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 <trace/events/writeback.h>
42 * Sleep at most 200ms at a time in balance_dirty_pages().
44 #define MAX_PAUSE max(HZ/5, 1)
47 * Try to keep balance_dirty_pages() call intervals higher than this many pages
48 * by raising pause time to max_pause when falls below it.
50 #define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
53 * Estimate write bandwidth at 200ms intervals.
55 #define BANDWIDTH_INTERVAL max(HZ/5, 1)
57 #define RATELIMIT_CALC_SHIFT 10
60 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
61 * will look to see if it needs to force writeback or throttling.
63 static long ratelimit_pages = 32;
65 /* The following parameters are exported via /proc/sys/vm */
68 * Start background writeback (via writeback threads) at this percentage
70 int dirty_background_ratio = 10;
73 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
74 * dirty_background_ratio * the amount of dirtyable memory
76 unsigned long dirty_background_bytes;
79 * free highmem will not be subtracted from the total free memory
80 * for calculating free ratios if vm_highmem_is_dirtyable is true
82 int vm_highmem_is_dirtyable;
85 * The generator of dirty data starts writeback at this percentage
87 int vm_dirty_ratio = 20;
90 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
91 * vm_dirty_ratio * the amount of dirtyable memory
93 unsigned long vm_dirty_bytes;
96 * The interval between `kupdate'-style writebacks
98 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
100 EXPORT_SYMBOL_GPL(dirty_writeback_interval);
103 * The longest time for which data is allowed to remain dirty
105 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
108 * Flag that makes the machine dump writes/reads and block dirtyings.
113 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
114 * a full sync is triggered after this time elapses without any disk activity.
118 EXPORT_SYMBOL(laptop_mode);
120 /* End of sysctl-exported parameters */
122 unsigned long global_dirty_limit;
125 * Scale the writeback cache size proportional to the relative writeout speeds.
127 * We do this by keeping a floating proportion between BDIs, based on page
128 * writeback completions [end_page_writeback()]. Those devices that write out
129 * pages fastest will get the larger share, while the slower will get a smaller
132 * We use page writeout completions because we are interested in getting rid of
133 * dirty pages. Having them written out is the primary goal.
135 * We introduce a concept of time, a period over which we measure these events,
136 * because demand can/will vary over time. The length of this period itself is
137 * measured in page writeback completions.
140 static struct fprop_global writeout_completions;
142 static void writeout_period(unsigned long t);
143 /* Timer for aging of writeout_completions */
144 static struct timer_list writeout_period_timer =
145 TIMER_DEFERRED_INITIALIZER(writeout_period, 0, 0);
146 static unsigned long writeout_period_time = 0;
149 * Length of period for aging writeout fractions of bdis. This is an
150 * arbitrarily chosen number. The longer the period, the slower fractions will
151 * reflect changes in current writeout rate.
153 #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
156 * Work out the current dirty-memory clamping and background writeout
159 * The main aim here is to lower them aggressively if there is a lot of mapped
160 * memory around. To avoid stressing page reclaim with lots of unreclaimable
161 * pages. It is better to clamp down on writers than to start swapping, and
162 * performing lots of scanning.
164 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
166 * We don't permit the clamping level to fall below 5% - that is getting rather
169 * We make sure that the background writeout level is below the adjusted
174 * In a memory zone, there is a certain amount of pages we consider
175 * available for the page cache, which is essentially the number of
176 * free and reclaimable pages, minus some zone reserves to protect
177 * lowmem and the ability to uphold the zone's watermarks without
178 * requiring writeback.
180 * This number of dirtyable pages is the base value of which the
181 * user-configurable dirty ratio is the effictive number of pages that
182 * are allowed to be actually dirtied. Per individual zone, or
183 * globally by using the sum of dirtyable pages over all zones.
185 * Because the user is allowed to specify the dirty limit globally as
186 * absolute number of bytes, calculating the per-zone dirty limit can
187 * require translating the configured limit into a percentage of
188 * global dirtyable memory first.
192 * zone_dirtyable_memory - number of dirtyable pages in a zone
195 * Returns the zone's number of pages potentially available for dirty
196 * page cache. This is the base value for the per-zone dirty limits.
198 static unsigned long zone_dirtyable_memory(struct zone *zone)
200 unsigned long nr_pages;
202 nr_pages = zone_page_state(zone, NR_FREE_PAGES);
203 nr_pages -= min(nr_pages, zone->dirty_balance_reserve);
205 nr_pages += zone_reclaimable_pages(zone);
210 static unsigned long highmem_dirtyable_memory(unsigned long total)
212 #ifdef CONFIG_HIGHMEM
216 for_each_node_state(node, N_HIGH_MEMORY) {
217 struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
219 x += zone_dirtyable_memory(z);
222 * Unreclaimable memory (kernel memory or anonymous memory
223 * without swap) can bring down the dirtyable pages below
224 * the zone's dirty balance reserve and the above calculation
225 * will underflow. However we still want to add in nodes
226 * which are below threshold (negative values) to get a more
227 * accurate calculation but make sure that the total never
234 * Make sure that the number of highmem pages is never larger
235 * than the number of the total dirtyable memory. This can only
236 * occur in very strange VM situations but we want to make sure
237 * that this does not occur.
239 return min(x, total);
246 * global_dirtyable_memory - number of globally dirtyable pages
248 * Returns the global number of pages potentially available for dirty
249 * page cache. This is the base value for the global dirty limits.
251 static unsigned long global_dirtyable_memory(void)
255 x = global_page_state(NR_FREE_PAGES);
256 x -= min(x, dirty_balance_reserve);
258 x += global_reclaimable_pages();
260 if (!vm_highmem_is_dirtyable)
261 x -= highmem_dirtyable_memory(x);
263 /* Subtract min_free_kbytes */
264 x -= min_t(unsigned long, x, min_free_kbytes >> (PAGE_SHIFT - 10));
266 return x + 1; /* Ensure that we never return 0 */
270 * global_dirty_limits - background-writeback and dirty-throttling thresholds
272 * Calculate the dirty thresholds based on sysctl parameters
273 * - vm.dirty_background_ratio or vm.dirty_background_bytes
274 * - vm.dirty_ratio or vm.dirty_bytes
275 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
278 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
280 unsigned long background;
282 unsigned long uninitialized_var(available_memory);
283 struct task_struct *tsk;
285 if (!vm_dirty_bytes || !dirty_background_bytes)
286 available_memory = global_dirtyable_memory();
289 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
291 dirty = (vm_dirty_ratio * available_memory) / 100;
293 if (dirty_background_bytes)
294 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
296 background = (dirty_background_ratio * available_memory) / 100;
298 if (background >= dirty)
299 background = dirty / 2;
301 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
302 background += background / 4;
305 *pbackground = background;
307 trace_global_dirty_state(background, dirty);
311 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
314 * Returns the maximum number of dirty pages allowed in a zone, based
315 * on the zone's dirtyable memory.
317 static unsigned long zone_dirty_limit(struct zone *zone)
319 unsigned long zone_memory = zone_dirtyable_memory(zone);
320 struct task_struct *tsk = current;
324 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
325 zone_memory / global_dirtyable_memory();
327 dirty = vm_dirty_ratio * zone_memory / 100;
329 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
336 * zone_dirty_ok - tells whether a zone is within its dirty limits
337 * @zone: the zone to check
339 * Returns %true when the dirty pages in @zone are within the zone's
340 * dirty limit, %false if the limit is exceeded.
