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 <trace/events/writeback.h>
40 * Sleep at most 200ms at a time in balance_dirty_pages().
42 #define MAX_PAUSE max(HZ/5, 1)
45 * Try to keep balance_dirty_pages() call intervals higher than this many pages
46 * by raising pause time to max_pause when falls below it.
48 #define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
51 * Estimate write bandwidth at 200ms intervals.
53 #define BANDWIDTH_INTERVAL max(HZ/5, 1)
55 #define RATELIMIT_CALC_SHIFT 10
58 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
59 * will look to see if it needs to force writeback or throttling.
61 static long ratelimit_pages = 32;
63 /* The following parameters are exported via /proc/sys/vm */
66 * Start background writeback (via writeback threads) at this percentage
68 int dirty_background_ratio = 10;
71 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
72 * dirty_background_ratio * the amount of dirtyable memory
74 unsigned long dirty_background_bytes;
77 * free highmem will not be subtracted from the total free memory
78 * for calculating free ratios if vm_highmem_is_dirtyable is true
80 int vm_highmem_is_dirtyable;
83 * The generator of dirty data starts writeback at this percentage
85 int vm_dirty_ratio = 20;
88 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
89 * vm_dirty_ratio * the amount of dirtyable memory
91 unsigned long vm_dirty_bytes;
94 * The interval between `kupdate'-style writebacks
96 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
98 EXPORT_SYMBOL_GPL(dirty_writeback_interval);
101 * The longest time for which data is allowed to remain dirty
103 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
106 * Flag that makes the machine dump writes/reads and block dirtyings.
111 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
112 * a full sync is triggered after this time elapses without any disk activity.
116 EXPORT_SYMBOL(laptop_mode);
118 /* End of sysctl-exported parameters */
120 unsigned long global_dirty_limit;
123 * Scale the writeback cache size proportional to the relative writeout speeds.
125 * We do this by keeping a floating proportion between BDIs, based on page
126 * writeback completions [end_page_writeback()]. Those devices that write out
127 * pages fastest will get the larger share, while the slower will get a smaller
130 * We use page writeout completions because we are interested in getting rid of
131 * dirty pages. Having them written out is the primary goal.
133 * We introduce a concept of time, a period over which we measure these events,
134 * because demand can/will vary over time. The length of this period itself is
135 * measured in page writeback completions.
138 static struct prop_descriptor vm_completions;
141 * Work out the current dirty-memory clamping and background writeout
144 * The main aim here is to lower them aggressively if there is a lot of mapped
145 * memory around. To avoid stressing page reclaim with lots of unreclaimable
146 * pages. It is better to clamp down on writers than to start swapping, and
147 * performing lots of scanning.
149 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
151 * We don't permit the clamping level to fall below 5% - that is getting rather
154 * We make sure that the background writeout level is below the adjusted
159 * In a memory zone, there is a certain amount of pages we consider
160 * available for the page cache, which is essentially the number of
161 * free and reclaimable pages, minus some zone reserves to protect
162 * lowmem and the ability to uphold the zone's watermarks without
163 * requiring writeback.
165 * This number of dirtyable pages is the base value of which the
166 * user-configurable dirty ratio is the effictive number of pages that
167 * are allowed to be actually dirtied. Per individual zone, or
168 * globally by using the sum of dirtyable pages over all zones.
170 * Because the user is allowed to specify the dirty limit globally as
171 * absolute number of bytes, calculating the per-zone dirty limit can
172 * require translating the configured limit into a percentage of
173 * global dirtyable memory first.
176 static unsigned long highmem_dirtyable_memory(unsigned long total)
178 #ifdef CONFIG_HIGHMEM
182 for_each_node_state(node, N_HIGH_MEMORY) {
183 unsigned long nr_pages;
185 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
187 nr_pages = zone_page_state(z, NR_FREE_PAGES) +
188 zone_reclaimable_pages(z);
190 * make sure that the number of pages for this node
191 * is never "negative".
193 nr_pages -= min(nr_pages, z->dirty_balance_reserve);
197 * Make sure that the number of highmem pages is never larger
198 * than the number of the total dirtyable memory. This can only
199 * occur in very strange VM situations but we want to make sure
200 * that this does not occur.
202 return min(x, total);
209 * global_dirtyable_memory - number of globally dirtyable pages
211 * Returns the global number of pages potentially available for dirty
212 * page cache. This is the base value for the global dirty limits.
214 unsigned long global_dirtyable_memory(void)
218 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
219 x -= min(x, dirty_balance_reserve);
221 if (!vm_highmem_is_dirtyable)
222 x -= min(x, highmem_dirtyable_memory(x));
224 return x + 1; /* Ensure that we never return 0 */
228 * global_dirty_limits - background-writeback and dirty-throttling thresholds
230 * Calculate the dirty thresholds based on sysctl parameters
231 * - vm.dirty_background_ratio or vm.dirty_background_bytes
232 * - vm.dirty_ratio or vm.dirty_bytes
233 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
236 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
238 unsigned long background;
240 unsigned long uninitialized_var(available_memory);
241 struct task_struct *tsk;
243 if (!vm_dirty_bytes || !dirty_background_bytes)
244 available_memory = global_dirtyable_memory();
247 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
249 dirty = (vm_dirty_ratio * available_memory) / 100;
251 if (dirty_background_bytes)
252 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
254 background = (dirty_background_ratio * available_memory) / 100;
256 if (background >= dirty)
257 background = dirty / 2;
259 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
260 background += background / 4;
263 *pbackground = background;
265 trace_global_dirty_state(background, dirty);
269 * zone_dirtyable_memory - number of dirtyable pages in a zone
272 * Returns the zone's number of pages potentially available for dirty
273 * page cache. This is the base value for the per-zone dirty limits.
275 static unsigned long zone_dirtyable_memory(struct zone *zone)
278 * The effective global number of dirtyable pages may exclude
279 * highmem as a big-picture measure to keep the ratio between
280 * dirty memory and lowmem reasonable.
282 * But this function is purely about the individual zone and a
283 * highmem zone can hold its share of dirty pages, so we don't
284 * care about vm_highmem_is_dirtyable here.
286 return zone_page_state(zone, NR_FREE_PAGES) +
287 zone_reclaimable_pages(zone) -
288 zone->dirty_balance_reserve;
292 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
295 * Returns the maximum number of dirty pages allowed in a zone, based
296 * on the zone's dirtyable memory.
298 static unsigned long zone_dirty_limit(struct zone *zone)
300 unsigned long zone_memory = zone_dirtyable_memory(zone);
301 struct task_struct *tsk = current;
305 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
306 zone_memory / global_dirtyable_memory();
308 dirty = vm_dirty_ratio * zone_memory / 100;
310 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
317 * zone_dirty_ok - tells whether a zone is within its dirty limits
318 * @zone: the zone to check
320 * Returns %true when the dirty pages in @zone are within the zone's
321 * dirty limit, %false if the limit is exceeded.
