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>
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 * Estimate write bandwidth at 200ms intervals.
47 #define BANDWIDTH_INTERVAL max(HZ/5, 1)
49 #define RATELIMIT_CALC_SHIFT 10
52 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
53 * will look to see if it needs to force writeback or throttling.
55 static long ratelimit_pages = 32;
57 /* The following parameters are exported via /proc/sys/vm */
60 * Start background writeback (via writeback threads) at this percentage
62 int dirty_background_ratio = 10;
65 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
66 * dirty_background_ratio * the amount of dirtyable memory
68 unsigned long dirty_background_bytes;
71 * free highmem will not be subtracted from the total free memory
72 * for calculating free ratios if vm_highmem_is_dirtyable is true
74 int vm_highmem_is_dirtyable;
77 * The generator of dirty data starts writeback at this percentage
79 int vm_dirty_ratio = 20;
82 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
83 * vm_dirty_ratio * the amount of dirtyable memory
85 unsigned long vm_dirty_bytes;
88 * The interval between `kupdate'-style writebacks
90 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
93 * The longest time for which data is allowed to remain dirty
95 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
98 * Flag that makes the machine dump writes/reads and block dirtyings.
103 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
104 * a full sync is triggered after this time elapses without any disk activity.
108 EXPORT_SYMBOL(laptop_mode);
110 /* End of sysctl-exported parameters */
112 unsigned long global_dirty_limit;
115 * Scale the writeback cache size proportional to the relative writeout speeds.
117 * We do this by keeping a floating proportion between BDIs, based on page
118 * writeback completions [end_page_writeback()]. Those devices that write out
119 * pages fastest will get the larger share, while the slower will get a smaller
122 * We use page writeout completions because we are interested in getting rid of
123 * dirty pages. Having them written out is the primary goal.
125 * We introduce a concept of time, a period over which we measure these events,
126 * because demand can/will vary over time. The length of this period itself is
127 * measured in page writeback completions.
130 static struct prop_descriptor vm_completions;
133 * Work out the current dirty-memory clamping and background writeout
136 * The main aim here is to lower them aggressively if there is a lot of mapped
137 * memory around. To avoid stressing page reclaim with lots of unreclaimable
138 * pages. It is better to clamp down on writers than to start swapping, and
139 * performing lots of scanning.
141 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
143 * We don't permit the clamping level to fall below 5% - that is getting rather
146 * We make sure that the background writeout level is below the adjusted
151 * In a memory zone, there is a certain amount of pages we consider
152 * available for the page cache, which is essentially the number of
153 * free and reclaimable pages, minus some zone reserves to protect
154 * lowmem and the ability to uphold the zone's watermarks without
155 * requiring writeback.
157 * This number of dirtyable pages is the base value of which the
158 * user-configurable dirty ratio is the effictive number of pages that
159 * are allowed to be actually dirtied. Per individual zone, or
160 * globally by using the sum of dirtyable pages over all zones.
162 * Because the user is allowed to specify the dirty limit globally as
163 * absolute number of bytes, calculating the per-zone dirty limit can
164 * require translating the configured limit into a percentage of
165 * global dirtyable memory first.
168 static unsigned long highmem_dirtyable_memory(unsigned long total)
170 #ifdef CONFIG_HIGHMEM
174 for_each_node_state(node, N_HIGH_MEMORY) {
176 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
178 x += zone_page_state(z, NR_FREE_PAGES) +
179 zone_reclaimable_pages(z) - z->dirty_balance_reserve;
182 * Make sure that the number of highmem pages is never larger
183 * than the number of the total dirtyable memory. This can only
184 * occur in very strange VM situations but we want to make sure
185 * that this does not occur.
187 return min(x, total);
194 * global_dirtyable_memory - number of globally dirtyable pages
196 * Returns the global number of pages potentially available for dirty
197 * page cache. This is the base value for the global dirty limits.