342 bool zone_dirty_ok(struct zone *zone)
344 unsigned long limit = zone_dirty_limit(zone);
346 return zone_page_state(zone, NR_FILE_DIRTY) +
347 zone_page_state(zone, NR_UNSTABLE_NFS) +
348 zone_page_state(zone, NR_WRITEBACK) <= limit;
351 int dirty_background_ratio_handler(struct ctl_table *table, int write,
352 void __user *buffer, size_t *lenp,
357 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
358 if (ret == 0 && write)
359 dirty_background_bytes = 0;
363 int dirty_background_bytes_handler(struct ctl_table *table, int write,
364 void __user *buffer, size_t *lenp,
369 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
370 if (ret == 0 && write)
371 dirty_background_ratio = 0;
375 int dirty_ratio_handler(struct ctl_table *table, int write,
376 void __user *buffer, size_t *lenp,
379 int old_ratio = vm_dirty_ratio;
382 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
383 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
384 writeback_set_ratelimit();
390 int dirty_bytes_handler(struct ctl_table *table, int write,
391 void __user *buffer, size_t *lenp,
394 unsigned long old_bytes = vm_dirty_bytes;
397 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
398 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
399 writeback_set_ratelimit();
405 static unsigned long wp_next_time(unsigned long cur_time)
407 cur_time += VM_COMPLETIONS_PERIOD_LEN;
408 /* 0 has a special meaning... */
415 * Increment the BDI's writeout completion count and the global writeout
416 * completion count. Called from test_clear_page_writeback().
418 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
420 __inc_bdi_stat(bdi, BDI_WRITTEN);
421 __fprop_inc_percpu_max(&writeout_completions, &bdi->completions,
423 /* First event after period switching was turned off? */
424 if (!unlikely(writeout_period_time)) {
426 * We can race with other __bdi_writeout_inc calls here but
427 * it does not cause any harm since the resulting time when
428 * timer will fire and what is in writeout_period_time will be
431 writeout_period_time = wp_next_time(jiffies);
432 mod_timer(&writeout_period_timer, writeout_period_time);
436 void bdi_writeout_inc(struct backing_dev_info *bdi)
440 local_irq_save(flags);
441 __bdi_writeout_inc(bdi);
442 local_irq_restore(flags);
444 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
447 * Obtain an accurate fraction of the BDI's portion.
449 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
450 long *numerator, long *denominator)
452 fprop_fraction_percpu(&writeout_completions, &bdi->completions,
453 numerator, denominator);
457 * On idle system, we can be called long after we scheduled because we use
458 * deferred timers so count with missed periods.
460 static void writeout_period(unsigned long t)
462 int miss_periods = (jiffies - writeout_period_time) /
463 VM_COMPLETIONS_PERIOD_LEN;
465 if (fprop_new_period(&writeout_completions, miss_periods + 1)) {
466 writeout_period_time = wp_next_time(writeout_period_time +
467 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
468 mod_timer(&writeout_period_timer, writeout_period_time);
471 * Aging has zeroed all fractions. Stop wasting CPU on period
474 writeout_period_time = 0;
479 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
480 * registered backing devices, which, for obvious reasons, can not
483 static unsigned int bdi_min_ratio;
485 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
489 spin_lock_bh(&bdi_lock);
490 if (min_ratio > bdi->max_ratio) {
493 min_ratio -= bdi->min_ratio;
494 if (bdi_min_ratio + min_ratio < 100) {
495 bdi_min_ratio += min_ratio;
496 bdi->min_ratio += min_ratio;
501 spin_unlock_bh(&bdi_lock);
506 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
513 spin_lock_bh(&bdi_lock);
514 if (bdi->min_ratio > max_ratio) {
517 bdi->max_ratio = max_ratio;
518 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
520 spin_unlock_bh(&bdi_lock);
524 EXPORT_SYMBOL(bdi_set_max_ratio);
526 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
527 unsigned long bg_thresh)
529 return (thresh + bg_thresh) / 2;
532 static unsigned long hard_dirty_limit(unsigned long thresh)
534 return max(thresh, global_dirty_limit);
538 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
539 * @bdi: the backing_dev_info to query
540 * @dirty: global dirty limit in pages
542 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
543 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
545 * Note that balance_dirty_pages() will only seriously take it as a hard limit
546 * when sleeping max_pause per page is not enough to keep the dirty pages under
547 * control. For example, when the device is completely stalled due to some error
548 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
549 * In the other normal situations, it acts more gently by throttling the tasks
550 * more (rather than completely block them) when the bdi dirty pages go high.
552 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
553 * - starving fast devices
554 * - piling up dirty pages (that will take long time to sync) on slow devices
556 * The bdi's share of dirty limit will be adapting to its throughput and
557 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
559 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
562 long numerator, denominator;
565 * Calculate this BDI's share of the dirty ratio.
567 bdi_writeout_fraction(bdi, &numerator, &denominator);
569 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
570 bdi_dirty *= numerator;
571 do_div(bdi_dirty, denominator);
573 bdi_dirty += (dirty * bdi->min_ratio) / 100;
574 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
575 bdi_dirty = dirty * bdi->max_ratio / 100;
581 * Dirty position control.
583 * (o) global/bdi setpoints
585 * We want the dirty pages be balanced around the global/bdi setpoints.
586 * When the number of dirty pages is higher/lower than the setpoint, the
587 * dirty position control ratio (and hence task dirty ratelimit) will be
588 * decreased/increased to bring the dirty pages back to the setpoint.
590 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
592 * if (dirty < setpoint) scale up pos_ratio
593 * if (dirty > setpoint) scale down pos_ratio
595 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
596 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
598 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
600 * (o) global control line
604 * | |<===== global dirty control scope ======>|
612 * 1.0 ................................*
618 * 0 +------------.------------------.----------------------*------------->
619 * freerun^ setpoint^ limit^ dirty pages
621 * (o) bdi control line
629 * | * |<=========== span ============>|
630 * 1.0 .......................*
642 * 1/4 ...............................................* * * * * * * * * * * *
646 * 0 +----------------------.-------------------------------.------------->
647 * bdi_setpoint^ x_intercept^
649 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
650 * be smoothly throttled down to normal if it starts high in situations like
651 * - start writing to a slow SD card and a fast disk at the same time. The SD
652 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
653 * - the bdi dirty thresh drops quickly due to change of JBOD workload
655 static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
656 unsigned long thresh,
657 unsigned long bg_thresh,
659 unsigned long bdi_thresh,
660 unsigned long bdi_dirty)
662 unsigned long write_bw = bdi->avg_write_bandwidth;
663 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
664 unsigned long limit = hard_dirty_limit(thresh);
665 unsigned long x_intercept;
666 unsigned long setpoint; /* dirty pages' target balance point */
667 unsigned long bdi_setpoint;
669 long long pos_ratio; /* for scaling up/down the rate limit */
672 if (unlikely(dirty >= limit))
679 * f(dirty) := 1.0 + (----------------)
682 * it's a 3rd order polynomial that subjects to
684 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
685 * (2) f(setpoint) = 1.0 => the balance point
686 * (3) f(limit) = 0 => the hard limit
687 * (4) df/dx <= 0 => negative feedback control
688 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
689 * => fast response on large errors; small oscillation near setpoint
691 setpoint = (freerun + limit) / 2;
692 x = div_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
693 limit - setpoint + 1);
695 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
696 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
697 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
700 * We have computed basic pos_ratio above based on global situation. If
701 * the bdi is over/under its share of dirty pages, we want to scale
702 * pos_ratio further down/up. That is done by the following mechanism.
