4 * Copyright (C) 2002, Linus Torvalds.
5 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Contains functions related to writing back dirty pages at the
10 * 10Apr2002 Andrew Morton
14 #include <linux/kernel.h>
15 #include <linux/export.h>
16 #include <linux/spinlock.h>
19 #include <linux/swap.h>
20 #include <linux/slab.h>
21 #include <linux/pagemap.h>
22 #include <linux/writeback.h>
23 #include <linux/init.h>
24 #include <linux/backing-dev.h>
25 #include <linux/task_io_accounting_ops.h>
26 #include <linux/blkdev.h>
27 #include <linux/mpage.h>
28 #include <linux/rmap.h>
29 #include <linux/percpu.h>
30 #include <linux/notifier.h>
31 #include <linux/smp.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/syscalls.h>
35 #include <linux/buffer_head.h> /* __set_page_dirty_buffers */
36 #include <linux/pagevec.h>
37 #include <linux/timer.h>
38 #include <linux/sched/rt.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 makes the machine dump writes/reads and block dirtyings.
116 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
117 * a full sync is triggered after this time elapses without any disk activity.
121 EXPORT_SYMBOL(laptop_mode);
123 /* End of sysctl-exported parameters */
125 struct wb_domain global_wb_domain;
127 /* consolidated parameters for balance_dirty_pages() and its subroutines */
128 struct dirty_throttle_control {
129 #ifdef CONFIG_CGROUP_WRITEBACK
130 struct wb_domain *dom;
131 struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */
133 struct bdi_writeback *wb;
134 struct fprop_local_percpu *wb_completions;
136 unsigned long avail; /* dirtyable */
137 unsigned long dirty; /* file_dirty + write + nfs */
138 unsigned long thresh; /* dirty threshold */
139 unsigned long bg_thresh; /* dirty background threshold */
141 unsigned long wb_dirty; /* per-wb counterparts */
142 unsigned long wb_thresh;
143 unsigned long wb_bg_thresh;
145 unsigned long pos_ratio;
148 #define DTC_INIT_COMMON(__wb) .wb = (__wb), \
149 .wb_completions = &(__wb)->completions
152 * Length of period for aging writeout fractions of bdis. This is an
153 * arbitrarily chosen number. The longer the period, the slower fractions will
154 * reflect changes in current writeout rate.
156 #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
158 #ifdef CONFIG_CGROUP_WRITEBACK
160 #define GDTC_INIT(__wb) .dom = &global_wb_domain, \
161 DTC_INIT_COMMON(__wb)
162 #define GDTC_INIT_NO_WB .dom = &global_wb_domain
164 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
169 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
174 static void wb_min_max_ratio(struct bdi_writeback *wb,
175 unsigned long *minp, unsigned long *maxp)
177 unsigned long this_bw = wb->avg_write_bandwidth;
178 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
179 unsigned long long min = wb->bdi->min_ratio;
180 unsigned long long max = wb->bdi->max_ratio;
183 * @wb may already be clean by the time control reaches here and
184 * the total may not include its bw.
186 if (this_bw < tot_bw) {
201 #else /* CONFIG_CGROUP_WRITEBACK */
203 #define GDTC_INIT(__wb) DTC_INIT_COMMON(__wb)
204 #define GDTC_INIT_NO_WB
206 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
208 return &global_wb_domain;
211 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
216 static void wb_min_max_ratio(struct bdi_writeback *wb,
217 unsigned long *minp, unsigned long *maxp)
219 *minp = wb->bdi->min_ratio;
220 *maxp = wb->bdi->max_ratio;
223 #endif /* CONFIG_CGROUP_WRITEBACK */
226 * In a memory zone, there is a certain amount of pages we consider
227 * available for the page cache, which is essentially the number of
228 * free and reclaimable pages, minus some zone reserves to protect
229 * lowmem and the ability to uphold the zone's watermarks without
230 * requiring writeback.
232 * This number of dirtyable pages is the base value of which the
233 * user-configurable dirty ratio is the effictive number of pages that
234 * are allowed to be actually dirtied. Per individual zone, or
235 * globally by using the sum of dirtyable pages over all zones.
237 * Because the user is allowed to specify the dirty limit globally as
238 * absolute number of bytes, calculating the per-zone dirty limit can
239 * require translating the configured limit into a percentage of
240 * global dirtyable memory first.
244 * zone_dirtyable_memory - number of dirtyable pages in a zone
247 * Returns the zone's number of pages potentially available for dirty
248 * page cache. This is the base value for the per-zone dirty limits.
250 static unsigned long zone_dirtyable_memory(struct zone *zone)
252 unsigned long nr_pages;
254 nr_pages = zone_page_state(zone, NR_FREE_PAGES);
255 nr_pages -= min(nr_pages, zone->dirty_balance_reserve);
257 nr_pages += zone_page_state(zone, NR_INACTIVE_FILE);
258 nr_pages += zone_page_state(zone, NR_ACTIVE_FILE);
263 static unsigned long highmem_dirtyable_memory(unsigned long total)
265 #ifdef CONFIG_HIGHMEM
269 for_each_node_state(node, N_HIGH_MEMORY) {
270 struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
272 x += zone_dirtyable_memory(z);
275 * Unreclaimable memory (kernel memory or anonymous memory
276 * without swap) can bring down the dirtyable pages below
277 * the zone's dirty balance reserve and the above calculation
278 * will underflow. However we still want to add in nodes
279 * which are below threshold (negative values) to get a more
280 * accurate calculation but make sure that the total never
287 * Make sure that the number of highmem pages is never larger
288 * than the number of the total dirtyable memory. This can only
289 * occur in very strange VM situations but we want to make sure
290 * that this does not occur.
292 return min(x, total);
299 * global_dirtyable_memory - number of globally dirtyable pages
301 * Returns the global number of pages potentially available for dirty
302 * page cache. This is the base value for the global dirty limits.
304 static unsigned long global_dirtyable_memory(void)
308 x = global_page_state(NR_FREE_PAGES);
309 x -= min(x, dirty_balance_reserve);
311 x += global_page_state(NR_INACTIVE_FILE);
312 x += global_page_state(NR_ACTIVE_FILE);
314 if (!vm_highmem_is_dirtyable)
315 x -= highmem_dirtyable_memory(x);
317 return x + 1; /* Ensure that we never return 0 */
321 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
322 * @dtc: dirty_throttle_control of interest
324 * Calculate @dtc->thresh and ->bg_thresh considering
325 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
326 * must ensure that @dtc->avail is set before calling this function. The
327 * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
330 static void domain_dirty_limits(struct dirty_throttle_control *dtc)
332 const unsigned long available_memory = dtc->avail;
333 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
334 unsigned long bytes = vm_dirty_bytes;
335 unsigned long bg_bytes = dirty_background_bytes;
336 unsigned long ratio = vm_dirty_ratio;
337 unsigned long bg_ratio = dirty_background_ratio;
338 unsigned long thresh;
339 unsigned long bg_thresh;
340 struct task_struct *tsk;
342 /* gdtc is !NULL iff @dtc is for memcg domain */
344 unsigned long global_avail = gdtc->avail;
347 * The byte settings can't be applied directly to memcg
348 * domains. Convert them to ratios by scaling against
349 * globally available memory.
352 ratio = min(DIV_ROUND_UP(bytes, PAGE_SIZE) * 100 /
353 global_avail, 100UL);
355 bg_ratio = min(DIV_ROUND_UP(bg_bytes, PAGE_SIZE) * 100 /
356 global_avail, 100UL);
357 bytes = bg_bytes = 0;
361 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
363 thresh = (ratio * available_memory) / 100;
366 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
368 bg_thresh = (bg_ratio * available_memory) / 100;
370 if (bg_thresh >= thresh)
371 bg_thresh = thresh / 2;
373 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
374 bg_thresh += bg_thresh / 4;
375 thresh += thresh / 4;
377 dtc->thresh = thresh;
378 dtc->bg_thresh = bg_thresh;
380 /* we should eventually report the domain in the TP */
382 trace_global_dirty_state(bg_thresh, thresh);
386 * global_dirty_limits - background-writeback and dirty-throttling thresholds
387 * @pbackground: out parameter for bg_thresh
388 * @pdirty: out parameter for thresh
390 * Calculate bg_thresh and thresh for global_wb_domain. See
391 * domain_dirty_limits() for details.
