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 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 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 int vm_highmem_is_dirtyable;
88 * The generator of dirty data starts writeback at this percentage
90 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 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);
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;
494 int dirty_background_ratio_handler(struct ctl_table *table, int write,
495 void *buffer, size_t *lenp, loff_t *ppos)
499 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
500 if (ret == 0 && write)
501 dirty_background_bytes = 0;
505 int dirty_background_bytes_handler(struct ctl_table *table, int write,
506 void *buffer, size_t *lenp, loff_t *ppos)
510 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
511 if (ret == 0 && write)
512 dirty_background_ratio = 0;
516 int dirty_ratio_handler(struct ctl_table *table, int write, void *buffer,
517 size_t *lenp, loff_t *ppos)
519 int old_ratio = vm_dirty_ratio;
522 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
523 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
524 writeback_set_ratelimit();
530 int dirty_bytes_handler(struct ctl_table *table, int write,
531 void *buffer, size_t *lenp, loff_t *ppos)
533 unsigned long old_bytes = vm_dirty_bytes;
536 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
537 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
538 writeback_set_ratelimit();
544 static unsigned long wp_next_time(unsigned long cur_time)
546 cur_time += VM_COMPLETIONS_PERIOD_LEN;
547 /* 0 has a special meaning... */
553 static void wb_domain_writeout_add(struct wb_domain *dom,
554 struct fprop_local_percpu *completions,
555 unsigned int max_prop_frac, long nr)
557 __fprop_add_percpu_max(&dom->completions, completions,
559 /* First event after period switching was turned off? */
560 if (unlikely(!dom->period_time)) {
562 * We can race with other __bdi_writeout_inc calls here but
563 * it does not cause any harm since the resulting time when
564 * timer will fire and what is in writeout_period_time will be
567 dom->period_time = wp_next_time(jiffies);
568 mod_timer(&dom->period_timer, dom->period_time);
573 * Increment @wb's writeout completion count and the global writeout
574 * completion count. Called from __folio_end_writeback().
576 static inline void __wb_writeout_add(struct bdi_writeback *wb, long nr)
578 struct wb_domain *cgdom;
580 wb_stat_mod(wb, WB_WRITTEN, nr);
581 wb_domain_writeout_add(&global_wb_domain, &wb->completions,
582 wb->bdi->max_prop_frac, nr);
584 cgdom = mem_cgroup_wb_domain(wb);
586 wb_domain_writeout_add(cgdom, wb_memcg_completions(wb),
587 wb->bdi->max_prop_frac, nr);
590 void wb_writeout_inc(struct bdi_writeback *wb)
594 local_irq_save(flags);
595 __wb_writeout_add(wb, 1);
596 local_irq_restore(flags);
598 EXPORT_SYMBOL_GPL(wb_writeout_inc);
601 * On idle system, we can be called long after we scheduled because we use
602 * deferred timers so count with missed periods.
604 static void writeout_period(struct timer_list *t)
606 struct wb_domain *dom = from_timer(dom, t, period_timer);
607 int miss_periods = (jiffies - dom->period_time) /
608 VM_COMPLETIONS_PERIOD_LEN;
610 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
611 dom->period_time = wp_next_time(dom->period_time +
612 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
613 mod_timer(&dom->period_timer, dom->period_time);
616 * Aging has zeroed all fractions. Stop wasting CPU on period
619 dom->period_time = 0;
623 int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
625 memset(dom, 0, sizeof(*dom));
627 spin_lock_init(&dom->lock);
629 timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
631 dom->dirty_limit_tstamp = jiffies;
633 return fprop_global_init(&dom->completions, gfp);
636 #ifdef CONFIG_CGROUP_WRITEBACK
637 void wb_domain_exit(struct wb_domain *dom)
639 del_timer_sync(&dom->period_timer);
640 fprop_global_destroy(&dom->completions);
645 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
646 * registered backing devices, which, for obvious reasons, can not
649 static unsigned int bdi_min_ratio;
651 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
656 spin_lock_bh(&bdi_lock);
657 if (min_ratio > bdi->max_ratio) {
660 if (min_ratio < bdi->min_ratio) {
661 delta = bdi->min_ratio - min_ratio;
662 bdi_min_ratio -= delta;
663 bdi->min_ratio = min_ratio;
665 delta = min_ratio - bdi->min_ratio;
666 if (bdi_min_ratio + delta < 100) {
667 bdi_min_ratio += delta;
668 bdi->min_ratio = min_ratio;
674 spin_unlock_bh(&bdi_lock);
679 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
686 spin_lock_bh(&bdi_lock);
687 if (bdi->min_ratio > max_ratio) {
690 bdi->max_ratio = max_ratio;
691 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
693 spin_unlock_bh(&bdi_lock);
697 EXPORT_SYMBOL(bdi_set_max_ratio);
699 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
700 unsigned long bg_thresh)
702 return (thresh + bg_thresh) / 2;
705 static unsigned long hard_dirty_limit(struct wb_domain *dom,
706 unsigned long thresh)
708 return max(thresh, dom->dirty_limit);
712 * Memory which can be further allocated to a memcg domain is capped by
713 * system-wide clean memory excluding the amount being used in the domain.
715 static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
716 unsigned long filepages, unsigned long headroom)
718 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
719 unsigned long clean = filepages - min(filepages, mdtc->dirty);
720 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
721 unsigned long other_clean = global_clean - min(global_clean, clean);
723 mdtc->avail = filepages + min(headroom, other_clean);
727 * __wb_calc_thresh - @wb's share of dirty throttling threshold
728 * @dtc: dirty_throttle_context of interest
730 * Note that balance_dirty_pages() will only seriously take it as a hard limit
731 * when sleeping max_pause per page is not enough to keep the dirty pages under
732 * control. For example, when the device is completely stalled due to some error
733 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
734 * In the other normal situations, it acts more gently by throttling the tasks
735 * more (rather than completely block them) when the wb dirty pages go high.
737 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
738 * - starving fast devices
739 * - piling up dirty pages (that will take long time to sync) on slow devices
741 * The wb's share of dirty limit will be adapting to its throughput and
742 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
744 * Return: @wb's dirty limit in pages. The term "dirty" in the context of
745 * dirty balancing includes all PG_dirty and PG_writeback pages.
747 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
749 struct wb_domain *dom = dtc_dom(dtc);
750 unsigned long thresh = dtc->thresh;
752 unsigned long numerator, denominator;
753 unsigned long wb_min_ratio, wb_max_ratio;
756 * Calculate this BDI's share of the thresh ratio.
758 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
759 &numerator, &denominator);
761 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
762 wb_thresh *= numerator;
763 wb_thresh = div64_ul(wb_thresh, denominator);
765 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
767 wb_thresh += (thresh * wb_min_ratio) / 100;
768 if (wb_thresh > (thresh * wb_max_ratio) / 100)
769 wb_thresh = thresh * wb_max_ratio / 100;
774 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
776 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
778 return __wb_calc_thresh(&gdtc);
783 * f(dirty) := 1.0 + (----------------)
786 * it's a 3rd order polynomial that subjects to
788 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
789 * (2) f(setpoint) = 1.0 => the balance point
790 * (3) f(limit) = 0 => the hard limit
791 * (4) df/dx <= 0 => negative feedback control
792 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
793 * => fast response on large errors; small oscillation near setpoint
795 static long long pos_ratio_polynom(unsigned long setpoint,
802 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
803 (limit - setpoint) | 1);
805 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
806 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
807 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
809 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
813 * Dirty position control.
815 * (o) global/bdi setpoints
817 * We want the dirty pages be balanced around the global/wb setpoints.
818 * When the number of dirty pages is higher/lower than the setpoint, the
819 * dirty position control ratio (and hence task dirty ratelimit) will be
820 * decreased/increased to bring the dirty pages back to the setpoint.
822 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
824 * if (dirty < setpoint) scale up pos_ratio
825 * if (dirty > setpoint) scale down pos_ratio
827 * if (wb_dirty < wb_setpoint) scale up pos_ratio
828 * if (wb_dirty > wb_setpoint) scale down pos_ratio
830 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
832 * (o) global control line
836 * | |<===== global dirty control scope ======>|
844 * 1.0 ................................*
850 * 0 +------------.------------------.----------------------*------------->
851 * freerun^ setpoint^ limit^ dirty pages
853 * (o) wb control line
861 * | * |<=========== span ============>|
862 * 1.0 .......................*
874 * 1/4 ...............................................* * * * * * * * * * * *
878 * 0 +----------------------.-------------------------------.------------->
879 * wb_setpoint^ x_intercept^
881 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
882 * be smoothly throttled down to normal if it starts high in situations like
883 * - start writing to a slow SD card and a fast disk at the same time. The SD
884 * card's wb_dirty may rush to many times higher than wb_setpoint.