708 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
710 * x_intercept - bdi_dirty
711 * := --------------------------
712 * x_intercept - bdi_setpoint
714 * The main bdi control line is a linear function that subjects to
716 * (1) f(bdi_setpoint) = 1.0
717 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
718 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
720 * For single bdi case, the dirty pages are observed to fluctuate
721 * regularly within range
722 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
723 * for various filesystems, where (2) can yield in a reasonable 12.5%
724 * fluctuation range for pos_ratio.
726 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
727 * own size, so move the slope over accordingly and choose a slope that
728 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
730 if (unlikely(bdi_thresh > thresh))
733 * It's very possible that bdi_thresh is close to 0 not because the
734 * device is slow, but that it has remained inactive for long time.
735 * Honour such devices a reasonable good (hopefully IO efficient)
736 * threshold, so that the occasional writes won't be blocked and active
737 * writes can rampup the threshold quickly.
739 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
741 * scale global setpoint to bdi's:
742 * bdi_setpoint = setpoint * bdi_thresh / thresh
744 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
745 bdi_setpoint = setpoint * (u64)x >> 16;
747 * Use span=(8*write_bw) in single bdi case as indicated by
748 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
750 * bdi_thresh thresh - bdi_thresh
751 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
754 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
755 x_intercept = bdi_setpoint + span;
757 if (bdi_dirty < x_intercept - span / 4) {
758 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
759 x_intercept - bdi_setpoint + 1);
764 * bdi reserve area, safeguard against dirty pool underrun and disk idle
765 * It may push the desired control point of global dirty pages higher
768 x_intercept = bdi_thresh / 2;
769 if (bdi_dirty < x_intercept) {
770 if (bdi_dirty > x_intercept / 8)
771 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
779 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
780 unsigned long elapsed,
781 unsigned long written)
783 const unsigned long period = roundup_pow_of_two(3 * HZ);
784 unsigned long avg = bdi->avg_write_bandwidth;
785 unsigned long old = bdi->write_bandwidth;
789 * bw = written * HZ / elapsed
791 * bw * elapsed + write_bandwidth * (period - elapsed)
792 * write_bandwidth = ---------------------------------------------------
795 bw = written - bdi->written_stamp;
797 if (unlikely(elapsed > period)) {
802 bw += (u64)bdi->write_bandwidth * (period - elapsed);
803 bw >>= ilog2(period);
806 * one more level of smoothing, for filtering out sudden spikes
808 if (avg > old && old >= (unsigned long)bw)
809 avg -= (avg - old) >> 3;
811 if (avg < old && old <= (unsigned long)bw)
812 avg += (old - avg) >> 3;
815 bdi->write_bandwidth = bw;
816 bdi->avg_write_bandwidth = avg;
820 * The global dirtyable memory and dirty threshold could be suddenly knocked
821 * down by a large amount (eg. on the startup of KVM in a swapless system).
822 * This may throw the system into deep dirty exceeded state and throttle
823 * heavy/light dirtiers alike. To retain good responsiveness, maintain
824 * global_dirty_limit for tracking slowly down to the knocked down dirty
827 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
829 unsigned long limit = global_dirty_limit;
832 * Follow up in one step.
834 if (limit < thresh) {
840 * Follow down slowly. Use the higher one as the target, because thresh
841 * may drop below dirty. This is exactly the reason to introduce
842 * global_dirty_limit which is guaranteed to lie above the dirty pages.
844 thresh = max(thresh, dirty);
845 if (limit > thresh) {
846 limit -= (limit - thresh) >> 5;
851 global_dirty_limit = limit;
854 static void global_update_bandwidth(unsigned long thresh,
858 static DEFINE_SPINLOCK(dirty_lock);
859 static unsigned long update_time;
862 * check locklessly first to optimize away locking for the most time
864 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
867 spin_lock(&dirty_lock);
868 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
869 update_dirty_limit(thresh, dirty);
872 spin_unlock(&dirty_lock);
876 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
878 * Normal bdi tasks will be curbed at or below it in long term.
879 * Obviously it should be around (write_bw / N) when there are N dd tasks.
881 static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
882 unsigned long thresh,
883 unsigned long bg_thresh,
885 unsigned long bdi_thresh,
886 unsigned long bdi_dirty,
887 unsigned long dirtied,
888 unsigned long elapsed)
890 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
891 unsigned long limit = hard_dirty_limit(thresh);
892 unsigned long setpoint = (freerun + limit) / 2;
893 unsigned long write_bw = bdi->avg_write_bandwidth;
894 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
895 unsigned long dirty_rate;
896 unsigned long task_ratelimit;
897 unsigned long balanced_dirty_ratelimit;
898 unsigned long pos_ratio;
903 * The dirty rate will match the writeout rate in long term, except
904 * when dirty pages are truncated by userspace or re-dirtied by FS.
906 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
908 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
909 bdi_thresh, bdi_dirty);
911 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
913 task_ratelimit = (u64)dirty_ratelimit *
914 pos_ratio >> RATELIMIT_CALC_SHIFT;
915 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
918 * A linear estimation of the "balanced" throttle rate. The theory is,
919 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
920 * dirty_rate will be measured to be (N * task_ratelimit). So the below
921 * formula will yield the balanced rate limit (write_bw / N).
923 * Note that the expanded form is not a pure rate feedback:
924 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
925 * but also takes pos_ratio into account:
926 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
928 * (1) is not realistic because pos_ratio also takes part in balancing
929 * the dirty rate. Consider the state
930 * pos_ratio = 0.5 (3)
931 * rate = 2 * (write_bw / N) (4)
932 * If (1) is used, it will stuck in that state! Because each dd will
934 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
936 * dirty_rate = N * task_ratelimit = write_bw (6)
937 * put (6) into (1) we get
938 * rate_(i+1) = rate_(i) (7)
940 * So we end up using (2) to always keep
941 * rate_(i+1) ~= (write_bw / N) (8)
942 * regardless of the value of pos_ratio. As long as (8) is satisfied,
943 * pos_ratio is able to drive itself to 1.0, which is not only where
944 * the dirty count meet the setpoint, but also where the slope of
945 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
947 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
950 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
952 if (unlikely(balanced_dirty_ratelimit > write_bw))
953 balanced_dirty_ratelimit = write_bw;
956 * We could safely do this and return immediately:
958 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
960 * However to get a more stable dirty_ratelimit, the below elaborated
961 * code makes use of task_ratelimit to filter out singular points and
962 * limit the step size.
964 * The below code essentially only uses the relative value of
966 * task_ratelimit - dirty_ratelimit
967 * = (pos_ratio - 1) * dirty_ratelimit
969 * which reflects the direction and size of dirty position error.
