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 * Unreclaimable memory (kernel memory or anonymous memory
328 * without swap) can bring down the dirtyable pages below
329 * the zone's dirty balance reserve and the above calculation
330 * will underflow. However we still want to add in nodes
331 * which are below threshold (negative values) to get a more
332 * accurate calculation but make sure that the total never
339 * Make sure that the number of highmem pages is never larger
340 * than the number of the total dirtyable memory. This can only
341 * occur in very strange VM situations but we want to make sure
342 * that this does not occur.
344 return min(x, total);
351 * global_dirtyable_memory - number of globally dirtyable pages
353 * Return: the global number of pages potentially available for dirty
354 * page cache. This is the base value for the global dirty limits.
356 static unsigned long global_dirtyable_memory(void)
360 x = global_zone_page_state(NR_FREE_PAGES);
362 * Pages reserved for the kernel should not be considered
363 * dirtyable, to prevent a situation where reclaim has to
364 * clean pages in order to balance the zones.
366 x -= min(x, totalreserve_pages);
368 x += global_node_page_state(NR_INACTIVE_FILE);
369 x += global_node_page_state(NR_ACTIVE_FILE);
371 if (!vm_highmem_is_dirtyable)
372 x -= highmem_dirtyable_memory(x);
374 return x + 1; /* Ensure that we never return 0 */
378 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
379 * @dtc: dirty_throttle_control of interest
381 * Calculate @dtc->thresh and ->bg_thresh considering
382 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
383 * must ensure that @dtc->avail is set before calling this function. The
384 * dirty limits will be lifted by 1/4 for real-time tasks.
386 static void domain_dirty_limits(struct dirty_throttle_control *dtc)
388 const unsigned long available_memory = dtc->avail;
389 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
390 unsigned long bytes = vm_dirty_bytes;
391 unsigned long bg_bytes = dirty_background_bytes;
392 /* convert ratios to per-PAGE_SIZE for higher precision */
393 unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
394 unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
395 unsigned long thresh;
396 unsigned long bg_thresh;
397 struct task_struct *tsk;
399 /* gdtc is !NULL iff @dtc is for memcg domain */
401 unsigned long global_avail = gdtc->avail;
404 * The byte settings can't be applied directly to memcg
405 * domains. Convert them to ratios by scaling against
406 * globally available memory. As the ratios are in
407 * per-PAGE_SIZE, they can be obtained by dividing bytes by
411 ratio = min(DIV_ROUND_UP(bytes, global_avail),
414 bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
416 bytes = bg_bytes = 0;
420 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
422 thresh = (ratio * available_memory) / PAGE_SIZE;
425 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
427 bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
429 if (bg_thresh >= thresh)
430 bg_thresh = thresh / 2;
433 bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
434 thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
436 dtc->thresh = thresh;
437 dtc->bg_thresh = bg_thresh;
439 /* we should eventually report the domain in the TP */
441 trace_global_dirty_state(bg_thresh, thresh);
445 * global_dirty_limits - background-writeback and dirty-throttling thresholds
446 * @pbackground: out parameter for bg_thresh
447 * @pdirty: out parameter for thresh
449 * Calculate bg_thresh and thresh for global_wb_domain. See
450 * domain_dirty_limits() for details.
452 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
454 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
456 gdtc.avail = global_dirtyable_memory();
457 domain_dirty_limits(&gdtc);
459 *pbackground = gdtc.bg_thresh;
460 *pdirty = gdtc.thresh;
464 * node_dirty_limit - maximum number of dirty pages allowed in a node
467 * Return: the maximum number of dirty pages allowed in a node, based
468 * on the node's dirtyable memory.
470 static unsigned long node_dirty_limit(struct pglist_data *pgdat)
472 unsigned long node_memory = node_dirtyable_memory(pgdat);
473 struct task_struct *tsk = current;
477 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
478 node_memory / global_dirtyable_memory();
480 dirty = vm_dirty_ratio * node_memory / 100;
489 * node_dirty_ok - tells whether a node is within its dirty limits
490 * @pgdat: the node to check
492 * Return: %true when the dirty pages in @pgdat are within the node's
493 * dirty limit, %false if the limit is exceeded.
495 bool node_dirty_ok(struct pglist_data *pgdat)
497 unsigned long limit = node_dirty_limit(pgdat);
498 unsigned long nr_pages = 0;
500 nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
501 nr_pages += node_page_state(pgdat, NR_WRITEBACK);
503 return nr_pages <= limit;
506 int dirty_background_ratio_handler(struct ctl_table *table, int write,
507 void *buffer, size_t *lenp, loff_t *ppos)
511 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
512 if (ret == 0 && write)
513 dirty_background_bytes = 0;
517 int dirty_background_bytes_handler(struct ctl_table *table, int write,
518 void *buffer, size_t *lenp, loff_t *ppos)
522 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
523 if (ret == 0 && write)
524 dirty_background_ratio = 0;
528 int dirty_ratio_handler(struct ctl_table *table, int write, void *buffer,
529 size_t *lenp, loff_t *ppos)
531 int old_ratio = vm_dirty_ratio;
534 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
535 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
536 writeback_set_ratelimit();
542 int dirty_bytes_handler(struct ctl_table *table, int write,
543 void *buffer, size_t *lenp, loff_t *ppos)
545 unsigned long old_bytes = vm_dirty_bytes;
548 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
549 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
550 writeback_set_ratelimit();
556 static unsigned long wp_next_time(unsigned long cur_time)
558 cur_time += VM_COMPLETIONS_PERIOD_LEN;
559 /* 0 has a special meaning... */
565 static void wb_domain_writeout_inc(struct wb_domain *dom,
566 struct fprop_local_percpu *completions,
567 unsigned int max_prop_frac)
569 __fprop_inc_percpu_max(&dom->completions, completions,
571 /* First event after period switching was turned off? */
572 if (unlikely(!dom->period_time)) {
574 * We can race with other __bdi_writeout_inc calls here but
575 * it does not cause any harm since the resulting time when
576 * timer will fire and what is in writeout_period_time will be
579 dom->period_time = wp_next_time(jiffies);
580 mod_timer(&dom->period_timer, dom->period_time);
585 * Increment @wb's writeout completion count and the global writeout
586 * completion count. Called from test_clear_page_writeback().
588 static inline void __wb_writeout_inc(struct bdi_writeback *wb)
590 struct wb_domain *cgdom;
592 inc_wb_stat(wb, WB_WRITTEN);
593 wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
594 wb->bdi->max_prop_frac);
596 cgdom = mem_cgroup_wb_domain(wb);
598 wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
599 wb->bdi->max_prop_frac);
602 void wb_writeout_inc(struct bdi_writeback *wb)
606 local_irq_save(flags);
607 __wb_writeout_inc(wb);
608 local_irq_restore(flags);
610 EXPORT_SYMBOL_GPL(wb_writeout_inc);
613 * On idle system, we can be called long after we scheduled because we use
614 * deferred timers so count with missed periods.
616 static void writeout_period(struct timer_list *t)
618 struct wb_domain *dom = from_timer(dom, t, period_timer);
619 int miss_periods = (jiffies - dom->period_time) /
620 VM_COMPLETIONS_PERIOD_LEN;
622 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
623 dom->period_time = wp_next_time(dom->period_time +
624 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
625 mod_timer(&dom->period_timer, dom->period_time);
628 * Aging has zeroed all fractions. Stop wasting CPU on period
631 dom->period_time = 0;
635 int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
637 memset(dom, 0, sizeof(*dom));
639 spin_lock_init(&dom->lock);
641 timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
643 dom->dirty_limit_tstamp = jiffies;
645 return fprop_global_init(&dom->completions, gfp);
648 #ifdef CONFIG_CGROUP_WRITEBACK
649 void wb_domain_exit(struct wb_domain *dom)
651 del_timer_sync(&dom->period_timer);
652 fprop_global_destroy(&dom->completions);
657 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
658 * registered backing devices, which, for obvious reasons, can not
661 static unsigned int bdi_min_ratio;
663 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
667 spin_lock_bh(&bdi_lock);
668 if (min_ratio > bdi->max_ratio) {
671 min_ratio -= bdi->min_ratio;
672 if (bdi_min_ratio + min_ratio < 100) {
673 bdi_min_ratio += min_ratio;
674 bdi->min_ratio += min_ratio;
679 spin_unlock_bh(&bdi_lock);
684 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
691 spin_lock_bh(&bdi_lock);
692 if (bdi->min_ratio > max_ratio) {
695 bdi->max_ratio = max_ratio;
696 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
698 spin_unlock_bh(&bdi_lock);
702 EXPORT_SYMBOL(bdi_set_max_ratio);
704 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
705 unsigned long bg_thresh)
707 return (thresh + bg_thresh) / 2;
710 static unsigned long hard_dirty_limit(struct wb_domain *dom,
711 unsigned long thresh)
713 return max(thresh, dom->dirty_limit);
717 * Memory which can be further allocated to a memcg domain is capped by
718 * system-wide clean memory excluding the amount being used in the domain.