393 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
395 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
397 gdtc.avail = global_dirtyable_memory();
398 domain_dirty_limits(&gdtc);
400 *pbackground = gdtc.bg_thresh;
401 *pdirty = gdtc.thresh;
405 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
408 * Returns the maximum number of dirty pages allowed in a zone, based
409 * on the zone's dirtyable memory.
411 static unsigned long zone_dirty_limit(struct zone *zone)
413 unsigned long zone_memory = zone_dirtyable_memory(zone);
414 struct task_struct *tsk = current;
418 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
419 zone_memory / global_dirtyable_memory();
421 dirty = vm_dirty_ratio * zone_memory / 100;
423 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
430 * zone_dirty_ok - tells whether a zone is within its dirty limits
431 * @zone: the zone to check
433 * Returns %true when the dirty pages in @zone are within the zone's
434 * dirty limit, %false if the limit is exceeded.
436 bool zone_dirty_ok(struct zone *zone)
438 unsigned long limit = zone_dirty_limit(zone);
440 return zone_page_state(zone, NR_FILE_DIRTY) +
441 zone_page_state(zone, NR_UNSTABLE_NFS) +
442 zone_page_state(zone, NR_WRITEBACK) <= limit;
445 int dirty_background_ratio_handler(struct ctl_table *table, int write,
446 void __user *buffer, size_t *lenp,
451 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
452 if (ret == 0 && write)
453 dirty_background_bytes = 0;
457 int dirty_background_bytes_handler(struct ctl_table *table, int write,
458 void __user *buffer, size_t *lenp,
463 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
464 if (ret == 0 && write)
465 dirty_background_ratio = 0;
469 int dirty_ratio_handler(struct ctl_table *table, int write,
470 void __user *buffer, size_t *lenp,
473 int old_ratio = vm_dirty_ratio;
476 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
477 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
478 writeback_set_ratelimit();
484 int dirty_bytes_handler(struct ctl_table *table, int write,
485 void __user *buffer, size_t *lenp,
488 unsigned long old_bytes = vm_dirty_bytes;
491 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
492 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
493 writeback_set_ratelimit();
499 static unsigned long wp_next_time(unsigned long cur_time)
501 cur_time += VM_COMPLETIONS_PERIOD_LEN;
502 /* 0 has a special meaning... */
508 static void wb_domain_writeout_inc(struct wb_domain *dom,
509 struct fprop_local_percpu *completions,
510 unsigned int max_prop_frac)
512 __fprop_inc_percpu_max(&dom->completions, completions,
514 /* First event after period switching was turned off? */
515 if (!unlikely(dom->period_time)) {
517 * We can race with other __bdi_writeout_inc calls here but
518 * it does not cause any harm since the resulting time when
519 * timer will fire and what is in writeout_period_time will be
522 dom->period_time = wp_next_time(jiffies);
523 mod_timer(&dom->period_timer, dom->period_time);
528 * Increment @wb's writeout completion count and the global writeout
529 * completion count. Called from test_clear_page_writeback().
531 static inline void __wb_writeout_inc(struct bdi_writeback *wb)
533 __inc_wb_stat(wb, WB_WRITTEN);
534 wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
535 wb->bdi->max_prop_frac);
538 void wb_writeout_inc(struct bdi_writeback *wb)
542 local_irq_save(flags);
543 __wb_writeout_inc(wb);
544 local_irq_restore(flags);
546 EXPORT_SYMBOL_GPL(wb_writeout_inc);
549 * On idle system, we can be called long after we scheduled because we use
550 * deferred timers so count with missed periods.
552 static void writeout_period(unsigned long t)
554 struct wb_domain *dom = (void *)t;
555 int miss_periods = (jiffies - dom->period_time) /
556 VM_COMPLETIONS_PERIOD_LEN;
558 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
559 dom->period_time = wp_next_time(dom->period_time +
560 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
561 mod_timer(&dom->period_timer, dom->period_time);
564 * Aging has zeroed all fractions. Stop wasting CPU on period
567 dom->period_time = 0;
571 int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
573 memset(dom, 0, sizeof(*dom));
575 spin_lock_init(&dom->lock);
577 init_timer_deferrable(&dom->period_timer);
578 dom->period_timer.function = writeout_period;
579 dom->period_timer.data = (unsigned long)dom;
581 dom->dirty_limit_tstamp = jiffies;
583 return fprop_global_init(&dom->completions, gfp);
587 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
588 * registered backing devices, which, for obvious reasons, can not
591 static unsigned int bdi_min_ratio;
593 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
597 spin_lock_bh(&bdi_lock);
598 if (min_ratio > bdi->max_ratio) {
601 min_ratio -= bdi->min_ratio;
602 if (bdi_min_ratio + min_ratio < 100) {
603 bdi_min_ratio += min_ratio;
604 bdi->min_ratio += min_ratio;
609 spin_unlock_bh(&bdi_lock);
614 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
621 spin_lock_bh(&bdi_lock);
622 if (bdi->min_ratio > max_ratio) {
625 bdi->max_ratio = max_ratio;
626 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
628 spin_unlock_bh(&bdi_lock);
632 EXPORT_SYMBOL(bdi_set_max_ratio);
634 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
635 unsigned long bg_thresh)
637 return (thresh + bg_thresh) / 2;
640 static unsigned long hard_dirty_limit(struct wb_domain *dom,
641 unsigned long thresh)
643 return max(thresh, dom->dirty_limit);
647 * __wb_calc_thresh - @wb's share of dirty throttling threshold
648 * @dtc: dirty_throttle_context of interest
650 * Returns @wb's dirty limit in pages. The term "dirty" in the context of
651 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
653 * Note that balance_dirty_pages() will only seriously take it as a hard limit
654 * when sleeping max_pause per page is not enough to keep the dirty pages under
655 * control. For example, when the device is completely stalled due to some error
656 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
657 * In the other normal situations, it acts more gently by throttling the tasks
658 * more (rather than completely block them) when the wb dirty pages go high.
660 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
661 * - starving fast devices
662 * - piling up dirty pages (that will take long time to sync) on slow devices
664 * The wb's share of dirty limit will be adapting to its throughput and
665 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
667 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
669 struct wb_domain *dom = dtc_dom(dtc);
670 unsigned long thresh = dtc->thresh;
672 long numerator, denominator;
673 unsigned long wb_min_ratio, wb_max_ratio;
676 * Calculate this BDI's share of the thresh ratio.
678 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
679 &numerator, &denominator);
681 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
682 wb_thresh *= numerator;
683 do_div(wb_thresh, denominator);
685 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
687 wb_thresh += (thresh * wb_min_ratio) / 100;
688 if (wb_thresh > (thresh * wb_max_ratio) / 100)
689 wb_thresh = thresh * wb_max_ratio / 100;
694 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
696 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
698 return __wb_calc_thresh(&gdtc);
703 * f(dirty) := 1.0 + (----------------)
706 * it's a 3rd order polynomial that subjects to
708 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
709 * (2) f(setpoint) = 1.0 => the balance point
710 * (3) f(limit) = 0 => the hard limit
711 * (4) df/dx <= 0 => negative feedback control
712 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
713 * => fast response on large errors; small oscillation near setpoint
715 static long long pos_ratio_polynom(unsigned long setpoint,
722 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
723 limit - setpoint + 1);
725 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
726 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
727 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
729 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
733 * Dirty position control.
735 * (o) global/bdi setpoints
737 * We want the dirty pages be balanced around the global/wb setpoints.
738 * When the number of dirty pages is higher/lower than the setpoint, the
739 * dirty position control ratio (and hence task dirty ratelimit) will be
740 * decreased/increased to bring the dirty pages back to the setpoint.
742 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
744 * if (dirty < setpoint) scale up pos_ratio
745 * if (dirty > setpoint) scale down pos_ratio
747 * if (wb_dirty < wb_setpoint) scale up pos_ratio
748 * if (wb_dirty > wb_setpoint) scale down pos_ratio
750 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
752 * (o) global control line
756 * | |<===== global dirty control scope ======>|
764 * 1.0 ................................*
770 * 0 +------------.------------------.----------------------*------------->
771 * freerun^ setpoint^ limit^ dirty pages
773 * (o) wb control line
781 * | * |<=========== span ============>|
782 * 1.0 .......................*
794 * 1/4 ...............................................* * * * * * * * * * * *
798 * 0 +----------------------.-------------------------------.------------->
799 * wb_setpoint^ x_intercept^
801 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
802 * be smoothly throttled down to normal if it starts high in situations like
803 * - start writing to a slow SD card and a fast disk at the same time. The SD
804 * card's wb_dirty may rush to many times higher than wb_setpoint.
