thresh(gdtc->wb, gdtc->bg_thresh);
2022 if (thresh < 2 * wb_stat_error())
2023 reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
2025 reclaimable = wb_stat(wb, WB_RECLAIMABLE);
2027 if (reclaimable > thresh)
2031 unsigned long filepages, headroom, writeback;
2033 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
2035 mdtc_calc_avail(mdtc, filepages, headroom);
2036 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
2038 if (mdtc->dirty > mdtc->bg_thresh)
2041 thresh = wb_calc_thresh(mdtc->wb, mdtc->bg_thresh);
2042 if (thresh < 2 * wb_stat_error())
2043 reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
2045 reclaimable = wb_stat(wb, WB_RECLAIMABLE);
2047 if (reclaimable > thresh)
2054 #ifdef CONFIG_SYSCTL
2056 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
2058 static int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
2059 void *buffer, size_t *length, loff_t *ppos)
2061 unsigned int old_interval = dirty_writeback_interval;
2064 ret = proc_dointvec(table, write, buffer, length, ppos);
2067 * Writing 0 to dirty_writeback_interval will disable periodic writeback
2068 * and a different non-zero value will wakeup the writeback threads.
2069 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
2070 * iterate over all bdis and wbs.
2071 * The reason we do this is to make the change take effect immediately.
2073 if (!ret && write && dirty_writeback_interval &&
2074 dirty_writeback_interval != old_interval)
2075 wakeup_flusher_threads(WB_REASON_PERIODIC);
2081 void laptop_mode_timer_fn(struct timer_list *t)
2083 struct backing_dev_info *backing_dev_info =
2084 from_timer(backing_dev_info, t, laptop_mode_wb_timer);
2086 wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
2090 * We've spun up the disk and we're in laptop mode: schedule writeback
2091 * of all dirty data a few seconds from now. If the flush is already scheduled
2092 * then push it back - the user is still using the disk.
2094 void laptop_io_completion(struct backing_dev_info *info)
2096 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2100 * We're in laptop mode and we've just synced. The sync's writes will have
2101 * caused another writeback to be scheduled by laptop_io_completion.
2102 * Nothing needs to be written back anymore, so we unschedule the writeback.
2104 void laptop_sync_completion(void)
2106 struct backing_dev_info *bdi;
2110 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2111 del_timer(&bdi->laptop_mode_wb_timer);
2117 * If ratelimit_pages is too high then we can get into dirty-data overload
2118 * if a large number of processes all perform writes at the same time.
2120 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2121 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2125 void writeback_set_ratelimit(void)
2127 struct wb_domain *dom = &global_wb_domain;
2128 unsigned long background_thresh;
2129 unsigned long dirty_thresh;
2131 global_dirty_limits(&background_thresh, &dirty_thresh);
2132 dom->dirty_limit = dirty_thresh;
2133 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2134 if (ratelimit_pages < 16)
2135 ratelimit_pages = 16;
2138 static int page_writeback_cpu_online(unsigned int cpu)
2140 writeback_set_ratelimit();
2144 #ifdef CONFIG_SYSCTL
2146 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
2147 static const unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
2149 static struct ctl_table vm_page_writeback_sysctls[] = {
2151 .procname = "dirty_background_ratio",
2152 .data = &dirty_background_ratio,
2153 .maxlen = sizeof(dirty_background_ratio),
2155 .proc_handler = dirty_background_ratio_handler,
2156 .extra1 = SYSCTL_ZERO,
2157 .extra2 = SYSCTL_ONE_HUNDRED,
2160 .procname = "dirty_background_bytes",
2161 .data = &dirty_background_bytes,
2162 .maxlen = sizeof(dirty_background_bytes),
2164 .proc_handler = dirty_background_bytes_handler,
2165 .extra1 = SYSCTL_LONG_ONE,
2168 .procname = "dirty_ratio",
2169 .data = &vm_dirty_ratio,
2170 .maxlen = sizeof(vm_dirty_ratio),
2172 .proc_handler = dirty_ratio_handler,
2173 .extra1 = SYSCTL_ZERO,
2174 .extra2 = SYSCTL_ONE_HUNDRED,
2177 .procname = "dirty_bytes",
2178 .data = &vm_dirty_bytes,
2179 .maxlen = sizeof(vm_dirty_bytes),