199 unsigned long global_dirtyable_memory(void)
203 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages() -
204 dirty_balance_reserve;
206 if (!vm_highmem_is_dirtyable)
207 x -= highmem_dirtyable_memory(x);
209 return x + 1; /* Ensure that we never return 0 */
213 * global_dirty_limits - background-writeback and dirty-throttling thresholds
215 * Calculate the dirty thresholds based on sysctl parameters
216 * - vm.dirty_background_ratio or vm.dirty_background_bytes
217 * - vm.dirty_ratio or vm.dirty_bytes
218 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
221 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
223 unsigned long background;
225 unsigned long uninitialized_var(available_memory);
226 struct task_struct *tsk;
228 if (!vm_dirty_bytes || !dirty_background_bytes)
229 available_memory = global_dirtyable_memory();
232 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
234 dirty = (vm_dirty_ratio * available_memory) / 100;
236 if (dirty_background_bytes)
237 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
239 background = (dirty_background_ratio * available_memory) / 100;
241 if (background >= dirty)
242 background = dirty / 2;
244 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
245 background += background / 4;
248 *pbackground = background;
250 trace_global_dirty_state(background, dirty);
254 * zone_dirtyable_memory - number of dirtyable pages in a zone
257 * Returns the zone's number of pages potentially available for dirty
258 * page cache. This is the base value for the per-zone dirty limits.
260 static unsigned long zone_dirtyable_memory(struct zone *zone)
263 * The effective global number of dirtyable pages may exclude
264 * highmem as a big-picture measure to keep the ratio between
265 * dirty memory and lowmem reasonable.
267 * But this function is purely about the individual zone and a
268 * highmem zone can hold its share of dirty pages, so we don't
269 * care about vm_highmem_is_dirtyable here.
271 return zone_page_state(zone, NR_FREE_PAGES) +
272 zone_reclaimable_pages(zone) -
273 zone->dirty_balance_reserve;
277 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
280 * Returns the maximum number of dirty pages allowed in a zone, based
281 * on the zone's dirtyable memory.
283 static unsigned long zone_dirty_limit(struct zone *zone)
285 unsigned long zone_memory = zone_dirtyable_memory(zone);
286 struct task_struct *tsk = current;
290 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
291 zone_memory / global_dirtyable_memory();
293 dirty = vm_dirty_ratio * zone_memory / 100;
295 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
302 * zone_dirty_ok - tells whether a zone is within its dirty limits
303 * @zone: the zone to check
305 * Returns %true when the dirty pages in @zone are within the zone's
306 * dirty limit, %false if the limit is exceeded.
308 bool zone_dirty_ok(struct zone *zone)
310 unsigned long limit = zone_dirty_limit(zone);
312 return zone_page_state(zone, NR_FILE_DIRTY) +
313 zone_page_state(zone, NR_UNSTABLE_NFS) +
314 zone_page_state(zone, NR_WRITEBACK) <= limit;
318 * couple the period to the dirty_ratio:
320 * period/2 ~ roundup_pow_of_two(dirty limit)
322 static int calc_period_shift(void)
324 unsigned long dirty_total;
327 dirty_total = vm_dirty_bytes / PAGE_SIZE;
329 dirty_total = (vm_dirty_ratio * global_dirtyable_memory()) /
331 return 2 + ilog2(dirty_total - 1);
335 * update the period when the dirty threshold changes.
337 static void update_completion_period(void)
339 int shift = calc_period_shift();
340 prop_change_shift(&vm_completions, shift);
342 writeback_set_ratelimit();
345 int dirty_background_ratio_handler(struct ctl_table *table, int write,
346 void __user *buffer, size_t *lenp,
351 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
352 if (ret == 0 && write)
353 dirty_background_bytes = 0;
357 int dirty_background_bytes_handler(struct ctl_table *table, int write,
358 void __user *buffer, size_t *lenp,
363 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
364 if (ret == 0 && write)
365 dirty_background_ratio = 0;
369 int dirty_ratio_handler(struct ctl_table *table, int write,
370 void __user *buffer, size_t *lenp,
373 int old_ratio = vm_dirty_ratio;
376 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
377 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
378 update_completion_period();
384 int dirty_bytes_handler(struct ctl_table *table, int write,
385 void __user *buffer, size_t *lenp,
388 unsigned long old_bytes = vm_dirty_bytes;
391 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
392 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
393 update_completion_period();
400 * Increment the BDI's writeout completion count and the global writeout
401 * completion count. Called from test_clear_page_writeback().
403 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
405 __inc_bdi_stat(bdi, BDI_WRITTEN);
406 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
410 void bdi_writeout_inc(struct backing_dev_info *bdi)
414 local_irq_save(flags);
415 __bdi_writeout_inc(bdi);
416 local_irq_restore(flags);
418 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
421 * Obtain an accurate fraction of the BDI's portion.