805 * - the wb dirty thresh drops quickly due to change of JBOD workload
807 static void wb_position_ratio(struct dirty_throttle_control *dtc)
809 struct bdi_writeback *wb = dtc->wb;
810 unsigned long write_bw = wb->avg_write_bandwidth;
811 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
812 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
813 unsigned long wb_thresh = dtc->wb_thresh;
814 unsigned long x_intercept;
815 unsigned long setpoint; /* dirty pages' target balance point */
816 unsigned long wb_setpoint;
818 long long pos_ratio; /* for scaling up/down the rate limit */
823 if (unlikely(dtc->dirty >= limit))
829 * See comment for pos_ratio_polynom().
831 setpoint = (freerun + limit) / 2;
832 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
835 * The strictlimit feature is a tool preventing mistrusted filesystems
836 * from growing a large number of dirty pages before throttling. For
837 * such filesystems balance_dirty_pages always checks wb counters
838 * against wb limits. Even if global "nr_dirty" is under "freerun".
839 * This is especially important for fuse which sets bdi->max_ratio to
840 * 1% by default. Without strictlimit feature, fuse writeback may
841 * consume arbitrary amount of RAM because it is accounted in
842 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
844 * Here, in wb_position_ratio(), we calculate pos_ratio based on
845 * two values: wb_dirty and wb_thresh. Let's consider an example:
846 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
847 * limits are set by default to 10% and 20% (background and throttle).
848 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
849 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
850 * about ~6K pages (as the average of background and throttle wb
851 * limits). The 3rd order polynomial will provide positive feedback if
852 * wb_dirty is under wb_setpoint and vice versa.
854 * Note, that we cannot use global counters in these calculations
855 * because we want to throttle process writing to a strictlimit wb
856 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
857 * in the example above).
859 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
860 long long wb_pos_ratio;
862 if (dtc->wb_dirty < 8) {
863 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
864 2 << RATELIMIT_CALC_SHIFT);
868 if (dtc->wb_dirty >= wb_thresh)
871 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
874 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
877 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
881 * Typically, for strictlimit case, wb_setpoint << setpoint
882 * and pos_ratio >> wb_pos_ratio. In the other words global
883 * state ("dirty") is not limiting factor and we have to
884 * make decision based on wb counters. But there is an
885 * important case when global pos_ratio should get precedence:
886 * global limits are exceeded (e.g. due to activities on other
887 * wb's) while given strictlimit wb is below limit.
889 * "pos_ratio * wb_pos_ratio" would work for the case above,
890 * but it would look too non-natural for the case of all
891 * activity in the system coming from a single strictlimit wb
892 * with bdi->max_ratio == 100%.
894 * Note that min() below somewhat changes the dynamics of the
895 * control system. Normally, pos_ratio value can be well over 3
896 * (when globally we are at freerun and wb is well below wb
897 * setpoint). Now the maximum pos_ratio in the same situation
898 * is 2. We might want to tweak this if we observe the control
899 * system is too slow to adapt.
901 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
906 * We have computed basic pos_ratio above based on global situation. If
907 * the wb is over/under its share of dirty pages, we want to scale
908 * pos_ratio further down/up. That is done by the following mechanism.
914 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
916 * x_intercept - wb_dirty
917 * := --------------------------
918 * x_intercept - wb_setpoint
920 * The main wb control line is a linear function that subjects to
922 * (1) f(wb_setpoint) = 1.0
923 * (2) k = - 1 / (8 * write_bw) (in single wb case)
924 * or equally: x_intercept = wb_setpoint + 8 * write_bw
926 * For single wb case, the dirty pages are observed to fluctuate
927 * regularly within range
928 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
929 * for various filesystems, where (2) can yield in a reasonable 12.5%
930 * fluctuation range for pos_ratio.
932 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
933 * own size, so move the slope over accordingly and choose a slope that
934 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
936 if (unlikely(wb_thresh > dtc->thresh))
937 wb_thresh = dtc->thresh;
939 * It's very possible that wb_thresh is close to 0 not because the
940 * device is slow, but that it has remained inactive for long time.
941 * Honour such devices a reasonable good (hopefully IO efficient)
942 * threshold, so that the occasional writes won't be blocked and active
943 * writes can rampup the threshold quickly.
945 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
947 * scale global setpoint to wb's:
948 * wb_setpoint = setpoint * wb_thresh / thresh
950 x = div_u64((u64)wb_thresh << 16, dtc->thresh + 1);
951 wb_setpoint = setpoint * (u64)x >> 16;
953 * Use span=(8*write_bw) in single wb case as indicated by
954 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
956 * wb_thresh thresh - wb_thresh
957 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
960 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
961 x_intercept = wb_setpoint + span;
963 if (dtc->wb_dirty < x_intercept - span / 4) {
964 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
965 x_intercept - wb_setpoint + 1);
970 * wb reserve area, safeguard against dirty pool underrun and disk idle
971 * It may push the desired control point of global dirty pages higher
974 x_intercept = wb_thresh / 2;
975 if (dtc->wb_dirty < x_intercept) {
976 if (dtc->wb_dirty > x_intercept / 8)
977 pos_ratio = div_u64(pos_ratio * x_intercept,
983 dtc->pos_ratio = pos_ratio;
986 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
987 unsigned long elapsed,
988 unsigned long written)
990 const unsigned long period = roundup_pow_of_two(3 * HZ);
991 unsigned long avg = wb->avg_write_bandwidth;
992 unsigned long old = wb->write_bandwidth;
996 * bw = written * HZ / elapsed
998 * bw * elapsed + write_bandwidth * (period - elapsed)
999 * write_bandwidth = ---------------------------------------------------
1002 * @written may have decreased due to account_page_redirty().
1003 * Avoid underflowing @bw calculation.
1005 bw = written - min(written, wb->written_stamp);
1007 if (unlikely(elapsed > period)) {
1008 do_div(bw, elapsed);
1012 bw += (u64)wb->write_bandwidth * (period - elapsed);
1013 bw >>= ilog2(period);
1016 * one more level of smoothing, for filtering out sudden spikes
1018 if (avg > old && old >= (unsigned long)bw)
1019 avg -= (avg - old) >> 3;
1021 if (avg < old && old <= (unsigned long)bw)
1022 avg += (old - avg) >> 3;
1025 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1026 avg = max(avg, 1LU);
1027 if (wb_has_dirty_io(wb)) {
1028 long delta = avg - wb->avg_write_bandwidth;
1029 WARN_ON_ONCE(atomic_long_add_return(delta,
1030 &wb->bdi->tot_write_bandwidth) <= 0);
1032 wb->write_bandwidth = bw;
1033 wb->avg_write_bandwidth = avg;
1036 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1038 struct wb_domain *dom = dtc_dom(dtc);
1039 unsigned long thresh = dtc->thresh;
1040 unsigned long limit = dom->dirty_limit;
1043 * Follow up in one step.
1045 if (limit < thresh) {
1051 * Follow down slowly. Use the higher one as the target, because thresh
1052 * may drop below dirty. This is exactly the reason to introduce
1053 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1055 thresh = max(thresh, dtc->dirty);
1056 if (limit > thresh) {
1057 limit -= (limit - thresh) >> 5;
1062 dom->dirty_limit = limit;
1065 static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
1068 struct wb_domain *dom = dtc_dom(dtc);
1071 * check locklessly first to optimize away locking for the most time
1073 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1076 spin_lock(&dom->lock);
1077 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1078 update_dirty_limit(dtc);
1079 dom->dirty_limit_tstamp = now;
1081 spin_unlock(&dom->lock);
1085 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1087 * Normal wb tasks will be curbed at or below it in long term.
1088 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1090 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1091 unsigned long dirtied,
1092 unsigned long elapsed)
1094 struct bdi_writeback *wb = dtc->wb;
1095 unsigned long dirty = dtc->dirty;
1096 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1097 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1098 unsigned long setpoint = (freerun + limit) / 2;
1099 unsigned long write_bw = wb->avg_write_bandwidth;
1100 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1101 unsigned long dirty_rate;
1102 unsigned long task_ratelimit;
1103 unsigned long balanced_dirty_ratelimit;
1108 * The dirty rate will match the writeout rate in long term, except
1109 * when dirty pages are truncated by userspace or re-dirtied by FS.