885 * - the wb dirty thresh drops quickly due to change of JBOD workload
887 static void wb_position_ratio(struct dirty_throttle_control *dtc)
889 struct bdi_writeback *wb = dtc->wb;
890 unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth);
891 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
892 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
893 unsigned long wb_thresh = dtc->wb_thresh;
894 unsigned long x_intercept;
895 unsigned long setpoint; /* dirty pages' target balance point */
896 unsigned long wb_setpoint;
898 long long pos_ratio; /* for scaling up/down the rate limit */
903 if (unlikely(dtc->dirty >= limit))
909 * See comment for pos_ratio_polynom().
911 setpoint = (freerun + limit) / 2;
912 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
915 * The strictlimit feature is a tool preventing mistrusted filesystems
916 * from growing a large number of dirty pages before throttling. For
917 * such filesystems balance_dirty_pages always checks wb counters
918 * against wb limits. Even if global "nr_dirty" is under "freerun".
919 * This is especially important for fuse which sets bdi->max_ratio to
920 * 1% by default. Without strictlimit feature, fuse writeback may
921 * consume arbitrary amount of RAM because it is accounted in
922 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
924 * Here, in wb_position_ratio(), we calculate pos_ratio based on
925 * two values: wb_dirty and wb_thresh. Let's consider an example:
926 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
927 * limits are set by default to 10% and 20% (background and throttle).
928 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
929 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
930 * about ~6K pages (as the average of background and throttle wb
931 * limits). The 3rd order polynomial will provide positive feedback if
932 * wb_dirty is under wb_setpoint and vice versa.
934 * Note, that we cannot use global counters in these calculations
935 * because we want to throttle process writing to a strictlimit wb
936 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
937 * in the example above).
939 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
940 long long wb_pos_ratio;
942 if (dtc->wb_dirty < 8) {
943 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
944 2 << RATELIMIT_CALC_SHIFT);
948 if (dtc->wb_dirty >= wb_thresh)
951 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
954 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
957 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
961 * Typically, for strictlimit case, wb_setpoint << setpoint
962 * and pos_ratio >> wb_pos_ratio. In the other words global
963 * state ("dirty") is not limiting factor and we have to
964 * make decision based on wb counters. But there is an
965 * important case when global pos_ratio should get precedence:
966 * global limits are exceeded (e.g. due to activities on other
967 * wb's) while given strictlimit wb is below limit.
969 * "pos_ratio * wb_pos_ratio" would work for the case above,
970 * but it would look too non-natural for the case of all
971 * activity in the system coming from a single strictlimit wb
972 * with bdi->max_ratio == 100%.
974 * Note that min() below somewhat changes the dynamics of the
975 * control system. Normally, pos_ratio value can be well over 3
976 * (when globally we are at freerun and wb is well below wb
977 * setpoint). Now the maximum pos_ratio in the same situation
978 * is 2. We might want to tweak this if we observe the control
979 * system is too slow to adapt.
981 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
986 * We have computed basic pos_ratio above based on global situation. If
987 * the wb is over/under its share of dirty pages, we want to scale
988 * pos_ratio further down/up. That is done by the following mechanism.
994 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
996 * x_intercept - wb_dirty
997 * := --------------------------
998 * x_intercept - wb_setpoint
1000 * The main wb control line is a linear function that subjects to
1002 * (1) f(wb_setpoint) = 1.0
1003 * (2) k = - 1 / (8 * write_bw) (in single wb case)
1004 * or equally: x_intercept = wb_setpoint + 8 * write_bw
1006 * For single wb case, the dirty pages are observed to fluctuate
1007 * regularly within range
1008 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1009 * for various filesystems, where (2) can yield in a reasonable 12.5%
1010 * fluctuation range for pos_ratio.
1012 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1013 * own size, so move the slope over accordingly and choose a slope that
1014 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1016 if (unlikely(wb_thresh > dtc->thresh))
1017 wb_thresh = dtc->thresh;
1019 * It's very possible that wb_thresh is close to 0 not because the
1020 * device is slow, but that it has remained inactive for long time.
1021 * Honour such devices a reasonable good (hopefully IO efficient)
1022 * threshold, so that the occasional writes won't be blocked and active
1023 * writes can rampup the threshold quickly.
1025 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1027 * scale global setpoint to wb's:
1028 * wb_setpoint = setpoint * wb_thresh / thresh
1030 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1031 wb_setpoint = setpoint * (u64)x >> 16;
1033 * Use span=(8*write_bw) in single wb case as indicated by
1034 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1036 * wb_thresh thresh - wb_thresh
1037 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1040 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1041 x_intercept = wb_setpoint + span;
1043 if (dtc->wb_dirty < x_intercept - span / 4) {
1044 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1045 (x_intercept - wb_setpoint) | 1);
1050 * wb reserve area, safeguard against dirty pool underrun and disk idle
1051 * It may push the desired control point of global dirty pages higher
1054 x_intercept = wb_thresh / 2;
1055 if (dtc->wb_dirty < x_intercept) {
1056 if (dtc->wb_dirty > x_intercept / 8)
1057 pos_ratio = div_u64(pos_ratio * x_intercept,
1063 dtc->pos_ratio = pos_ratio;
1066 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1067 unsigned long elapsed,
1068 unsigned long written)
1070 const unsigned long period = roundup_pow_of_two(3 * HZ);
1071 unsigned long avg = wb->avg_write_bandwidth;
1072 unsigned long old = wb->write_bandwidth;
1076 * bw = written * HZ / elapsed
1078 * bw * elapsed + write_bandwidth * (period - elapsed)
1079 * write_bandwidth = ---------------------------------------------------
1082 * @written may have decreased due to folio_account_redirty().
1083 * Avoid underflowing @bw calculation.
1085 bw = written - min(written, wb->written_stamp);
1087 if (unlikely(elapsed > period)) {
1088 bw = div64_ul(bw, elapsed);
1092 bw += (u64)wb->write_bandwidth * (period - elapsed);
1093 bw >>= ilog2(period);
1096 * one more level of smoothing, for filtering out sudden spikes
1098 if (avg > old && old >= (unsigned long)bw)
1099 avg -= (avg - old) >> 3;
1101 if (avg < old && old <= (unsigned long)bw)
1102 avg += (old - avg) >> 3;
1105 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1106 avg = max(avg, 1LU);
1107 if (wb_has_dirty_io(wb)) {
1108 long delta = avg - wb->avg_write_bandwidth;
1109 WARN_ON_ONCE(atomic_long_add_return(delta,
1110 &wb->bdi->tot_write_bandwidth) <= 0);
1112 wb->write_bandwidth = bw;
1113 WRITE_ONCE(wb->avg_write_bandwidth, avg);
1116 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1118 struct wb_domain *dom = dtc_dom(dtc);
1119 unsigned long thresh = dtc->thresh;
1120 unsigned long limit = dom->dirty_limit;
1123 * Follow up in one step.
1125 if (limit < thresh) {
1131 * Follow down slowly. Use the higher one as the target, because thresh
1132 * may drop below dirty. This is exactly the reason to introduce
1133 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1135 thresh = max(thresh, dtc->dirty);
1136 if (limit > thresh) {
1137 limit -= (limit - thresh) >> 5;
1142 dom->dirty_limit = limit;
1145 static void domain_update_dirty_limit(struct dirty_throttle_control *dtc,
1148 struct wb_domain *dom = dtc_dom(dtc);
1151 * check locklessly first to optimize away locking for the most time
1153 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1156 spin_lock(&dom->lock);
1157 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1158 update_dirty_limit(dtc);
1159 dom->dirty_limit_tstamp = now;
1161 spin_unlock(&dom->lock);
1165 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1167 * Normal wb tasks will be curbed at or below it in long term.
1168 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1170 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1171 unsigned long dirtied,
1172 unsigned long elapsed)
1174 struct bdi_writeback *wb = dtc->wb;
1175 unsigned long dirty = dtc->dirty;
1176 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1177 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1178 unsigned long setpoint = (freerun + limit) / 2;
1179 unsigned long write_bw = wb->avg_write_bandwidth;
1180 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1181 unsigned long dirty_rate;
1182 unsigned long task_ratelimit;
1183 unsigned long balanced_dirty_ratelimit;
1186 unsigned long shift;
1189 * The dirty rate will match the writeout rate in long term, except
1190 * when dirty pages are truncated by userspace or re-dirtied by FS.