720 static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
721 unsigned long filepages, unsigned long headroom)
723 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
724 unsigned long clean = filepages - min(filepages, mdtc->dirty);
725 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
726 unsigned long other_clean = global_clean - min(global_clean, clean);
728 mdtc->avail = filepages + min(headroom, other_clean);
732 * __wb_calc_thresh - @wb's share of dirty throttling threshold
733 * @dtc: dirty_throttle_context of interest
735 * Note that balance_dirty_pages() will only seriously take it as a hard limit
736 * when sleeping max_pause per page is not enough to keep the dirty pages under
737 * control. For example, when the device is completely stalled due to some error
738 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
739 * In the other normal situations, it acts more gently by throttling the tasks
740 * more (rather than completely block them) when the wb dirty pages go high.
742 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
743 * - starving fast devices
744 * - piling up dirty pages (that will take long time to sync) on slow devices
746 * The wb's share of dirty limit will be adapting to its throughput and
747 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
749 * Return: @wb's dirty limit in pages. The term "dirty" in the context of
750 * dirty balancing includes all PG_dirty and PG_writeback pages.
752 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
754 struct wb_domain *dom = dtc_dom(dtc);
755 unsigned long thresh = dtc->thresh;
757 unsigned long numerator, denominator;
758 unsigned long wb_min_ratio, wb_max_ratio;
761 * Calculate this BDI's share of the thresh ratio.
763 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
764 &numerator, &denominator);
766 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
767 wb_thresh *= numerator;
768 wb_thresh = div64_ul(wb_thresh, denominator);
770 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
772 wb_thresh += (thresh * wb_min_ratio) / 100;
773 if (wb_thresh > (thresh * wb_max_ratio) / 100)
774 wb_thresh = thresh * wb_max_ratio / 100;
779 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
781 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
783 return __wb_calc_thresh(&gdtc);
788 * f(dirty) := 1.0 + (----------------)
791 * it's a 3rd order polynomial that subjects to
793 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
794 * (2) f(setpoint) = 1.0 => the balance point
795 * (3) f(limit) = 0 => the hard limit
796 * (4) df/dx <= 0 => negative feedback control
797 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
798 * => fast response on large errors; small oscillation near setpoint
800 static long long pos_ratio_polynom(unsigned long setpoint,
807 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
808 (limit - setpoint) | 1);
810 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
811 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
812 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
814 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
818 * Dirty position control.
820 * (o) global/bdi setpoints
822 * We want the dirty pages be balanced around the global/wb setpoints.
823 * When the number of dirty pages is higher/lower than the setpoint, the
824 * dirty position control ratio (and hence task dirty ratelimit) will be
825 * decreased/increased to bring the dirty pages back to the setpoint.
827 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
829 * if (dirty < setpoint) scale up pos_ratio
830 * if (dirty > setpoint) scale down pos_ratio
832 * if (wb_dirty < wb_setpoint) scale up pos_ratio
833 * if (wb_dirty > wb_setpoint) scale down pos_ratio
835 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
837 * (o) global control line
841 * | |<===== global dirty control scope ======>|
849 * 1.0 ................................*
855 * 0 +------------.------------------.----------------------*------------->
856 * freerun^ setpoint^ limit^ dirty pages
858 * (o) wb control line
866 * | * |<=========== span ============>|
867 * 1.0 .......................*
879 * 1/4 ...............................................* * * * * * * * * * * *
883 * 0 +----------------------.-------------------------------.------------->
884 * wb_setpoint^ x_intercept^
886 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
887 * be smoothly throttled down to normal if it starts high in situations like
888 * - start writing to a slow SD card and a fast disk at the same time. The SD
889 * card's wb_dirty may rush to many times higher than wb_setpoint.
890 * - the wb dirty thresh drops quickly due to change of JBOD workload
892 static void wb_position_ratio(struct dirty_throttle_control *dtc)
894 struct bdi_writeback *wb = dtc->wb;
895 unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth);
896 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
897 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
898 unsigned long wb_thresh = dtc->wb_thresh;
899 unsigned long x_intercept;
900 unsigned long setpoint; /* dirty pages' target balance point */
901 unsigned long wb_setpoint;
903 long long pos_ratio; /* for scaling up/down the rate limit */
908 if (unlikely(dtc->dirty >= limit))
914 * See comment for pos_ratio_polynom().
916 setpoint = (freerun + limit) / 2;
917 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
920 * The strictlimit feature is a tool preventing mistrusted filesystems
921 * from growing a large number of dirty pages before throttling. For
922 * such filesystems balance_dirty_pages always checks wb counters
923 * against wb limits. Even if global "nr_dirty" is under "freerun".
924 * This is especially important for fuse which sets bdi->max_ratio to
925 * 1% by default. Without strictlimit feature, fuse writeback may
926 * consume arbitrary amount of RAM because it is accounted in
927 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
929 * Here, in wb_position_ratio(), we calculate pos_ratio based on
930 * two values: wb_dirty and wb_thresh. Let's consider an example:
931 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
932 * limits are set by default to 10% and 20% (background and throttle).
933 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
934 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
935 * about ~6K pages (as the average of background and throttle wb
936 * limits). The 3rd order polynomial will provide positive feedback if
937 * wb_dirty is under wb_setpoint and vice versa.
939 * Note, that we cannot use global counters in these calculations
940 * because we want to throttle process writing to a strictlimit wb
941 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
942 * in the example above).
944 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
945 long long wb_pos_ratio;
947 if (dtc->wb_dirty < 8) {
948 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
949 2 << RATELIMIT_CALC_SHIFT);
953 if (dtc->wb_dirty >= wb_thresh)
956 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
959 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
962 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
966 * Typically, for strictlimit case, wb_setpoint << setpoint
967 * and pos_ratio >> wb_pos_ratio. In the other words global
968 * state ("dirty") is not limiting factor and we have to
969 * make decision based on wb counters. But there is an
970 * important case when global pos_ratio should get precedence:
971 * global limits are exceeded (e.g. due to activities on other
972 * wb's) while given strictlimit wb is below limit.
974 * "pos_ratio * wb_pos_ratio" would work for the case above,
975 * but it would look too non-natural for the case of all
976 * activity in the system coming from a single strictlimit wb
977 * with bdi->max_ratio == 100%.
979 * Note that min() below somewhat changes the dynamics of the
980 * control system. Normally, pos_ratio value can be well over 3
981 * (when globally we are at freerun and wb is well below wb
982 * setpoint). Now the maximum pos_ratio in the same situation
983 * is 2. We might want to tweak this if we observe the control
984 * system is too slow to adapt.
986 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
991 * We have computed basic pos_ratio above based on global situation. If
992 * the wb is over/under its share of dirty pages, we want to scale
993 * pos_ratio further down/up. That is done by the following mechanism.
999 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1001 * x_intercept - wb_dirty
1002 * := --------------------------
1003 * x_intercept - wb_setpoint
1005 * The main wb control line is a linear function that subjects to
1007 * (1) f(wb_setpoint) = 1.0
1008 * (2) k = - 1 / (8 * write_bw) (in single wb case)
1009 * or equally: x_intercept = wb_setpoint + 8 * write_bw
1011 * For single wb case, the dirty pages are observed to fluctuate
1012 * regularly within range
1013 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1014 * for various filesystems, where (2) can yield in a reasonable 12.5%
1015 * fluctuation range for pos_ratio.
1017 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1018 * own size, so move the slope over accordingly and choose a slope that
1019 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1021 if (unlikely(wb_thresh > dtc->thresh))
1022 wb_thresh = dtc->thresh;
1024 * It's very possible that wb_thresh is close to 0 not because the
1025 * device is slow, but that it has remained inactive for long time.
1026 * Honour such devices a reasonable good (hopefully IO efficient)
1027 * threshold, so that the occasional writes won't be blocked and active
1028 * writes can rampup the threshold quickly.