2181 .proc_handler = dirty_bytes_handler,
2182 .extra1 = (void *)&dirty_bytes_min,
2185 .procname = "dirty_writeback_centisecs",
2186 .data = &dirty_writeback_interval,
2187 .maxlen = sizeof(dirty_writeback_interval),
2189 .proc_handler = dirty_writeback_centisecs_handler,
2192 .procname = "dirty_expire_centisecs",
2193 .data = &dirty_expire_interval,
2194 .maxlen = sizeof(dirty_expire_interval),
2196 .proc_handler = proc_dointvec_minmax,
2197 .extra1 = SYSCTL_ZERO,
2199 #ifdef CONFIG_HIGHMEM
2201 .procname = "highmem_is_dirtyable",
2202 .data = &vm_highmem_is_dirtyable,
2203 .maxlen = sizeof(vm_highmem_is_dirtyable),
2205 .proc_handler = proc_dointvec_minmax,
2206 .extra1 = SYSCTL_ZERO,
2207 .extra2 = SYSCTL_ONE,
2211 .procname = "laptop_mode",
2212 .data = &laptop_mode,
2213 .maxlen = sizeof(laptop_mode),
2215 .proc_handler = proc_dointvec_jiffies,
2222 * Called early on to tune the page writeback dirty limits.
2224 * We used to scale dirty pages according to how total memory
2225 * related to pages that could be allocated for buffers.
2227 * However, that was when we used "dirty_ratio" to scale with
2228 * all memory, and we don't do that any more. "dirty_ratio"
2229 * is now applied to total non-HIGHPAGE memory, and as such we can't
2230 * get into the old insane situation any more where we had
2231 * large amounts of dirty pages compared to a small amount of
2232 * non-HIGHMEM memory.
2234 * But we might still want to scale the dirty_ratio by how
2235 * much memory the box has..
2237 void __init page_writeback_init(void)
2239 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2241 cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2242 page_writeback_cpu_online, NULL);
2243 cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2244 page_writeback_cpu_online);
2245 #ifdef CONFIG_SYSCTL
2246 register_sysctl_init("vm", vm_page_writeback_sysctls);
2251 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2252 * @mapping: address space structure to write
2253 * @start: starting page index
2254 * @end: ending page index (inclusive)
2256 * This function scans the page range from @start to @end (inclusive) and tags
2257 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2258 * that write_cache_pages (or whoever calls this function) will then use
2259 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2260 * used to avoid livelocking of writeback by a process steadily creating new
2261 * dirty pages in the file (thus it is important for this function to be quick
2262 * so that it can tag pages faster than a dirtying process can create them).
2264 void tag_pages_for_writeback(struct address_space *mapping,
2265 pgoff_t start, pgoff_t end)
2267 XA_STATE(xas, &mapping->i_pages, start);
2268 unsigned int tagged = 0;
2272 xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2273 xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2274 if (++tagged % XA_CHECK_SCHED)
2278 xas_unlock_irq(&xas);
2282 xas_unlock_irq(&xas);
2284 EXPORT_SYMBOL(tag_pages_for_writeback);
2287 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2288 * @mapping: address space structure to write
2289 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2290 * @writepage: function called for each page
2291 * @data: data passed to writepage function
2293 * If a page is already under I/O, write_cache_pages() skips it, even
2294 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2295 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2296 * and msync() need to guarantee that all the data which was dirty at the time
2297 * the call was made get new I/O started against them. If wbc->sync_mode is
2298 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2299 * existing IO to complete.
2301 * To avoid livelocks (when other process dirties new pages), we first tag
2302 * pages which should be written back with TOWRITE tag and only then start
2303 * writing them. For data-integrity sync we have to be careful so that we do
2304 * not miss some pages (e.g., because some other process has cleared TOWRITE
2305 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2306 * by the process clearing the DIRTY tag (and submitting the page for IO).