323 bool zone_dirty_ok(struct zone *zone)
325 unsigned long limit = zone_dirty_limit(zone);
327 return zone_page_state(zone, NR_FILE_DIRTY) +
328 zone_page_state(zone, NR_UNSTABLE_NFS) +
329 zone_page_state(zone, NR_WRITEBACK) <= limit;
333 * couple the period to the dirty_ratio:
335 * period/2 ~ roundup_pow_of_two(dirty limit)
337 static int calc_period_shift(void)
339 unsigned long dirty_total;
342 dirty_total = vm_dirty_bytes / PAGE_SIZE;
344 dirty_total = (vm_dirty_ratio * global_dirtyable_memory()) /
346 return 2 + ilog2(dirty_total - 1);
350 * update the period when the dirty threshold changes.
352 static void update_completion_period(void)
354 int shift = calc_period_shift();
355 prop_change_shift(&vm_completions, shift);
357 writeback_set_ratelimit();
360 int dirty_background_ratio_handler(struct ctl_table *table, int write,
361 void __user *buffer, size_t *lenp,
366 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
367 if (ret == 0 && write)
368 dirty_background_bytes = 0;
372 int dirty_background_bytes_handler(struct ctl_table *table, int write,
373 void __user *buffer, size_t *lenp,
378 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
379 if (ret == 0 && write)
380 dirty_background_ratio = 0;
384 int dirty_ratio_handler(struct ctl_table *table, int write,
385 void __user *buffer, size_t *lenp,
388 int old_ratio = vm_dirty_ratio;
391 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
392 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
393 update_completion_period();
399 int dirty_bytes_handler(struct ctl_table *table, int write,
400 void __user *buffer, size_t *lenp,
403 unsigned long old_bytes = vm_dirty_bytes;
406 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
407 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
408 update_completion_period();
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 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
425 void bdi_writeout_inc(struct backing_dev_info *bdi)
429 local_irq_save(flags);
430 __bdi_writeout_inc(bdi);
431 local_irq_restore(flags);
433 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
436 * Obtain an accurate fraction of the BDI's portion.
438 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
439 long *numerator, long *denominator)
441 prop_fraction_percpu(&vm_completions, &bdi->completions,
442 numerator, denominator);
446 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
447 * registered backing devices, which, for obvious reasons, can not
450 static unsigned int bdi_min_ratio;
452 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
456 spin_lock_bh(&bdi_lock);
457 if (min_ratio > bdi->max_ratio) {
460 min_ratio -= bdi->min_ratio;
461 if (bdi_min_ratio + min_ratio < 100) {
462 bdi_min_ratio += min_ratio;
463 bdi->min_ratio += min_ratio;
468 spin_unlock_bh(&bdi_lock);
473 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
480 spin_lock_bh(&bdi_lock);
481 if (bdi->min_ratio > max_ratio) {
484 bdi->max_ratio = max_ratio;
485 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
487 spin_unlock_bh(&bdi_lock);
491 EXPORT_SYMBOL(bdi_set_max_ratio);
493 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
494 unsigned long bg_thresh)
496 return (thresh + bg_thresh) / 2;
499 static unsigned long hard_dirty_limit(unsigned long thresh)
501 return max(thresh, global_dirty_limit);
505 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
506 * @bdi: the backing_dev_info to query
507 * @dirty: global dirty limit in pages
509 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
510 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
512 * Note that balance_dirty_pages() will only seriously take it as a hard limit
513 * when sleeping max_pause per page is not enough to keep the dirty pages under
514 * control. For example, when the device is completely stalled due to some error
515 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
516 * In the other normal situations, it acts more gently by throttling the tasks
517 * more (rather than completely block them) when the bdi dirty pages go high.
519 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
520 * - starving fast devices
521 * - piling up dirty pages (that will take long time to sync) on slow devices
523 * The bdi's share of dirty limit will be adapting to its throughput and
524 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
526 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
529 long numerator, denominator;
532 * Calculate this BDI's share of the dirty ratio.
534 bdi_writeout_fraction(bdi, &numerator, &denominator);
536 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
537 bdi_dirty *= numerator;
538 do_div(bdi_dirty, denominator);
540 bdi_dirty += (dirty * bdi->min_ratio) / 100;
541 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
542 bdi_dirty = dirty * bdi->max_ratio / 100;
548 * Dirty position control.
550 * (o) global/bdi setpoints
552 * We want the dirty pages be balanced around the global/bdi setpoints.
553 * When the number of dirty pages is higher/lower than the setpoint, the
554 * dirty position control ratio (and hence task dirty ratelimit) will be
555 * decreased/increased to bring the dirty pages back to the setpoint.
557 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
559 * if (dirty < setpoint) scale up pos_ratio
560 * if (dirty > setpoint) scale down pos_ratio
562 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
563 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
565 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
567 * (o) global control line
571 * | |<===== global dirty control scope ======>|
579 * 1.0 ................................*
585 * 0 +------------.------------------.----------------------*------------->
586 * freerun^ setpoint^ limit^ dirty pages
588 * (o) bdi control line
596 * | * |<=========== span ============>|
597 * 1.0 .......................*
609 * 1/4 ...............................................* * * * * * * * * * * *
613 * 0 +----------------------.-------------------------------.------------->
614 * bdi_setpoint^ x_intercept^
616 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
617 * be smoothly throttled down to normal if it starts high in situations like
618 * - start writing to a slow SD card and a fast disk at the same time. The SD
619 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
620 * - the bdi dirty thresh drops quickly due to change of JBOD workload
622 static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
623 unsigned long thresh,
624 unsigned long bg_thresh,
626 unsigned long bdi_thresh,
627 unsigned long bdi_dirty)
629 unsigned long write_bw = bdi->avg_write_bandwidth;
630 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
631 unsigned long limit = hard_dirty_limit(thresh);
632 unsigned long x_intercept;
633 unsigned long setpoint; /* dirty pages' target balance point */
634 unsigned long bdi_setpoint;
636 long long pos_ratio; /* for scaling up/down the rate limit */
639 if (unlikely(dirty >= limit))
646 * f(dirty) := 1.0 + (----------------)
649 * it's a 3rd order polynomial that subjects to
651 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
652 * (2) f(setpoint) = 1.0 => the balance point
653 * (3) f(limit) = 0 => the hard limit
654 * (4) df/dx <= 0 => negative feedback control
655 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
656 * => fast response on large errors; small oscillation near setpoint
658 setpoint = (freerun + limit) / 2;
659 x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
660 limit - setpoint + 1);
662 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
663 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
664 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
667 * We have computed basic pos_ratio above based on global situation. If
668 * the bdi is over/under its share of dirty pages, we want to scale
669 * pos_ratio further down/up. That is done by the following mechanism.