1030 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1032 * scale global setpoint to wb's:
1033 * wb_setpoint = setpoint * wb_thresh / thresh
1035 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1036 wb_setpoint = setpoint * (u64)x >> 16;
1038 * Use span=(8*write_bw) in single wb case as indicated by
1039 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1041 * wb_thresh thresh - wb_thresh
1042 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1045 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1046 x_intercept = wb_setpoint + span;
1048 if (dtc->wb_dirty < x_intercept - span / 4) {
1049 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1050 (x_intercept - wb_setpoint) | 1);
1055 * wb reserve area, safeguard against dirty pool underrun and disk idle
1056 * It may push the desired control point of global dirty pages higher
1059 x_intercept = wb_thresh / 2;
1060 if (dtc->wb_dirty < x_intercept) {
1061 if (dtc->wb_dirty > x_intercept / 8)
1062 pos_ratio = div_u64(pos_ratio * x_intercept,
1068 dtc->pos_ratio = pos_ratio;
1071 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1072 unsigned long elapsed,
1073 unsigned long written)
1075 const unsigned long period = roundup_pow_of_two(3 * HZ);
1076 unsigned long avg = wb->avg_write_bandwidth;
1077 unsigned long old = wb->write_bandwidth;
1081 * bw = written * HZ / elapsed
1083 * bw * elapsed + write_bandwidth * (period - elapsed)
1084 * write_bandwidth = ---------------------------------------------------
1087 * @written may have decreased due to account_page_redirty().
1088 * Avoid underflowing @bw calculation.
1090 bw = written - min(written, wb->written_stamp);
1092 if (unlikely(elapsed > period)) {
1093 bw = div64_ul(bw, elapsed);
1097 bw += (u64)wb->write_bandwidth * (period - elapsed);
1098 bw >>= ilog2(period);
1101 * one more level of smoothing, for filtering out sudden spikes
1103 if (avg > old && old >= (unsigned long)bw)
1104 avg -= (avg - old) >> 3;
1106 if (avg < old && old <= (unsigned long)bw)
1107 avg += (old - avg) >> 3;
1110 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1111 avg = max(avg, 1LU);
1112 if (wb_has_dirty_io(wb)) {
1113 long delta = avg - wb->avg_write_bandwidth;
1114 WARN_ON_ONCE(atomic_long_add_return(delta,
1115 &wb->bdi->tot_write_bandwidth) <= 0);
1117 wb->write_bandwidth = bw;
1118 WRITE_ONCE(wb->avg_write_bandwidth, avg);
1121 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1123 struct wb_domain *dom = dtc_dom(dtc);
1124 unsigned long thresh = dtc->thresh;
1125 unsigned long limit = dom->dirty_limit;
1128 * Follow up in one step.
1130 if (limit < thresh) {
1136 * Follow down slowly. Use the higher one as the target, because thresh
1137 * may drop below dirty. This is exactly the reason to introduce
1138 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1140 thresh = max(thresh, dtc->dirty);
1141 if (limit > thresh) {
1142 limit -= (limit - thresh) >> 5;
1147 dom->dirty_limit = limit;
1150 static void domain_update_dirty_limit(struct dirty_throttle_control *dtc,
1153 struct wb_domain *dom = dtc_dom(dtc);
1156 * check locklessly first to optimize away locking for the most time
1158 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1161 spin_lock(&dom->lock);
1162 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1163 update_dirty_limit(dtc);
1164 dom->dirty_limit_tstamp = now;
1166 spin_unlock(&dom->lock);
1170 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1172 * Normal wb tasks will be curbed at or below it in long term.
1173 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1175 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1176 unsigned long dirtied,
1177 unsigned long elapsed)
1179 struct bdi_writeback *wb = dtc->wb;
1180 unsigned long dirty = dtc->dirty;
1181 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1182 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1183 unsigned long setpoint = (freerun + limit) / 2;
1184 unsigned long write_bw = wb->avg_write_bandwidth;
1185 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1186 unsigned long dirty_rate;
1187 unsigned long task_ratelimit;
1188 unsigned long balanced_dirty_ratelimit;
1191 unsigned long shift;
1194 * The dirty rate will match the writeout rate in long term, except
1195 * when dirty pages are truncated by userspace or re-dirtied by FS.
1197 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1200 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1202 task_ratelimit = (u64)dirty_ratelimit *
1203 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1204 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1207 * A linear estimation of the "balanced" throttle rate. The theory is,
1208 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1209 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1210 * formula will yield the balanced rate limit (write_bw / N).
1212 * Note that the expanded form is not a pure rate feedback:
1213 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1214 * but also takes pos_ratio into account:
1215 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1217 * (1) is not realistic because pos_ratio also takes part in balancing
1218 * the dirty rate. Consider the state
1219 * pos_ratio = 0.5 (3)
1220 * rate = 2 * (write_bw / N) (4)
1221 * If (1) is used, it will stuck in that state! Because each dd will
1223 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1225 * dirty_rate = N * task_ratelimit = write_bw (6)
1226 * put (6) into (1) we get
1227 * rate_(i+1) = rate_(i) (7)
1229 * So we end up using (2) to always keep
1230 * rate_(i+1) ~= (write_bw / N) (8)
1231 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1232 * pos_ratio is able to drive itself to 1.0, which is not only where
1233 * the dirty count meet the setpoint, but also where the slope of
1234 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1236 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1239 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1241 if (unlikely(balanced_dirty_ratelimit > write_bw))
1242 balanced_dirty_ratelimit = write_bw;
1245 * We could safely do this and return immediately:
1247 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1249 * However to get a more stable dirty_ratelimit, the below elaborated
1250 * code makes use of task_ratelimit to filter out singular points and
1251 * limit the step size.
1253 * The below code essentially only uses the relative value of
1255 * task_ratelimit - dirty_ratelimit
1256 * = (pos_ratio - 1) * dirty_ratelimit
1258 * which reflects the direction and size of dirty position error.
1262 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1263 * task_ratelimit is on the same side of dirty_ratelimit, too.
1265 * - dirty_ratelimit > balanced_dirty_ratelimit
1266 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1267 * lowering dirty_ratelimit will help meet both the position and rate
1268 * control targets. Otherwise, don't update dirty_ratelimit if it will
1269 * only help meet the rate target. After all, what the users ultimately
1270 * feel and care are stable dirty rate and small position error.
1272 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1273 * and filter out the singular points of balanced_dirty_ratelimit. Which
1274 * keeps jumping around randomly and can even leap far away at times
1275 * due to the small 200ms estimation period of dirty_rate (we want to
1276 * keep that period small to reduce time lags).
1281 * For strictlimit case, calculations above were based on wb counters
1282 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1283 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1284 * Hence, to calculate "step" properly, we have to use wb_dirty as
1285 * "dirty" and wb_setpoint as "setpoint".
1287 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1288 * it's possible that wb_thresh is close to zero due to inactivity
1289 * of backing device.
1291 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1292 dirty = dtc->wb_dirty;
1293 if (dtc->wb_dirty < 8)
1294 setpoint = dtc->wb_dirty + 1;
1296 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1299 if (dirty < setpoint) {
1300 x = min3(wb->balanced_dirty_ratelimit,
1301 balanced_dirty_ratelimit, task_ratelimit);
1302 if (dirty_ratelimit < x)
1303 step = x - dirty_ratelimit;
1305 x = max3(wb->balanced_dirty_ratelimit,
1306 balanced_dirty_ratelimit, task_ratelimit);
1307 if (dirty_ratelimit > x)
1308 step = dirty_ratelimit - x;
1312 * Don't pursue 100% rate matching. It's impossible since the balanced
1313 * rate itself is constantly fluctuating. So decrease the track speed
1314 * when it gets close to the target. Helps eliminate pointless tremors.
1316 shift = dirty_ratelimit / (2 * step + 1);
1317 if (shift < BITS_PER_LONG)
1318 step = DIV_ROUND_UP(step >> shift, 8);
1322 if (dirty_ratelimit < balanced_dirty_ratelimit)
1323 dirty_ratelimit += step;
1325 dirty_ratelimit -= step;
1327 WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL));
1328 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1330 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1333 static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1334 struct dirty_throttle_control *mdtc,
1335 bool update_ratelimit)
1337 struct bdi_writeback *wb = gdtc->wb;
1338 unsigned long now = jiffies;
1339 unsigned long elapsed;
1340 unsigned long dirtied;
1341 unsigned long written;
1343 spin_lock(&wb->list_lock);
1346 * Lockless checks for elapsed time are racy and delayed update after
1347 * IO completion doesn't do it at all (to make sure written pages are
1348 * accounted reasonably quickly). Make sure elapsed >= 1 to avoid
1351 elapsed = max(now - wb->bw_time_stamp, 1UL);
1352 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1353 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1355 if (update_ratelimit) {
1356 domain_update_dirty_limit(gdtc, now);
1357 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1360 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1361 * compiler has no way to figure that out. Help it.