973 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
974 * task_ratelimit is on the same side of dirty_ratelimit, too.
976 * - dirty_ratelimit > balanced_dirty_ratelimit
977 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
978 * lowering dirty_ratelimit will help meet both the position and rate
979 * control targets. Otherwise, don't update dirty_ratelimit if it will
980 * only help meet the rate target. After all, what the users ultimately
981 * feel and care are stable dirty rate and small position error.
983 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
984 * and filter out the singular points of balanced_dirty_ratelimit. Which
985 * keeps jumping around randomly and can even leap far away at times
986 * due to the small 200ms estimation period of dirty_rate (we want to
987 * keep that period small to reduce time lags).
990 if (dirty < setpoint) {
991 x = min(bdi->balanced_dirty_ratelimit,
992 min(balanced_dirty_ratelimit, task_ratelimit));
993 if (dirty_ratelimit < x)
994 step = x - dirty_ratelimit;
996 x = max(bdi->balanced_dirty_ratelimit,
997 max(balanced_dirty_ratelimit, task_ratelimit));
998 if (dirty_ratelimit > x)
999 step = dirty_ratelimit - x;
1003 * Don't pursue 100% rate matching. It's impossible since the balanced
1004 * rate itself is constantly fluctuating. So decrease the track speed
1005 * when it gets close to the target. Helps eliminate pointless tremors.
1007 step >>= dirty_ratelimit / (2 * step + 1);
1009 * Limit the tracking speed to avoid overshooting.
1011 step = (step + 7) / 8;
1013 if (dirty_ratelimit < balanced_dirty_ratelimit)
1014 dirty_ratelimit += step;
1016 dirty_ratelimit -= step;
1018 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1019 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1021 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
1024 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
1025 unsigned long thresh,
1026 unsigned long bg_thresh,
1027 unsigned long dirty,
1028 unsigned long bdi_thresh,
1029 unsigned long bdi_dirty,
1030 unsigned long start_time)
1032 unsigned long now = jiffies;
1033 unsigned long elapsed = now - bdi->bw_time_stamp;
1034 unsigned long dirtied;
1035 unsigned long written;
1038 * rate-limit, only update once every 200ms.
1040 if (elapsed < BANDWIDTH_INTERVAL)
1043 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
1044 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
1047 * Skip quiet periods when disk bandwidth is under-utilized.
1048 * (at least 1s idle time between two flusher runs)
1050 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
1054 global_update_bandwidth(thresh, dirty, now);
1055 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1056 bdi_thresh, bdi_dirty,
1059 bdi_update_write_bandwidth(bdi, elapsed, written);
1062 bdi->dirtied_stamp = dirtied;
1063 bdi->written_stamp = written;
1064 bdi->bw_time_stamp = now;
1067 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
1068 unsigned long thresh,
1069 unsigned long bg_thresh,
1070 unsigned long dirty,
1071 unsigned long bdi_thresh,
1072 unsigned long bdi_dirty,
1073 unsigned long start_time)
1075 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1077 spin_lock(&bdi->wb.list_lock);
1078 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1079 bdi_thresh, bdi_dirty, start_time);
1080 spin_unlock(&bdi->wb.list_lock);
1084 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1085 * will look to see if it needs to start dirty throttling.
1087 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1088 * global_page_state() too often. So scale it near-sqrt to the safety margin
1089 * (the number of pages we may dirty without exceeding the dirty limits).
1091 static unsigned long dirty_poll_interval(unsigned long dirty,
1092 unsigned long thresh)
1095 return 1UL << (ilog2(thresh - dirty) >> 1);
1100 static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
1101 unsigned long bdi_dirty)
1103 unsigned long bw = bdi->avg_write_bandwidth;
1107 * Limit pause time for small memory systems. If sleeping for too long
1108 * time, a small pool of dirty/writeback pages may go empty and disk go
1111 * 8 serves as the safety ratio.
1113 t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1116 return min_t(unsigned long, t, MAX_PAUSE);
1119 static long bdi_min_pause(struct backing_dev_info *bdi,
1121 unsigned long task_ratelimit,
1122 unsigned long dirty_ratelimit,
1123 int *nr_dirtied_pause)
1125 long hi = ilog2(bdi->avg_write_bandwidth);
1126 long lo = ilog2(bdi->dirty_ratelimit);
1127 long t; /* target pause */
1128 long pause; /* estimated next pause */
1129 int pages; /* target nr_dirtied_pause */
1131 /* target for 10ms pause on 1-dd case */
1132 t = max(1, HZ / 100);
1135 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1138 * (N * 10ms) on 2^N concurrent tasks.
1141 t += (hi - lo) * (10 * HZ) / 1024;
1144 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1145 * on the much more stable dirty_ratelimit. However the next pause time
1146 * will be computed based on task_ratelimit and the two rate limits may
1147 * depart considerably at some time. Especially if task_ratelimit goes
1148 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1149 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1150 * result task_ratelimit won't be executed faithfully, which could
1151 * eventually bring down dirty_ratelimit.
1153 * We apply two rules to fix it up:
1154 * 1) try to estimate the next pause time and if necessary, use a lower
1155 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1156 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1157 * 2) limit the target pause time to max_pause/2, so that the normal
1158 * small fluctuations of task_ratelimit won't trigger rule (1) and
1159 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1161 t = min(t, 1 + max_pause / 2);
1162 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1165 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1166 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1167 * When the 16 consecutive reads are often interrupted by some dirty
1168 * throttling pause during the async writes, cfq will go into idles
1169 * (deadline is fine). So push nr_dirtied_pause as high as possible
1170 * until reaches DIRTY_POLL_THRESH=32 pages.
1172 if (pages < DIRTY_POLL_THRESH) {
1174 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1175 if (pages > DIRTY_POLL_THRESH) {
1176 pages = DIRTY_POLL_THRESH;
1177 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1181 pause = HZ * pages / (task_ratelimit + 1);
1182 if (pause > max_pause) {
1184 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1187 *nr_dirtied_pause = pages;
1189 * The minimal pause time will normally be half the target pause time.
1191 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1195 * balance_dirty_pages() must be called by processes which are generating dirty
1196 * data. It looks at the number of dirty pages in the machine and will force
1197 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1198 * If we're over `background_thresh' then the writeback threads are woken to
1199 * perform some writeout.
1201 static void balance_dirty_pages(struct address_space *mapping,
1202 unsigned long pages_dirtied)
1204 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1205 unsigned long bdi_reclaimable;
1206 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1207 unsigned long bdi_dirty;
1208 unsigned long freerun;
1209 unsigned long background_thresh;
1210 unsigned long dirty_thresh;
1211 unsigned long bdi_thresh;
1216 int nr_dirtied_pause;
1217 bool dirty_exceeded = false;
1218 unsigned long task_ratelimit;
1219 unsigned long dirty_ratelimit;
1220 unsigned long pos_ratio;
1221 struct backing_dev_info *bdi = mapping->backing_dev_info;
1222 unsigned long start_time = jiffies;
1225 unsigned long now = jiffies;
1228 * Unstable writes are a feature of certain networked
1229 * filesystems (i.e. NFS) in which data may have been
1230 * written to the server's write cache, but has not yet
1231 * been flushed to permanent storage.