675 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
677 * x_intercept - bdi_dirty
678 * := --------------------------
679 * x_intercept - bdi_setpoint
681 * The main bdi control line is a linear function that subjects to
683 * (1) f(bdi_setpoint) = 1.0
684 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
685 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
687 * For single bdi case, the dirty pages are observed to fluctuate
688 * regularly within range
689 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
690 * for various filesystems, where (2) can yield in a reasonable 12.5%
691 * fluctuation range for pos_ratio.
693 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
694 * own size, so move the slope over accordingly and choose a slope that
695 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
697 if (unlikely(bdi_thresh > thresh))
700 * It's very possible that bdi_thresh is close to 0 not because the
701 * device is slow, but that it has remained inactive for long time.
702 * Honour such devices a reasonable good (hopefully IO efficient)
703 * threshold, so that the occasional writes won't be blocked and active
704 * writes can rampup the threshold quickly.
706 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
708 * scale global setpoint to bdi's:
709 * bdi_setpoint = setpoint * bdi_thresh / thresh
711 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
712 bdi_setpoint = setpoint * (u64)x >> 16;
714 * Use span=(8*write_bw) in single bdi case as indicated by
715 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
717 * bdi_thresh thresh - bdi_thresh
718 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
721 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
722 x_intercept = bdi_setpoint + span;
724 if (bdi_dirty < x_intercept - span / 4) {
725 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
726 x_intercept - bdi_setpoint + 1);
731 * bdi reserve area, safeguard against dirty pool underrun and disk idle
732 * It may push the desired control point of global dirty pages higher
735 x_intercept = bdi_thresh / 2;
736 if (bdi_dirty < x_intercept) {
737 if (bdi_dirty > x_intercept / 8)
738 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
746 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
747 unsigned long elapsed,
748 unsigned long written)
750 const unsigned long period = roundup_pow_of_two(3 * HZ);
751 unsigned long avg = bdi->avg_write_bandwidth;
752 unsigned long old = bdi->write_bandwidth;
756 * bw = written * HZ / elapsed
758 * bw * elapsed + write_bandwidth * (period - elapsed)
759 * write_bandwidth = ---------------------------------------------------
762 bw = written - bdi->written_stamp;
764 if (unlikely(elapsed > period)) {
769 bw += (u64)bdi->write_bandwidth * (period - elapsed);
770 bw >>= ilog2(period);
773 * one more level of smoothing, for filtering out sudden spikes
775 if (avg > old && old >= (unsigned long)bw)
776 avg -= (avg - old) >> 3;
778 if (avg < old && old <= (unsigned long)bw)
779 avg += (old - avg) >> 3;
782 bdi->write_bandwidth = bw;
783 bdi->avg_write_bandwidth = avg;
787 * The global dirtyable memory and dirty threshold could be suddenly knocked
788 * down by a large amount (eg. on the startup of KVM in a swapless system).
789 * This may throw the system into deep dirty exceeded state and throttle
790 * heavy/light dirtiers alike. To retain good responsiveness, maintain
791 * global_dirty_limit for tracking slowly down to the knocked down dirty
794 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
796 unsigned long limit = global_dirty_limit;
799 * Follow up in one step.
801 if (limit < thresh) {
807 * Follow down slowly. Use the higher one as the target, because thresh
808 * may drop below dirty. This is exactly the reason to introduce
809 * global_dirty_limit which is guaranteed to lie above the dirty pages.
811 thresh = max(thresh, dirty);
812 if (limit > thresh) {
813 limit -= (limit - thresh) >> 5;
818 global_dirty_limit = limit;
821 static void global_update_bandwidth(unsigned long thresh,
825 static DEFINE_SPINLOCK(dirty_lock);
826 static unsigned long update_time;
829 * check locklessly first to optimize away locking for the most time
831 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
834 spin_lock(&dirty_lock);
835 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
836 update_dirty_limit(thresh, dirty);
839 spin_unlock(&dirty_lock);
843 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
845 * Normal bdi tasks will be curbed at or below it in long term.
846 * Obviously it should be around (write_bw / N) when there are N dd tasks.
848 static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
849 unsigned long thresh,
850 unsigned long bg_thresh,
852 unsigned long bdi_thresh,
853 unsigned long bdi_dirty,
854 unsigned long dirtied,
855 unsigned long elapsed)
857 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
858 unsigned long limit = hard_dirty_limit(thresh);
859 unsigned long setpoint = (freerun + limit) / 2;
860 unsigned long write_bw = bdi->avg_write_bandwidth;
861 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
862 unsigned long dirty_rate;
863 unsigned long task_ratelimit;
864 unsigned long balanced_dirty_ratelimit;
865 unsigned long pos_ratio;
870 * The dirty rate will match the writeout rate in long term, except
871 * when dirty pages are truncated by userspace or re-dirtied by FS.
873 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
875 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
876 bdi_thresh, bdi_dirty);
878 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
880 task_ratelimit = (u64)dirty_ratelimit *
881 pos_ratio >> RATELIMIT_CALC_SHIFT;
882 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
885 * A linear estimation of the "balanced" throttle rate. The theory is,
886 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
887 * dirty_rate will be measured to be (N * task_ratelimit). So the below
888 * formula will yield the balanced rate limit (write_bw / N).
890 * Note that the expanded form is not a pure rate feedback:
891 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
892 * but also takes pos_ratio into account:
893 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
895 * (1) is not realistic because pos_ratio also takes part in balancing
896 * the dirty rate. Consider the state
897 * pos_ratio = 0.5 (3)
898 * rate = 2 * (write_bw / N) (4)
899 * If (1) is used, it will stuck in that state! Because each dd will
901 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
903 * dirty_rate = N * task_ratelimit = write_bw (6)
904 * put (6) into (1) we get
905 * rate_(i+1) = rate_(i) (7)
907 * So we end up using (2) to always keep
908 * rate_(i+1) ~= (write_bw / N) (8)
909 * regardless of the value of pos_ratio. As long as (8) is satisfied,
910 * pos_ratio is able to drive itself to 1.0, which is not only where
911 * the dirty count meet the setpoint, but also where the slope of
912 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
914 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
917 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
919 if (unlikely(balanced_dirty_ratelimit > write_bw))
920 balanced_dirty_ratelimit = write_bw;
923 * We could safely do this and return immediately:
925 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
927 * However to get a more stable dirty_ratelimit, the below elaborated
928 * code makes use of task_ratelimit to filter out sigular points and
929 * limit the step size.
931 * The below code essentially only uses the relative value of
933 * task_ratelimit - dirty_ratelimit
934 * = (pos_ratio - 1) * dirty_ratelimit
936 * which reflects the direction and size of dirty position error.
940 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
941 * task_ratelimit is on the same side of dirty_ratelimit, too.
943 * - dirty_ratelimit > balanced_dirty_ratelimit
944 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
945 * lowering dirty_ratelimit will help meet both the position and rate
946 * control targets. Otherwise, don't update dirty_ratelimit if it will
947 * only help meet the rate target. After all, what the users ultimately
948 * feel and care are stable dirty rate and small position error.