1192 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1195 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1197 task_ratelimit = (u64)dirty_ratelimit *
1198 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1199 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1202 * A linear estimation of the "balanced" throttle rate. The theory is,
1203 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1204 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1205 * formula will yield the balanced rate limit (write_bw / N).
1207 * Note that the expanded form is not a pure rate feedback:
1208 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1209 * but also takes pos_ratio into account:
1210 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1212 * (1) is not realistic because pos_ratio also takes part in balancing
1213 * the dirty rate. Consider the state
1214 * pos_ratio = 0.5 (3)
1215 * rate = 2 * (write_bw / N) (4)
1216 * If (1) is used, it will stuck in that state! Because each dd will
1218 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1220 * dirty_rate = N * task_ratelimit = write_bw (6)
1221 * put (6) into (1) we get
1222 * rate_(i+1) = rate_(i) (7)
1224 * So we end up using (2) to always keep
1225 * rate_(i+1) ~= (write_bw / N) (8)
1226 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1227 * pos_ratio is able to drive itself to 1.0, which is not only where
1228 * the dirty count meet the setpoint, but also where the slope of
1229 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1231 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1234 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1236 if (unlikely(balanced_dirty_ratelimit > write_bw))
1237 balanced_dirty_ratelimit = write_bw;
1240 * We could safely do this and return immediately:
1242 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1244 * However to get a more stable dirty_ratelimit, the below elaborated
1245 * code makes use of task_ratelimit to filter out singular points and
1246 * limit the step size.
1248 * The below code essentially only uses the relative value of
1250 * task_ratelimit - dirty_ratelimit
1251 * = (pos_ratio - 1) * dirty_ratelimit
1253 * which reflects the direction and size of dirty position error.
1257 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1258 * task_ratelimit is on the same side of dirty_ratelimit, too.
1260 * - dirty_ratelimit > balanced_dirty_ratelimit
1261 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1262 * lowering dirty_ratelimit will help meet both the position and rate
1263 * control targets. Otherwise, don't update dirty_ratelimit if it will
1264 * only help meet the rate target. After all, what the users ultimately
1265 * feel and care are stable dirty rate and small position error.
1267 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1268 * and filter out the singular points of balanced_dirty_ratelimit. Which
1269 * keeps jumping around randomly and can even leap far away at times
1270 * due to the small 200ms estimation period of dirty_rate (we want to
1271 * keep that period small to reduce time lags).
1276 * For strictlimit case, calculations above were based on wb counters
1277 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1278 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1279 * Hence, to calculate "step" properly, we have to use wb_dirty as
1280 * "dirty" and wb_setpoint as "setpoint".
1282 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1283 * it's possible that wb_thresh is close to zero due to inactivity
1284 * of backing device.
1286 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1287 dirty = dtc->wb_dirty;
1288 if (dtc->wb_dirty < 8)
1289 setpoint = dtc->wb_dirty + 1;
1291 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1294 if (dirty < setpoint) {
1295 x = min3(wb->balanced_dirty_ratelimit,
1296 balanced_dirty_ratelimit, task_ratelimit);
1297 if (dirty_ratelimit < x)
1298 step = x - dirty_ratelimit;
1300 x = max3(wb->balanced_dirty_ratelimit,
1301 balanced_dirty_ratelimit, task_ratelimit);
1302 if (dirty_ratelimit > x)
1303 step = dirty_ratelimit - x;
1307 * Don't pursue 100% rate matching. It's impossible since the balanced
1308 * rate itself is constantly fluctuating. So decrease the track speed
1309 * when it gets close to the target. Helps eliminate pointless tremors.
1311 shift = dirty_ratelimit / (2 * step + 1);
1312 if (shift < BITS_PER_LONG)
1313 step = DIV_ROUND_UP(step >> shift, 8);
1317 if (dirty_ratelimit < balanced_dirty_ratelimit)
1318 dirty_ratelimit += step;
1320 dirty_ratelimit -= step;
1322 WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL));
1323 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1325 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1328 static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1329 struct dirty_throttle_control *mdtc,
1330 bool update_ratelimit)
1332 struct bdi_writeback *wb = gdtc->wb;
1333 unsigned long now = jiffies;
1334 unsigned long elapsed;
1335 unsigned long dirtied;
1336 unsigned long written;
1338 spin_lock(&wb->list_lock);
1341 * Lockless checks for elapsed time are racy and delayed update after
1342 * IO completion doesn't do it at all (to make sure written pages are
1343 * accounted reasonably quickly). Make sure elapsed >= 1 to avoid
1346 elapsed = max(now - wb->bw_time_stamp, 1UL);
1347 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1348 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1350 if (update_ratelimit) {
1351 domain_update_dirty_limit(gdtc, now);
1352 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1355 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1356 * compiler has no way to figure that out. Help it.
1358 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1359 domain_update_dirty_limit(mdtc, now);
1360 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1363 wb_update_write_bandwidth(wb, elapsed, written);
1365 wb->dirtied_stamp = dirtied;
1366 wb->written_stamp = written;
1367 WRITE_ONCE(wb->bw_time_stamp, now);
1368 spin_unlock(&wb->list_lock);
1371 void wb_update_bandwidth(struct bdi_writeback *wb)
1373 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1375 __wb_update_bandwidth(&gdtc, NULL, false);
1378 /* Interval after which we consider wb idle and don't estimate bandwidth */
1379 #define WB_BANDWIDTH_IDLE_JIF (HZ)
1381 static void wb_bandwidth_estimate_start(struct bdi_writeback *wb)
1383 unsigned long now = jiffies;
1384 unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp);
1386 if (elapsed > WB_BANDWIDTH_IDLE_JIF &&
1387 !atomic_read(&wb->writeback_inodes)) {
1388 spin_lock(&wb->list_lock);
1389 wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED);
1390 wb->written_stamp = wb_stat(wb, WB_WRITTEN);
1391 WRITE_ONCE(wb->bw_time_stamp, now);
1392 spin_unlock(&wb->list_lock);
1397 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1398 * will look to see if it needs to start dirty throttling.
1400 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1401 * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1402 * (the number of pages we may dirty without exceeding the dirty limits).
1404 static unsigned long dirty_poll_interval(unsigned long dirty,
1405 unsigned long thresh)
1408 return 1UL << (ilog2(thresh - dirty) >> 1);
1413 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1414 unsigned long wb_dirty)
1416 unsigned long bw = READ_ONCE(wb->avg_write_bandwidth);
1420 * Limit pause time for small memory systems. If sleeping for too long
1421 * time, a small pool of dirty/writeback pages may go empty and disk go
1424 * 8 serves as the safety ratio.
1426 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1429 return min_t(unsigned long, t, MAX_PAUSE);
1432 static long wb_min_pause(struct bdi_writeback *wb,
1434 unsigned long task_ratelimit,
1435 unsigned long dirty_ratelimit,
1436 int *nr_dirtied_pause)
1438 long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth));
1439 long lo = ilog2(READ_ONCE(wb->dirty_ratelimit));
1440 long t; /* target pause */
1441 long pause; /* estimated next pause */
1442 int pages; /* target nr_dirtied_pause */
1444 /* target for 10ms pause on 1-dd case */
1445 t = max(1, HZ / 100);
1448 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1451 * (N * 10ms) on 2^N concurrent tasks.
1454 t += (hi - lo) * (10 * HZ) / 1024;
1457 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1458 * on the much more stable dirty_ratelimit. However the next pause time
1459 * will be computed based on task_ratelimit and the two rate limits may
1460 * depart considerably at some time. Especially if task_ratelimit goes
1461 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1462 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1463 * result task_ratelimit won't be executed faithfully, which could
1464 * eventually bring down dirty_ratelimit.
1466 * We apply two rules to fix it up:
1467 * 1) try to estimate the next pause time and if necessary, use a lower
1468 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1469 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1470 * 2) limit the target pause time to max_pause/2, so that the normal
1471 * small fluctuations of task_ratelimit won't trigger rule (1) and
1472 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1474 t = min(t, 1 + max_pause / 2);
1475 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1478 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1479 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1480 * When the 16 consecutive reads are often interrupted by some dirty
1481 * throttling pause during the async writes, cfq will go into idles
1482 * (deadline is fine). So push nr_dirtied_pause as high as possible
1483 * until reaches DIRTY_POLL_THRESH=32 pages.