1363 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1364 domain_update_dirty_limit(mdtc, now);
1365 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1368 wb_update_write_bandwidth(wb, elapsed, written);
1370 wb->dirtied_stamp = dirtied;
1371 wb->written_stamp = written;
1372 WRITE_ONCE(wb->bw_time_stamp, now);
1373 spin_unlock(&wb->list_lock);
1376 void wb_update_bandwidth(struct bdi_writeback *wb)
1378 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1380 __wb_update_bandwidth(&gdtc, NULL, false);
1383 /* Interval after which we consider wb idle and don't estimate bandwidth */
1384 #define WB_BANDWIDTH_IDLE_JIF (HZ)
1386 static void wb_bandwidth_estimate_start(struct bdi_writeback *wb)
1388 unsigned long now = jiffies;
1389 unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp);
1391 if (elapsed > WB_BANDWIDTH_IDLE_JIF &&
1392 !atomic_read(&wb->writeback_inodes)) {
1393 spin_lock(&wb->list_lock);
1394 wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED);
1395 wb->written_stamp = wb_stat(wb, WB_WRITTEN);
1396 WRITE_ONCE(wb->bw_time_stamp, now);
1397 spin_unlock(&wb->list_lock);
1402 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1403 * will look to see if it needs to start dirty throttling.
1405 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1406 * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1407 * (the number of pages we may dirty without exceeding the dirty limits).
1409 static unsigned long dirty_poll_interval(unsigned long dirty,
1410 unsigned long thresh)
1413 return 1UL << (ilog2(thresh - dirty) >> 1);
1418 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1419 unsigned long wb_dirty)
1421 unsigned long bw = READ_ONCE(wb->avg_write_bandwidth);
1425 * Limit pause time for small memory systems. If sleeping for too long
1426 * time, a small pool of dirty/writeback pages may go empty and disk go
1429 * 8 serves as the safety ratio.
1431 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1434 return min_t(unsigned long, t, MAX_PAUSE);
1437 static long wb_min_pause(struct bdi_writeback *wb,
1439 unsigned long task_ratelimit,
1440 unsigned long dirty_ratelimit,
1441 int *nr_dirtied_pause)
1443 long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth));
1444 long lo = ilog2(READ_ONCE(wb->dirty_ratelimit));
1445 long t; /* target pause */
1446 long pause; /* estimated next pause */
1447 int pages; /* target nr_dirtied_pause */
1449 /* target for 10ms pause on 1-dd case */
1450 t = max(1, HZ / 100);
1453 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1456 * (N * 10ms) on 2^N concurrent tasks.
1459 t += (hi - lo) * (10 * HZ) / 1024;
1462 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1463 * on the much more stable dirty_ratelimit. However the next pause time
1464 * will be computed based on task_ratelimit and the two rate limits may
1465 * depart considerably at some time. Especially if task_ratelimit goes
1466 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1467 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1468 * result task_ratelimit won't be executed faithfully, which could
1469 * eventually bring down dirty_ratelimit.
1471 * We apply two rules to fix it up:
1472 * 1) try to estimate the next pause time and if necessary, use a lower
1473 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1474 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1475 * 2) limit the target pause time to max_pause/2, so that the normal
1476 * small fluctuations of task_ratelimit won't trigger rule (1) and
1477 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1479 t = min(t, 1 + max_pause / 2);
1480 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1483 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1484 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1485 * When the 16 consecutive reads are often interrupted by some dirty
1486 * throttling pause during the async writes, cfq will go into idles
1487 * (deadline is fine). So push nr_dirtied_pause as high as possible
1488 * until reaches DIRTY_POLL_THRESH=32 pages.
1490 if (pages < DIRTY_POLL_THRESH) {
1492 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1493 if (pages > DIRTY_POLL_THRESH) {
1494 pages = DIRTY_POLL_THRESH;
1495 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1499 pause = HZ * pages / (task_ratelimit + 1);
1500 if (pause > max_pause) {
1502 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1505 *nr_dirtied_pause = pages;
1507 * The minimal pause time will normally be half the target pause time.
1509 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1512 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1514 struct bdi_writeback *wb = dtc->wb;
1515 unsigned long wb_reclaimable;
1518 * wb_thresh is not treated as some limiting factor as
1519 * dirty_thresh, due to reasons
1520 * - in JBOD setup, wb_thresh can fluctuate a lot
1521 * - in a system with HDD and USB key, the USB key may somehow
1522 * go into state (wb_dirty >> wb_thresh) either because
1523 * wb_dirty starts high, or because wb_thresh drops low.
1524 * In this case we don't want to hard throttle the USB key
1525 * dirtiers for 100 seconds until wb_dirty drops under
1526 * wb_thresh. Instead the auxiliary wb control line in
1527 * wb_position_ratio() will let the dirtier task progress
1528 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1530 dtc->wb_thresh = __wb_calc_thresh(dtc);
1531 dtc->wb_bg_thresh = dtc->thresh ?
1532 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1535 * In order to avoid the stacked BDI deadlock we need
1536 * to ensure we accurately count the 'dirty' pages when
1537 * the threshold is low.
1539 * Otherwise it would be possible to get thresh+n pages
1540 * reported dirty, even though there are thresh-m pages
1541 * actually dirty; with m+n sitting in the percpu
1544 if (dtc->wb_thresh < 2 * wb_stat_error()) {
1545 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1546 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1548 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1549 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1554 * balance_dirty_pages() must be called by processes which are generating dirty
1555 * data. It looks at the number of dirty pages in the machine and will force
1556 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1557 * If we're over `background_thresh' then the writeback threads are woken to
1558 * perform some writeout.
1560 static void balance_dirty_pages(struct bdi_writeback *wb,
1561 unsigned long pages_dirtied)
1563 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1564 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1565 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1566 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1568 struct dirty_throttle_control *sdtc;
1569 unsigned long nr_reclaimable; /* = file_dirty */
1574 int nr_dirtied_pause;
1575 bool dirty_exceeded = false;
1576 unsigned long task_ratelimit;
1577 unsigned long dirty_ratelimit;
1578 struct backing_dev_info *bdi = wb->bdi;
1579 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1580 unsigned long start_time = jiffies;
1583 unsigned long now = jiffies;
1584 unsigned long dirty, thresh, bg_thresh;
1585 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1586 unsigned long m_thresh = 0;
1587 unsigned long m_bg_thresh = 0;
1589 nr_reclaimable = global_node_page_state(NR_FILE_DIRTY);
1590 gdtc->avail = global_dirtyable_memory();
1591 gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1593 domain_dirty_limits(gdtc);
1595 if (unlikely(strictlimit)) {
1596 wb_dirty_limits(gdtc);
1598 dirty = gdtc->wb_dirty;
1599 thresh = gdtc->wb_thresh;
1600 bg_thresh = gdtc->wb_bg_thresh;
1602 dirty = gdtc->dirty;
1603 thresh = gdtc->thresh;
1604 bg_thresh = gdtc->bg_thresh;
1608 unsigned long filepages, headroom, writeback;
1611 * If @wb belongs to !root memcg, repeat the same
1612 * basic calculations for the memcg domain.
1614 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1615 &mdtc->dirty, &writeback);
1616 mdtc->dirty += writeback;
1617 mdtc_calc_avail(mdtc, filepages, headroom);
1619 domain_dirty_limits(mdtc);
1621 if (unlikely(strictlimit)) {
1622 wb_dirty_limits(mdtc);
1623 m_dirty = mdtc->wb_dirty;
1624 m_thresh = mdtc->wb_thresh;
1625 m_bg_thresh = mdtc->wb_bg_thresh;
1627 m_dirty = mdtc->dirty;
1628 m_thresh = mdtc->thresh;
1629 m_bg_thresh = mdtc->bg_thresh;
1634 * Throttle it only when the background writeback cannot
1635 * catch-up. This avoids (excessively) small writeouts
1636 * when the wb limits are ramping up in case of !strictlimit.
1638 * In strictlimit case make decision based on the wb counters
1639 * and limits. Small writeouts when the wb limits are ramping
1640 * up are the price we consciously pay for strictlimit-ing.