950 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
951 * and filter out the sigular points of balanced_dirty_ratelimit. Which
952 * keeps jumping around randomly and can even leap far away at times
953 * due to the small 200ms estimation period of dirty_rate (we want to
954 * keep that period small to reduce time lags).
957 if (dirty < setpoint) {
958 x = min(bdi->balanced_dirty_ratelimit,
959 min(balanced_dirty_ratelimit, task_ratelimit));
960 if (dirty_ratelimit < x)
961 step = x - dirty_ratelimit;
963 x = max(bdi->balanced_dirty_ratelimit,
964 max(balanced_dirty_ratelimit, task_ratelimit));
965 if (dirty_ratelimit > x)
966 step = dirty_ratelimit - x;
970 * Don't pursue 100% rate matching. It's impossible since the balanced
971 * rate itself is constantly fluctuating. So decrease the track speed
972 * when it gets close to the target. Helps eliminate pointless tremors.
974 step >>= dirty_ratelimit / (2 * step + 1);
976 * Limit the tracking speed to avoid overshooting.
978 step = (step + 7) / 8;
980 if (dirty_ratelimit < balanced_dirty_ratelimit)
981 dirty_ratelimit += step;
983 dirty_ratelimit -= step;
985 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
986 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
988 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
991 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
992 unsigned long thresh,
993 unsigned long bg_thresh,
995 unsigned long bdi_thresh,
996 unsigned long bdi_dirty,
997 unsigned long start_time)
999 unsigned long now = jiffies;
1000 unsigned long elapsed = now - bdi->bw_time_stamp;
1001 unsigned long dirtied;
1002 unsigned long written;
1005 * rate-limit, only update once every 200ms.
1007 if (elapsed < BANDWIDTH_INTERVAL)
1010 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
1011 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
1014 * Skip quiet periods when disk bandwidth is under-utilized.
1015 * (at least 1s idle time between two flusher runs)
1017 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
1021 global_update_bandwidth(thresh, dirty, now);
1022 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1023 bdi_thresh, bdi_dirty,
1026 bdi_update_write_bandwidth(bdi, elapsed, written);
1029 bdi->dirtied_stamp = dirtied;
1030 bdi->written_stamp = written;
1031 bdi->bw_time_stamp = now;
1034 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
1035 unsigned long thresh,
1036 unsigned long bg_thresh,
1037 unsigned long dirty,
1038 unsigned long bdi_thresh,
1039 unsigned long bdi_dirty,
1040 unsigned long start_time)
1042 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1044 spin_lock(&bdi->wb.list_lock);
1045 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1046 bdi_thresh, bdi_dirty, start_time);
1047 spin_unlock(&bdi->wb.list_lock);
1051 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
1052 * will look to see if it needs to start dirty throttling.
1054 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1055 * global_page_state() too often. So scale it near-sqrt to the safety margin
1056 * (the number of pages we may dirty without exceeding the dirty limits).
1058 static unsigned long dirty_poll_interval(unsigned long dirty,
1059 unsigned long thresh)
1062 return 1UL << (ilog2(thresh - dirty) >> 1);
1067 static long bdi_max_pause(struct backing_dev_info *bdi,
1068 unsigned long bdi_dirty)
1070 long bw = bdi->avg_write_bandwidth;
1074 * Limit pause time for small memory systems. If sleeping for too long
1075 * time, a small pool of dirty/writeback pages may go empty and disk go
1078 * 8 serves as the safety ratio.
1080 t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1083 return min_t(long, t, MAX_PAUSE);
1086 static long bdi_min_pause(struct backing_dev_info *bdi,
1088 unsigned long task_ratelimit,
1089 unsigned long dirty_ratelimit,
1090 int *nr_dirtied_pause)
1092 long hi = ilog2(bdi->avg_write_bandwidth);
1093 long lo = ilog2(bdi->dirty_ratelimit);
1094 long t; /* target pause */
1095 long pause; /* estimated next pause */
1096 int pages; /* target nr_dirtied_pause */
1098 /* target for 10ms pause on 1-dd case */
1099 t = max(1, HZ / 100);
1102 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1105 * (N * 10ms) on 2^N concurrent tasks.
1108 t += (hi - lo) * (10 * HZ) / 1024;
1111 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1112 * on the much more stable dirty_ratelimit. However the next pause time
1113 * will be computed based on task_ratelimit and the two rate limits may
1114 * depart considerably at some time. Especially if task_ratelimit goes
1115 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1116 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1117 * result task_ratelimit won't be executed faithfully, which could
1118 * eventually bring down dirty_ratelimit.
1120 * We apply two rules to fix it up:
1121 * 1) try to estimate the next pause time and if necessary, use a lower
1122 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1123 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1124 * 2) limit the target pause time to max_pause/2, so that the normal
1125 * small fluctuations of task_ratelimit won't trigger rule (1) and
1126 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1128 t = min(t, 1 + max_pause / 2);
1129 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1132 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1133 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1134 * When the 16 consecutive reads are often interrupted by some dirty
1135 * throttling pause during the async writes, cfq will go into idles
1136 * (deadline is fine). So push nr_dirtied_pause as high as possible
1137 * until reaches DIRTY_POLL_THRESH=32 pages.
1139 if (pages < DIRTY_POLL_THRESH) {
1141 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1142 if (pages > DIRTY_POLL_THRESH) {
1143 pages = DIRTY_POLL_THRESH;
1144 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1148 pause = HZ * pages / (task_ratelimit + 1);
1149 if (pause > max_pause) {
1151 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1154 *nr_dirtied_pause = pages;
1156 * The minimal pause time will normally be half the target pause time.
1158 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1162 * balance_dirty_pages() must be called by processes which are generating dirty
1163 * data. It looks at the number of dirty pages in the machine and will force
1164 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1165 * If we're over `background_thresh' then the writeback threads are woken to
1166 * perform some writeout.
1168 static void balance_dirty_pages(struct address_space *mapping,
1169 unsigned long pages_dirtied)
1171 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1172 unsigned long bdi_reclaimable;
1173 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1174 unsigned long bdi_dirty;
1175 unsigned long freerun;
1176 unsigned long background_thresh;
1177 unsigned long dirty_thresh;
1178 unsigned long bdi_thresh;
1183 int nr_dirtied_pause;
1184 bool dirty_exceeded = false;
1185 unsigned long task_ratelimit;
1186 unsigned long dirty_ratelimit;
1187 unsigned long pos_ratio;
1188 struct backing_dev_info *bdi = mapping->backing_dev_info;
1189 unsigned long start_time = jiffies;
1192 unsigned long now = jiffies;
1195 * Unstable writes are a feature of certain networked
1196 * filesystems (i.e. NFS) in which data may have been
1197 * written to the server's write cache, but has not yet
1198 * been flushed to permanent storage.
1200 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1201 global_page_state(NR_UNSTABLE_NFS);
1202 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1204 global_dirty_limits(&background_thresh, &dirty_thresh);
1207 * Throttle it only when the background writeback cannot
1208 * catch-up. This avoids (excessively) small writeouts
1209 * when the bdi limits are ramping up.