1485 if (pages < DIRTY_POLL_THRESH) {
1487 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1488 if (pages > DIRTY_POLL_THRESH) {
1489 pages = DIRTY_POLL_THRESH;
1490 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1494 pause = HZ * pages / (task_ratelimit + 1);
1495 if (pause > max_pause) {
1497 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1500 *nr_dirtied_pause = pages;
1502 * The minimal pause time will normally be half the target pause time.
1504 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1507 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1509 struct bdi_writeback *wb = dtc->wb;
1510 unsigned long wb_reclaimable;
1513 * wb_thresh is not treated as some limiting factor as
1514 * dirty_thresh, due to reasons
1515 * - in JBOD setup, wb_thresh can fluctuate a lot
1516 * - in a system with HDD and USB key, the USB key may somehow
1517 * go into state (wb_dirty >> wb_thresh) either because
1518 * wb_dirty starts high, or because wb_thresh drops low.
1519 * In this case we don't want to hard throttle the USB key
1520 * dirtiers for 100 seconds until wb_dirty drops under
1521 * wb_thresh. Instead the auxiliary wb control line in
1522 * wb_position_ratio() will let the dirtier task progress
1523 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1525 dtc->wb_thresh = __wb_calc_thresh(dtc);
1526 dtc->wb_bg_thresh = dtc->thresh ?
1527 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1530 * In order to avoid the stacked BDI deadlock we need
1531 * to ensure we accurately count the 'dirty' pages when
1532 * the threshold is low.
1534 * Otherwise it would be possible to get thresh+n pages
1535 * reported dirty, even though there are thresh-m pages
1536 * actually dirty; with m+n sitting in the percpu
1539 if (dtc->wb_thresh < 2 * wb_stat_error()) {
1540 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1541 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1543 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1544 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1549 * balance_dirty_pages() must be called by processes which are generating dirty
1550 * data. It looks at the number of dirty pages in the machine and will force
1551 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1552 * If we're over `background_thresh' then the writeback threads are woken to
1553 * perform some writeout.
1555 static void balance_dirty_pages(struct bdi_writeback *wb,
1556 unsigned long pages_dirtied)
1558 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1559 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1560 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1561 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1563 struct dirty_throttle_control *sdtc;
1564 unsigned long nr_reclaimable; /* = file_dirty */
1569 int nr_dirtied_pause;
1570 bool dirty_exceeded = false;
1571 unsigned long task_ratelimit;
1572 unsigned long dirty_ratelimit;
1573 struct backing_dev_info *bdi = wb->bdi;
1574 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1575 unsigned long start_time = jiffies;
1578 unsigned long now = jiffies;
1579 unsigned long dirty, thresh, bg_thresh;
1580 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1581 unsigned long m_thresh = 0;
1582 unsigned long m_bg_thresh = 0;
1584 nr_reclaimable = global_node_page_state(NR_FILE_DIRTY);
1585 gdtc->avail = global_dirtyable_memory();
1586 gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1588 domain_dirty_limits(gdtc);
1590 if (unlikely(strictlimit)) {
1591 wb_dirty_limits(gdtc);
1593 dirty = gdtc->wb_dirty;
1594 thresh = gdtc->wb_thresh;
1595 bg_thresh = gdtc->wb_bg_thresh;
1597 dirty = gdtc->dirty;
1598 thresh = gdtc->thresh;
1599 bg_thresh = gdtc->bg_thresh;
1603 unsigned long filepages, headroom, writeback;
1606 * If @wb belongs to !root memcg, repeat the same
1607 * basic calculations for the memcg domain.
1609 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1610 &mdtc->dirty, &writeback);
1611 mdtc->dirty += writeback;
1612 mdtc_calc_avail(mdtc, filepages, headroom);
1614 domain_dirty_limits(mdtc);
1616 if (unlikely(strictlimit)) {
1617 wb_dirty_limits(mdtc);
1618 m_dirty = mdtc->wb_dirty;
1619 m_thresh = mdtc->wb_thresh;
1620 m_bg_thresh = mdtc->wb_bg_thresh;
1622 m_dirty = mdtc->dirty;
1623 m_thresh = mdtc->thresh;
1624 m_bg_thresh = mdtc->bg_thresh;
1629 * Throttle it only when the background writeback cannot
1630 * catch-up. This avoids (excessively) small writeouts
1631 * when the wb limits are ramping up in case of !strictlimit.
1633 * In strictlimit case make decision based on the wb counters
1634 * and limits. Small writeouts when the wb limits are ramping
1635 * up are the price we consciously pay for strictlimit-ing.
1637 * If memcg domain is in effect, @dirty should be under
1638 * both global and memcg freerun ceilings.
1640 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1642 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1644 unsigned long m_intv;
1647 intv = dirty_poll_interval(dirty, thresh);
1650 current->dirty_paused_when = now;
1651 current->nr_dirtied = 0;
1653 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1654 current->nr_dirtied_pause = min(intv, m_intv);
1658 if (unlikely(!writeback_in_progress(wb)))
1659 wb_start_background_writeback(wb);
1661 mem_cgroup_flush_foreign(wb);
1664 * Calculate global domain's pos_ratio and select the
1665 * global dtc by default.
1668 wb_dirty_limits(gdtc);
1670 if ((current->flags & PF_LOCAL_THROTTLE) &&
1672 dirty_freerun_ceiling(gdtc->wb_thresh,
1673 gdtc->wb_bg_thresh))
1675 * LOCAL_THROTTLE tasks must not be throttled
1676 * when below the per-wb freerun ceiling.
1681 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1682 ((gdtc->dirty > gdtc->thresh) || strictlimit);
1684 wb_position_ratio(gdtc);
1689 * If memcg domain is in effect, calculate its
1690 * pos_ratio. @wb should satisfy constraints from
1691 * both global and memcg domains. Choose the one
1692 * w/ lower pos_ratio.
1695 wb_dirty_limits(mdtc);
1697 if ((current->flags & PF_LOCAL_THROTTLE) &&
1699 dirty_freerun_ceiling(mdtc->wb_thresh,
1700 mdtc->wb_bg_thresh))
1702 * LOCAL_THROTTLE tasks must not be
1703 * throttled when below the per-wb
1708 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1709 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1711 wb_position_ratio(mdtc);
1712 if (mdtc->pos_ratio < gdtc->pos_ratio)
1716 if (dirty_exceeded && !wb->dirty_exceeded)
1717 wb->dirty_exceeded = 1;
1719 if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
1720 BANDWIDTH_INTERVAL))
1721 __wb_update_bandwidth(gdtc, mdtc, true);
1723 /* throttle according to the chosen dtc */
1724 dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit);
1725 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1726 RATELIMIT_CALC_SHIFT;
1727 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1728 min_pause = wb_min_pause(wb, max_pause,
1729 task_ratelimit, dirty_ratelimit,
1732 if (unlikely(task_ratelimit == 0)) {
1737 period = HZ * pages_dirtied / task_ratelimit;
1739 if (current->dirty_paused_when)
1740 pause -= now - current->dirty_paused_when;
1742 * For less than 1s think time (ext3/4 may block the dirtier
1743 * for up to 800ms from time to time on 1-HDD; so does xfs,
1744 * however at much less frequency), try to compensate it in
1745 * future periods by updating the virtual time; otherwise just
1746 * do a reset, as it may be a light dirtier.
1748 if (pause < min_pause) {
1749 trace_balance_dirty_pages(wb,
1762 current->dirty_paused_when = now;
1763 current->nr_dirtied = 0;
1764 } else if (period) {
1765 current->dirty_paused_when += period;
1766 current->nr_dirtied = 0;
1767 } else if (current->nr_dirtied_pause <= pages_dirtied)
1768 current->nr_dirtied_pause += pages_dirtied;
1771 if (unlikely(pause > max_pause)) {
1772 /* for occasional dropped task_ratelimit */
1773 now += min(pause - max_pause, max_pause);
1778 trace_balance_dirty_pages(wb,
1790 __set_current_state(TASK_KILLABLE);
1791 wb->dirty_sleep = now;
1792 io_schedule_timeout(pause);
1794 current->dirty_paused_when = now + pause;
1795 current->nr_dirtied = 0;
1796 current->nr_dirtied_pause = nr_dirtied_pause;
1799 * This is typically equal to (dirty < thresh) and can also
1800 * keep "1000+ dd on a slow USB stick" under control.