1111 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1114 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1116 task_ratelimit = (u64)dirty_ratelimit *
1117 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1118 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1121 * A linear estimation of the "balanced" throttle rate. The theory is,
1122 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1123 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1124 * formula will yield the balanced rate limit (write_bw / N).
1126 * Note that the expanded form is not a pure rate feedback:
1127 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1128 * but also takes pos_ratio into account:
1129 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1131 * (1) is not realistic because pos_ratio also takes part in balancing
1132 * the dirty rate. Consider the state
1133 * pos_ratio = 0.5 (3)
1134 * rate = 2 * (write_bw / N) (4)
1135 * If (1) is used, it will stuck in that state! Because each dd will
1137 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1139 * dirty_rate = N * task_ratelimit = write_bw (6)
1140 * put (6) into (1) we get
1141 * rate_(i+1) = rate_(i) (7)
1143 * So we end up using (2) to always keep
1144 * rate_(i+1) ~= (write_bw / N) (8)
1145 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1146 * pos_ratio is able to drive itself to 1.0, which is not only where
1147 * the dirty count meet the setpoint, but also where the slope of
1148 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1150 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1153 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1155 if (unlikely(balanced_dirty_ratelimit > write_bw))
1156 balanced_dirty_ratelimit = write_bw;
1159 * We could safely do this and return immediately:
1161 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1163 * However to get a more stable dirty_ratelimit, the below elaborated
1164 * code makes use of task_ratelimit to filter out singular points and
1165 * limit the step size.
1167 * The below code essentially only uses the relative value of
1169 * task_ratelimit - dirty_ratelimit
1170 * = (pos_ratio - 1) * dirty_ratelimit
1172 * which reflects the direction and size of dirty position error.
1176 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1177 * task_ratelimit is on the same side of dirty_ratelimit, too.
1179 * - dirty_ratelimit > balanced_dirty_ratelimit
1180 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1181 * lowering dirty_ratelimit will help meet both the position and rate
1182 * control targets. Otherwise, don't update dirty_ratelimit if it will
1183 * only help meet the rate target. After all, what the users ultimately
1184 * feel and care are stable dirty rate and small position error.
1186 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1187 * and filter out the singular points of balanced_dirty_ratelimit. Which
1188 * keeps jumping around randomly and can even leap far away at times
1189 * due to the small 200ms estimation period of dirty_rate (we want to
1190 * keep that period small to reduce time lags).
1195 * For strictlimit case, calculations above were based on wb counters
1196 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1197 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1198 * Hence, to calculate "step" properly, we have to use wb_dirty as
1199 * "dirty" and wb_setpoint as "setpoint".
1201 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1202 * it's possible that wb_thresh is close to zero due to inactivity
1203 * of backing device.
1205 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1206 dirty = dtc->wb_dirty;
1207 if (dtc->wb_dirty < 8)
1208 setpoint = dtc->wb_dirty + 1;
1210 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1213 if (dirty < setpoint) {
1214 x = min3(wb->balanced_dirty_ratelimit,
1215 balanced_dirty_ratelimit, task_ratelimit);
1216 if (dirty_ratelimit < x)
1217 step = x - dirty_ratelimit;
1219 x = max3(wb->balanced_dirty_ratelimit,
1220 balanced_dirty_ratelimit, task_ratelimit);
1221 if (dirty_ratelimit > x)
1222 step = dirty_ratelimit - x;
1226 * Don't pursue 100% rate matching. It's impossible since the balanced
1227 * rate itself is constantly fluctuating. So decrease the track speed
1228 * when it gets close to the target. Helps eliminate pointless tremors.
1230 step >>= dirty_ratelimit / (2 * step + 1);
1232 * Limit the tracking speed to avoid overshooting.
1234 step = (step + 7) / 8;
1236 if (dirty_ratelimit < balanced_dirty_ratelimit)
1237 dirty_ratelimit += step;
1239 dirty_ratelimit -= step;
1241 wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1242 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1244 trace_bdi_dirty_ratelimit(wb->bdi, dirty_rate, task_ratelimit);
1247 static void __wb_update_bandwidth(struct dirty_throttle_control *dtc,
1248 unsigned long start_time,
1249 bool update_ratelimit)
1251 struct bdi_writeback *wb = dtc->wb;
1252 unsigned long now = jiffies;
1253 unsigned long elapsed = now - wb->bw_time_stamp;
1254 unsigned long dirtied;
1255 unsigned long written;
1257 lockdep_assert_held(&wb->list_lock);
1260 * rate-limit, only update once every 200ms.
1262 if (elapsed < BANDWIDTH_INTERVAL)
1265 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1266 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1269 * Skip quiet periods when disk bandwidth is under-utilized.
1270 * (at least 1s idle time between two flusher runs)
1272 if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1275 if (update_ratelimit) {
1276 domain_update_bandwidth(dtc, now);
1277 wb_update_dirty_ratelimit(dtc, dirtied, elapsed);
1279 wb_update_write_bandwidth(wb, elapsed, written);
1282 wb->dirtied_stamp = dirtied;
1283 wb->written_stamp = written;
1284 wb->bw_time_stamp = now;
1287 void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1289 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1291 __wb_update_bandwidth(&gdtc, start_time, false);
1295 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1296 * will look to see if it needs to start dirty throttling.
1298 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1299 * global_page_state() too often. So scale it near-sqrt to the safety margin
1300 * (the number of pages we may dirty without exceeding the dirty limits).
1302 static unsigned long dirty_poll_interval(unsigned long dirty,
1303 unsigned long thresh)
1306 return 1UL << (ilog2(thresh - dirty) >> 1);
1311 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1312 unsigned long wb_dirty)
1314 unsigned long bw = wb->avg_write_bandwidth;
1318 * Limit pause time for small memory systems. If sleeping for too long
1319 * time, a small pool of dirty/writeback pages may go empty and disk go
1322 * 8 serves as the safety ratio.
1324 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1327 return min_t(unsigned long, t, MAX_PAUSE);
1330 static long wb_min_pause(struct bdi_writeback *wb,
1332 unsigned long task_ratelimit,
1333 unsigned long dirty_ratelimit,
1334 int *nr_dirtied_pause)
1336 long hi = ilog2(wb->avg_write_bandwidth);
1337 long lo = ilog2(wb->dirty_ratelimit);
1338 long t; /* target pause */
1339 long pause; /* estimated next pause */
1340 int pages; /* target nr_dirtied_pause */
1342 /* target for 10ms pause on 1-dd case */
1343 t = max(1, HZ / 100);
1346 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1349 * (N * 10ms) on 2^N concurrent tasks.
1352 t += (hi - lo) * (10 * HZ) / 1024;
1355 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1356 * on the much more stable dirty_ratelimit. However the next pause time
1357 * will be computed based on task_ratelimit and the two rate limits may
1358 * depart considerably at some time. Especially if task_ratelimit goes
1359 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1360 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1361 * result task_ratelimit won't be executed faithfully, which could
1362 * eventually bring down dirty_ratelimit.
1364 * We apply two rules to fix it up:
1365 * 1) try to estimate the next pause time and if necessary, use a lower
1366 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1367 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1368 * 2) limit the target pause time to max_pause/2, so that the normal
1369 * small fluctuations of task_ratelimit won't trigger rule (1) and
1370 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1372 t = min(t, 1 + max_pause / 2);
1373 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1376 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1377 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1378 * When the 16 consecutive reads are often interrupted by some dirty
1379 * throttling pause during the async writes, cfq will go into idles
1380 * (deadline is fine). So push nr_dirtied_pause as high as possible
1381 * until reaches DIRTY_POLL_THRESH=32 pages.
1383 if (pages < DIRTY_POLL_THRESH) {
1385 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1386 if (pages > DIRTY_POLL_THRESH) {
1387 pages = DIRTY_POLL_THRESH;
1388 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1392 pause = HZ * pages / (task_ratelimit + 1);
1393 if (pause > max_pause) {
1395 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1398 *nr_dirtied_pause = pages;
1400 * The minimal pause time will normally be half the target pause time.