1211 freerun = dirty_freerun_ceiling(dirty_thresh,
1213 if (nr_dirty <= freerun) {
1214 current->dirty_paused_when = now;
1215 current->nr_dirtied = 0;
1216 current->nr_dirtied_pause =
1217 dirty_poll_interval(nr_dirty, dirty_thresh);
1221 if (unlikely(!writeback_in_progress(bdi)))
1222 bdi_start_background_writeback(bdi);
1225 * bdi_thresh is not treated as some limiting factor as
1226 * dirty_thresh, due to reasons
1227 * - in JBOD setup, bdi_thresh can fluctuate a lot
1228 * - in a system with HDD and USB key, the USB key may somehow
1229 * go into state (bdi_dirty >> bdi_thresh) either because
1230 * bdi_dirty starts high, or because bdi_thresh drops low.
1231 * In this case we don't want to hard throttle the USB key
1232 * dirtiers for 100 seconds until bdi_dirty drops under
1233 * bdi_thresh. Instead the auxiliary bdi control line in
1234 * bdi_position_ratio() will let the dirtier task progress
1235 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1237 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1240 * In order to avoid the stacked BDI deadlock we need
1241 * to ensure we accurately count the 'dirty' pages when
1242 * the threshold is low.
1244 * Otherwise it would be possible to get thresh+n pages
1245 * reported dirty, even though there are thresh-m pages
1246 * actually dirty; with m+n sitting in the percpu
1249 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1250 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1251 bdi_dirty = bdi_reclaimable +
1252 bdi_stat_sum(bdi, BDI_WRITEBACK);
1254 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1255 bdi_dirty = bdi_reclaimable +
1256 bdi_stat(bdi, BDI_WRITEBACK);
1259 dirty_exceeded = (bdi_dirty > bdi_thresh) &&
1260 (nr_dirty > dirty_thresh);
1261 if (dirty_exceeded && !bdi->dirty_exceeded)
1262 bdi->dirty_exceeded = 1;
1264 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1265 nr_dirty, bdi_thresh, bdi_dirty,
1268 dirty_ratelimit = bdi->dirty_ratelimit;
1269 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1270 background_thresh, nr_dirty,
1271 bdi_thresh, bdi_dirty);
1272 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1273 RATELIMIT_CALC_SHIFT;
1274 max_pause = bdi_max_pause(bdi, bdi_dirty);
1275 min_pause = bdi_min_pause(bdi, max_pause,
1276 task_ratelimit, dirty_ratelimit,
1279 if (unlikely(task_ratelimit == 0)) {
1284 period = HZ * pages_dirtied / task_ratelimit;
1286 if (current->dirty_paused_when)
1287 pause -= now - current->dirty_paused_when;
1289 * For less than 1s think time (ext3/4 may block the dirtier
1290 * for up to 800ms from time to time on 1-HDD; so does xfs,
1291 * however at much less frequency), try to compensate it in
1292 * future periods by updating the virtual time; otherwise just
1293 * do a reset, as it may be a light dirtier.
1295 if (pause < min_pause) {
1296 trace_balance_dirty_pages(bdi,
1309 current->dirty_paused_when = now;
1310 current->nr_dirtied = 0;
1311 } else if (period) {
1312 current->dirty_paused_when += period;
1313 current->nr_dirtied = 0;
1314 } else if (current->nr_dirtied_pause <= pages_dirtied)
1315 current->nr_dirtied_pause += pages_dirtied;
1318 if (unlikely(pause > max_pause)) {
1319 /* for occasional dropped task_ratelimit */
1320 now += min(pause - max_pause, max_pause);
1325 trace_balance_dirty_pages(bdi,
1337 __set_current_state(TASK_KILLABLE);
1338 io_schedule_timeout(pause);
1340 current->dirty_paused_when = now + pause;
1341 current->nr_dirtied = 0;
1342 current->nr_dirtied_pause = nr_dirtied_pause;
1345 * This is typically equal to (nr_dirty < dirty_thresh) and can
1346 * also keep "1000+ dd on a slow USB stick" under control.
1352 * In the case of an unresponding NFS server and the NFS dirty
1353 * pages exceeds dirty_thresh, give the other good bdi's a pipe
1354 * to go through, so that tasks on them still remain responsive.
1356 * In theory 1 page is enough to keep the comsumer-producer
1357 * pipe going: the flusher cleans 1 page => the task dirties 1
1358 * more page. However bdi_dirty has accounting errors. So use
1359 * the larger and more IO friendly bdi_stat_error.
1361 if (bdi_dirty <= bdi_stat_error(bdi))
1364 if (fatal_signal_pending(current))
1368 if (!dirty_exceeded && bdi->dirty_exceeded)
1369 bdi->dirty_exceeded = 0;
1371 if (writeback_in_progress(bdi))
1375 * In laptop mode, we wait until hitting the higher threshold before
1376 * starting background writeout, and then write out all the way down
1377 * to the lower threshold. So slow writers cause minimal disk activity.
1379 * In normal mode, we start background writeout at the lower
1380 * background_thresh, to keep the amount of dirty memory low.
1385 if (nr_reclaimable > background_thresh)
1386 bdi_start_background_writeback(bdi);
1389 void set_page_dirty_balance(struct page *page, int page_mkwrite)
1391 if (set_page_dirty(page) || page_mkwrite) {
1392 struct address_space *mapping = page_mapping(page);
1395 balance_dirty_pages_ratelimited(mapping);
1399 static DEFINE_PER_CPU(int, bdp_ratelimits);
1402 * Normal tasks are throttled by
1404 * dirty tsk->nr_dirtied_pause pages;
1405 * take a snap in balance_dirty_pages();
1407 * However there is a worst case. If every task exit immediately when dirtied
1408 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1409 * called to throttle the page dirties. The solution is to save the not yet
1410 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1411 * randomly into the running tasks. This works well for the above worst case,
1412 * as the new task will pick up and accumulate the old task's leaked dirty
1413 * count and eventually get throttled.
1415 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1418 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
1419 * @mapping: address_space which was dirtied
1420 * @nr_pages_dirtied: number of pages which the caller has just dirtied
1422 * Processes which are dirtying memory should call in here once for each page
1423 * which was newly dirtied. The function will periodically check the system's
1424 * dirty state and will initiate writeback if needed.
1426 * On really big machines, get_writeback_state is expensive, so try to avoid
1427 * calling it too often (ratelimiting). But once we're over the dirty memory
1428 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1429 * from overshooting the limit by (ratelimit_pages) each.
1431 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1432 unsigned long nr_pages_dirtied)
1434 struct backing_dev_info *bdi = mapping->backing_dev_info;
1438 if (!bdi_cap_account_dirty(bdi))
1441 ratelimit = current->nr_dirtied_pause;
1442 if (bdi->dirty_exceeded)
1443 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1447 * This prevents one CPU to accumulate too many dirtied pages without
1448 * calling into balance_dirty_pages(), which can happen when there are
1449 * 1000+ tasks, all of them start dirtying pages at exactly the same
1450 * time, hence all honoured too large initial task->nr_dirtied_pause.