1806 * In the case of an unresponsive NFS server and the NFS dirty
1807 * pages exceeds dirty_thresh, give the other good wb's a pipe
1808 * to go through, so that tasks on them still remain responsive.
1810 * In theory 1 page is enough to keep the consumer-producer
1811 * pipe going: the flusher cleans 1 page => the task dirties 1
1812 * more page. However wb_dirty has accounting errors. So use
1813 * the larger and more IO friendly wb_stat_error.
1815 if (sdtc->wb_dirty <= wb_stat_error())
1818 if (fatal_signal_pending(current))
1822 if (!dirty_exceeded && wb->dirty_exceeded)
1823 wb->dirty_exceeded = 0;
1825 if (writeback_in_progress(wb))
1829 * In laptop mode, we wait until hitting the higher threshold before
1830 * starting background writeout, and then write out all the way down
1831 * to the lower threshold. So slow writers cause minimal disk activity.
1833 * In normal mode, we start background writeout at the lower
1834 * background_thresh, to keep the amount of dirty memory low.
1839 if (nr_reclaimable > gdtc->bg_thresh)
1840 wb_start_background_writeback(wb);
1843 static DEFINE_PER_CPU(int, bdp_ratelimits);
1846 * Normal tasks are throttled by
1848 * dirty tsk->nr_dirtied_pause pages;
1849 * take a snap in balance_dirty_pages();
1851 * However there is a worst case. If every task exit immediately when dirtied
1852 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1853 * called to throttle the page dirties. The solution is to save the not yet
1854 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1855 * randomly into the running tasks. This works well for the above worst case,
1856 * as the new task will pick up and accumulate the old task's leaked dirty
1857 * count and eventually get throttled.
1859 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1862 * balance_dirty_pages_ratelimited - balance dirty memory state
1863 * @mapping: address_space which was dirtied
1865 * Processes which are dirtying memory should call in here once for each page
1866 * which was newly dirtied. The function will periodically check the system's
1867 * dirty state and will initiate writeback if needed.
1869 * Once we're over the dirty memory limit we decrease the ratelimiting
1870 * by a lot, to prevent individual processes from overshooting the limit
1871 * by (ratelimit_pages) each.
1873 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1875 struct inode *inode = mapping->host;
1876 struct backing_dev_info *bdi = inode_to_bdi(inode);
1877 struct bdi_writeback *wb = NULL;
1881 if (!(bdi->capabilities & BDI_CAP_WRITEBACK))
1884 if (inode_cgwb_enabled(inode))
1885 wb = wb_get_create_current(bdi, GFP_KERNEL);
1889 ratelimit = current->nr_dirtied_pause;
1890 if (wb->dirty_exceeded)
1891 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1895 * This prevents one CPU to accumulate too many dirtied pages without
1896 * calling into balance_dirty_pages(), which can happen when there are
1897 * 1000+ tasks, all of them start dirtying pages at exactly the same
1898 * time, hence all honoured too large initial task->nr_dirtied_pause.
1900 p = this_cpu_ptr(&bdp_ratelimits);
1901 if (unlikely(current->nr_dirtied >= ratelimit))
1903 else if (unlikely(*p >= ratelimit_pages)) {
1908 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1909 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1910 * the dirty throttling and livelock other long-run dirtiers.
1912 p = this_cpu_ptr(&dirty_throttle_leaks);
1913 if (*p > 0 && current->nr_dirtied < ratelimit) {
1914 unsigned long nr_pages_dirtied;
1915 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1916 *p -= nr_pages_dirtied;
1917 current->nr_dirtied += nr_pages_dirtied;
1921 if (unlikely(current->nr_dirtied >= ratelimit))
1922 balance_dirty_pages(wb, current->nr_dirtied);
1926 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1929 * wb_over_bg_thresh - does @wb need to be written back?
1930 * @wb: bdi_writeback of interest
1932 * Determines whether background writeback should keep writing @wb or it's
1935 * Return: %true if writeback should continue.
1937 bool wb_over_bg_thresh(struct bdi_writeback *wb)
1939 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1940 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1941 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1942 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1944 unsigned long reclaimable;
1945 unsigned long thresh;
1948 * Similar to balance_dirty_pages() but ignores pages being written
1949 * as we're trying to decide whether to put more under writeback.
1951 gdtc->avail = global_dirtyable_memory();
1952 gdtc->dirty = global_node_page_state(NR_FILE_DIRTY);
1953 domain_dirty_limits(gdtc);
1955 if (gdtc->dirty > gdtc->bg_thresh)
1958 thresh = wb_calc_thresh(gdtc->wb, gdtc->bg_thresh);
1959 if (thresh < 2 * wb_stat_error())
1960 reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1962 reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1964 if (reclaimable > thresh)
1968 unsigned long filepages, headroom, writeback;
1970 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1972 mdtc_calc_avail(mdtc, filepages, headroom);
1973 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1975 if (mdtc->dirty > mdtc->bg_thresh)
1978 thresh = wb_calc_thresh(mdtc->wb, mdtc->bg_thresh);
1979 if (thresh < 2 * wb_stat_error())
1980 reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1982 reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1984 if (reclaimable > thresh)
1992 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1994 int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1995 void *buffer, size_t *length, loff_t *ppos)
1997 unsigned int old_interval = dirty_writeback_interval;
2000 ret = proc_dointvec(table, write, buffer, length, ppos);
2003 * Writing 0 to dirty_writeback_interval will disable periodic writeback
2004 * and a different non-zero value will wakeup the writeback threads.
2005 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
2006 * iterate over all bdis and wbs.
2007 * The reason we do this is to make the change take effect immediately.
2009 if (!ret && write && dirty_writeback_interval &&
2010 dirty_writeback_interval != old_interval)
2011 wakeup_flusher_threads(WB_REASON_PERIODIC);
2016 void laptop_mode_timer_fn(struct timer_list *t)
2018 struct backing_dev_info *backing_dev_info =
2019 from_timer(backing_dev_info, t, laptop_mode_wb_timer);
2021 wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
2025 * We've spun up the disk and we're in laptop mode: schedule writeback
2026 * of all dirty data a few seconds from now. If the flush is already scheduled
2027 * then push it back - the user is still using the disk.
2029 void laptop_io_completion(struct backing_dev_info *info)
2031 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2035 * We're in laptop mode and we've just synced. The sync's writes will have
2036 * caused another writeback to be scheduled by laptop_io_completion.
2037 * Nothing needs to be written back anymore, so we unschedule the writeback.
2039 void laptop_sync_completion(void)
2041 struct backing_dev_info *bdi;
2045 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2046 del_timer(&bdi->laptop_mode_wb_timer);
2052 * If ratelimit_pages is too high then we can get into dirty-data overload
2053 * if a large number of processes all perform writes at the same time.
2055 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2056 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2060 void writeback_set_ratelimit(void)
2062 struct wb_domain *dom = &global_wb_domain;
2063 unsigned long background_thresh;
2064 unsigned long dirty_thresh;
2066 global_dirty_limits(&background_thresh, &dirty_thresh);
2067 dom->dirty_limit = dirty_thresh;
2068 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2069 if (ratelimit_pages < 16)
2070 ratelimit_pages = 16;
2073 static int page_writeback_cpu_online(unsigned int cpu)
2075 writeback_set_ratelimit();
2080 * Called early on to tune the page writeback dirty limits.
2082 * We used to scale dirty pages according to how total memory
2083 * related to pages that could be allocated for buffers.
2085 * However, that was when we used "dirty_ratio" to scale with
2086 * all memory, and we don't do that any more. "dirty_ratio"
2087 * is now applied to total non-HIGHPAGE memory, and as such we can't
2088 * get into the old insane situation any more where we had
2089 * large amounts of dirty pages compared to a small amount of
2090 * non-HIGHMEM memory.
2092 * But we might still want to scale the dirty_ratio by how
2093 * much memory the box has..
2095 void __init page_writeback_init(void)
2097 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2099 cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2100 page_writeback_cpu_online, NULL);
2101 cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2102 page_writeback_cpu_online);
2106 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2107 * @mapping: address space structure to write
2108 * @start: starting page index
2109 * @end: ending page index (inclusive)
2111 * This function scans the page range from @start to @end (inclusive) and tags
2112 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2113 * that write_cache_pages (or whoever calls this function) will then use
2114 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2115 * used to avoid livelocking of writeback by a process steadily creating new
2116 * dirty pages in the file (thus it is important for this function to be quick
2117 * so that it can tag pages faster than a dirtying process can create them).