423 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
424 long *numerator, long *denominator)
426 prop_fraction_percpu(&vm_completions, &bdi->completions,
427 numerator, denominator);
431 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
432 * registered backing devices, which, for obvious reasons, can not
435 static unsigned int bdi_min_ratio;
437 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
441 spin_lock_bh(&bdi_lock);
442 if (min_ratio > bdi->max_ratio) {
445 min_ratio -= bdi->min_ratio;
446 if (bdi_min_ratio + min_ratio < 100) {
447 bdi_min_ratio += min_ratio;
448 bdi->min_ratio += min_ratio;
453 spin_unlock_bh(&bdi_lock);
458 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
465 spin_lock_bh(&bdi_lock);
466 if (bdi->min_ratio > max_ratio) {
469 bdi->max_ratio = max_ratio;
470 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
472 spin_unlock_bh(&bdi_lock);
476 EXPORT_SYMBOL(bdi_set_max_ratio);
478 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
479 unsigned long bg_thresh)
481 return (thresh + bg_thresh) / 2;
484 static unsigned long hard_dirty_limit(unsigned long thresh)
486 return max(thresh, global_dirty_limit);
490 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
491 * @bdi: the backing_dev_info to query
492 * @dirty: global dirty limit in pages
494 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
495 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
497 * Note that balance_dirty_pages() will only seriously take it as a hard limit
498 * when sleeping max_pause per page is not enough to keep the dirty pages under
499 * control. For example, when the device is completely stalled due to some error
500 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
501 * In the other normal situations, it acts more gently by throttling the tasks
502 * more (rather than completely block them) when the bdi dirty pages go high.
504 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
505 * - starving fast devices
506 * - piling up dirty pages (that will take long time to sync) on slow devices
508 * The bdi's share of dirty limit will be adapting to its throughput and
509 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
511 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
514 long numerator, denominator;
517 * Calculate this BDI's share of the dirty ratio.
519 bdi_writeout_fraction(bdi, &numerator, &denominator);
521 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
522 bdi_dirty *= numerator;
523 do_div(bdi_dirty, denominator);
525 bdi_dirty += (dirty * bdi->min_ratio) / 100;
526 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
527 bdi_dirty = dirty * bdi->max_ratio / 100;
533 * Dirty position control.
535 * (o) global/bdi setpoints
537 * We want the dirty pages be balanced around the global/bdi setpoints.
538 * When the number of dirty pages is higher/lower than the setpoint, the
539 * dirty position control ratio (and hence task dirty ratelimit) will be
540 * decreased/increased to bring the dirty pages back to the setpoint.
542 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
544 * if (dirty < setpoint) scale up pos_ratio
545 * if (dirty > setpoint) scale down pos_ratio
547 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
548 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
550 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
552 * (o) global control line
556 * | |<===== global dirty control scope ======>|
564 * 1.0 ................................*
570 * 0 +------------.------------------.----------------------*------------->
571 * freerun^ setpoint^ limit^ dirty pages
573 * (o) bdi control line
581 * | * |<=========== span ============>|
582 * 1.0 .......................*
594 * 1/4 ...............................................* * * * * * * * * * * *
598 * 0 +----------------------.-------------------------------.------------->
599 * bdi_setpoint^ x_intercept^
601 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
602 * be smoothly throttled down to normal if it starts high in situations like
603 * - start writing to a slow SD card and a fast disk at the same time. The SD
604 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
605 * - the bdi dirty thresh drops quickly due to change of JBOD workload
607 static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
608 unsigned long thresh,
609 unsigned long bg_thresh,
611 unsigned long bdi_thresh,
612 unsigned long bdi_dirty)
614 unsigned long write_bw = bdi->avg_write_bandwidth;
615 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
616 unsigned long limit = hard_dirty_limit(thresh);
617 unsigned long x_intercept;
618 unsigned long setpoint; /* dirty pages' target balance point */
619 unsigned long bdi_setpoint;
621 long long pos_ratio; /* for scaling up/down the rate limit */
624 if (unlikely(dirty >= limit))
631 * f(dirty) := 1.0 + (----------------)
634 * it's a 3rd order polynomial that subjects to
636 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
637 * (2) f(setpoint) = 1.0 => the balance point
638 * (3) f(limit) = 0 => the hard limit
639 * (4) df/dx <= 0 => negative feedback control
640 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
641 * => fast response on large errors; small oscillation near setpoint
643 setpoint = (freerun + limit) / 2;
644 x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
645 limit - setpoint + 1);
647 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
648 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
649 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
652 * We have computed basic pos_ratio above based on global situation. If
653 * the bdi is over/under its share of dirty pages, we want to scale
654 * pos_ratio further down/up. That is done by the following mechanism.