1452 p = &__get_cpu_var(bdp_ratelimits);
1453 if (unlikely(current->nr_dirtied >= ratelimit))
1455 else if (unlikely(*p >= ratelimit_pages)) {
1460 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1461 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1462 * the dirty throttling and livelock other long-run dirtiers.
1464 p = &__get_cpu_var(dirty_throttle_leaks);
1465 if (*p > 0 && current->nr_dirtied < ratelimit) {
1466 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1467 *p -= nr_pages_dirtied;
1468 current->nr_dirtied += nr_pages_dirtied;
1472 if (unlikely(current->nr_dirtied >= ratelimit))
1473 balance_dirty_pages(mapping, current->nr_dirtied);
1475 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
1477 void throttle_vm_writeout(gfp_t gfp_mask)
1479 unsigned long background_thresh;
1480 unsigned long dirty_thresh;
1483 global_dirty_limits(&background_thresh, &dirty_thresh);
1484 dirty_thresh = hard_dirty_limit(dirty_thresh);
1487 * Boost the allowable dirty threshold a bit for page
1488 * allocators so they don't get DoS'ed by heavy writers
1490 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1492 if (global_page_state(NR_UNSTABLE_NFS) +
1493 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1495 congestion_wait(BLK_RW_ASYNC, HZ/10);
1498 * The caller might hold locks which can prevent IO completion
1499 * or progress in the filesystem. So we cannot just sit here
1500 * waiting for IO to complete.
1502 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1508 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1510 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1511 void __user *buffer, size_t *length, loff_t *ppos)
1513 proc_dointvec(table, write, buffer, length, ppos);
1514 bdi_arm_supers_timer();
1519 void laptop_mode_timer_fn(unsigned long data)
1521 struct request_queue *q = (struct request_queue *)data;
1522 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1523 global_page_state(NR_UNSTABLE_NFS);
1526 * We want to write everything out, not just down to the dirty
1529 if (bdi_has_dirty_io(&q->backing_dev_info))
1530 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1531 WB_REASON_LAPTOP_TIMER);
1535 * We've spun up the disk and we're in laptop mode: schedule writeback
1536 * of all dirty data a few seconds from now. If the flush is already scheduled
1537 * then push it back - the user is still using the disk.
1539 void laptop_io_completion(struct backing_dev_info *info)
1541 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1545 * We're in laptop mode and we've just synced. The sync's writes will have
1546 * caused another writeback to be scheduled by laptop_io_completion.
1547 * Nothing needs to be written back anymore, so we unschedule the writeback.
1549 void laptop_sync_completion(void)
1551 struct backing_dev_info *bdi;
1555 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1556 del_timer(&bdi->laptop_mode_wb_timer);
1563 * If ratelimit_pages is too high then we can get into dirty-data overload
1564 * if a large number of processes all perform writes at the same time.
1565 * If it is too low then SMP machines will call the (expensive)
1566 * get_writeback_state too often.
1568 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1569 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1573 void writeback_set_ratelimit(void)
1575 unsigned long background_thresh;
1576 unsigned long dirty_thresh;
1577 global_dirty_limits(&background_thresh, &dirty_thresh);
1578 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1579 if (ratelimit_pages < 16)
1580 ratelimit_pages = 16;
1583 static int __cpuinit
1584 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1586 writeback_set_ratelimit();
1590 static struct notifier_block __cpuinitdata ratelimit_nb = {
1591 .notifier_call = ratelimit_handler,
1596 * Called early on to tune the page writeback dirty limits.
1598 * We used to scale dirty pages according to how total memory
1599 * related to pages that could be allocated for buffers (by
1600 * comparing nr_free_buffer_pages() to vm_total_pages.
1602 * However, that was when we used "dirty_ratio" to scale with
1603 * all memory, and we don't do that any more. "dirty_ratio"
1604 * is now applied to total non-HIGHPAGE memory (by subtracting
1605 * totalhigh_pages from vm_total_pages), and as such we can't
1606 * get into the old insane situation any more where we had
1607 * large amounts of dirty pages compared to a small amount of
1608 * non-HIGHMEM memory.
1610 * But we might still want to scale the dirty_ratio by how
1611 * much memory the box has..
1613 void __init page_writeback_init(void)
1617 writeback_set_ratelimit();
1618 register_cpu_notifier(&ratelimit_nb);
1620 shift = calc_period_shift();
1621 prop_descriptor_init(&vm_completions, shift);
1625 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1626 * @mapping: address space structure to write
1627 * @start: starting page index
1628 * @end: ending page index (inclusive)
1630 * This function scans the page range from @start to @end (inclusive) and tags
1631 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1632 * that write_cache_pages (or whoever calls this function) will then use
1633 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1634 * used to avoid livelocking of writeback by a process steadily creating new
1635 * dirty pages in the file (thus it is important for this function to be quick
1636 * so that it can tag pages faster than a dirtying process can create them).
1639 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1641 void tag_pages_for_writeback(struct address_space *mapping,
1642 pgoff_t start, pgoff_t end)
1644 #define WRITEBACK_TAG_BATCH 4096
1645 unsigned long tagged;
1648 spin_lock_irq(&mapping->tree_lock);
1649 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1650 &start, end, WRITEBACK_TAG_BATCH,
1651 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1652 spin_unlock_irq(&mapping->tree_lock);
1653 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1655 /* We check 'start' to handle wrapping when end == ~0UL */
1656 } while (tagged >= WRITEBACK_TAG_BATCH && start);
1658 EXPORT_SYMBOL(tag_pages_for_writeback);
1661 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1662 * @mapping: address space structure to write
1663 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1664 * @writepage: function called for each page
1665 * @data: data passed to writepage function
1667 * If a page is already under I/O, write_cache_pages() skips it, even
1668 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1669 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1670 * and msync() need to guarantee that all the data which was dirty at the time
1671 * the call was made get new I/O started against them. If wbc->sync_mode is
1672 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1673 * existing IO to complete.
1675 * To avoid livelocks (when other process dirties new pages), we first tag
1676 * pages which should be written back with TOWRITE tag and only then start
1677 * writing them. For data-integrity sync we have to be careful so that we do
1678 * not miss some pages (e.g., because some other process has cleared TOWRITE
1679 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1680 * by the process clearing the DIRTY tag (and submitting the page for IO).