2308 * To avoid deadlocks between range_cyclic writeback and callers that hold
2309 * pages in PageWriteback to aggregate IO until write_cache_pages() returns,
2310 * we do not loop back to the start of the file. Doing so causes a page
2311 * lock/page writeback access order inversion - we should only ever lock
2312 * multiple pages in ascending page->index order, and looping back to the start
2313 * of the file violates that rule and causes deadlocks.
2315 * Return: %0 on success, negative error code otherwise
2317 int write_cache_pages(struct address_space *mapping,
2318 struct writeback_control *wbc, writepage_t writepage,
2324 struct pagevec pvec;
2327 pgoff_t end; /* Inclusive */
2329 int range_whole = 0;
2332 pagevec_init(&pvec);
2333 if (wbc->range_cyclic) {
2334 index = mapping->writeback_index; /* prev offset */
2337 index = wbc->range_start >> PAGE_SHIFT;
2338 end = wbc->range_end >> PAGE_SHIFT;
2339 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2342 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
2343 tag_pages_for_writeback(mapping, index, end);
2344 tag = PAGECACHE_TAG_TOWRITE;
2346 tag = PAGECACHE_TAG_DIRTY;
2349 while (!done && (index <= end)) {
2352 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2357 for (i = 0; i < nr_pages; i++) {
2358 struct page *page = pvec.pages[i];
2360 done_index = page->index;
2365 * Page truncated or invalidated. We can freely skip it
2366 * then, even for data integrity operations: the page
2367 * has disappeared concurrently, so there could be no
2368 * real expectation of this data integrity operation
2369 * even if there is now a new, dirty page at the same
2370 * pagecache address.
2372 if (unlikely(page->mapping != mapping)) {
2378 if (!PageDirty(page)) {
2379 /* someone wrote it for us */
2380 goto continue_unlock;
2383 if (PageWriteback(page)) {
2384 if (wbc->sync_mode != WB_SYNC_NONE)
2385 wait_on_page_writeback(page);
2387 goto continue_unlock;
2390 BUG_ON(PageWriteback(page));
2391 if (!clear_page_dirty_for_io(page))
2392 goto continue_unlock;
2394 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2395 error = (*writepage)(page, wbc, data);
2396 if (unlikely(error)) {
2398 * Handle errors according to the type of
2399 * writeback. There's no need to continue for
2400 * background writeback. Just push done_index
2401 * past this page so media errors won't choke
2402 * writeout for the entire file. For integrity
2403 * writeback, we must process the entire dirty
2404 * set regardless of errors because the fs may
2405 * still have state to clear for each page. In
2406 * that case we continue processing and return
2409 if (error == AOP_WRITEPAGE_ACTIVATE) {
2412 } else if (wbc->sync_mode != WB_SYNC_ALL) {
2414 done_index = page->index + 1;
2423 * We stop writing back only if we are not doing
2424 * integrity sync. In case of integrity sync we have to
2425 * keep going until we have written all the pages
2426 * we tagged for writeback prior to entering this loop.
2428 if (--wbc->nr_to_write <= 0 &&
2429 wbc->sync_mode == WB_SYNC_NONE) {
2434 pagevec_release(&pvec);
2439 * If we hit the last page and there is more work to be done: wrap
2440 * back the index back to the start of the file for the next
2441 * time we are called.
2443 if (wbc->range_cyclic && !done)
2445 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2446 mapping->writeback_index = done_index;
2450 EXPORT_SYMBOL(write_cache_pages);
2453 * Function used by generic_writepages to call the real writepage
2454 * function and set the mapping flags on error
2456 static int __writepage(struct page *page, struct writeback_control *wbc,
2459 struct address_space *mapping = data;
2460 int ret = mapping->a_ops->writepage(page, wbc);
2461 mapping_set_error(mapping, ret);
2466 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2467 * @mapping: address space structure to write
2468 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2470 * This is a library function, which implements the writepages()
2471 * address_space_operation.