660 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
662 * x_intercept - bdi_dirty
663 * := --------------------------
664 * x_intercept - bdi_setpoint
666 * The main bdi control line is a linear function that subjects to
668 * (1) f(bdi_setpoint) = 1.0
669 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
670 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
672 * For single bdi case, the dirty pages are observed to fluctuate
673 * regularly within range
674 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
675 * for various filesystems, where (2) can yield in a reasonable 12.5%
676 * fluctuation range for pos_ratio.
678 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
679 * own size, so move the slope over accordingly and choose a slope that
680 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
682 if (unlikely(bdi_thresh > thresh))
685 * It's very possible that bdi_thresh is close to 0 not because the
686 * device is slow, but that it has remained inactive for long time.
687 * Honour such devices a reasonable good (hopefully IO efficient)
688 * threshold, so that the occasional writes won't be blocked and active
689 * writes can rampup the threshold quickly.
691 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
693 * scale global setpoint to bdi's:
694 * bdi_setpoint = setpoint * bdi_thresh / thresh
696 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
697 bdi_setpoint = setpoint * (u64)x >> 16;
699 * Use span=(8*write_bw) in single bdi case as indicated by
700 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
702 * bdi_thresh thresh - bdi_thresh
703 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
706 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
707 x_intercept = bdi_setpoint + span;
709 if (bdi_dirty < x_intercept - span / 4) {
710 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
711 x_intercept - bdi_setpoint + 1);
716 * bdi reserve area, safeguard against dirty pool underrun and disk idle
717 * It may push the desired control point of global dirty pages higher
720 x_intercept = bdi_thresh / 2;
721 if (bdi_dirty < x_intercept) {
722 if (bdi_dirty > x_intercept / 8)
723 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
731 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
732 unsigned long elapsed,
733 unsigned long written)
735 const unsigned long period = roundup_pow_of_two(3 * HZ);
736 unsigned long avg = bdi->avg_write_bandwidth;
737 unsigned long old = bdi->write_bandwidth;
741 * bw = written * HZ / elapsed
743 * bw * elapsed + write_bandwidth * (period - elapsed)
744 * write_bandwidth = ---------------------------------------------------
747 bw = written - bdi->written_stamp;
749 if (unlikely(elapsed > period)) {
754 bw += (u64)bdi->write_bandwidth * (period - elapsed);
755 bw >>= ilog2(period);
758 * one more level of smoothing, for filtering out sudden spikes
760 if (avg > old && old >= (unsigned long)bw)
761 avg -= (avg - old) >> 3;
763 if (avg < old && old <= (unsigned long)bw)
764 avg += (old - avg) >> 3;
767 bdi->write_bandwidth = bw;
768 bdi->avg_write_bandwidth = avg;
772 * The global dirtyable memory and dirty threshold could be suddenly knocked
773 * down by a large amount (eg. on the startup of KVM in a swapless system).
774 * This may throw the system into deep dirty exceeded state and throttle
775 * heavy/light dirtiers alike. To retain good responsiveness, maintain
776 * global_dirty_limit for tracking slowly down to the knocked down dirty
779 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
781 unsigned long limit = global_dirty_limit;
784 * Follow up in one step.
786 if (limit < thresh) {
792 * Follow down slowly. Use the higher one as the target, because thresh
793 * may drop below dirty. This is exactly the reason to introduce
794 * global_dirty_limit which is guaranteed to lie above the dirty pages.
796 thresh = max(thresh, dirty);
797 if (limit > thresh) {
798 limit -= (limit - thresh) >> 5;
803 global_dirty_limit = limit;
806 static void global_update_bandwidth(unsigned long thresh,
810 static DEFINE_SPINLOCK(dirty_lock);
811 static unsigned long update_time;
814 * check locklessly first to optimize away locking for the most time
816 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
819 spin_lock(&dirty_lock);
820 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
821 update_dirty_limit(thresh, dirty);
824 spin_unlock(&dirty_lock);
828 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
830 * Normal bdi tasks will be curbed at or below it in long term.
831 * Obviously it should be around (write_bw / N) when there are N dd tasks.