1402 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1405 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1407 struct bdi_writeback *wb = dtc->wb;
1408 unsigned long wb_reclaimable;
1411 * wb_thresh is not treated as some limiting factor as
1412 * dirty_thresh, due to reasons
1413 * - in JBOD setup, wb_thresh can fluctuate a lot
1414 * - in a system with HDD and USB key, the USB key may somehow
1415 * go into state (wb_dirty >> wb_thresh) either because
1416 * wb_dirty starts high, or because wb_thresh drops low.
1417 * In this case we don't want to hard throttle the USB key
1418 * dirtiers for 100 seconds until wb_dirty drops under
1419 * wb_thresh. Instead the auxiliary wb control line in
1420 * wb_position_ratio() will let the dirtier task progress
1421 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1423 dtc->wb_thresh = __wb_calc_thresh(dtc);
1424 dtc->wb_bg_thresh = dtc->thresh ?
1425 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1428 * In order to avoid the stacked BDI deadlock we need
1429 * to ensure we accurately count the 'dirty' pages when
1430 * the threshold is low.
1432 * Otherwise it would be possible to get thresh+n pages
1433 * reported dirty, even though there are thresh-m pages
1434 * actually dirty; with m+n sitting in the percpu
1437 if (dtc->wb_thresh < 2 * wb_stat_error(wb)) {
1438 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1439 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1441 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1442 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1447 * balance_dirty_pages() must be called by processes which are generating dirty
1448 * data. It looks at the number of dirty pages in the machine and will force
1449 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1450 * If we're over `background_thresh' then the writeback threads are woken to
1451 * perform some writeout.
1453 static void balance_dirty_pages(struct address_space *mapping,
1454 struct bdi_writeback *wb,
1455 unsigned long pages_dirtied)
1457 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1458 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1459 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1464 int nr_dirtied_pause;
1465 bool dirty_exceeded = false;
1466 unsigned long task_ratelimit;
1467 unsigned long dirty_ratelimit;
1468 struct backing_dev_info *bdi = wb->bdi;
1469 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1470 unsigned long start_time = jiffies;
1473 unsigned long now = jiffies;
1474 unsigned long dirty, thresh, bg_thresh;
1477 * Unstable writes are a feature of certain networked
1478 * filesystems (i.e. NFS) in which data may have been
1479 * written to the server's write cache, but has not yet
1480 * been flushed to permanent storage.
1482 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1483 global_page_state(NR_UNSTABLE_NFS);
1484 gdtc->avail = global_dirtyable_memory();
1485 gdtc->dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1487 domain_dirty_limits(gdtc);
1489 if (unlikely(strictlimit)) {
1490 wb_dirty_limits(gdtc);
1492 dirty = gdtc->wb_dirty;
1493 thresh = gdtc->wb_thresh;
1494 bg_thresh = gdtc->wb_bg_thresh;
1496 dirty = gdtc->dirty;
1497 thresh = gdtc->thresh;
1498 bg_thresh = gdtc->bg_thresh;
1502 * Throttle it only when the background writeback cannot
1503 * catch-up. This avoids (excessively) small writeouts
1504 * when the wb limits are ramping up in case of !strictlimit.
1506 * In strictlimit case make decision based on the wb counters
1507 * and limits. Small writeouts when the wb limits are ramping
1508 * up are the price we consciously pay for strictlimit-ing.
1510 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh)) {
1511 current->dirty_paused_when = now;
1512 current->nr_dirtied = 0;
1513 current->nr_dirtied_pause =
1514 dirty_poll_interval(dirty, thresh);
1518 if (unlikely(!writeback_in_progress(wb)))
1519 wb_start_background_writeback(wb);
1522 wb_dirty_limits(gdtc);
1524 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1525 ((gdtc->dirty > gdtc->thresh) || strictlimit);
1527 wb_position_ratio(gdtc);
1529 if (dirty_exceeded && !wb->dirty_exceeded)
1530 wb->dirty_exceeded = 1;
1532 if (time_is_before_jiffies(wb->bw_time_stamp +
1533 BANDWIDTH_INTERVAL)) {
1534 spin_lock(&wb->list_lock);
1535 __wb_update_bandwidth(gdtc, start_time, true);
1536 spin_unlock(&wb->list_lock);
1539 dirty_ratelimit = wb->dirty_ratelimit;
1540 task_ratelimit = ((u64)dirty_ratelimit * gdtc->pos_ratio) >>
1541 RATELIMIT_CALC_SHIFT;
1542 max_pause = wb_max_pause(wb, gdtc->wb_dirty);
1543 min_pause = wb_min_pause(wb, max_pause,
1544 task_ratelimit, dirty_ratelimit,
1547 if (unlikely(task_ratelimit == 0)) {
1552 period = HZ * pages_dirtied / task_ratelimit;
1554 if (current->dirty_paused_when)
1555 pause -= now - current->dirty_paused_when;
1557 * For less than 1s think time (ext3/4 may block the dirtier
1558 * for up to 800ms from time to time on 1-HDD; so does xfs,
1559 * however at much less frequency), try to compensate it in
1560 * future periods by updating the virtual time; otherwise just
1561 * do a reset, as it may be a light dirtier.
1563 if (pause < min_pause) {
1564 trace_balance_dirty_pages(bdi,
1577 current->dirty_paused_when = now;
1578 current->nr_dirtied = 0;
1579 } else if (period) {
1580 current->dirty_paused_when += period;
1581 current->nr_dirtied = 0;
1582 } else if (current->nr_dirtied_pause <= pages_dirtied)
1583 current->nr_dirtied_pause += pages_dirtied;
1586 if (unlikely(pause > max_pause)) {
1587 /* for occasional dropped task_ratelimit */
1588 now += min(pause - max_pause, max_pause);
1593 trace_balance_dirty_pages(bdi,
1605 __set_current_state(TASK_KILLABLE);
1606 io_schedule_timeout(pause);
1608 current->dirty_paused_when = now + pause;
1609 current->nr_dirtied = 0;
1610 current->nr_dirtied_pause = nr_dirtied_pause;
1613 * This is typically equal to (dirty < thresh) and can also
1614 * keep "1000+ dd on a slow USB stick" under control.
1620 * In the case of an unresponding NFS server and the NFS dirty
1621 * pages exceeds dirty_thresh, give the other good wb's a pipe
1622 * to go through, so that tasks on them still remain responsive.
1624 * In theory 1 page is enough to keep the comsumer-producer
1625 * pipe going: the flusher cleans 1 page => the task dirties 1
1626 * more page. However wb_dirty has accounting errors. So use
1627 * the larger and more IO friendly wb_stat_error.
1629 if (gdtc->wb_dirty <= wb_stat_error(wb))
1632 if (fatal_signal_pending(current))
1636 if (!dirty_exceeded && wb->dirty_exceeded)
1637 wb->dirty_exceeded = 0;
1639 if (writeback_in_progress(wb))
1643 * In laptop mode, we wait until hitting the higher threshold before
1644 * starting background writeout, and then write out all the way down
1645 * to the lower threshold. So slow writers cause minimal disk activity.
1647 * In normal mode, we start background writeout at the lower
1648 * background_thresh, to keep the amount of dirty memory low.
1653 if (nr_reclaimable > gdtc->bg_thresh)
1654 wb_start_background_writeback(wb);
1657 static DEFINE_PER_CPU(int, bdp_ratelimits);
1660 * Normal tasks are throttled by
1662 * dirty tsk->nr_dirtied_pause pages;
1663 * take a snap in balance_dirty_pages();
1665 * However there is a worst case. If every task exit immediately when dirtied
1666 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1667 * called to throttle the page dirties. The solution is to save the not yet
1668 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1669 * randomly into the running tasks. This works well for the above worst case,
1670 * as the new task will pick up and accumulate the old task's leaked dirty
1671 * count and eventually get throttled.
1673 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1676 * balance_dirty_pages_ratelimited - balance dirty memory state
1677 * @mapping: address_space which was dirtied
1679 * Processes which are dirtying memory should call in here once for each page
1680 * which was newly dirtied. The function will periodically check the system's
1681 * dirty state and will initiate writeback if needed.
1683 * On really big machines, get_writeback_state is expensive, so try to avoid
1684 * calling it too often (ratelimiting). But once we're over the dirty memory
1685 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1686 * from overshooting the limit by (ratelimit_pages) each.