2473 * Return: %0 on success, negative error code otherwise
2475 int generic_writepages(struct address_space *mapping,
2476 struct writeback_control *wbc)
2478 struct blk_plug plug;
2481 /* deal with chardevs and other special file */
2482 if (!mapping->a_ops->writepage)
2485 blk_start_plug(&plug);
2486 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2487 blk_finish_plug(&plug);
2491 EXPORT_SYMBOL(generic_writepages);
2493 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2496 struct bdi_writeback *wb;
2498 if (wbc->nr_to_write <= 0)
2500 wb = inode_to_wb_wbc(mapping->host, wbc);
2501 wb_bandwidth_estimate_start(wb);
2503 if (mapping->a_ops->writepages)
2504 ret = mapping->a_ops->writepages(mapping, wbc);
2506 ret = generic_writepages(mapping, wbc);
2507 if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2511 * Lacking an allocation context or the locality or writeback
2512 * state of any of the inode's pages, throttle based on
2513 * writeback activity on the local node. It's as good a
2516 reclaim_throttle(NODE_DATA(numa_node_id()),
2517 VMSCAN_THROTTLE_WRITEBACK);
2520 * Usually few pages are written by now from those we've just submitted
2521 * but if there's constant writeback being submitted, this makes sure
2522 * writeback bandwidth is updated once in a while.
2524 if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
2525 BANDWIDTH_INTERVAL))
2526 wb_update_bandwidth(wb);
2531 * folio_write_one - write out a single folio and wait on I/O.
2532 * @folio: The folio to write.
2534 * The folio must be locked by the caller and will be unlocked upon return.
2536 * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2539 * Return: %0 on success, negative error code otherwise
2541 int folio_write_one(struct folio *folio)
2543 struct address_space *mapping = folio->mapping;
2545 struct writeback_control wbc = {
2546 .sync_mode = WB_SYNC_ALL,
2547 .nr_to_write = folio_nr_pages(folio),
2550 BUG_ON(!folio_test_locked(folio));
2552 folio_wait_writeback(folio);
2554 if (folio_clear_dirty_for_io(folio)) {
2556 ret = mapping->a_ops->writepage(&folio->page, &wbc);
2558 folio_wait_writeback(folio);
2561 folio_unlock(folio);
2565 ret = filemap_check_errors(mapping);
2568 EXPORT_SYMBOL(folio_write_one);
2571 * For address_spaces which do not use buffers nor write back.
2573 bool noop_dirty_folio(struct address_space *mapping, struct folio *folio)
2575 if (!folio_test_dirty(folio))
2576 return !folio_test_set_dirty(folio);
2579 EXPORT_SYMBOL(noop_dirty_folio);
2582 * Helper function for set_page_dirty family.
2584 * Caller must hold lock_page_memcg().
2586 * NOTE: This relies on being atomic wrt interrupts.
2588 static void folio_account_dirtied(struct folio *folio,
2589 struct address_space *mapping)
2591 struct inode *inode = mapping->host;
2593 trace_writeback_dirty_folio(folio, mapping);
2595 if (mapping_can_writeback(mapping)) {
2596 struct bdi_writeback *wb;
2597 long nr = folio_nr_pages(folio);
2599 inode_attach_wb(inode, &folio->page);
2600 wb = inode_to_wb(inode);
2602 __lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, nr);
2603 __zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
2604 __node_stat_mod_folio(folio, NR_DIRTIED, nr);
2605 wb_stat_mod(wb, WB_RECLAIMABLE, nr);
2606 wb_stat_mod(wb, WB_DIRTIED, nr);
2607 task_io_account_write(nr * PAGE_SIZE);
2608 current->nr_dirtied += nr;
2609 __this_cpu_add(bdp_ratelimits, nr);
2611 mem_cgroup_track_foreign_dirty(folio, wb);
2616 * Helper function for deaccounting dirty page without writeback.
2618 * Caller must hold lock_page_memcg().
2620 void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb)
2622 long nr = folio_nr_pages(folio);
2624 lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2625 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2626 wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2627 task_io_account_cancelled_write(nr * PAGE_SIZE);
2631 * Mark the folio dirty, and set it dirty in the page cache, and mark
2634 * If warn is true, then emit a warning if the folio is not uptodate and has
2635 * not been truncated.
2637 * The caller must hold lock_page_memcg(). Most callers have the folio
2638 * locked. A few have the folio blocked from truncation through other
2639 * means (eg zap_page_range() has it mapped and is holding the page table
2640 * lock). This can also be called from mark_buffer_dirty(), which I
2641 * cannot prove is always protected against truncate.