2119 void tag_pages_for_writeback(struct address_space *mapping,
2120 pgoff_t start, pgoff_t end)
2122 XA_STATE(xas, &mapping->i_pages, start);
2123 unsigned int tagged = 0;
2127 xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2128 xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2129 if (++tagged % XA_CHECK_SCHED)
2133 xas_unlock_irq(&xas);
2137 xas_unlock_irq(&xas);
2139 EXPORT_SYMBOL(tag_pages_for_writeback);
2142 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2143 * @mapping: address space structure to write
2144 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2145 * @writepage: function called for each page
2146 * @data: data passed to writepage function
2148 * If a page is already under I/O, write_cache_pages() skips it, even
2149 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2150 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2151 * and msync() need to guarantee that all the data which was dirty at the time
2152 * the call was made get new I/O started against them. If wbc->sync_mode is
2153 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2154 * existing IO to complete.
2156 * To avoid livelocks (when other process dirties new pages), we first tag
2157 * pages which should be written back with TOWRITE tag and only then start
2158 * writing them. For data-integrity sync we have to be careful so that we do
2159 * not miss some pages (e.g., because some other process has cleared TOWRITE
2160 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2161 * by the process clearing the DIRTY tag (and submitting the page for IO).
2163 * To avoid deadlocks between range_cyclic writeback and callers that hold
2164 * pages in PageWriteback to aggregate IO until write_cache_pages() returns,
2165 * we do not loop back to the start of the file. Doing so causes a page
2166 * lock/page writeback access order inversion - we should only ever lock
2167 * multiple pages in ascending page->index order, and looping back to the start
2168 * of the file violates that rule and causes deadlocks.
2170 * Return: %0 on success, negative error code otherwise
2172 int write_cache_pages(struct address_space *mapping,
2173 struct writeback_control *wbc, writepage_t writepage,
2179 struct pagevec pvec;
2182 pgoff_t end; /* Inclusive */
2184 int range_whole = 0;
2187 pagevec_init(&pvec);
2188 if (wbc->range_cyclic) {
2189 index = mapping->writeback_index; /* prev offset */
2192 index = wbc->range_start >> PAGE_SHIFT;
2193 end = wbc->range_end >> PAGE_SHIFT;
2194 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2197 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
2198 tag_pages_for_writeback(mapping, index, end);
2199 tag = PAGECACHE_TAG_TOWRITE;
2201 tag = PAGECACHE_TAG_DIRTY;
2204 while (!done && (index <= end)) {
2207 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2212 for (i = 0; i < nr_pages; i++) {
2213 struct page *page = pvec.pages[i];
2215 done_index = page->index;
2220 * Page truncated or invalidated. We can freely skip it
2221 * then, even for data integrity operations: the page
2222 * has disappeared concurrently, so there could be no
2223 * real expectation of this data integrity operation
2224 * even if there is now a new, dirty page at the same
2225 * pagecache address.
2227 if (unlikely(page->mapping != mapping)) {
2233 if (!PageDirty(page)) {
2234 /* someone wrote it for us */
2235 goto continue_unlock;
2238 if (PageWriteback(page)) {
2239 if (wbc->sync_mode != WB_SYNC_NONE)
2240 wait_on_page_writeback(page);
2242 goto continue_unlock;
2245 BUG_ON(PageWriteback(page));
2246 if (!clear_page_dirty_for_io(page))
2247 goto continue_unlock;
2249 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2250 error = (*writepage)(page, wbc, data);
2251 if (unlikely(error)) {
2253 * Handle errors according to the type of
2254 * writeback. There's no need to continue for
2255 * background writeback. Just push done_index
2256 * past this page so media errors won't choke
2257 * writeout for the entire file. For integrity
2258 * writeback, we must process the entire dirty
2259 * set regardless of errors because the fs may
2260 * still have state to clear for each page. In
2261 * that case we continue processing and return
2264 if (error == AOP_WRITEPAGE_ACTIVATE) {
2267 } else if (wbc->sync_mode != WB_SYNC_ALL) {
2269 done_index = page->index + 1;
2278 * We stop writing back only if we are not doing
2279 * integrity sync. In case of integrity sync we have to
2280 * keep going until we have written all the pages
2281 * we tagged for writeback prior to entering this loop.
2283 if (--wbc->nr_to_write <= 0 &&
2284 wbc->sync_mode == WB_SYNC_NONE) {
2289 pagevec_release(&pvec);
2294 * If we hit the last page and there is more work to be done: wrap
2295 * back the index back to the start of the file for the next
2296 * time we are called.
2298 if (wbc->range_cyclic && !done)
2300 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2301 mapping->writeback_index = done_index;
2305 EXPORT_SYMBOL(write_cache_pages);
2308 * Function used by generic_writepages to call the real writepage
2309 * function and set the mapping flags on error
2311 static int __writepage(struct page *page, struct writeback_control *wbc,
2314 struct address_space *mapping = data;
2315 int ret = mapping->a_ops->writepage(page, wbc);
2316 mapping_set_error(mapping, ret);
2321 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2322 * @mapping: address space structure to write
2323 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2325 * This is a library function, which implements the writepages()
2326 * address_space_operation.
2328 * Return: %0 on success, negative error code otherwise
2330 int generic_writepages(struct address_space *mapping,
2331 struct writeback_control *wbc)
2333 struct blk_plug plug;
2336 /* deal with chardevs and other special file */
2337 if (!mapping->a_ops->writepage)
2340 blk_start_plug(&plug);
2341 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2342 blk_finish_plug(&plug);
2346 EXPORT_SYMBOL(generic_writepages);
2348 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2351 struct bdi_writeback *wb;
2353 if (wbc->nr_to_write <= 0)
2355 wb = inode_to_wb_wbc(mapping->host, wbc);
2356 wb_bandwidth_estimate_start(wb);
2358 if (mapping->a_ops->writepages)
2359 ret = mapping->a_ops->writepages(mapping, wbc);
2361 ret = generic_writepages(mapping, wbc);
2362 if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2366 * Lacking an allocation context or the locality or writeback
2367 * state of any of the inode's pages, throttle based on
2368 * writeback activity on the local node. It's as good a
2371 reclaim_throttle(NODE_DATA(numa_node_id()),
2372 VMSCAN_THROTTLE_WRITEBACK);
2375 * Usually few pages are written by now from those we've just submitted
2376 * but if there's constant writeback being submitted, this makes sure
2377 * writeback bandwidth is updated once in a while.
2379 if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
2380 BANDWIDTH_INTERVAL))
2381 wb_update_bandwidth(wb);
2386 * folio_write_one - write out a single folio and wait on I/O.
2387 * @folio: The folio to write.
2389 * The folio must be locked by the caller and will be unlocked upon return.
2391 * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2394 * Return: %0 on success, negative error code otherwise
2396 int folio_write_one(struct folio *folio)
2398 struct address_space *mapping = folio->mapping;
2400 struct writeback_control wbc = {
2401 .sync_mode = WB_SYNC_ALL,
2402 .nr_to_write = folio_nr_pages(folio),
2405 BUG_ON(!folio_test_locked(folio));
2407 folio_wait_writeback(folio);
2409 if (folio_clear_dirty_for_io(folio)) {
2411 ret = mapping->a_ops->writepage(&folio->page, &wbc);
2413 folio_wait_writeback(folio);
2416 folio_unlock(folio);
2420 ret = filemap_check_errors(mapping);
2423 EXPORT_SYMBOL(folio_write_one);
2426 * For address_spaces which do not use buffers nor write back.
2428 bool noop_dirty_folio(struct address_space *mapping, struct folio *folio)
2430 if (!folio_test_dirty(folio))
2431 return !folio_test_set_dirty(folio);
2434 EXPORT_SYMBOL(noop_dirty_folio);
2437 * Helper function for set_page_dirty family.
2439 * Caller must hold lock_page_memcg().
2441 * NOTE: This relies on being atomic wrt interrupts.
2443 static void folio_account_dirtied(struct folio *folio,
2444 struct address_space *mapping)
2446 struct inode *inode = mapping->host;
2448 trace_writeback_dirty_folio(folio, mapping);
2450 if (mapping_can_writeback(mapping)) {
2451 struct bdi_writeback *wb;
2452 long nr = folio_nr_pages(folio);
2454 inode_attach_wb(inode, &folio->page);
2455 wb = inode_to_wb(inode);
2457 __lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, nr);
2458 __zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
2459 __node_stat_mod_folio(folio, NR_DIRTIED, nr);
2460 wb_stat_mod(wb, WB_RECLAIMABLE, nr);
2461 wb_stat_mod(wb, WB_DIRTIED, nr);
2462 task_io_account_write(nr * PAGE_SIZE);
2463 current->nr_dirtied += nr;
2464 __this_cpu_add(bdp_ratelimits, nr);
2466 mem_cgroup_track_foreign_dirty(folio, wb);
2471 * Helper function for deaccounting dirty page without writeback.