1688 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1690 struct inode *inode = mapping->host;
1691 struct backing_dev_info *bdi = inode_to_bdi(inode);
1692 struct bdi_writeback *wb = NULL;
1696 if (!bdi_cap_account_dirty(bdi))
1699 if (inode_cgwb_enabled(inode))
1700 wb = wb_get_create_current(bdi, GFP_KERNEL);
1704 ratelimit = current->nr_dirtied_pause;
1705 if (wb->dirty_exceeded)
1706 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1710 * This prevents one CPU to accumulate too many dirtied pages without
1711 * calling into balance_dirty_pages(), which can happen when there are
1712 * 1000+ tasks, all of them start dirtying pages at exactly the same
1713 * time, hence all honoured too large initial task->nr_dirtied_pause.
1715 p = this_cpu_ptr(&bdp_ratelimits);
1716 if (unlikely(current->nr_dirtied >= ratelimit))
1718 else if (unlikely(*p >= ratelimit_pages)) {
1723 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1724 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1725 * the dirty throttling and livelock other long-run dirtiers.
1727 p = this_cpu_ptr(&dirty_throttle_leaks);
1728 if (*p > 0 && current->nr_dirtied < ratelimit) {
1729 unsigned long nr_pages_dirtied;
1730 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1731 *p -= nr_pages_dirtied;
1732 current->nr_dirtied += nr_pages_dirtied;
1736 if (unlikely(current->nr_dirtied >= ratelimit))
1737 balance_dirty_pages(mapping, wb, current->nr_dirtied);
1741 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1744 * wb_over_bg_thresh - does @wb need to be written back?
1745 * @wb: bdi_writeback of interest
1747 * Determines whether background writeback should keep writing @wb or it's
1748 * clean enough. Returns %true if writeback should continue.
1750 bool wb_over_bg_thresh(struct bdi_writeback *wb)
1752 unsigned long background_thresh, dirty_thresh;
1754 global_dirty_limits(&background_thresh, &dirty_thresh);
1756 if (global_page_state(NR_FILE_DIRTY) +
1757 global_page_state(NR_UNSTABLE_NFS) > background_thresh)
1760 if (wb_stat(wb, WB_RECLAIMABLE) > wb_calc_thresh(wb, background_thresh))
1766 void throttle_vm_writeout(gfp_t gfp_mask)
1768 unsigned long background_thresh;
1769 unsigned long dirty_thresh;
1772 global_dirty_limits(&background_thresh, &dirty_thresh);
1773 dirty_thresh = hard_dirty_limit(&global_wb_domain, dirty_thresh);
1776 * Boost the allowable dirty threshold a bit for page
1777 * allocators so they don't get DoS'ed by heavy writers
1779 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1781 if (global_page_state(NR_UNSTABLE_NFS) +
1782 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1784 congestion_wait(BLK_RW_ASYNC, HZ/10);
1787 * The caller might hold locks which can prevent IO completion
1788 * or progress in the filesystem. So we cannot just sit here
1789 * waiting for IO to complete.
1791 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1797 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1799 int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1800 void __user *buffer, size_t *length, loff_t *ppos)
1802 proc_dointvec(table, write, buffer, length, ppos);
1807 void laptop_mode_timer_fn(unsigned long data)
1809 struct request_queue *q = (struct request_queue *)data;
1810 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1811 global_page_state(NR_UNSTABLE_NFS);
1812 struct bdi_writeback *wb;
1813 struct wb_iter iter;
1816 * We want to write everything out, not just down to the dirty
1819 if (!bdi_has_dirty_io(&q->backing_dev_info))
1822 bdi_for_each_wb(wb, &q->backing_dev_info, &iter, 0)
1823 if (wb_has_dirty_io(wb))
1824 wb_start_writeback(wb, nr_pages, true,
1825 WB_REASON_LAPTOP_TIMER);
1829 * We've spun up the disk and we're in laptop mode: schedule writeback
1830 * of all dirty data a few seconds from now. If the flush is already scheduled
1831 * then push it back - the user is still using the disk.
1833 void laptop_io_completion(struct backing_dev_info *info)
1835 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1839 * We're in laptop mode and we've just synced. The sync's writes will have
1840 * caused another writeback to be scheduled by laptop_io_completion.
1841 * Nothing needs to be written back anymore, so we unschedule the writeback.
1843 void laptop_sync_completion(void)
1845 struct backing_dev_info *bdi;
1849 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1850 del_timer(&bdi->laptop_mode_wb_timer);
1857 * If ratelimit_pages is too high then we can get into dirty-data overload
1858 * if a large number of processes all perform writes at the same time.
1859 * If it is too low then SMP machines will call the (expensive)
1860 * get_writeback_state too often.
1862 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1863 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1867 void writeback_set_ratelimit(void)
1869 struct wb_domain *dom = &global_wb_domain;
1870 unsigned long background_thresh;
1871 unsigned long dirty_thresh;
1873 global_dirty_limits(&background_thresh, &dirty_thresh);
1874 dom->dirty_limit = dirty_thresh;
1875 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1876 if (ratelimit_pages < 16)
1877 ratelimit_pages = 16;
1881 ratelimit_handler(struct notifier_block *self, unsigned long action,
1885 switch (action & ~CPU_TASKS_FROZEN) {
1888 writeback_set_ratelimit();
1895 static struct notifier_block ratelimit_nb = {
1896 .notifier_call = ratelimit_handler,
1901 * Called early on to tune the page writeback dirty limits.
1903 * We used to scale dirty pages according to how total memory
1904 * related to pages that could be allocated for buffers (by
1905 * comparing nr_free_buffer_pages() to vm_total_pages.
1907 * However, that was when we used "dirty_ratio" to scale with
1908 * all memory, and we don't do that any more. "dirty_ratio"
1909 * is now applied to total non-HIGHPAGE memory (by subtracting
1910 * totalhigh_pages from vm_total_pages), and as such we can't
1911 * get into the old insane situation any more where we had
1912 * large amounts of dirty pages compared to a small amount of
1913 * non-HIGHMEM memory.
1915 * But we might still want to scale the dirty_ratio by how
1916 * much memory the box has..
1918 void __init page_writeback_init(void)
1920 writeback_set_ratelimit();
1921 register_cpu_notifier(&ratelimit_nb);
1923 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
1927 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1928 * @mapping: address space structure to write
1929 * @start: starting page index
1930 * @end: ending page index (inclusive)
1932 * This function scans the page range from @start to @end (inclusive) and tags
1933 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1934 * that write_cache_pages (or whoever calls this function) will then use
1935 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1936 * used to avoid livelocking of writeback by a process steadily creating new
1937 * dirty pages in the file (thus it is important for this function to be quick
1938 * so that it can tag pages faster than a dirtying process can create them).
1941 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1943 void tag_pages_for_writeback(struct address_space *mapping,
1944 pgoff_t start, pgoff_t end)
1946 #define WRITEBACK_TAG_BATCH 4096
1947 unsigned long tagged;
1950 spin_lock_irq(&mapping->tree_lock);
1951 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1952 &start, end, WRITEBACK_TAG_BATCH,
1953 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1954 spin_unlock_irq(&mapping->tree_lock);
1955 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1957 /* We check 'start' to handle wrapping when end == ~0UL */
1958 } while (tagged >= WRITEBACK_TAG_BATCH && start);
1960 EXPORT_SYMBOL(tag_pages_for_writeback);
1963 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1964 * @mapping: address space structure to write
1965 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1966 * @writepage: function called for each page
1967 * @data: data passed to writepage function
1969 * If a page is already under I/O, write_cache_pages() skips it, even
1970 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1971 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1972 * and msync() need to guarantee that all the data which was dirty at the time
1973 * the call was made get new I/O started against them. If wbc->sync_mode is
1974 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1975 * existing IO to complete.
1977 * To avoid livelocks (when other process dirties new pages), we first tag
1978 * pages which should be written back with TOWRITE tag and only then start
1979 * writing them. For data-integrity sync we have to be careful so that we do
1980 * not miss some pages (e.g., because some other process has cleared TOWRITE
1981 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1982 * by the process clearing the DIRTY tag (and submitting the page for IO).