2643 void __folio_mark_dirty(struct folio *folio, struct address_space *mapping,
2646 unsigned long flags;
2648 xa_lock_irqsave(&mapping->i_pages, flags);
2649 if (folio->mapping) { /* Race with truncate? */
2650 WARN_ON_ONCE(warn && !folio_test_uptodate(folio));
2651 folio_account_dirtied(folio, mapping);
2652 __xa_set_mark(&mapping->i_pages, folio_index(folio),
2653 PAGECACHE_TAG_DIRTY);
2655 xa_unlock_irqrestore(&mapping->i_pages, flags);
2659 * filemap_dirty_folio - Mark a folio dirty for filesystems which do not use buffer_heads.
2660 * @mapping: Address space this folio belongs to.
2661 * @folio: Folio to be marked as dirty.
2663 * Filesystems which do not use buffer heads should call this function
2664 * from their set_page_dirty address space operation. It ignores the
2665 * contents of folio_get_private(), so if the filesystem marks individual
2666 * blocks as dirty, the filesystem should handle that itself.
2668 * This is also sometimes used by filesystems which use buffer_heads when
2669 * a single buffer is being dirtied: we want to set the folio dirty in
2670 * that case, but not all the buffers. This is a "bottom-up" dirtying,
2671 * whereas block_dirty_folio() is a "top-down" dirtying.
2673 * The caller must ensure this doesn't race with truncation. Most will
2674 * simply hold the folio lock, but e.g. zap_pte_range() calls with the
2675 * folio mapped and the pte lock held, which also locks out truncation.
2677 bool filemap_dirty_folio(struct address_space *mapping, struct folio *folio)
2679 folio_memcg_lock(folio);
2680 if (folio_test_set_dirty(folio)) {
2681 folio_memcg_unlock(folio);
2685 __folio_mark_dirty(folio, mapping, !folio_test_private(folio));
2686 folio_memcg_unlock(folio);
2688 if (mapping->host) {
2689 /* !PageAnon && !swapper_space */
2690 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2694 EXPORT_SYMBOL(filemap_dirty_folio);
2697 * folio_account_redirty - Manually account for redirtying a page.
2698 * @folio: The folio which is being redirtied.
2700 * Most filesystems should call folio_redirty_for_writepage() instead
2701 * of this fuction. If your filesystem is doing writeback outside the
2702 * context of a writeback_control(), it can call this when redirtying
2703 * a folio, to de-account the dirty counters (NR_DIRTIED, WB_DIRTIED,
2704 * tsk->nr_dirtied), so that they match the written counters (NR_WRITTEN,
2705 * WB_WRITTEN) in long term. The mismatches will lead to systematic errors
2706 * in balanced_dirty_ratelimit and the dirty pages position control.
2708 void folio_account_redirty(struct folio *folio)
2710 struct address_space *mapping = folio->mapping;
2712 if (mapping && mapping_can_writeback(mapping)) {
2713 struct inode *inode = mapping->host;
2714 struct bdi_writeback *wb;
2715 struct wb_lock_cookie cookie = {};
2716 long nr = folio_nr_pages(folio);
2718 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2719 current->nr_dirtied -= nr;
2720 node_stat_mod_folio(folio, NR_DIRTIED, -nr);
2721 wb_stat_mod(wb, WB_DIRTIED, -nr);
2722 unlocked_inode_to_wb_end(inode, &cookie);
2725 EXPORT_SYMBOL(folio_account_redirty);
2728 * folio_redirty_for_writepage - Decline to write a dirty folio.
2729 * @wbc: The writeback control.
2730 * @folio: The folio.
2732 * When a writepage implementation decides that it doesn't want to write
2733 * @folio for some reason, it should call this function, unlock @folio and
2736 * Return: True if we redirtied the folio. False if someone else dirtied
2739 bool folio_redirty_for_writepage(struct writeback_control *wbc,
2740 struct folio *folio)
2743 long nr = folio_nr_pages(folio);
2745 wbc->pages_skipped += nr;
2746 ret = filemap_dirty_folio(folio->mapping, folio);
2747 folio_account_redirty(folio);
2751 EXPORT_SYMBOL(folio_redirty_for_writepage);
2754 * folio_mark_dirty - Mark a folio as being modified.