1642 * If memcg domain is in effect, @dirty should be under
1643 * both global and memcg freerun ceilings.
1645 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1647 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1649 unsigned long m_intv;
1652 intv = dirty_poll_interval(dirty, thresh);
1655 current->dirty_paused_when = now;
1656 current->nr_dirtied = 0;
1658 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1659 current->nr_dirtied_pause = min(intv, m_intv);
1663 if (unlikely(!writeback_in_progress(wb)))
1664 wb_start_background_writeback(wb);
1666 mem_cgroup_flush_foreign(wb);
1669 * Calculate global domain's pos_ratio and select the
1670 * global dtc by default.
1673 wb_dirty_limits(gdtc);
1675 if ((current->flags & PF_LOCAL_THROTTLE) &&
1677 dirty_freerun_ceiling(gdtc->wb_thresh,
1678 gdtc->wb_bg_thresh))
1680 * LOCAL_THROTTLE tasks must not be throttled
1681 * when below the per-wb freerun ceiling.
1686 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1687 ((gdtc->dirty > gdtc->thresh) || strictlimit);
1689 wb_position_ratio(gdtc);
1694 * If memcg domain is in effect, calculate its
1695 * pos_ratio. @wb should satisfy constraints from
1696 * both global and memcg domains. Choose the one
1697 * w/ lower pos_ratio.
1700 wb_dirty_limits(mdtc);
1702 if ((current->flags & PF_LOCAL_THROTTLE) &&
1704 dirty_freerun_ceiling(mdtc->wb_thresh,
1705 mdtc->wb_bg_thresh))
1707 * LOCAL_THROTTLE tasks must not be
1708 * throttled when below the per-wb
1713 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1714 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1716 wb_position_ratio(mdtc);
1717 if (mdtc->pos_ratio < gdtc->pos_ratio)
1721 if (dirty_exceeded && !wb->dirty_exceeded)
1722 wb->dirty_exceeded = 1;
1724 if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
1725 BANDWIDTH_INTERVAL))
1726 __wb_update_bandwidth(gdtc, mdtc, true);
1728 /* throttle according to the chosen dtc */
1729 dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit);
1730 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1731 RATELIMIT_CALC_SHIFT;
1732 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1733 min_pause = wb_min_pause(wb, max_pause,
1734 task_ratelimit, dirty_ratelimit,
1737 if (unlikely(task_ratelimit == 0)) {
1742 period = HZ * pages_dirtied / task_ratelimit;
1744 if (current->dirty_paused_when)
1745 pause -= now - current->dirty_paused_when;
1747 * For less than 1s think time (ext3/4 may block the dirtier
1748 * for up to 800ms from time to time on 1-HDD; so does xfs,
1749 * however at much less frequency), try to compensate it in
1750 * future periods by updating the virtual time; otherwise just
1751 * do a reset, as it may be a light dirtier.
1753 if (pause < min_pause) {
1754 trace_balance_dirty_pages(wb,
1767 current->dirty_paused_when = now;
1768 current->nr_dirtied = 0;
1769 } else if (period) {
1770 current->dirty_paused_when += period;
1771 current->nr_dirtied = 0;
1772 } else if (current->nr_dirtied_pause <= pages_dirtied)
1773 current->nr_dirtied_pause += pages_dirtied;
1776 if (unlikely(pause > max_pause)) {
1777 /* for occasional dropped task_ratelimit */
1778 now += min(pause - max_pause, max_pause);
1783 trace_balance_dirty_pages(wb,
1795 __set_current_state(TASK_KILLABLE);
1796 wb->dirty_sleep = now;
1797 io_schedule_timeout(pause);
1799 current->dirty_paused_when = now + pause;
1800 current->nr_dirtied = 0;
1801 current->nr_dirtied_pause = nr_dirtied_pause;
1804 * This is typically equal to (dirty < thresh) and can also
1805 * keep "1000+ dd on a slow USB stick" under control.
1811 * In the case of an unresponsive NFS server and the NFS dirty
1812 * pages exceeds dirty_thresh, give the other good wb's a pipe
1813 * to go through, so that tasks on them still remain responsive.
1815 * In theory 1 page is enough to keep the consumer-producer
1816 * pipe going: the flusher cleans 1 page => the task dirties 1
1817 * more page. However wb_dirty has accounting errors. So use
1818 * the larger and more IO friendly wb_stat_error.
1820 if (sdtc->wb_dirty <= wb_stat_error())
1823 if (fatal_signal_pending(current))
1827 if (!dirty_exceeded && wb->dirty_exceeded)
1828 wb->dirty_exceeded = 0;
1830 if (writeback_in_progress(wb))
1834 * In laptop mode, we wait until hitting the higher threshold before
1835 * starting background writeout, and then write out all the way down
1836 * to the lower threshold. So slow writers cause minimal disk activity.
1838 * In normal mode, we start background writeout at the lower
1839 * background_thresh, to keep the amount of dirty memory low.
1844 if (nr_reclaimable > gdtc->bg_thresh)
1845 wb_start_background_writeback(wb);
1848 static DEFINE_PER_CPU(int, bdp_ratelimits);
1851 * Normal tasks are throttled by
1853 * dirty tsk->nr_dirtied_pause pages;
1854 * take a snap in balance_dirty_pages();
1856 * However there is a worst case. If every task exit immediately when dirtied
1857 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1858 * called to throttle the page dirties. The solution is to save the not yet
1859 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1860 * randomly into the running tasks. This works well for the above worst case,
1861 * as the new task will pick up and accumulate the old task's leaked dirty
1862 * count and eventually get throttled.
1864 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1867 * balance_dirty_pages_ratelimited - balance dirty memory state
1868 * @mapping: address_space which was dirtied
1870 * Processes which are dirtying memory should call in here once for each page
1871 * which was newly dirtied. The function will periodically check the system's
1872 * dirty state and will initiate writeback if needed.
1874 * Once we're over the dirty memory limit we decrease the ratelimiting
1875 * by a lot, to prevent individual processes from overshooting the limit
1876 * by (ratelimit_pages) each.
1878 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1880 struct inode *inode = mapping->host;
1881 struct backing_dev_info *bdi = inode_to_bdi(inode);
1882 struct bdi_writeback *wb = NULL;
1886 if (!(bdi->capabilities & BDI_CAP_WRITEBACK))
1889 if (inode_cgwb_enabled(inode))
1890 wb = wb_get_create_current(bdi, GFP_KERNEL);
1894 ratelimit = current->nr_dirtied_pause;
1895 if (wb->dirty_exceeded)
1896 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1900 * This prevents one CPU to accumulate too many dirtied pages without
1901 * calling into balance_dirty_pages(), which can happen when there are
1902 * 1000+ tasks, all of them start dirtying pages at exactly the same
1903 * time, hence all honoured too large initial task->nr_dirtied_pause.
1905 p = this_cpu_ptr(&bdp_ratelimits);
1906 if (unlikely(current->nr_dirtied >= ratelimit))
1908 else if (unlikely(*p >= ratelimit_pages)) {
1913 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1914 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1915 * the dirty throttling and livelock other long-run dirtiers.
1917 p = this_cpu_ptr(&dirty_throttle_leaks);
1918 if (*p > 0 && current->nr_dirtied < ratelimit) {
1919 unsigned long nr_pages_dirtied;
1920 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1921 *p -= nr_pages_dirtied;
1922 current->nr_dirtied += nr_pages_dirtied;
1926 if (unlikely(current->nr_dirtied >= ratelimit))
1927 balance_dirty_pages(wb, current->nr_dirtied);
1931 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1934 * wb_over_bg_thresh - does @wb need to be written back?
1935 * @wb: bdi_writeback of interest
1937 * Determines whether background writeback should keep writing @wb or it's
1940 * Return: %true if writeback should continue.
1942 bool wb_over_bg_thresh(struct bdi_writeback *wb)
1944 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1945 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1946 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1947 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1949 unsigned long reclaimable;
1950 unsigned long thresh;
1953 * Similar to balance_dirty_pages() but ignores pages being written
1954 * as we're trying to decide whether to put more under writeback.