1984 int write_cache_pages(struct address_space *mapping,
1985 struct writeback_control *wbc, writepage_t writepage,
1990 struct pagevec pvec;
1992 pgoff_t uninitialized_var(writeback_index);
1994 pgoff_t end; /* Inclusive */
1997 int range_whole = 0;
2000 pagevec_init(&pvec, 0);
2001 if (wbc->range_cyclic) {
2002 writeback_index = mapping->writeback_index; /* prev offset */
2003 index = writeback_index;
2010 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2011 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2012 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2014 cycled = 1; /* ignore range_cyclic tests */
2016 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2017 tag = PAGECACHE_TAG_TOWRITE;
2019 tag = PAGECACHE_TAG_DIRTY;
2021 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2022 tag_pages_for_writeback(mapping, index, end);
2024 while (!done && (index <= end)) {
2027 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2028 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2032 for (i = 0; i < nr_pages; i++) {
2033 struct page *page = pvec.pages[i];
2036 * At this point, the page may be truncated or
2037 * invalidated (changing page->mapping to NULL), or
2038 * even swizzled back from swapper_space to tmpfs file
2039 * mapping. However, page->index will not change
2040 * because we have a reference on the page.
2042 if (page->index > end) {
2044 * can't be range_cyclic (1st pass) because
2045 * end == -1 in that case.
2051 done_index = page->index;
2056 * Page truncated or invalidated. We can freely skip it
2057 * then, even for data integrity operations: the page
2058 * has disappeared concurrently, so there could be no
2059 * real expectation of this data interity operation
2060 * even if there is now a new, dirty page at the same
2061 * pagecache address.
2063 if (unlikely(page->mapping != mapping)) {
2069 if (!PageDirty(page)) {
2070 /* someone wrote it for us */
2071 goto continue_unlock;
2074 if (PageWriteback(page)) {
2075 if (wbc->sync_mode != WB_SYNC_NONE)
2076 wait_on_page_writeback(page);
2078 goto continue_unlock;
2081 BUG_ON(PageWriteback(page));
2082 if (!clear_page_dirty_for_io(page))
2083 goto continue_unlock;
2085 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2086 ret = (*writepage)(page, wbc, data);
2087 if (unlikely(ret)) {
2088 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2093 * done_index is set past this page,
2094 * so media errors will not choke
2095 * background writeout for the entire
2096 * file. This has consequences for
2097 * range_cyclic semantics (ie. it may
2098 * not be suitable for data integrity
2101 done_index = page->index + 1;
2108 * We stop writing back only if we are not doing
2109 * integrity sync. In case of integrity sync we have to
2110 * keep going until we have written all the pages
2111 * we tagged for writeback prior to entering this loop.
2113 if (--wbc->nr_to_write <= 0 &&
2114 wbc->sync_mode == WB_SYNC_NONE) {
2119 pagevec_release(&pvec);
2122 if (!cycled && !done) {
2125 * We hit the last page and there is more work to be done: wrap
2126 * back to the start of the file
2130 end = writeback_index - 1;
2133 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2134 mapping->writeback_index = done_index;
2138 EXPORT_SYMBOL(write_cache_pages);
2141 * Function used by generic_writepages to call the real writepage
2142 * function and set the mapping flags on error
2144 static int __writepage(struct page *page, struct writeback_control *wbc,
2147 struct address_space *mapping = data;
2148 int ret = mapping->a_ops->writepage(page, wbc);
2149 mapping_set_error(mapping, ret);
2154 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2155 * @mapping: address space structure to write
2156 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2158 * This is a library function, which implements the writepages()
2159 * address_space_operation.
2161 int generic_writepages(struct address_space *mapping,
2162 struct writeback_control *wbc)
2164 struct blk_plug plug;
2167 /* deal with chardevs and other special file */
2168 if (!mapping->a_ops->writepage)
2171 blk_start_plug(&plug);
2172 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2173 blk_finish_plug(&plug);
2177 EXPORT_SYMBOL(generic_writepages);
2179 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2183 if (wbc->nr_to_write <= 0)
2185 if (mapping->a_ops->writepages)
2186 ret = mapping->a_ops->writepages(mapping, wbc);
2188 ret = generic_writepages(mapping, wbc);
2193 * write_one_page - write out a single page and optionally wait on I/O
2194 * @page: the page to write
2195 * @wait: if true, wait on writeout
2197 * The page must be locked by the caller and will be unlocked upon return.
2199 * write_one_page() returns a negative error code if I/O failed.
2201 int write_one_page(struct page *page, int wait)
2203 struct address_space *mapping = page->mapping;
2205 struct writeback_control wbc = {
2206 .sync_mode = WB_SYNC_ALL,
2210 BUG_ON(!PageLocked(page));
2213 wait_on_page_writeback(page);
2215 if (clear_page_dirty_for_io(page)) {
2216 page_cache_get(page);
2217 ret = mapping->a_ops->writepage(page, &wbc);
2218 if (ret == 0 && wait) {
2219 wait_on_page_writeback(page);
2220 if (PageError(page))
2223 page_cache_release(page);
2229 EXPORT_SYMBOL(write_one_page);
2232 * For address_spaces which do not use buffers nor write back.
2234 int __set_page_dirty_no_writeback(struct page *page)
2236 if (!PageDirty(page))
2237 return !TestSetPageDirty(page);
2242 * Helper function for set_page_dirty family.
2244 * Caller must hold mem_cgroup_begin_page_stat().
2246 * NOTE: This relies on being atomic wrt interrupts.
2248 void account_page_dirtied(struct page *page, struct address_space *mapping,
2249 struct mem_cgroup *memcg)
2251 struct inode *inode = mapping->host;
2253 trace_writeback_dirty_page(page, mapping);
2255 if (mapping_cap_account_dirty(mapping)) {
2256 struct bdi_writeback *wb;
2258 inode_attach_wb(inode, page);
2259 wb = inode_to_wb(inode);
2261 mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
2262 __inc_zone_page_state(page, NR_FILE_DIRTY);
2263 __inc_zone_page_state(page, NR_DIRTIED);
2264 __inc_wb_stat(wb, WB_RECLAIMABLE);
2265 __inc_wb_stat(wb, WB_DIRTIED);
2266 task_io_account_write(PAGE_CACHE_SIZE);
2267 current->nr_dirtied++;
2268 this_cpu_inc(bdp_ratelimits);
2271 EXPORT_SYMBOL(account_page_dirtied);
2274 * Helper function for deaccounting dirty page without writeback.
2276 * Caller must hold mem_cgroup_begin_page_stat().
2278 void account_page_cleaned(struct page *page, struct address_space *mapping,
2279 struct mem_cgroup *memcg)
2281 if (mapping_cap_account_dirty(mapping)) {
2282 mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
2283 dec_zone_page_state(page, NR_FILE_DIRTY);
2284 dec_wb_stat(inode_to_wb(mapping->host), WB_RECLAIMABLE);
2285 task_io_account_cancelled_write(PAGE_CACHE_SIZE);
2290 * For address_spaces which do not use buffers. Just tag the page as dirty in
2293 * This is also used when a single buffer is being dirtied: we want to set the
2294 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2295 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2297 * The caller must ensure this doesn't race with truncation. Most will simply
2298 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2299 * the pte lock held, which also locks out truncation.
2301 int __set_page_dirty_nobuffers(struct page *page)
2303 struct mem_cgroup *memcg;
2305 memcg = mem_cgroup_begin_page_stat(page);
2306 if (!TestSetPageDirty(page)) {
2307 struct address_space *mapping = page_mapping(page);
2308 unsigned long flags;
2311 mem_cgroup_end_page_stat(memcg);
2315 spin_lock_irqsave(&mapping->tree_lock, flags);
2316 BUG_ON(page_mapping(page) != mapping);
2317 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2318 account_page_dirtied(page, mapping, memcg);
2319 radix_tree_tag_set(&mapping->page_tree, page_index(page),
2320 PAGECACHE_TAG_DIRTY);
2321 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2322 mem_cgroup_end_page_stat(memcg);
2324 if (mapping->host) {
2325 /* !PageAnon && !swapper_space */
2326 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2330 mem_cgroup_end_page_stat(memcg);
2333 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2336 * Call this whenever redirtying a page, to de-account the dirty counters
2337 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2338 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2339 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2342 void account_page_redirty(struct page *page)
2344 struct address_space *mapping = page->mapping;
2346 if (mapping && mapping_cap_account_dirty(mapping)) {
2347 struct bdi_writeback *wb = inode_to_wb(mapping->host);
2349 current->nr_dirtied--;
2350 dec_zone_page_state(page, NR_DIRTIED);
2351 dec_wb_stat(wb, WB_DIRTIED);
2354 EXPORT_SYMBOL(account_page_redirty);
2357 * When a writepage implementation decides that it doesn't want to write this
2358 * page for some reason, it should redirty the locked page via
2359 * redirty_page_for_writepage() and it should then unlock the page and return 0
2361 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2365 wbc->pages_skipped++;
2366 ret = __set_page_dirty_nobuffers(page);
2367 account_page_redirty(page);
2370 EXPORT_SYMBOL(redirty_page_for_writepage);
2375 * For pages with a mapping this should be done under the page lock
2376 * for the benefit of asynchronous memory errors who prefer a consistent
2377 * dirty state. This rule can be broken in some special cases,
2378 * but should be better not to.