2755 * @folio: The folio.
2757 * The folio may not be truncated while this function is running.
2758 * Holding the folio lock is sufficient to prevent truncation, but some
2759 * callers cannot acquire a sleeping lock. These callers instead hold
2760 * the page table lock for a page table which contains at least one page
2761 * in this folio. Truncation will block on the page table lock as it
2762 * unmaps pages before removing the folio from its mapping.
2764 * Return: True if the folio was newly dirtied, false if it was already dirty.
2766 bool folio_mark_dirty(struct folio *folio)
2768 struct address_space *mapping = folio_mapping(folio);
2770 if (likely(mapping)) {
2772 * readahead/lru_deactivate_page could remain
2773 * PG_readahead/PG_reclaim due to race with folio_end_writeback
2774 * About readahead, if the folio is written, the flags would be
2775 * reset. So no problem.
2776 * About lru_deactivate_page, if the folio is redirtied,
2777 * the flag will be reset. So no problem. but if the
2778 * folio is used by readahead it will confuse readahead
2779 * and make it restart the size rampup process. But it's
2780 * a trivial problem.
2782 if (folio_test_reclaim(folio))
2783 folio_clear_reclaim(folio);
2784 return mapping->a_ops->dirty_folio(mapping, folio);
2787 return noop_dirty_folio(mapping, folio);
2789 EXPORT_SYMBOL(folio_mark_dirty);
2792 * set_page_dirty() is racy if the caller has no reference against
2793 * page->mapping->host, and if the page is unlocked. This is because another
2794 * CPU could truncate the page off the mapping and then free the mapping.
2796 * Usually, the page _is_ locked, or the caller is a user-space process which
2797 * holds a reference on the inode by having an open file.
2799 * In other cases, the page should be locked before running set_page_dirty().
2801 int set_page_dirty_lock(struct page *page)
2806 ret = set_page_dirty(page);
2810 EXPORT_SYMBOL(set_page_dirty_lock);
2813 * This cancels just the dirty bit on the kernel page itself, it does NOT
2814 * actually remove dirty bits on any mmap's that may be around. It also
2815 * leaves the page tagged dirty, so any sync activity will still find it on
2816 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2817 * look at the dirty bits in the VM.
2819 * Doing this should *normally* only ever be done when a page is truncated,
2820 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2821 * this when it notices that somebody has cleaned out all the buffers on a
2822 * page without actually doing it through the VM. Can you say "ext3 is
2823 * horribly ugly"? Thought you could.
2825 void __folio_cancel_dirty(struct folio *folio)
2827 struct address_space *mapping = folio_mapping(folio);
2829 if (mapping_can_writeback(mapping)) {
2830 struct inode *inode = mapping->host;
2831 struct bdi_writeback *wb;
2832 struct wb_lock_cookie cookie = {};
2834 folio_memcg_lock(folio);
2835 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2837 if (folio_test_clear_dirty(folio))
2838 folio_account_cleaned(folio, wb);
2840 unlocked_inode_to_wb_end(inode, &cookie);
2841 folio_memcg_unlock(folio);
2843 folio_clear_dirty(folio);
2846 EXPORT_SYMBOL(__folio_cancel_dirty);
2849 * Clear a folio's dirty flag, while caring for dirty memory accounting.
2850 * Returns true if the folio was previously dirty.
2852 * This is for preparing to put the folio under writeout. We leave
2853 * the folio tagged as dirty in the xarray so that a concurrent
2854 * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk.
2855 * The ->writepage implementation will run either folio_start_writeback()
2856 * or folio_mark_dirty(), at which stage we bring the folio's dirty flag
2857 * and xarray dirty tag back into sync.
2859 * This incoherency between the folio's dirty flag and xarray tag is
2860 * unfortunate, but it only exists while the folio is locked.