1233 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1234 global_page_state(NR_UNSTABLE_NFS);
1235 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1237 global_dirty_limits(&background_thresh, &dirty_thresh);
1240 * Throttle it only when the background writeback cannot
1241 * catch-up. This avoids (excessively) small writeouts
1242 * when the bdi limits are ramping up.
1244 freerun = dirty_freerun_ceiling(dirty_thresh,
1246 if (nr_dirty <= freerun) {
1247 current->dirty_paused_when = now;
1248 current->nr_dirtied = 0;
1249 current->nr_dirtied_pause =
1250 dirty_poll_interval(nr_dirty, dirty_thresh);
1254 if (unlikely(!writeback_in_progress(bdi)))
1255 bdi_start_background_writeback(bdi);
1258 * bdi_thresh is not treated as some limiting factor as
1259 * dirty_thresh, due to reasons
1260 * - in JBOD setup, bdi_thresh can fluctuate a lot
1261 * - in a system with HDD and USB key, the USB key may somehow
1262 * go into state (bdi_dirty >> bdi_thresh) either because
1263 * bdi_dirty starts high, or because bdi_thresh drops low.
1264 * In this case we don't want to hard throttle the USB key
1265 * dirtiers for 100 seconds until bdi_dirty drops under
1266 * bdi_thresh. Instead the auxiliary bdi control line in
1267 * bdi_position_ratio() will let the dirtier task progress
1268 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1270 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1273 * In order to avoid the stacked BDI deadlock we need
1274 * to ensure we accurately count the 'dirty' pages when
1275 * the threshold is low.
1277 * Otherwise it would be possible to get thresh+n pages
1278 * reported dirty, even though there are thresh-m pages
1279 * actually dirty; with m+n sitting in the percpu
1282 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1283 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1284 bdi_dirty = bdi_reclaimable +
1285 bdi_stat_sum(bdi, BDI_WRITEBACK);
1287 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1288 bdi_dirty = bdi_reclaimable +
1289 bdi_stat(bdi, BDI_WRITEBACK);
1292 dirty_exceeded = (bdi_dirty > bdi_thresh) &&
1293 (nr_dirty > dirty_thresh);
1294 if (dirty_exceeded && !bdi->dirty_exceeded)
1295 bdi->dirty_exceeded = 1;
1297 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1298 nr_dirty, bdi_thresh, bdi_dirty,
1301 dirty_ratelimit = bdi->dirty_ratelimit;
1302 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1303 background_thresh, nr_dirty,
1304 bdi_thresh, bdi_dirty);
1305 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1306 RATELIMIT_CALC_SHIFT;
1307 max_pause = bdi_max_pause(bdi, bdi_dirty);
1308 min_pause = bdi_min_pause(bdi, max_pause,
1309 task_ratelimit, dirty_ratelimit,
1312 if (unlikely(task_ratelimit == 0)) {
1317 period = HZ * pages_dirtied / task_ratelimit;
1319 if (current->dirty_paused_when)
1320 pause -= now - current->dirty_paused_when;
1322 * For less than 1s think time (ext3/4 may block the dirtier
1323 * for up to 800ms from time to time on 1-HDD; so does xfs,
1324 * however at much less frequency), try to compensate it in
1325 * future periods by updating the virtual time; otherwise just
1326 * do a reset, as it may be a light dirtier.
1328 if (pause < min_pause) {
1329 trace_balance_dirty_pages(bdi,
1342 current->dirty_paused_when = now;
1343 current->nr_dirtied = 0;
1344 } else if (period) {
1345 current->dirty_paused_when += period;
1346 current->nr_dirtied = 0;
1347 } else if (current->nr_dirtied_pause <= pages_dirtied)
1348 current->nr_dirtied_pause += pages_dirtied;
1351 if (unlikely(pause > max_pause)) {
1352 /* for occasional dropped task_ratelimit */
1353 now += min(pause - max_pause, max_pause);
1358 trace_balance_dirty_pages(bdi,
1370 __set_current_state(TASK_KILLABLE);
1371 io_schedule_timeout(pause);
1373 current->dirty_paused_when = now + pause;
1374 current->nr_dirtied = 0;
1375 current->nr_dirtied_pause = nr_dirtied_pause;
1378 * This is typically equal to (nr_dirty < dirty_thresh) and can
1379 * also keep "1000+ dd on a slow USB stick" under control.
1385 * In the case of an unresponding NFS server and the NFS dirty
1386 * pages exceeds dirty_thresh, give the other good bdi's a pipe
1387 * to go through, so that tasks on them still remain responsive.
1389 * In theory 1 page is enough to keep the comsumer-producer
1390 * pipe going: the flusher cleans 1 page => the task dirties 1
1391 * more page. However bdi_dirty has accounting errors. So use
1392 * the larger and more IO friendly bdi_stat_error.
1394 if (bdi_dirty <= bdi_stat_error(bdi))
1397 if (fatal_signal_pending(current))
1401 if (!dirty_exceeded && bdi->dirty_exceeded)
1402 bdi->dirty_exceeded = 0;
1404 if (writeback_in_progress(bdi))
1408 * In laptop mode, we wait until hitting the higher threshold before
1409 * starting background writeout, and then write out all the way down
1410 * to the lower threshold. So slow writers cause minimal disk activity.
1412 * In normal mode, we start background writeout at the lower
1413 * background_thresh, to keep the amount of dirty memory low.
1418 if (nr_reclaimable > background_thresh)
1419 bdi_start_background_writeback(bdi);
1422 void set_page_dirty_balance(struct page *page, int page_mkwrite)
1424 if (set_page_dirty(page) || page_mkwrite) {
1425 struct address_space *mapping = page_mapping(page);
1428 balance_dirty_pages_ratelimited(mapping);
1432 static DEFINE_PER_CPU(int, bdp_ratelimits);
1435 * Normal tasks are throttled by
1437 * dirty tsk->nr_dirtied_pause pages;
1438 * take a snap in balance_dirty_pages();
1440 * However there is a worst case. If every task exit immediately when dirtied
1441 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1442 * called to throttle the page dirties. The solution is to save the not yet
1443 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1444 * randomly into the running tasks. This works well for the above worst case,
1445 * as the new task will pick up and accumulate the old task's leaked dirty
1446 * count and eventually get throttled.
1448 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1451 * balance_dirty_pages_ratelimited - balance dirty memory state
1452 * @mapping: address_space which was dirtied
1454 * Processes which are dirtying memory should call in here once for each page
1455 * which was newly dirtied. The function will periodically check the system's
1456 * dirty state and will initiate writeback if needed.
1458 * On really big machines, get_writeback_state is expensive, so try to avoid
1459 * calling it too often (ratelimiting). But once we're over the dirty memory
1460 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1461 * from overshooting the limit by (ratelimit_pages) each.
1463 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1465 struct backing_dev_info *bdi = mapping->backing_dev_info;
1469 if (!bdi_cap_account_dirty(bdi))
1472 ratelimit = current->nr_dirtied_pause;
1473 if (bdi->dirty_exceeded)
1474 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1478 * This prevents one CPU to accumulate too many dirtied pages without
1479 * calling into balance_dirty_pages(), which can happen when there are
1480 * 1000+ tasks, all of them start dirtying pages at exactly the same
1481 * time, hence all honoured too large initial task->nr_dirtied_pause.