2380 * If the mapping doesn't provide a set_page_dirty a_op, then
2381 * just fall through and assume that it wants buffer_heads.
2383 int set_page_dirty(struct page *page)
2385 struct address_space *mapping = page_mapping(page);
2387 if (likely(mapping)) {
2388 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2390 * readahead/lru_deactivate_page could remain
2391 * PG_readahead/PG_reclaim due to race with end_page_writeback
2392 * About readahead, if the page is written, the flags would be
2393 * reset. So no problem.
2394 * About lru_deactivate_page, if the page is redirty, the flag
2395 * will be reset. So no problem. but if the page is used by readahead
2396 * it will confuse readahead and make it restart the size rampup
2397 * process. But it's a trivial problem.
2399 if (PageReclaim(page))
2400 ClearPageReclaim(page);
2403 spd = __set_page_dirty_buffers;
2405 return (*spd)(page);
2407 if (!PageDirty(page)) {
2408 if (!TestSetPageDirty(page))
2413 EXPORT_SYMBOL(set_page_dirty);
2416 * set_page_dirty() is racy if the caller has no reference against
2417 * page->mapping->host, and if the page is unlocked. This is because another
2418 * CPU could truncate the page off the mapping and then free the mapping.
2420 * Usually, the page _is_ locked, or the caller is a user-space process which
2421 * holds a reference on the inode by having an open file.
2423 * In other cases, the page should be locked before running set_page_dirty().
2425 int set_page_dirty_lock(struct page *page)
2430 ret = set_page_dirty(page);
2434 EXPORT_SYMBOL(set_page_dirty_lock);
2437 * This cancels just the dirty bit on the kernel page itself, it does NOT
2438 * actually remove dirty bits on any mmap's that may be around. It also
2439 * leaves the page tagged dirty, so any sync activity will still find it on
2440 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2441 * look at the dirty bits in the VM.
2443 * Doing this should *normally* only ever be done when a page is truncated,
2444 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2445 * this when it notices that somebody has cleaned out all the buffers on a
2446 * page without actually doing it through the VM. Can you say "ext3 is
2447 * horribly ugly"? Thought you could.
2449 void cancel_dirty_page(struct page *page)
2451 struct address_space *mapping = page_mapping(page);
2453 if (mapping_cap_account_dirty(mapping)) {
2454 struct mem_cgroup *memcg;
2456 memcg = mem_cgroup_begin_page_stat(page);
2458 if (TestClearPageDirty(page))
2459 account_page_cleaned(page, mapping, memcg);
2461 mem_cgroup_end_page_stat(memcg);
2463 ClearPageDirty(page);
2466 EXPORT_SYMBOL(cancel_dirty_page);
2469 * Clear a page's dirty flag, while caring for dirty memory accounting.
2470 * Returns true if the page was previously dirty.
2472 * This is for preparing to put the page under writeout. We leave the page
2473 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2474 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2475 * implementation will run either set_page_writeback() or set_page_dirty(),
2476 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2479 * This incoherency between the page's dirty flag and radix-tree tag is
2480 * unfortunate, but it only exists while the page is locked.
2482 int clear_page_dirty_for_io(struct page *page)
2484 struct address_space *mapping = page_mapping(page);
2485 struct mem_cgroup *memcg;
2488 BUG_ON(!PageLocked(page));
2490 if (mapping && mapping_cap_account_dirty(mapping)) {
2492 * Yes, Virginia, this is indeed insane.
2494 * We use this sequence to make sure that
2495 * (a) we account for dirty stats properly
2496 * (b) we tell the low-level filesystem to
2497 * mark the whole page dirty if it was
2498 * dirty in a pagetable. Only to then
2499 * (c) clean the page again and return 1 to
2500 * cause the writeback.
2502 * This way we avoid all nasty races with the
2503 * dirty bit in multiple places and clearing
2504 * them concurrently from different threads.
2506 * Note! Normally the "set_page_dirty(page)"
2507 * has no effect on the actual dirty bit - since
2508 * that will already usually be set. But we
2509 * need the side effects, and it can help us
2512 * We basically use the page "master dirty bit"
2513 * as a serialization point for all the different
2514 * threads doing their things.
2516 if (page_mkclean(page))
2517 set_page_dirty(page);
2519 * We carefully synchronise fault handlers against
2520 * installing a dirty pte and marking the page dirty
2521 * at this point. We do this by having them hold the
2522 * page lock while dirtying the page, and pages are
2523 * always locked coming in here, so we get the desired
2526 memcg = mem_cgroup_begin_page_stat(page);
2527 if (TestClearPageDirty(page)) {
2528 mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
2529 dec_zone_page_state(page, NR_FILE_DIRTY);
2530 dec_wb_stat(inode_to_wb(mapping->host), WB_RECLAIMABLE);
2533 mem_cgroup_end_page_stat(memcg);
2536 return TestClearPageDirty(page);
2538 EXPORT_SYMBOL(clear_page_dirty_for_io);
2540 int test_clear_page_writeback(struct page *page)
2542 struct address_space *mapping = page_mapping(page);
2543 struct mem_cgroup *memcg;
2546 memcg = mem_cgroup_begin_page_stat(page);
2548 struct inode *inode = mapping->host;
2549 struct backing_dev_info *bdi = inode_to_bdi(inode);
2550 unsigned long flags;
2552 spin_lock_irqsave(&mapping->tree_lock, flags);
2553 ret = TestClearPageWriteback(page);
2555 radix_tree_tag_clear(&mapping->page_tree,
2557 PAGECACHE_TAG_WRITEBACK);
2558 if (bdi_cap_account_writeback(bdi)) {
2559 struct bdi_writeback *wb = inode_to_wb(inode);
2561 __dec_wb_stat(wb, WB_WRITEBACK);
2562 __wb_writeout_inc(wb);
2565 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2567 ret = TestClearPageWriteback(page);
2570 mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
2571 dec_zone_page_state(page, NR_WRITEBACK);
2572 inc_zone_page_state(page, NR_WRITTEN);
2574 mem_cgroup_end_page_stat(memcg);
2578 int __test_set_page_writeback(struct page *page, bool keep_write)
2580 struct address_space *mapping = page_mapping(page);
2581 struct mem_cgroup *memcg;
2584 memcg = mem_cgroup_begin_page_stat(page);
2586 struct inode *inode = mapping->host;
2587 struct backing_dev_info *bdi = inode_to_bdi(inode);
2588 unsigned long flags;
2590 spin_lock_irqsave(&mapping->tree_lock, flags);
2591 ret = TestSetPageWriteback(page);
2593 radix_tree_tag_set(&mapping->page_tree,
2595 PAGECACHE_TAG_WRITEBACK);
2596 if (bdi_cap_account_writeback(bdi))
2597 __inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2599 if (!PageDirty(page))
2600 radix_tree_tag_clear(&mapping->page_tree,
2602 PAGECACHE_TAG_DIRTY);
2604 radix_tree_tag_clear(&mapping->page_tree,
2606 PAGECACHE_TAG_TOWRITE);
2607 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2609 ret = TestSetPageWriteback(page);
2612 mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
2613 inc_zone_page_state(page, NR_WRITEBACK);
2615 mem_cgroup_end_page_stat(memcg);
2619 EXPORT_SYMBOL(__test_set_page_writeback);
2622 * Return true if any of the pages in the mapping are marked with the
2625 int mapping_tagged(struct address_space *mapping, int tag)
2627 return radix_tree_tagged(&mapping->page_tree, tag);
2629 EXPORT_SYMBOL(mapping_tagged);
2632 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2633 * @page: The page to wait on.
2635 * This function determines if the given page is related to a backing device
2636 * that requires page contents to be held stable during writeback. If so, then
2637 * it will wait for any pending writeback to complete.
2639 void wait_for_stable_page(struct page *page)
2641 if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2642 wait_on_page_writeback(page);
2644 EXPORT_SYMBOL_GPL(wait_for_stable_page);