2473 * Caller must hold lock_page_memcg().
2475 void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb)
2477 long nr = folio_nr_pages(folio);
2479 lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2480 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2481 wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2482 task_io_account_cancelled_write(nr * PAGE_SIZE);
2486 * Mark the folio dirty, and set it dirty in the page cache, and mark
2489 * If warn is true, then emit a warning if the folio is not uptodate and has
2490 * not been truncated.
2492 * The caller must hold lock_page_memcg(). Most callers have the folio
2493 * locked. A few have the folio blocked from truncation through other
2494 * means (eg zap_page_range() has it mapped and is holding the page table
2495 * lock). This can also be called from mark_buffer_dirty(), which I
2496 * cannot prove is always protected against truncate.
2498 void __folio_mark_dirty(struct folio *folio, struct address_space *mapping,
2501 unsigned long flags;
2503 xa_lock_irqsave(&mapping->i_pages, flags);
2504 if (folio->mapping) { /* Race with truncate? */
2505 WARN_ON_ONCE(warn && !folio_test_uptodate(folio));
2506 folio_account_dirtied(folio, mapping);
2507 __xa_set_mark(&mapping->i_pages, folio_index(folio),
2508 PAGECACHE_TAG_DIRTY);
2510 xa_unlock_irqrestore(&mapping->i_pages, flags);
2514 * filemap_dirty_folio - Mark a folio dirty for filesystems which do not use buffer_heads.
2515 * @mapping: Address space this folio belongs to.
2516 * @folio: Folio to be marked as dirty.
2518 * Filesystems which do not use buffer heads should call this function
2519 * from their set_page_dirty address space operation. It ignores the
2520 * contents of folio_get_private(), so if the filesystem marks individual
2521 * blocks as dirty, the filesystem should handle that itself.
2523 * This is also sometimes used by filesystems which use buffer_heads when
2524 * a single buffer is being dirtied: we want to set the folio dirty in
2525 * that case, but not all the buffers. This is a "bottom-up" dirtying,
2526 * whereas block_dirty_folio() is a "top-down" dirtying.
2528 * The caller must ensure this doesn't race with truncation. Most will
2529 * simply hold the folio lock, but e.g. zap_pte_range() calls with the
2530 * folio mapped and the pte lock held, which also locks out truncation.
2532 bool filemap_dirty_folio(struct address_space *mapping, struct folio *folio)
2534 folio_memcg_lock(folio);
2535 if (folio_test_set_dirty(folio)) {
2536 folio_memcg_unlock(folio);
2540 __folio_mark_dirty(folio, mapping, !folio_test_private(folio));
2541 folio_memcg_unlock(folio);
2543 if (mapping->host) {
2544 /* !PageAnon && !swapper_space */
2545 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2549 EXPORT_SYMBOL(filemap_dirty_folio);
2552 * folio_account_redirty - Manually account for redirtying a page.
2553 * @folio: The folio which is being redirtied.
2555 * Most filesystems should call folio_redirty_for_writepage() instead
2556 * of this fuction. If your filesystem is doing writeback outside the
2557 * context of a writeback_control(), it can call this when redirtying
2558 * a folio, to de-account the dirty counters (NR_DIRTIED, WB_DIRTIED,
2559 * tsk->nr_dirtied), so that they match the written counters (NR_WRITTEN,
2560 * WB_WRITTEN) in long term. The mismatches will lead to systematic errors
2561 * in balanced_dirty_ratelimit and the dirty pages position control.
2563 void folio_account_redirty(struct folio *folio)
2565 struct address_space *mapping = folio->mapping;
2567 if (mapping && mapping_can_writeback(mapping)) {
2568 struct inode *inode = mapping->host;
2569 struct bdi_writeback *wb;
2570 struct wb_lock_cookie cookie = {};
2571 long nr = folio_nr_pages(folio);
2573 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2574 current->nr_dirtied -= nr;
2575 node_stat_mod_folio(folio, NR_DIRTIED, -nr);
2576 wb_stat_mod(wb, WB_DIRTIED, -nr);
2577 unlocked_inode_to_wb_end(inode, &cookie);
2580 EXPORT_SYMBOL(folio_account_redirty);
2583 * folio_redirty_for_writepage - Decline to write a dirty folio.
2584 * @wbc: The writeback control.
2585 * @folio: The folio.
2587 * When a writepage implementation decides that it doesn't want to write
2588 * @folio for some reason, it should call this function, unlock @folio and
2591 * Return: True if we redirtied the folio. False if someone else dirtied
2594 bool folio_redirty_for_writepage(struct writeback_control *wbc,
2595 struct folio *folio)
2598 long nr = folio_nr_pages(folio);
2600 wbc->pages_skipped += nr;
2601 ret = filemap_dirty_folio(folio->mapping, folio);
2602 folio_account_redirty(folio);
2606 EXPORT_SYMBOL(folio_redirty_for_writepage);
2609 * folio_mark_dirty - Mark a folio as being modified.
2610 * @folio: The folio.
2612 * For folios with a mapping this should be done with the folio lock held
2613 * for the benefit of asynchronous memory errors who prefer a consistent
2614 * dirty state. This rule can be broken in some special cases,
2615 * but should be better not to.
2617 * Return: True if the folio was newly dirtied, false if it was already dirty.
2619 bool folio_mark_dirty(struct folio *folio)
2621 struct address_space *mapping = folio_mapping(folio);
2623 if (likely(mapping)) {
2625 * readahead/lru_deactivate_page could remain
2626 * PG_readahead/PG_reclaim due to race with folio_end_writeback
2627 * About readahead, if the folio is written, the flags would be
2628 * reset. So no problem.
2629 * About lru_deactivate_page, if the folio is redirtied,
2630 * the flag will be reset. So no problem. but if the
2631 * folio is used by readahead it will confuse readahead
2632 * and make it restart the size rampup process. But it's
2633 * a trivial problem.
2635 if (folio_test_reclaim(folio))
2636 folio_clear_reclaim(folio);
2637 return mapping->a_ops->dirty_folio(mapping, folio);
2640 return noop_dirty_folio(mapping, folio);
2642 EXPORT_SYMBOL(folio_mark_dirty);
2645 * set_page_dirty() is racy if the caller has no reference against
2646 * page->mapping->host, and if the page is unlocked. This is because another
2647 * CPU could truncate the page off the mapping and then free the mapping.
2649 * Usually, the page _is_ locked, or the caller is a user-space process which
2650 * holds a reference on the inode by having an open file.
2652 * In other cases, the page should be locked before running set_page_dirty().
2654 int set_page_dirty_lock(struct page *page)
2659 ret = set_page_dirty(page);
2663 EXPORT_SYMBOL(set_page_dirty_lock);
2666 * This cancels just the dirty bit on the kernel page itself, it does NOT
2667 * actually remove dirty bits on any mmap's that may be around. It also
2668 * leaves the page tagged dirty, so any sync activity will still find it on
2669 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2670 * look at the dirty bits in the VM.
2672 * Doing this should *normally* only ever be done when a page is truncated,
2673 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2674 * this when it notices that somebody has cleaned out all the buffers on a
2675 * page without actually doing it through the VM. Can you say "ext3 is
2676 * horribly ugly"? Thought you could.
2678 void __folio_cancel_dirty(struct folio *folio)
2680 struct address_space *mapping = folio_mapping(folio);
2682 if (mapping_can_writeback(mapping)) {
2683 struct inode *inode = mapping->host;
2684 struct bdi_writeback *wb;
2685 struct wb_lock_cookie cookie = {};
2687 folio_memcg_lock(folio);
2688 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2690 if (folio_test_clear_dirty(folio))
2691 folio_account_cleaned(folio, wb);
2693 unlocked_inode_to_wb_end(inode, &cookie);
2694 folio_memcg_unlock(folio);
2696 folio_clear_dirty(folio);
2699 EXPORT_SYMBOL(__folio_cancel_dirty);
2702 * Clear a folio's dirty flag, while caring for dirty memory accounting.
2703 * Returns true if the folio was previously dirty.