1956 gdtc->avail = global_dirtyable_memory();
1957 gdtc->dirty = global_node_page_state(NR_FILE_DIRTY);
1958 domain_dirty_limits(gdtc);
1960 if (gdtc->dirty > gdtc->bg_thresh)
1963 thresh = wb_calc_thresh(gdtc->wb, gdtc->bg_thresh);
1964 if (thresh < 2 * wb_stat_error())
1965 reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1967 reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1969 if (reclaimable > thresh)
1973 unsigned long filepages, headroom, writeback;
1975 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1977 mdtc_calc_avail(mdtc, filepages, headroom);
1978 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1980 if (mdtc->dirty > mdtc->bg_thresh)
1983 thresh = wb_calc_thresh(mdtc->wb, mdtc->bg_thresh);
1984 if (thresh < 2 * wb_stat_error())
1985 reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1987 reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1989 if (reclaimable > thresh)
1997 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1999 int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
2000 void *buffer, size_t *length, loff_t *ppos)
2002 unsigned int old_interval = dirty_writeback_interval;
2005 ret = proc_dointvec(table, write, buffer, length, ppos);
2008 * Writing 0 to dirty_writeback_interval will disable periodic writeback
2009 * and a different non-zero value will wakeup the writeback threads.
2010 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
2011 * iterate over all bdis and wbs.
2012 * The reason we do this is to make the change take effect immediately.
2014 if (!ret && write && dirty_writeback_interval &&
2015 dirty_writeback_interval != old_interval)
2016 wakeup_flusher_threads(WB_REASON_PERIODIC);
2022 void laptop_mode_timer_fn(struct timer_list *t)
2024 struct backing_dev_info *backing_dev_info =
2025 from_timer(backing_dev_info, t, laptop_mode_wb_timer);
2027 wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
2031 * We've spun up the disk and we're in laptop mode: schedule writeback
2032 * of all dirty data a few seconds from now. If the flush is already scheduled
2033 * then push it back - the user is still using the disk.
2035 void laptop_io_completion(struct backing_dev_info *info)
2037 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2041 * We're in laptop mode and we've just synced. The sync's writes will have
2042 * caused another writeback to be scheduled by laptop_io_completion.
2043 * Nothing needs to be written back anymore, so we unschedule the writeback.
2045 void laptop_sync_completion(void)
2047 struct backing_dev_info *bdi;
2051 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2052 del_timer(&bdi->laptop_mode_wb_timer);
2059 * If ratelimit_pages is too high then we can get into dirty-data overload
2060 * if a large number of processes all perform writes at the same time.
2062 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2063 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2067 void writeback_set_ratelimit(void)
2069 struct wb_domain *dom = &global_wb_domain;
2070 unsigned long background_thresh;
2071 unsigned long dirty_thresh;
2073 global_dirty_limits(&background_thresh, &dirty_thresh);
2074 dom->dirty_limit = dirty_thresh;
2075 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2076 if (ratelimit_pages < 16)
2077 ratelimit_pages = 16;
2080 static int page_writeback_cpu_online(unsigned int cpu)
2082 writeback_set_ratelimit();
2087 * Called early on to tune the page writeback dirty limits.
2089 * We used to scale dirty pages according to how total memory
2090 * related to pages that could be allocated for buffers.
2092 * However, that was when we used "dirty_ratio" to scale with
2093 * all memory, and we don't do that any more. "dirty_ratio"
2094 * is now applied to total non-HIGHPAGE memory, and as such we can't
2095 * get into the old insane situation any more where we had
2096 * large amounts of dirty pages compared to a small amount of
2097 * non-HIGHMEM memory.
2099 * But we might still want to scale the dirty_ratio by how
2100 * much memory the box has..
2102 void __init page_writeback_init(void)
2104 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2106 cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2107 page_writeback_cpu_online, NULL);
2108 cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2109 page_writeback_cpu_online);
2113 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2114 * @mapping: address space structure to write
2115 * @start: starting page index
2116 * @end: ending page index (inclusive)
2118 * This function scans the page range from @start to @end (inclusive) and tags
2119 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2120 * that write_cache_pages (or whoever calls this function) will then use
2121 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2122 * used to avoid livelocking of writeback by a process steadily creating new
2123 * dirty pages in the file (thus it is important for this function to be quick
2124 * so that it can tag pages faster than a dirtying process can create them).
2126 void tag_pages_for_writeback(struct address_space *mapping,
2127 pgoff_t start, pgoff_t end)
2129 XA_STATE(xas, &mapping->i_pages, start);
2130 unsigned int tagged = 0;
2134 xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2135 xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2136 if (++tagged % XA_CHECK_SCHED)
2140 xas_unlock_irq(&xas);
2144 xas_unlock_irq(&xas);
2146 EXPORT_SYMBOL(tag_pages_for_writeback);
2149 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2150 * @mapping: address space structure to write
2151 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2152 * @writepage: function called for each page
2153 * @data: data passed to writepage function
2155 * If a page is already under I/O, write_cache_pages() skips it, even
2156 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2157 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2158 * and msync() need to guarantee that all the data which was dirty at the time
2159 * the call was made get new I/O started against them. If wbc->sync_mode is
2160 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2161 * existing IO to complete.
2163 * To avoid livelocks (when other process dirties new pages), we first tag
2164 * pages which should be written back with TOWRITE tag and only then start
2165 * writing them. For data-integrity sync we have to be careful so that we do
2166 * not miss some pages (e.g., because some other process has cleared TOWRITE
2167 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2168 * by the process clearing the DIRTY tag (and submitting the page for IO).
2170 * To avoid deadlocks between range_cyclic writeback and callers that hold
2171 * pages in PageWriteback to aggregate IO until write_cache_pages() returns,
2172 * we do not loop back to the start of the file. Doing so causes a page
2173 * lock/page writeback access order inversion - we should only ever lock
2174 * multiple pages in ascending page->index order, and looping back to the start
2175 * of the file violates that rule and causes deadlocks.
2177 * Return: %0 on success, negative error code otherwise
2179 int write_cache_pages(struct address_space *mapping,
2180 struct writeback_control *wbc, writepage_t writepage,
2186 struct pagevec pvec;
2189 pgoff_t end; /* Inclusive */
2191 int range_whole = 0;
2194 pagevec_init(&pvec);
2195 if (wbc->range_cyclic) {
2196 index = mapping->writeback_index; /* prev offset */
2199 index = wbc->range_start >> PAGE_SHIFT;
2200 end = wbc->range_end >> PAGE_SHIFT;
2201 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2204 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
2205 tag_pages_for_writeback(mapping, index, end);
2206 tag = PAGECACHE_TAG_TOWRITE;
2208 tag = PAGECACHE_TAG_DIRTY;
2211 while (!done && (index <= end)) {
2214 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2219 for (i = 0; i < nr_pages; i++) {
2220 struct page *page = pvec.pages[i];
2222 done_index = page->index;
2227 * Page truncated or invalidated. We can freely skip it
2228 * then, even for data integrity operations: the page
2229 * has disappeared concurrently, so there could be no
2230 * real expectation of this data integrity operation
2231 * even if there is now a new, dirty page at the same
2232 * pagecache address.
2234 if (unlikely(page->mapping != mapping)) {
2240 if (!PageDirty(page)) {
2241 /* someone wrote it for us */
2242 goto continue_unlock;
2245 if (PageWriteback(page)) {
2246 if (wbc->sync_mode != WB_SYNC_NONE)
2247 wait_on_page_writeback(page);
2249 goto continue_unlock;
2252 BUG_ON(PageWriteback(page));
2253 if (!clear_page_dirty_for_io(page))
2254 goto continue_unlock;
2256 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2257 error = (*writepage)(page, wbc, data);
2258 if (unlikely(error)) {
2260 * Handle errors according to the type of
2261 * writeback. There's no need to continue for
2262 * background writeback. Just push done_index
2263 * past this page so media errors won't choke
2264 * writeout for the entire file. For integrity
2265 * writeback, we must process the entire dirty
2266 * set regardless of errors because the fs may
2267 * still have state to clear for each page. In
2268 * that case we continue processing and return
2271 if (error == AOP_WRITEPAGE_ACTIVATE) {
2274 } else if (wbc->sync_mode != WB_SYNC_ALL) {
2276 done_index = page->index + 1;
2285 * We stop writing back only if we are not doing
2286 * integrity sync. In case of integrity sync we have to
2287 * keep going until we have written all the pages
2288 * we tagged for writeback prior to entering this loop.
2290 if (--wbc->nr_to_write <= 0 &&
2291 wbc->sync_mode == WB_SYNC_NONE) {
2296 pagevec_release(&pvec);
2301 * If we hit the last page and there is more work to be done: wrap
2302 * back the index back to the start of the file for the next
2303 * time we are called.