1682 int write_cache_pages(struct address_space *mapping,
1683 struct writeback_control *wbc, writepage_t writepage,
1688 struct pagevec pvec;
1690 pgoff_t uninitialized_var(writeback_index);
1692 pgoff_t end; /* Inclusive */
1695 int range_whole = 0;
1698 pagevec_init(&pvec, 0);
1699 if (wbc->range_cyclic) {
1700 writeback_index = mapping->writeback_index; /* prev offset */
1701 index = writeback_index;
1708 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1709 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1710 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1712 cycled = 1; /* ignore range_cyclic tests */
1714 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1715 tag = PAGECACHE_TAG_TOWRITE;
1717 tag = PAGECACHE_TAG_DIRTY;
1719 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1720 tag_pages_for_writeback(mapping, index, end);
1722 while (!done && (index <= end)) {
1725 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1726 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1730 for (i = 0; i < nr_pages; i++) {
1731 struct page *page = pvec.pages[i];
1734 * At this point, the page may be truncated or
1735 * invalidated (changing page->mapping to NULL), or
1736 * even swizzled back from swapper_space to tmpfs file
1737 * mapping. However, page->index will not change
1738 * because we have a reference on the page.
1740 if (page->index > end) {
1742 * can't be range_cyclic (1st pass) because
1743 * end == -1 in that case.
1749 done_index = page->index;
1754 * Page truncated or invalidated. We can freely skip it
1755 * then, even for data integrity operations: the page
1756 * has disappeared concurrently, so there could be no
1757 * real expectation of this data interity operation
1758 * even if there is now a new, dirty page at the same
1759 * pagecache address.
1761 if (unlikely(page->mapping != mapping)) {
1767 if (!PageDirty(page)) {
1768 /* someone wrote it for us */
1769 goto continue_unlock;
1772 if (PageWriteback(page)) {
1773 if (wbc->sync_mode != WB_SYNC_NONE)
1774 wait_on_page_writeback(page);
1776 goto continue_unlock;
1779 BUG_ON(PageWriteback(page));
1780 if (!clear_page_dirty_for_io(page))
1781 goto continue_unlock;
1783 trace_wbc_writepage(wbc, mapping->backing_dev_info);
1784 ret = (*writepage)(page, wbc, data);
1785 if (unlikely(ret)) {
1786 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1791 * done_index is set past this page,
1792 * so media errors will not choke
1793 * background writeout for the entire
1794 * file. This has consequences for
1795 * range_cyclic semantics (ie. it may
1796 * not be suitable for data integrity
1799 done_index = page->index + 1;
1806 * We stop writing back only if we are not doing
1807 * integrity sync. In case of integrity sync we have to
1808 * keep going until we have written all the pages
1809 * we tagged for writeback prior to entering this loop.
1811 if (--wbc->nr_to_write <= 0 &&
1812 wbc->sync_mode == WB_SYNC_NONE) {
1817 pagevec_release(&pvec);
1820 if (!cycled && !done) {
1823 * We hit the last page and there is more work to be done: wrap
1824 * back to the start of the file
1828 end = writeback_index - 1;
1831 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1832 mapping->writeback_index = done_index;
1836 EXPORT_SYMBOL(write_cache_pages);
1839 * Function used by generic_writepages to call the real writepage
1840 * function and set the mapping flags on error
1842 static int __writepage(struct page *page, struct writeback_control *wbc,
1845 struct address_space *mapping = data;
1846 int ret = mapping->a_ops->writepage(page, wbc);
1847 mapping_set_error(mapping, ret);
1852 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1853 * @mapping: address space structure to write
1854 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1856 * This is a library function, which implements the writepages()
1857 * address_space_operation.
1859 int generic_writepages(struct address_space *mapping,
1860 struct writeback_control *wbc)
1862 struct blk_plug plug;
1865 /* deal with chardevs and other special file */
1866 if (!mapping->a_ops->writepage)
1869 blk_start_plug(&plug);
1870 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1871 blk_finish_plug(&plug);
1875 EXPORT_SYMBOL(generic_writepages);
1877 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1881 if (wbc->nr_to_write <= 0)
1883 if (mapping->a_ops->writepages)
1884 ret = mapping->a_ops->writepages(mapping, wbc);
1886 ret = generic_writepages(mapping, wbc);
1891 * write_one_page - write out a single page and optionally wait on I/O
1892 * @page: the page to write
1893 * @wait: if true, wait on writeout
1895 * The page must be locked by the caller and will be unlocked upon return.
1897 * write_one_page() returns a negative error code if I/O failed.
1899 int write_one_page(struct page *page, int wait)
1901 struct address_space *mapping = page->mapping;
1903 struct writeback_control wbc = {
1904 .sync_mode = WB_SYNC_ALL,
1908 BUG_ON(!PageLocked(page));
1911 wait_on_page_writeback(page);
1913 if (clear_page_dirty_for_io(page)) {
1914 page_cache_get(page);
1915 ret = mapping->a_ops->writepage(page, &wbc);
1916 if (ret == 0 && wait) {
1917 wait_on_page_writeback(page);
1918 if (PageError(page))
1921 page_cache_release(page);
1927 EXPORT_SYMBOL(write_one_page);
1930 * For address_spaces which do not use buffers nor write back.
1932 int __set_page_dirty_no_writeback(struct page *page)
1934 if (!PageDirty(page))
1935 return !TestSetPageDirty(page);
1940 * Helper function for set_page_dirty family.
1941 * NOTE: This relies on being atomic wrt interrupts.
1943 void account_page_dirtied(struct page *page, struct address_space *mapping)
1945 if (mapping_cap_account_dirty(mapping)) {
1946 __inc_zone_page_state(page, NR_FILE_DIRTY);
1947 __inc_zone_page_state(page, NR_DIRTIED);
1948 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1949 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1950 task_io_account_write(PAGE_CACHE_SIZE);
1951 current->nr_dirtied++;
1952 this_cpu_inc(bdp_ratelimits);
1955 EXPORT_SYMBOL(account_page_dirtied);
1958 * Helper function for set_page_writeback family.
1959 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1962 void account_page_writeback(struct page *page)
1964 inc_zone_page_state(page, NR_WRITEBACK);
1966 EXPORT_SYMBOL(account_page_writeback);
1969 * For address_spaces which do not use buffers. Just tag the page as dirty in
1972 * This is also used when a single buffer is being dirtied: we want to set the
1973 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1974 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1976 * Most callers have locked the page, which pins the address_space in memory.
1977 * But zap_pte_range() does not lock the page, however in that case the
1978 * mapping is pinned by the vma's ->vm_file reference.
1980 * We take care to handle the case where the page was truncated from the
1981 * mapping by re-checking page_mapping() inside tree_lock.