2705 * This is for preparing to put the folio under writeout. We leave
2706 * the folio tagged as dirty in the xarray so that a concurrent
2707 * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk.
2708 * The ->writepage implementation will run either folio_start_writeback()
2709 * or folio_mark_dirty(), at which stage we bring the folio's dirty flag
2710 * and xarray dirty tag back into sync.
2712 * This incoherency between the folio's dirty flag and xarray tag is
2713 * unfortunate, but it only exists while the folio is locked.
2715 bool folio_clear_dirty_for_io(struct folio *folio)
2717 struct address_space *mapping = folio_mapping(folio);
2720 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2722 if (mapping && mapping_can_writeback(mapping)) {
2723 struct inode *inode = mapping->host;
2724 struct bdi_writeback *wb;
2725 struct wb_lock_cookie cookie = {};
2728 * Yes, Virginia, this is indeed insane.
2730 * We use this sequence to make sure that
2731 * (a) we account for dirty stats properly
2732 * (b) we tell the low-level filesystem to
2733 * mark the whole folio dirty if it was
2734 * dirty in a pagetable. Only to then
2735 * (c) clean the folio again and return 1 to
2736 * cause the writeback.
2738 * This way we avoid all nasty races with the
2739 * dirty bit in multiple places and clearing
2740 * them concurrently from different threads.
2742 * Note! Normally the "folio_mark_dirty(folio)"
2743 * has no effect on the actual dirty bit - since
2744 * that will already usually be set. But we
2745 * need the side effects, and it can help us
2748 * We basically use the folio "master dirty bit"
2749 * as a serialization point for all the different
2750 * threads doing their things.
2752 if (folio_mkclean(folio))
2753 folio_mark_dirty(folio);
2755 * We carefully synchronise fault handlers against
2756 * installing a dirty pte and marking the folio dirty
2757 * at this point. We do this by having them hold the
2758 * page lock while dirtying the folio, and folios are
2759 * always locked coming in here, so we get the desired
2762 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2763 if (folio_test_clear_dirty(folio)) {
2764 long nr = folio_nr_pages(folio);
2765 lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2766 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2767 wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2770 unlocked_inode_to_wb_end(inode, &cookie);
2773 return folio_test_clear_dirty(folio);
2775 EXPORT_SYMBOL(folio_clear_dirty_for_io);
2777 static void wb_inode_writeback_start(struct bdi_writeback *wb)
2779 atomic_inc(&wb->writeback_inodes);
2782 static void wb_inode_writeback_end(struct bdi_writeback *wb)
2784 atomic_dec(&wb->writeback_inodes);
2786 * Make sure estimate of writeback throughput gets updated after
2787 * writeback completed. We delay the update by BANDWIDTH_INTERVAL
2788 * (which is the interval other bandwidth updates use for batching) so
2789 * that if multiple inodes end writeback at a similar time, they get
2790 * batched into one bandwidth update.
2792 queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL);
2795 bool __folio_end_writeback(struct folio *folio)
2797 long nr = folio_nr_pages(folio);
2798 struct address_space *mapping = folio_mapping(folio);
2801 folio_memcg_lock(folio);
2802 if (mapping && mapping_use_writeback_tags(mapping)) {
2803 struct inode *inode = mapping->host;
2804 struct backing_dev_info *bdi = inode_to_bdi(inode);
2805 unsigned long flags;
2807 xa_lock_irqsave(&mapping->i_pages, flags);
2808 ret = folio_test_clear_writeback(folio);
2810 __xa_clear_mark(&mapping->i_pages, folio_index(folio),
2811 PAGECACHE_TAG_WRITEBACK);
2812 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
2813 struct bdi_writeback *wb = inode_to_wb(inode);
2815 wb_stat_mod(wb, WB_WRITEBACK, -nr);
2816 __wb_writeout_add(wb, nr);
2817 if (!mapping_tagged(mapping,
2818 PAGECACHE_TAG_WRITEBACK))
2819 wb_inode_writeback_end(wb);
2823 if (mapping->host && !mapping_tagged(mapping,
2824 PAGECACHE_TAG_WRITEBACK))
2825 sb_clear_inode_writeback(mapping->host);
2827 xa_unlock_irqrestore(&mapping->i_pages, flags);
2829 ret = folio_test_clear_writeback(folio);
2832 lruvec_stat_mod_folio(folio, NR_WRITEBACK, -nr);
2833 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2834 node_stat_mod_folio(folio, NR_WRITTEN, nr);
2836 folio_memcg_unlock(folio);
2840 bool __folio_start_writeback(struct folio *folio, bool keep_write)
2842 long nr = folio_nr_pages(folio);
2843 struct address_space *mapping = folio_mapping(folio);
2847 folio_memcg_lock(folio);
2848 if (mapping && mapping_use_writeback_tags(mapping)) {
2849 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
2850 struct inode *inode = mapping->host;
2851 struct backing_dev_info *bdi = inode_to_bdi(inode);
2852 unsigned long flags;
2854 xas_lock_irqsave(&xas, flags);
2856 ret = folio_test_set_writeback(folio);
2860 on_wblist = mapping_tagged(mapping,
2861 PAGECACHE_TAG_WRITEBACK);
2863 xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
2864 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
2865 struct bdi_writeback *wb = inode_to_wb(inode);
2867 wb_stat_mod(wb, WB_WRITEBACK, nr);
2869 wb_inode_writeback_start(wb);
2873 * We can come through here when swapping
2874 * anonymous folios, so we don't necessarily
2875 * have an inode to track for sync.
2877 if (mapping->host && !on_wblist)
2878 sb_mark_inode_writeback(mapping->host);
2880 if (!folio_test_dirty(folio))
2881 xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
2883 xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
2884 xas_unlock_irqrestore(&xas, flags);
2886 ret = folio_test_set_writeback(folio);
2889 lruvec_stat_mod_folio(folio, NR_WRITEBACK, nr);
2890 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
2892 folio_memcg_unlock(folio);
2893 access_ret = arch_make_folio_accessible(folio);
2895 * If writeback has been triggered on a page that cannot be made
2896 * accessible, it is too late to recover here.
2898 VM_BUG_ON_FOLIO(access_ret != 0, folio);
2902 EXPORT_SYMBOL(__folio_start_writeback);
2905 * folio_wait_writeback - Wait for a folio to finish writeback.
2906 * @folio: The folio to wait for.
2908 * If the folio is currently being written back to storage, wait for the
2911 * Context: Sleeps. Must be called in process context and with
2912 * no spinlocks held. Caller should hold a reference on the folio.
2913 * If the folio is not locked, writeback may start again after writeback
2916 void folio_wait_writeback(struct folio *folio)
2918 while (folio_test_writeback(folio)) {
2919 trace_folio_wait_writeback(folio, folio_mapping(folio));
2920 folio_wait_bit(folio, PG_writeback);
2923 EXPORT_SYMBOL_GPL(folio_wait_writeback);
2926 * folio_wait_writeback_killable - Wait for a folio to finish writeback.
2927 * @folio: The folio to wait for.
2929 * If the folio is currently being written back to storage, wait for the
2930 * I/O to complete or a fatal signal to arrive.
2932 * Context: Sleeps. Must be called in process context and with
2933 * no spinlocks held. Caller should hold a reference on the folio.
2934 * If the folio is not locked, writeback may start again after writeback
2936 * Return: 0 on success, -EINTR if we get a fatal signal while waiting.
2938 int folio_wait_writeback_killable(struct folio *folio)
2940 while (folio_test_writeback(folio)) {
2941 trace_folio_wait_writeback(folio, folio_mapping(folio));
2942 if (folio_wait_bit_killable(folio, PG_writeback))
2948 EXPORT_SYMBOL_GPL(folio_wait_writeback_killable);
2951 * folio_wait_stable() - wait for writeback to finish, if necessary.
2952 * @folio: The folio to wait on.
2954 * This function determines if the given folio is related to a backing
2955 * device that requires folio contents to be held stable during writeback.
2956 * If so, then it will wait for any pending writeback to complete.
2958 * Context: Sleeps. Must be called in process context and with
2959 * no spinlocks held. Caller should hold a reference on the folio.
2960 * If the folio is not locked, writeback may start again after writeback
2963 void folio_wait_stable(struct folio *folio)
2965 if (folio_inode(folio)->i_sb->s_iflags & SB_I_STABLE_WRITES)
2966 folio_wait_writeback(folio);
2968 EXPORT_SYMBOL_GPL(folio_wait_stable);