2305 if (wbc->range_cyclic && !done)
2307 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2308 mapping->writeback_index = done_index;
2312 EXPORT_SYMBOL(write_cache_pages);
2315 * Function used by generic_writepages to call the real writepage
2316 * function and set the mapping flags on error
2318 static int __writepage(struct page *page, struct writeback_control *wbc,
2321 struct address_space *mapping = data;
2322 int ret = mapping->a_ops->writepage(page, wbc);
2323 mapping_set_error(mapping, ret);
2328 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2329 * @mapping: address space structure to write
2330 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2332 * This is a library function, which implements the writepages()
2333 * address_space_operation.
2335 * Return: %0 on success, negative error code otherwise
2337 int generic_writepages(struct address_space *mapping,
2338 struct writeback_control *wbc)
2340 struct blk_plug plug;
2343 /* deal with chardevs and other special file */
2344 if (!mapping->a_ops->writepage)
2347 blk_start_plug(&plug);
2348 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2349 blk_finish_plug(&plug);
2353 EXPORT_SYMBOL(generic_writepages);
2355 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2358 struct bdi_writeback *wb;
2360 if (wbc->nr_to_write <= 0)
2362 wb = inode_to_wb_wbc(mapping->host, wbc);
2363 wb_bandwidth_estimate_start(wb);
2365 if (mapping->a_ops->writepages)
2366 ret = mapping->a_ops->writepages(mapping, wbc);
2368 ret = generic_writepages(mapping, wbc);
2369 if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2372 congestion_wait(BLK_RW_ASYNC, HZ/50);
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 * write_one_page - write out a single page and wait on I/O
2387 * @page: the page to write
2389 * The page 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 write_one_page(struct page *page)
2398 struct address_space *mapping = page->mapping;
2400 struct writeback_control wbc = {
2401 .sync_mode = WB_SYNC_ALL,
2405 BUG_ON(!PageLocked(page));
2407 wait_on_page_writeback(page);
2409 if (clear_page_dirty_for_io(page)) {
2411 ret = mapping->a_ops->writepage(page, &wbc);
2413 wait_on_page_writeback(page);
2420 ret = filemap_check_errors(mapping);
2423 EXPORT_SYMBOL(write_one_page);
2426 * For address_spaces which do not use buffers nor write back.
2428 int __set_page_dirty_no_writeback(struct page *page)
2430 if (!PageDirty(page))
2431 return !TestSetPageDirty(page);
2434 EXPORT_SYMBOL(__set_page_dirty_no_writeback);
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 account_page_dirtied(struct page *page,
2444 struct address_space *mapping)
2446 struct inode *inode = mapping->host;
2448 trace_writeback_dirty_page(page, mapping);
2450 if (mapping_can_writeback(mapping)) {
2451 struct bdi_writeback *wb;
2453 inode_attach_wb(inode, page);
2454 wb = inode_to_wb(inode);
2456 __inc_lruvec_page_state(page, NR_FILE_DIRTY);
2457 __inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2458 __inc_node_page_state(page, NR_DIRTIED);
2459 inc_wb_stat(wb, WB_RECLAIMABLE);
2460 inc_wb_stat(wb, WB_DIRTIED);
2461 task_io_account_write(PAGE_SIZE);
2462 current->nr_dirtied++;
2463 __this_cpu_inc(bdp_ratelimits);
2465 mem_cgroup_track_foreign_dirty(page, wb);
2470 * Helper function for deaccounting dirty page without writeback.
2472 * Caller must hold lock_page_memcg().
2474 void account_page_cleaned(struct page *page, struct address_space *mapping,
2475 struct bdi_writeback *wb)
2477 if (mapping_can_writeback(mapping)) {
2478 dec_lruvec_page_state(page, NR_FILE_DIRTY);
2479 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2480 dec_wb_stat(wb, WB_RECLAIMABLE);
2481 task_io_account_cancelled_write(PAGE_SIZE);
2486 * Mark the page dirty, and set it dirty in the page cache, and mark the inode
2489 * If warn is true, then emit a warning if the page is not uptodate and has
2490 * not been truncated.
2492 * The caller must hold lock_page_memcg().
2494 void __set_page_dirty(struct page *page, struct address_space *mapping,
2497 unsigned long flags;
2499 xa_lock_irqsave(&mapping->i_pages, flags);
2500 if (page->mapping) { /* Race with truncate? */
2501 WARN_ON_ONCE(warn && !PageUptodate(page));
2502 account_page_dirtied(page, mapping);
2503 __xa_set_mark(&mapping->i_pages, page_index(page),
2504 PAGECACHE_TAG_DIRTY);
2506 xa_unlock_irqrestore(&mapping->i_pages, flags);
2510 * For address_spaces which do not use buffers. Just tag the page as dirty in
2513 * This is also used when a single buffer is being dirtied: we want to set the
2514 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2515 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2517 * The caller must ensure this doesn't race with truncation. Most will simply
2518 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2519 * the pte lock held, which also locks out truncation.
2521 int __set_page_dirty_nobuffers(struct page *page)
2523 lock_page_memcg(page);
2524 if (!TestSetPageDirty(page)) {
2525 struct address_space *mapping = page_mapping(page);
2528 unlock_page_memcg(page);
2531 __set_page_dirty(page, mapping, !PagePrivate(page));
2532 unlock_page_memcg(page);
2534 if (mapping->host) {
2535 /* !PageAnon && !swapper_space */
2536 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2540 unlock_page_memcg(page);
2543 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2546 * Call this whenever redirtying a page, to de-account the dirty counters
2547 * (NR_DIRTIED, WB_DIRTIED, tsk->nr_dirtied), so that they match the written
2548 * counters (NR_WRITTEN, WB_WRITTEN) in long term. The mismatches will lead to
2549 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2552 void account_page_redirty(struct page *page)
2554 struct address_space *mapping = page->mapping;
2556 if (mapping && mapping_can_writeback(mapping)) {
2557 struct inode *inode = mapping->host;
2558 struct bdi_writeback *wb;
2559 struct wb_lock_cookie cookie = {};
2561 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2562 current->nr_dirtied--;
2563 dec_node_page_state(page, NR_DIRTIED);
2564 dec_wb_stat(wb, WB_DIRTIED);
2565 unlocked_inode_to_wb_end(inode, &cookie);
2568 EXPORT_SYMBOL(account_page_redirty);
2571 * When a writepage implementation decides that it doesn't want to write this
2572 * page for some reason, it should redirty the locked page via
2573 * redirty_page_for_writepage() and it should then unlock the page and return 0
2575 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2579 wbc->pages_skipped++;
2580 ret = __set_page_dirty_nobuffers(page);
2581 account_page_redirty(page);
2584 EXPORT_SYMBOL(redirty_page_for_writepage);
2589 * For pages with a mapping this should be done under the page lock for the
2590 * benefit of asynchronous memory errors who prefer a consistent dirty state.
2591 * This rule can be broken in some special cases, but should be better not to.
2593 int set_page_dirty(struct page *page)
2595 struct address_space *mapping = page_mapping(page);
2597 page = compound_head(page);
2598 if (likely(mapping)) {
2600 * readahead/lru_deactivate_page could remain
2601 * PG_readahead/PG_reclaim due to race with end_page_writeback
2602 * About readahead, if the page is written, the flags would be
2603 * reset. So no problem.
2604 * About lru_deactivate_page, if the page is redirty, the flag
2605 * will be reset. So no problem. but if the page is used by readahead
2606 * it will confuse readahead and make it restart the size rampup
2607 * process. But it's a trivial problem.
2609 if (PageReclaim(page))
2610 ClearPageReclaim(page);
2611 return mapping->a_ops->set_page_dirty(page);
2613 if (!PageDirty(page)) {
2614 if (!TestSetPageDirty(page))
2619 EXPORT_SYMBOL(set_page_dirty);
2622 * set_page_dirty() is racy if the caller has no reference against
2623 * page->mapping->host, and if the page is unlocked. This is because another
2624 * CPU could truncate the page off the mapping and then free the mapping.
2626 * Usually, the page _is_ locked, or the caller is a user-space process which
2627 * holds a reference on the inode by having an open file.
2629 * In other cases, the page should be locked before running set_page_dirty().
2631 int set_page_dirty_lock(struct page *page)
2636 ret = set_page_dirty(page);
2640 EXPORT_SYMBOL(set_page_dirty_lock);
2643 * This cancels just the dirty bit on the kernel page itself, it does NOT
2644 * actually remove dirty bits on any mmap's that may be around. It also
2645 * leaves the page tagged dirty, so any sync activity will still find it on
2646 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2647 * look at the dirty bits in the VM.