1483 p = &__get_cpu_var(bdp_ratelimits);
1484 if (unlikely(current->nr_dirtied >= ratelimit))
1486 else if (unlikely(*p >= ratelimit_pages)) {
1491 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1492 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1493 * the dirty throttling and livelock other long-run dirtiers.
1495 p = &__get_cpu_var(dirty_throttle_leaks);
1496 if (*p > 0 && current->nr_dirtied < ratelimit) {
1497 unsigned long nr_pages_dirtied;
1498 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1499 *p -= nr_pages_dirtied;
1500 current->nr_dirtied += nr_pages_dirtied;
1504 if (unlikely(current->nr_dirtied >= ratelimit))
1505 balance_dirty_pages(mapping, current->nr_dirtied);
1507 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1509 void throttle_vm_writeout(gfp_t gfp_mask)
1511 unsigned long background_thresh;
1512 unsigned long dirty_thresh;
1515 global_dirty_limits(&background_thresh, &dirty_thresh);
1516 dirty_thresh = hard_dirty_limit(dirty_thresh);
1519 * Boost the allowable dirty threshold a bit for page
1520 * allocators so they don't get DoS'ed by heavy writers
1522 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1524 if (global_page_state(NR_UNSTABLE_NFS) +
1525 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1527 congestion_wait(BLK_RW_ASYNC, HZ/10);
1530 * The caller might hold locks which can prevent IO completion
1531 * or progress in the filesystem. So we cannot just sit here
1532 * waiting for IO to complete.
1534 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1540 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1542 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1543 void __user *buffer, size_t *length, loff_t *ppos)
1545 proc_dointvec(table, write, buffer, length, ppos);
1550 void laptop_mode_timer_fn(unsigned long data)
1552 struct request_queue *q = (struct request_queue *)data;
1553 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1554 global_page_state(NR_UNSTABLE_NFS);
1557 * We want to write everything out, not just down to the dirty
1560 if (bdi_has_dirty_io(&q->backing_dev_info))
1561 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1562 WB_REASON_LAPTOP_TIMER);
1566 * We've spun up the disk and we're in laptop mode: schedule writeback
1567 * of all dirty data a few seconds from now. If the flush is already scheduled
1568 * then push it back - the user is still using the disk.
1570 void laptop_io_completion(struct backing_dev_info *info)
1572 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1576 * We're in laptop mode and we've just synced. The sync's writes will have
1577 * caused another writeback to be scheduled by laptop_io_completion.
1578 * Nothing needs to be written back anymore, so we unschedule the writeback.
1580 void laptop_sync_completion(void)
1582 struct backing_dev_info *bdi;
1586 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1587 del_timer(&bdi->laptop_mode_wb_timer);
1594 * If ratelimit_pages is too high then we can get into dirty-data overload
1595 * if a large number of processes all perform writes at the same time.
1596 * If it is too low then SMP machines will call the (expensive)
1597 * get_writeback_state too often.
1599 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1600 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1604 void writeback_set_ratelimit(void)
1606 unsigned long background_thresh;
1607 unsigned long dirty_thresh;
1608 global_dirty_limits(&background_thresh, &dirty_thresh);
1609 global_dirty_limit = dirty_thresh;
1610 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1611 if (ratelimit_pages < 16)
1612 ratelimit_pages = 16;
1615 static int __cpuinit
1616 ratelimit_handler(struct notifier_block *self, unsigned long action,
1620 switch (action & ~CPU_TASKS_FROZEN) {
1623 writeback_set_ratelimit();
1630 static struct notifier_block __cpuinitdata ratelimit_nb = {
1631 .notifier_call = ratelimit_handler,
1636 * Called early on to tune the page writeback dirty limits.
1638 * We used to scale dirty pages according to how total memory
1639 * related to pages that could be allocated for buffers (by
1640 * comparing nr_free_buffer_pages() to vm_total_pages.
1642 * However, that was when we used "dirty_ratio" to scale with
1643 * all memory, and we don't do that any more. "dirty_ratio"
1644 * is now applied to total non-HIGHPAGE memory (by subtracting
1645 * totalhigh_pages from vm_total_pages), and as such we can't
1646 * get into the old insane situation any more where we had
1647 * large amounts of dirty pages compared to a small amount of
1648 * non-HIGHMEM memory.
1650 * But we might still want to scale the dirty_ratio by how
1651 * much memory the box has..
1653 void __init page_writeback_init(void)
1655 writeback_set_ratelimit();
1656 register_cpu_notifier(&ratelimit_nb);
1658 fprop_global_init(&writeout_completions);
1662 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1663 * @mapping: address space structure to write
1664 * @start: starting page index
1665 * @end: ending page index (inclusive)
1667 * This function scans the page range from @start to @end (inclusive) and tags
1668 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1669 * that write_cache_pages (or whoever calls this function) will then use
1670 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1671 * used to avoid livelocking of writeback by a process steadily creating new
1672 * dirty pages in the file (thus it is important for this function to be quick
1673 * so that it can tag pages faster than a dirtying process can create them).
1676 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1678 void tag_pages_for_writeback(struct address_space *mapping,
1679 pgoff_t start, pgoff_t end)
1681 #define WRITEBACK_TAG_BATCH 4096
1682 unsigned long tagged;
1685 spin_lock_irq(&mapping->tree_lock);
1686 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1687 &start, end, WRITEBACK_TAG_BATCH,
1688 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1689 spin_unlock_irq(&mapping->tree_lock);
1690 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1692 /* We check 'start' to handle wrapping when end == ~0UL */
1693 } while (tagged >= WRITEBACK_TAG_BATCH && start);
1695 EXPORT_SYMBOL(tag_pages_for_writeback);
1698 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1699 * @mapping: address space structure to write
1700 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1701 * @writepage: function called for each page
1702 * @data: data passed to writepage function
1704 * If a page is already under I/O, write_cache_pages() skips it, even
1705 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1706 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1707 * and msync() need to guarantee that all the data which was dirty at the time
1708 * the call was made get new I/O started against them. If wbc->sync_mode is
1709 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1710 * existing IO to complete.
1712 * To avoid livelocks (when other process dirties new pages), we first tag
1713 * pages which should be written back with TOWRITE tag and only then start
1714 * writing them. For data-integrity sync we have to be careful so that we do
1715 * not miss some pages (e.g., because some other process has cleared TOWRITE
1716 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1717 * by the process clearing the DIRTY tag (and submitting the page for IO).