1983 int __set_page_dirty_nobuffers(struct page *page)
1985 if (!TestSetPageDirty(page)) {
1986 struct address_space *mapping = page_mapping(page);
1987 struct address_space *mapping2;
1992 spin_lock_irq(&mapping->tree_lock);
1993 mapping2 = page_mapping(page);
1994 if (mapping2) { /* Race with truncate? */
1995 BUG_ON(mapping2 != mapping);
1996 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1997 account_page_dirtied(page, mapping);
1998 radix_tree_tag_set(&mapping->page_tree,
1999 page_index(page), PAGECACHE_TAG_DIRTY);
2001 spin_unlock_irq(&mapping->tree_lock);
2002 if (mapping->host) {
2003 /* !PageAnon && !swapper_space */
2004 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2010 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2013 * Call this whenever redirtying a page, to de-account the dirty counters
2014 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2015 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2016 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2019 void account_page_redirty(struct page *page)
2021 struct address_space *mapping = page->mapping;
2022 if (mapping && mapping_cap_account_dirty(mapping)) {
2023 current->nr_dirtied--;
2024 dec_zone_page_state(page, NR_DIRTIED);
2025 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2028 EXPORT_SYMBOL(account_page_redirty);
2031 * When a writepage implementation decides that it doesn't want to write this
2032 * page for some reason, it should redirty the locked page via
2033 * redirty_page_for_writepage() and it should then unlock the page and return 0
2035 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2037 wbc->pages_skipped++;
2038 account_page_redirty(page);
2039 return __set_page_dirty_nobuffers(page);
2041 EXPORT_SYMBOL(redirty_page_for_writepage);
2046 * For pages with a mapping this should be done under the page lock
2047 * for the benefit of asynchronous memory errors who prefer a consistent
2048 * dirty state. This rule can be broken in some special cases,
2049 * but should be better not to.
2051 * If the mapping doesn't provide a set_page_dirty a_op, then
2052 * just fall through and assume that it wants buffer_heads.
2054 int set_page_dirty(struct page *page)
2056 struct address_space *mapping = page_mapping(page);
2058 if (likely(mapping)) {
2059 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2061 * readahead/lru_deactivate_page could remain
2062 * PG_readahead/PG_reclaim due to race with end_page_writeback
2063 * About readahead, if the page is written, the flags would be
2064 * reset. So no problem.
2065 * About lru_deactivate_page, if the page is redirty, the flag
2066 * will be reset. So no problem. but if the page is used by readahead
2067 * it will confuse readahead and make it restart the size rampup
2068 * process. But it's a trivial problem.
2070 ClearPageReclaim(page);
2073 spd = __set_page_dirty_buffers;
2075 return (*spd)(page);
2077 if (!PageDirty(page)) {
2078 if (!TestSetPageDirty(page))
2083 EXPORT_SYMBOL(set_page_dirty);
2086 * set_page_dirty() is racy if the caller has no reference against
2087 * page->mapping->host, and if the page is unlocked. This is because another
2088 * CPU could truncate the page off the mapping and then free the mapping.
2090 * Usually, the page _is_ locked, or the caller is a user-space process which
2091 * holds a reference on the inode by having an open file.
2093 * In other cases, the page should be locked before running set_page_dirty().
2095 int set_page_dirty_lock(struct page *page)
2100 ret = set_page_dirty(page);
2104 EXPORT_SYMBOL(set_page_dirty_lock);
2107 * Clear a page's dirty flag, while caring for dirty memory accounting.
2108 * Returns true if the page was previously dirty.
2110 * This is for preparing to put the page under writeout. We leave the page
2111 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2112 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2113 * implementation will run either set_page_writeback() or set_page_dirty(),
2114 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2117 * This incoherency between the page's dirty flag and radix-tree tag is
2118 * unfortunate, but it only exists while the page is locked.
2120 int clear_page_dirty_for_io(struct page *page)
2122 struct address_space *mapping = page_mapping(page);
2124 BUG_ON(!PageLocked(page));
2126 if (mapping && mapping_cap_account_dirty(mapping)) {
2128 * Yes, Virginia, this is indeed insane.
2130 * We use this sequence to make sure that
2131 * (a) we account for dirty stats properly
2132 * (b) we tell the low-level filesystem to
2133 * mark the whole page dirty if it was
2134 * dirty in a pagetable. Only to then
2135 * (c) clean the page again and return 1 to
2136 * cause the writeback.
2138 * This way we avoid all nasty races with the
2139 * dirty bit in multiple places and clearing
2140 * them concurrently from different threads.
2142 * Note! Normally the "set_page_dirty(page)"
2143 * has no effect on the actual dirty bit - since
2144 * that will already usually be set. But we
2145 * need the side effects, and it can help us
2148 * We basically use the page "master dirty bit"
2149 * as a serialization point for all the different
2150 * threads doing their things.
2152 if (page_mkclean(page))
2153 set_page_dirty(page);
2155 * We carefully synchronise fault handlers against
2156 * installing a dirty pte and marking the page dirty
2157 * at this point. We do this by having them hold the
2158 * page lock at some point after installing their
2159 * pte, but before marking the page dirty.
2160 * Pages are always locked coming in here, so we get
2161 * the desired exclusion. See mm/memory.c:do_wp_page()
2162 * for more comments.
2164 if (TestClearPageDirty(page)) {
2165 dec_zone_page_state(page, NR_FILE_DIRTY);
2166 dec_bdi_stat(mapping->backing_dev_info,
2172 return TestClearPageDirty(page);
2174 EXPORT_SYMBOL(clear_page_dirty_for_io);
2176 int test_clear_page_writeback(struct page *page)
2178 struct address_space *mapping = page_mapping(page);
2182 struct backing_dev_info *bdi = mapping->backing_dev_info;
2183 unsigned long flags;
2185 spin_lock_irqsave(&mapping->tree_lock, flags);
2186 ret = TestClearPageWriteback(page);
2188 radix_tree_tag_clear(&mapping->page_tree,
2190 PAGECACHE_TAG_WRITEBACK);
2191 if (bdi_cap_account_writeback(bdi)) {
2192 __dec_bdi_stat(bdi, BDI_WRITEBACK);
2193 __bdi_writeout_inc(bdi);
2196 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2198 ret = TestClearPageWriteback(page);
2201 dec_zone_page_state(page, NR_WRITEBACK);
2202 inc_zone_page_state(page, NR_WRITTEN);
2207 int test_set_page_writeback(struct page *page)
2209 struct address_space *mapping = page_mapping(page);
2213 struct backing_dev_info *bdi = mapping->backing_dev_info;
2214 unsigned long flags;
2216 spin_lock_irqsave(&mapping->tree_lock, flags);
2217 ret = TestSetPageWriteback(page);
2219 radix_tree_tag_set(&mapping->page_tree,
2221 PAGECACHE_TAG_WRITEBACK);
2222 if (bdi_cap_account_writeback(bdi))
2223 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2225 if (!PageDirty(page))
2226 radix_tree_tag_clear(&mapping->page_tree,
2228 PAGECACHE_TAG_DIRTY);
2229 radix_tree_tag_clear(&mapping->page_tree,
2231 PAGECACHE_TAG_TOWRITE);
2232 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2234 ret = TestSetPageWriteback(page);
2237 account_page_writeback(page);
2241 EXPORT_SYMBOL(test_set_page_writeback);
2244 * Return true if any of the pages in the mapping are marked with the
2247 int mapping_tagged(struct address_space *mapping, int tag)
2249 return radix_tree_tagged(&mapping->page_tree, tag);
2251 EXPORT_SYMBOL(mapping_tagged);