2649 * Doing this should *normally* only ever be done when a page is truncated,
2650 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2651 * this when it notices that somebody has cleaned out all the buffers on a
2652 * page without actually doing it through the VM. Can you say "ext3 is
2653 * horribly ugly"? Thought you could.
2655 void __cancel_dirty_page(struct page *page)
2657 struct address_space *mapping = page_mapping(page);
2659 if (mapping_can_writeback(mapping)) {
2660 struct inode *inode = mapping->host;
2661 struct bdi_writeback *wb;
2662 struct wb_lock_cookie cookie = {};
2664 lock_page_memcg(page);
2665 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2667 if (TestClearPageDirty(page))
2668 account_page_cleaned(page, mapping, wb);
2670 unlocked_inode_to_wb_end(inode, &cookie);
2671 unlock_page_memcg(page);
2673 ClearPageDirty(page);
2676 EXPORT_SYMBOL(__cancel_dirty_page);
2679 * Clear a page's dirty flag, while caring for dirty memory accounting.
2680 * Returns true if the page was previously dirty.
2682 * This is for preparing to put the page under writeout. We leave the page
2683 * tagged as dirty in the xarray so that a concurrent write-for-sync
2684 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2685 * implementation will run either set_page_writeback() or set_page_dirty(),
2686 * at which stage we bring the page's dirty flag and xarray dirty tag
2689 * This incoherency between the page's dirty flag and xarray tag is
2690 * unfortunate, but it only exists while the page is locked.
2692 int clear_page_dirty_for_io(struct page *page)
2694 struct address_space *mapping = page_mapping(page);
2697 VM_BUG_ON_PAGE(!PageLocked(page), page);
2699 if (mapping && mapping_can_writeback(mapping)) {
2700 struct inode *inode = mapping->host;
2701 struct bdi_writeback *wb;
2702 struct wb_lock_cookie cookie = {};
2705 * Yes, Virginia, this is indeed insane.
2707 * We use this sequence to make sure that
2708 * (a) we account for dirty stats properly
2709 * (b) we tell the low-level filesystem to
2710 * mark the whole page dirty if it was
2711 * dirty in a pagetable. Only to then
2712 * (c) clean the page again and return 1 to
2713 * cause the writeback.
2715 * This way we avoid all nasty races with the
2716 * dirty bit in multiple places and clearing
2717 * them concurrently from different threads.
2719 * Note! Normally the "set_page_dirty(page)"
2720 * has no effect on the actual dirty bit - since
2721 * that will already usually be set. But we
2722 * need the side effects, and it can help us
2725 * We basically use the page "master dirty bit"
2726 * as a serialization point for all the different
2727 * threads doing their things.
2729 if (page_mkclean(page))
2730 set_page_dirty(page);
2732 * We carefully synchronise fault handlers against
2733 * installing a dirty pte and marking the page dirty
2734 * at this point. We do this by having them hold the
2735 * page lock while dirtying the page, and pages are
2736 * always locked coming in here, so we get the desired
2739 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2740 if (TestClearPageDirty(page)) {
2741 dec_lruvec_page_state(page, NR_FILE_DIRTY);
2742 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2743 dec_wb_stat(wb, WB_RECLAIMABLE);
2746 unlocked_inode_to_wb_end(inode, &cookie);
2749 return TestClearPageDirty(page);
2751 EXPORT_SYMBOL(clear_page_dirty_for_io);
2753 static void wb_inode_writeback_start(struct bdi_writeback *wb)
2755 atomic_inc(&wb->writeback_inodes);
2758 static void wb_inode_writeback_end(struct bdi_writeback *wb)
2760 atomic_dec(&wb->writeback_inodes);
2762 * Make sure estimate of writeback throughput gets updated after
2763 * writeback completed. We delay the update by BANDWIDTH_INTERVAL
2764 * (which is the interval other bandwidth updates use for batching) so
2765 * that if multiple inodes end writeback at a similar time, they get
2766 * batched into one bandwidth update.
2768 queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL);
2771 int test_clear_page_writeback(struct page *page)
2773 struct address_space *mapping = page_mapping(page);
2776 lock_page_memcg(page);
2777 if (mapping && mapping_use_writeback_tags(mapping)) {
2778 struct inode *inode = mapping->host;
2779 struct backing_dev_info *bdi = inode_to_bdi(inode);
2780 unsigned long flags;
2782 xa_lock_irqsave(&mapping->i_pages, flags);
2783 ret = TestClearPageWriteback(page);
2785 __xa_clear_mark(&mapping->i_pages, page_index(page),
2786 PAGECACHE_TAG_WRITEBACK);
2787 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
2788 struct bdi_writeback *wb = inode_to_wb(inode);
2790 dec_wb_stat(wb, WB_WRITEBACK);
2791 __wb_writeout_inc(wb);
2792 if (!mapping_tagged(mapping,
2793 PAGECACHE_TAG_WRITEBACK))
2794 wb_inode_writeback_end(wb);
2798 if (mapping->host && !mapping_tagged(mapping,
2799 PAGECACHE_TAG_WRITEBACK))
2800 sb_clear_inode_writeback(mapping->host);
2802 xa_unlock_irqrestore(&mapping->i_pages, flags);
2804 ret = TestClearPageWriteback(page);
2807 dec_lruvec_page_state(page, NR_WRITEBACK);
2808 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2809 inc_node_page_state(page, NR_WRITTEN);
2811 unlock_page_memcg(page);
2815 int __test_set_page_writeback(struct page *page, bool keep_write)
2817 struct address_space *mapping = page_mapping(page);
2818 int ret, access_ret;
2820 lock_page_memcg(page);
2821 if (mapping && mapping_use_writeback_tags(mapping)) {
2822 XA_STATE(xas, &mapping->i_pages, page_index(page));
2823 struct inode *inode = mapping->host;
2824 struct backing_dev_info *bdi = inode_to_bdi(inode);
2825 unsigned long flags;
2827 xas_lock_irqsave(&xas, flags);
2829 ret = TestSetPageWriteback(page);
2833 on_wblist = mapping_tagged(mapping,
2834 PAGECACHE_TAG_WRITEBACK);
2836 xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
2837 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
2838 struct bdi_writeback *wb = inode_to_wb(inode);
2840 inc_wb_stat(wb, WB_WRITEBACK);
2842 wb_inode_writeback_start(wb);
2846 * We can come through here when swapping anonymous
2847 * pages, so we don't necessarily have an inode to track
2850 if (mapping->host && !on_wblist)
2851 sb_mark_inode_writeback(mapping->host);
2853 if (!PageDirty(page))
2854 xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
2856 xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
2857 xas_unlock_irqrestore(&xas, flags);
2859 ret = TestSetPageWriteback(page);
2862 inc_lruvec_page_state(page, NR_WRITEBACK);
2863 inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2865 unlock_page_memcg(page);
2866 access_ret = arch_make_page_accessible(page);
2868 * If writeback has been triggered on a page that cannot be made
2869 * accessible, it is too late to recover here.
2871 VM_BUG_ON_PAGE(access_ret != 0, page);
2876 EXPORT_SYMBOL(__test_set_page_writeback);
2879 * Wait for a page to complete writeback
2881 void wait_on_page_writeback(struct page *page)
2883 while (PageWriteback(page)) {
2884 trace_wait_on_page_writeback(page, page_mapping(page));
2885 wait_on_page_bit(page, PG_writeback);
2888 EXPORT_SYMBOL_GPL(wait_on_page_writeback);
2891 * Wait for a page to complete writeback. Returns -EINTR if we get a
2892 * fatal signal while waiting.
2894 int wait_on_page_writeback_killable(struct page *page)
2896 while (PageWriteback(page)) {
2897 trace_wait_on_page_writeback(page, page_mapping(page));
2898 if (wait_on_page_bit_killable(page, PG_writeback))
2904 EXPORT_SYMBOL_GPL(wait_on_page_writeback_killable);
2907 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2908 * @page: The page to wait on.
2910 * This function determines if the given page is related to a backing device
2911 * that requires page contents to be held stable during writeback. If so, then
2912 * it will wait for any pending writeback to complete.
2914 void wait_for_stable_page(struct page *page)
2916 page = thp_head(page);
2917 if (page->mapping->host->i_sb->s_iflags & SB_I_STABLE_WRITES)
2918 wait_on_page_writeback(page);
2920 EXPORT_SYMBOL_GPL(wait_for_stable_page);