1719 int write_cache_pages(struct address_space *mapping,
1720 struct writeback_control *wbc, writepage_t writepage,
1725 struct pagevec pvec;
1727 pgoff_t uninitialized_var(writeback_index);
1729 pgoff_t end; /* Inclusive */
1732 int range_whole = 0;
1735 pagevec_init(&pvec, 0);
1736 if (wbc->range_cyclic) {
1737 writeback_index = mapping->writeback_index; /* prev offset */
1738 index = writeback_index;
1745 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1746 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1747 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1749 cycled = 1; /* ignore range_cyclic tests */
1751 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1752 tag = PAGECACHE_TAG_TOWRITE;
1754 tag = PAGECACHE_TAG_DIRTY;
1756 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1757 tag_pages_for_writeback(mapping, index, end);
1759 while (!done && (index <= end)) {
1762 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1763 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1767 for (i = 0; i < nr_pages; i++) {
1768 struct page *page = pvec.pages[i];
1771 * At this point, the page may be truncated or
1772 * invalidated (changing page->mapping to NULL), or
1773 * even swizzled back from swapper_space to tmpfs file
1774 * mapping. However, page->index will not change
1775 * because we have a reference on the page.
1777 if (page->index > end) {
1779 * can't be range_cyclic (1st pass) because
1780 * end == -1 in that case.
1786 done_index = page->index;
1791 * Page truncated or invalidated. We can freely skip it
1792 * then, even for data integrity operations: the page
1793 * has disappeared concurrently, so there could be no
1794 * real expectation of this data interity operation
1795 * even if there is now a new, dirty page at the same
1796 * pagecache address.
1798 if (unlikely(page->mapping != mapping)) {
1804 if (!PageDirty(page)) {
1805 /* someone wrote it for us */
1806 goto continue_unlock;
1809 if (PageWriteback(page)) {
1810 if (wbc->sync_mode != WB_SYNC_NONE)
1811 wait_on_page_writeback(page);
1813 goto continue_unlock;
1816 BUG_ON(PageWriteback(page));
1817 if (!clear_page_dirty_for_io(page))
1818 goto continue_unlock;
1820 trace_wbc_writepage(wbc, mapping->backing_dev_info);
1821 ret = (*writepage)(page, wbc, data);
1822 if (unlikely(ret)) {
1823 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1828 * done_index is set past this page,
1829 * so media errors will not choke
1830 * background writeout for the entire
1831 * file. This has consequences for
1832 * range_cyclic semantics (ie. it may
1833 * not be suitable for data integrity
1836 done_index = page->index + 1;
1843 * We stop writing back only if we are not doing
1844 * integrity sync. In case of integrity sync we have to
1845 * keep going until we have written all the pages
1846 * we tagged for writeback prior to entering this loop.
1848 if (--wbc->nr_to_write <= 0 &&
1849 wbc->sync_mode == WB_SYNC_NONE) {
1854 pagevec_release(&pvec);
1857 if (!cycled && !done) {
1860 * We hit the last page and there is more work to be done: wrap
1861 * back to the start of the file
1865 end = writeback_index - 1;
1868 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1869 mapping->writeback_index = done_index;
1873 EXPORT_SYMBOL(write_cache_pages);
1876 * Function used by generic_writepages to call the real writepage
1877 * function and set the mapping flags on error
1879 static int __writepage(struct page *page, struct writeback_control *wbc,
1882 struct address_space *mapping = data;
1883 int ret = mapping->a_ops->writepage(page, wbc);
1884 mapping_set_error(mapping, ret);
1889 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1890 * @mapping: address space structure to write
1891 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1893 * This is a library function, which implements the writepages()
1894 * address_space_operation.
1896 int generic_writepages(struct address_space *mapping,
1897 struct writeback_control *wbc)
1899 struct blk_plug plug;
1902 /* deal with chardevs and other special file */
1903 if (!mapping->a_ops->writepage)
1906 blk_start_plug(&plug);
1907 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1908 blk_finish_plug(&plug);
1912 EXPORT_SYMBOL(generic_writepages);
1914 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1918 if (wbc->nr_to_write <= 0)
1920 if (mapping->a_ops->writepages)
1921 ret = mapping->a_ops->writepages(mapping, wbc);
1923 ret = generic_writepages(mapping, wbc);
1928 * write_one_page - write out a single page and optionally wait on I/O
1929 * @page: the page to write
1930 * @wait: if true, wait on writeout
1932 * The page must be locked by the caller and will be unlocked upon return.
1934 * write_one_page() returns a negative error code if I/O failed.
1936 int write_one_page(struct page *page, int wait)
1938 struct address_space *mapping = page->mapping;
1940 struct writeback_control wbc = {
1941 .sync_mode = WB_SYNC_ALL,
1945 BUG_ON(!PageLocked(page));
1948 wait_on_page_writeback(page);
1950 if (clear_page_dirty_for_io(page)) {
1951 page_cache_get(page);
1952 ret = mapping->a_ops->writepage(page, &wbc);
1953 if (ret == 0 && wait) {
1954 wait_on_page_writeback(page);
1955 if (PageError(page))
1958 page_cache_release(page);
1964 EXPORT_SYMBOL(write_one_page);
1967 * For address_spaces which do not use buffers nor write back.
1969 int __set_page_dirty_no_writeback(struct page *page)
1971 if (!PageDirty(page))
1972 return !TestSetPageDirty(page);
1977 * Helper function for set_page_dirty family.
1978 * NOTE: This relies on being atomic wrt interrupts.
1980 void account_page_dirtied(struct page *page, struct address_space *mapping)
1982 trace_writeback_dirty_page(page, mapping);
1984 if (mapping_cap_account_dirty(mapping)) {
1985 __inc_zone_page_state(page, NR_FILE_DIRTY);
1986 __inc_zone_page_state(page, NR_DIRTIED);
1987 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1988 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1989 task_io_account_write(PAGE_CACHE_SIZE);
1990 current->nr_dirtied++;
1991 this_cpu_inc(bdp_ratelimits);
1994 EXPORT_SYMBOL(account_page_dirtied);
1997 * Helper function for set_page_writeback family.
1998 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
2001 void account_page_writeback(struct page *page)
2003 inc_zone_page_state(page, NR_WRITEBACK);
2005 EXPORT_SYMBOL(account_page_writeback);
2008 * For address_spaces which do not use buffers. Just tag the page as dirty in
2011 * This is also used when a single buffer is being dirtied: we want to set the
2012 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2013 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2015 * Most callers have locked the page, which pins the address_space in memory.
2016 * But zap_pte_range() does not lock the page, however in that case the
2017 * mapping is pinned by the vma's ->vm_file reference.
2019 * We take care to handle the case where the page was truncated from the
2020 * mapping by re-checking page_mapping() inside tree_lock.
2022 int __set_page_dirty_nobuffers(struct page *page)
2024 if (!TestSetPageDirty(page)) {
2025 struct address_space *mapping = page_mapping(page);
2026 struct address_space *mapping2;
2031 spin_lock_irq(&mapping->tree_lock);
2032 mapping2 = page_mapping(page);
2033 if (mapping2) { /* Race with truncate? */
2034 BUG_ON(mapping2 != mapping);
2035 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2036 account_page_dirtied(page, mapping);
2037 radix_tree_tag_set(&mapping->page_tree,
2038 page_index(page), PAGECACHE_TAG_DIRTY);
2040 spin_unlock_irq(&mapping->tree_lock);
2041 if (mapping->host) {
2042 /* !PageAnon && !swapper_space */
2043 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2049 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2052 * Call this whenever redirtying a page, to de-account the dirty counters
2053 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2054 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2055 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2058 void account_page_redirty(struct page *page)
2060 struct address_space *mapping = page->mapping;
2061 if (mapping && mapping_cap_account_dirty(mapping)) {
2062 current->nr_dirtied--;
2063 dec_zone_page_state(page, NR_DIRTIED);
2064 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2067 EXPORT_SYMBOL(account_page_redirty);
2070 * When a writepage implementation decides that it doesn't want to write this
2071 * page for some reason, it should redirty the locked page via
2072 * redirty_page_for_writepage() and it should then unlock the page and return 0
2074 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2076 wbc->pages_skipped++;
2077 account_page_redirty(page);
2078 return __set_page_dirty_nobuffers(page);
2080 EXPORT_SYMBOL(redirty_page_for_writepage);
2085 * For pages with a mapping this should be done under the page lock
2086 * for the benefit of asynchronous memory errors who prefer a consistent
2087 * dirty state. This rule can be broken in some special cases,
2088 * but should be better not to.