2862 bool folio_clear_dirty_for_io(struct folio *folio)
2864 struct address_space *mapping = folio_mapping(folio);
2867 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2869 if (mapping && mapping_can_writeback(mapping)) {
2870 struct inode *inode = mapping->host;
2871 struct bdi_writeback *wb;
2872 struct wb_lock_cookie cookie = {};
2875 * Yes, Virginia, this is indeed insane.
2877 * We use this sequence to make sure that
2878 * (a) we account for dirty stats properly
2879 * (b) we tell the low-level filesystem to
2880 * mark the whole folio dirty if it was
2881 * dirty in a pagetable. Only to then
2882 * (c) clean the folio again and return 1 to
2883 * cause the writeback.
2885 * This way we avoid all nasty races with the
2886 * dirty bit in multiple places and clearing
2887 * them concurrently from different threads.
2889 * Note! Normally the "folio_mark_dirty(folio)"
2890 * has no effect on the actual dirty bit - since
2891 * that will already usually be set. But we
2892 * need the side effects, and it can help us
2895 * We basically use the folio "master dirty bit"
2896 * as a serialization point for all the different
2897 * threads doing their things.
2899 if (folio_mkclean(folio))
2900 folio_mark_dirty(folio);
2902 * We carefully synchronise fault handlers against
2903 * installing a dirty pte and marking the folio dirty
2904 * at this point. We do this by having them hold the
2905 * page lock while dirtying the folio, and folios are
2906 * always locked coming in here, so we get the desired
2909 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2910 if (folio_test_clear_dirty(folio)) {
2911 long nr = folio_nr_pages(folio);
2912 lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2913 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2914 wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2917 unlocked_inode_to_wb_end(inode, &cookie);
2920 return folio_test_clear_dirty(folio);
2922 EXPORT_SYMBOL(folio_clear_dirty_for_io);
2924 static void wb_inode_writeback_start(struct bdi_writeback *wb)
2926 atomic_inc(&wb->writeback_inodes);
2929 static void wb_inode_writeback_end(struct bdi_writeback *wb)
2931 unsigned long flags;
2932 atomic_dec(&wb->writeback_inodes);
2934 * Make sure estimate of writeback throughput gets updated after
2935 * writeback completed. We delay the update by BANDWIDTH_INTERVAL
2936 * (which is the interval other bandwidth updates use for batching) so
2937 * that if multiple inodes end writeback at a similar time, they get
2938 * batched into one bandwidth update.
2940 spin_lock_irqsave(&wb->work_lock, flags);
2941 if (test_bit(WB_registered, &wb->state))
2942 queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL);
2943 spin_unlock_irqrestore(&wb->work_lock, flags);
2946 bool __folio_end_writeback(struct folio *folio)
2948 long nr = folio_nr_pages(folio);
2949 struct address_space *mapping = folio_mapping(folio);
2952 folio_memcg_lock(folio);
2953 if (mapping && mapping_use_writeback_tags(mapping)) {
2954 struct inode *inode = mapping->host;
2955 struct backing_dev_info *bdi = inode_to_bdi(inode);
2956 unsigned long flags;
2958 xa_lock_irqsave(&mapping->i_pages, flags);
2959 ret = folio_test_clear_writeback(folio);
2961 __xa_clear_mark(&mapping->i_pages, folio_index(folio),
2962 PAGECACHE_TAG_WRITEBACK);
2963 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
2964 struct bdi_writeback *wb = inode_to_wb(inode);
2966 wb_stat_mod(wb, WB_WRITEBACK, -nr);
2967 __wb_writeout_add(wb, nr);
2968 if (!mapping_tagged(mapping,
2969 PAGECACHE_TAG_WRITEBACK))
2970 wb_inode_writeback_end(wb);
2974 if (mapping->host && !mapping_tagged(mapping,
2975 PAGECACHE_TAG_WRITEBACK))
2976 sb_clear_inode_writeback(mapping->host);
2978 xa_unlock_irqrestore(&mapping->i_pages, flags);
2980 ret = folio_test_clear_writeback(folio);
2983 lruvec_stat_mod_folio(folio, NR_WRITEBACK, -nr);
2984 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2985 node_stat_mod_folio(folio, NR_WRITTEN, nr);
2987 folio_memcg_unlock(folio);
2991 bool __folio_start_writeback(struct folio *folio, bool keep_write)
2993 long nr = folio_nr_pages(folio);
2994 struct address_space *mapping = folio_mapping(folio);
2998 folio_memcg_lock(folio);
2999 if (mapping && mapping_use_writeback_tags(mapping)) {
3000 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
3001 struct inode *inode = mapping->host;
3002 struct backing_dev_info *bdi = inode_to_bdi(inode);
3003 unsigned long flags;
3005 xas_lock_irqsave(&xas, flags);
3007 ret = folio_test_set_writeback(folio);
3011 on_wblist = mapping_tagged(mapping,
3012 PAGECACHE_TAG_WRITEBACK);
3014 xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
3015 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
3016 struct bdi_writeback *wb = inode_to_wb(inode);
3018 wb_stat_mod(wb, WB_WRITEBACK, nr);
3020 wb_inode_writeback_start(wb);
3024 * We can come through here when swapping
3025 * anonymous folios, so we don't necessarily
3026 * have an inode to track for sync.