2090 * If the mapping doesn't provide a set_page_dirty a_op, then
2091 * just fall through and assume that it wants buffer_heads.
2093 int set_page_dirty(struct page *page)
2095 struct address_space *mapping = page_mapping(page);
2097 if (likely(mapping)) {
2098 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2100 * readahead/lru_deactivate_page could remain
2101 * PG_readahead/PG_reclaim due to race with end_page_writeback
2102 * About readahead, if the page is written, the flags would be
2103 * reset. So no problem.
2104 * About lru_deactivate_page, if the page is redirty, the flag
2105 * will be reset. So no problem. but if the page is used by readahead
2106 * it will confuse readahead and make it restart the size rampup
2107 * process. But it's a trivial problem.
2109 ClearPageReclaim(page);
2112 spd = __set_page_dirty_buffers;
2114 return (*spd)(page);
2116 if (!PageDirty(page)) {
2117 if (!TestSetPageDirty(page))
2122 EXPORT_SYMBOL(set_page_dirty);
2125 * set_page_dirty() is racy if the caller has no reference against
2126 * page->mapping->host, and if the page is unlocked. This is because another
2127 * CPU could truncate the page off the mapping and then free the mapping.
2129 * Usually, the page _is_ locked, or the caller is a user-space process which
2130 * holds a reference on the inode by having an open file.
2132 * In other cases, the page should be locked before running set_page_dirty().
2134 int set_page_dirty_lock(struct page *page)
2139 ret = set_page_dirty(page);
2143 EXPORT_SYMBOL(set_page_dirty_lock);
2146 * Clear a page's dirty flag, while caring for dirty memory accounting.
2147 * Returns true if the page was previously dirty.
2149 * This is for preparing to put the page under writeout. We leave the page
2150 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2151 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2152 * implementation will run either set_page_writeback() or set_page_dirty(),
2153 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2156 * This incoherency between the page's dirty flag and radix-tree tag is
2157 * unfortunate, but it only exists while the page is locked.
2159 int clear_page_dirty_for_io(struct page *page)
2161 struct address_space *mapping = page_mapping(page);
2163 BUG_ON(!PageLocked(page));
2165 if (mapping && mapping_cap_account_dirty(mapping)) {
2167 * Yes, Virginia, this is indeed insane.
2169 * We use this sequence to make sure that
2170 * (a) we account for dirty stats properly
2171 * (b) we tell the low-level filesystem to
2172 * mark the whole page dirty if it was
2173 * dirty in a pagetable. Only to then
2174 * (c) clean the page again and return 1 to
2175 * cause the writeback.
2177 * This way we avoid all nasty races with the
2178 * dirty bit in multiple places and clearing
2179 * them concurrently from different threads.
2181 * Note! Normally the "set_page_dirty(page)"
2182 * has no effect on the actual dirty bit - since
2183 * that will already usually be set. But we
2184 * need the side effects, and it can help us
2187 * We basically use the page "master dirty bit"
2188 * as a serialization point for all the different
2189 * threads doing their things.
2191 if (page_mkclean(page))
2192 set_page_dirty(page);
2194 * We carefully synchronise fault handlers against
2195 * installing a dirty pte and marking the page dirty
2196 * at this point. We do this by having them hold the
2197 * page lock at some point after installing their
2198 * pte, but before marking the page dirty.
2199 * Pages are always locked coming in here, so we get
2200 * the desired exclusion. See mm/memory.c:do_wp_page()
2201 * for more comments.
2203 if (TestClearPageDirty(page)) {
2204 dec_zone_page_state(page, NR_FILE_DIRTY);
2205 dec_bdi_stat(mapping->backing_dev_info,
2211 return TestClearPageDirty(page);
2213 EXPORT_SYMBOL(clear_page_dirty_for_io);
2215 int test_clear_page_writeback(struct page *page)
2217 struct address_space *mapping = page_mapping(page);
2221 struct backing_dev_info *bdi = mapping->backing_dev_info;
2222 unsigned long flags;
2224 spin_lock_irqsave(&mapping->tree_lock, flags);
2225 ret = TestClearPageWriteback(page);
2227 radix_tree_tag_clear(&mapping->page_tree,
2229 PAGECACHE_TAG_WRITEBACK);
2230 if (bdi_cap_account_writeback(bdi)) {
2231 __dec_bdi_stat(bdi, BDI_WRITEBACK);
2232 __bdi_writeout_inc(bdi);
2235 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2237 ret = TestClearPageWriteback(page);
2240 dec_zone_page_state(page, NR_WRITEBACK);
2241 inc_zone_page_state(page, NR_WRITTEN);
2246 int test_set_page_writeback(struct page *page)
2248 struct address_space *mapping = page_mapping(page);
2252 struct backing_dev_info *bdi = mapping->backing_dev_info;
2253 unsigned long flags;
2255 spin_lock_irqsave(&mapping->tree_lock, flags);
2256 ret = TestSetPageWriteback(page);
2258 radix_tree_tag_set(&mapping->page_tree,
2260 PAGECACHE_TAG_WRITEBACK);
2261 if (bdi_cap_account_writeback(bdi))
2262 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2264 if (!PageDirty(page))
2265 radix_tree_tag_clear(&mapping->page_tree,
2267 PAGECACHE_TAG_DIRTY);
2268 radix_tree_tag_clear(&mapping->page_tree,
2270 PAGECACHE_TAG_TOWRITE);
2271 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2273 ret = TestSetPageWriteback(page);
2276 account_page_writeback(page);
2280 EXPORT_SYMBOL(test_set_page_writeback);
2283 * Return true if any of the pages in the mapping are marked with the
2286 int mapping_tagged(struct address_space *mapping, int tag)
2288 return radix_tree_tagged(&mapping->page_tree, tag);
2290 EXPORT_SYMBOL(mapping_tagged);
2293 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2294 * @page: The page to wait on.
2296 * This function determines if the given page is related to a backing device
2297 * that requires page contents to be held stable during writeback. If so, then
2298 * it will wait for any pending writeback to complete.
2300 void wait_for_stable_page(struct page *page)
2302 struct address_space *mapping = page_mapping(page);
2303 struct backing_dev_info *bdi = mapping->backing_dev_info;
2305 if (!bdi_cap_stable_pages_required(bdi))
2308 wait_on_page_writeback(page);
2310 EXPORT_SYMBOL_GPL(wait_for_stable_page);