3028 if (mapping->host && !on_wblist)
3029 sb_mark_inode_writeback(mapping->host);
3031 if (!folio_test_dirty(folio))
3032 xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
3034 xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
3035 xas_unlock_irqrestore(&xas, flags);
3037 ret = folio_test_set_writeback(folio);
3040 lruvec_stat_mod_folio(folio, NR_WRITEBACK, nr);
3041 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
3043 folio_memcg_unlock(folio);
3044 access_ret = arch_make_folio_accessible(folio);
3046 * If writeback has been triggered on a page that cannot be made
3047 * accessible, it is too late to recover here.
3049 VM_BUG_ON_FOLIO(access_ret != 0, folio);
3053 EXPORT_SYMBOL(__folio_start_writeback);
3056 * folio_wait_writeback - Wait for a folio to finish writeback.
3057 * @folio: The folio to wait for.
3059 * If the folio is currently being written back to storage, wait for the
3062 * Context: Sleeps. Must be called in process context and with
3063 * no spinlocks held. Caller should hold a reference on the folio.
3064 * If the folio is not locked, writeback may start again after writeback
3067 void folio_wait_writeback(struct folio *folio)
3069 while (folio_test_writeback(folio)) {
3070 trace_folio_wait_writeback(folio, folio_mapping(folio));
3071 folio_wait_bit(folio, PG_writeback);
3074 EXPORT_SYMBOL_GPL(folio_wait_writeback);
3077 * folio_wait_writeback_killable - Wait for a folio to finish writeback.
3078 * @folio: The folio to wait for.
3080 * If the folio is currently being written back to storage, wait for the
3081 * I/O to complete or a fatal signal to arrive.
3083 * Context: Sleeps. Must be called in process context and with
3084 * no spinlocks held. Caller should hold a reference on the folio.
3085 * If the folio is not locked, writeback may start again after writeback
3087 * Return: 0 on success, -EINTR if we get a fatal signal while waiting.
3089 int folio_wait_writeback_killable(struct folio *folio)
3091 while (folio_test_writeback(folio)) {
3092 trace_folio_wait_writeback(folio, folio_mapping(folio));
3093 if (folio_wait_bit_killable(folio, PG_writeback))
3099 EXPORT_SYMBOL_GPL(folio_wait_writeback_killable);
3102 * folio_wait_stable() - wait for writeback to finish, if necessary.
3103 * @folio: The folio to wait on.
3105 * This function determines if the given folio is related to a backing
3106 * device that requires folio contents to be held stable during writeback.
3107 * If so, then it will wait for any pending writeback to complete.
3109 * Context: Sleeps. Must be called in process context and with
3110 * no spinlocks held. Caller should hold a reference on the folio.
3111 * If the folio is not locked, writeback may start again after writeback
3114 void folio_wait_stable(struct folio *folio)
3116 if (folio_inode(folio)->i_sb->s_iflags & SB_I_STABLE_WRITES)
3117 folio_wait_writeback(folio);
3119 EXPORT_SYMBOL_GPL(folio_wait_stable);