833 static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
834 unsigned long thresh,
835 unsigned long bg_thresh,
837 unsigned long bdi_thresh,
838 unsigned long bdi_dirty,
839 unsigned long dirtied,
840 unsigned long elapsed)
842 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
843 unsigned long limit = hard_dirty_limit(thresh);
844 unsigned long setpoint = (freerun + limit) / 2;
845 unsigned long write_bw = bdi->avg_write_bandwidth;
846 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
847 unsigned long dirty_rate;
848 unsigned long task_ratelimit;
849 unsigned long balanced_dirty_ratelimit;
850 unsigned long pos_ratio;
855 * The dirty rate will match the writeout rate in long term, except
856 * when dirty pages are truncated by userspace or re-dirtied by FS.
858 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
860 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
861 bdi_thresh, bdi_dirty);
863 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
865 task_ratelimit = (u64)dirty_ratelimit *
866 pos_ratio >> RATELIMIT_CALC_SHIFT;
867 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
870 * A linear estimation of the "balanced" throttle rate. The theory is,
871 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
872 * dirty_rate will be measured to be (N * task_ratelimit). So the below
873 * formula will yield the balanced rate limit (write_bw / N).
875 * Note that the expanded form is not a pure rate feedback:
876 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
877 * but also takes pos_ratio into account:
878 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
880 * (1) is not realistic because pos_ratio also takes part in balancing
881 * the dirty rate. Consider the state
882 * pos_ratio = 0.5 (3)
883 * rate = 2 * (write_bw / N) (4)
884 * If (1) is used, it will stuck in that state! Because each dd will
886 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
888 * dirty_rate = N * task_ratelimit = write_bw (6)
889 * put (6) into (1) we get
890 * rate_(i+1) = rate_(i) (7)
892 * So we end up using (2) to always keep
893 * rate_(i+1) ~= (write_bw / N) (8)
894 * regardless of the value of pos_ratio. As long as (8) is satisfied,
895 * pos_ratio is able to drive itself to 1.0, which is not only where
896 * the dirty count meet the setpoint, but also where the slope of
897 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
899 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
902 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
904 if (unlikely(balanced_dirty_ratelimit > write_bw))
905 balanced_dirty_ratelimit = write_bw;
908 * We could safely do this and return immediately:
910 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
912 * However to get a more stable dirty_ratelimit, the below elaborated
913 * code makes use of task_ratelimit to filter out sigular points and
914 * limit the step size.
916 * The below code essentially only uses the relative value of
918 * task_ratelimit - dirty_ratelimit
919 * = (pos_ratio - 1) * dirty_ratelimit
921 * which reflects the direction and size of dirty position error.
925 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
926 * task_ratelimit is on the same side of dirty_ratelimit, too.
928 * - dirty_ratelimit > balanced_dirty_ratelimit
929 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
930 * lowering dirty_ratelimit will help meet both the position and rate
931 * control targets. Otherwise, don't update dirty_ratelimit if it will
932 * only help meet the rate target. After all, what the users ultimately
933 * feel and care are stable dirty rate and small position error.
935 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
936 * and filter out the sigular points of balanced_dirty_ratelimit. Which
937 * keeps jumping around randomly and can even leap far away at times
938 * due to the small 200ms estimation period of dirty_rate (we want to
939 * keep that period small to reduce time lags).
942 if (dirty < setpoint) {
943 x = min(bdi->balanced_dirty_ratelimit,
944 min(balanced_dirty_ratelimit, task_ratelimit));
945 if (dirty_ratelimit < x)
946 step = x - dirty_ratelimit;
948 x = max(bdi->balanced_dirty_ratelimit,
949 max(balanced_dirty_ratelimit, task_ratelimit));
950 if (dirty_ratelimit > x)
951 step = dirty_ratelimit - x;
955 * Don't pursue 100% rate matching. It's impossible since the balanced
956 * rate itself is constantly fluctuating. So decrease the track speed
957 * when it gets close to the target. Helps eliminate pointless tremors.
959 step >>= dirty_ratelimit / (2 * step + 1);
961 * Limit the tracking speed to avoid overshooting.
963 step = (step + 7) / 8;
965 if (dirty_ratelimit < balanced_dirty_ratelimit)
966 dirty_ratelimit += step;
968 dirty_ratelimit -= step;
970 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
971 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
973 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
976 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
977 unsigned long thresh,
978 unsigned long bg_thresh,
980 unsigned long bdi_thresh,
981 unsigned long bdi_dirty,
982 unsigned long start_time)
984 unsigned long now = jiffies;
985 unsigned long elapsed = now - bdi->bw_time_stamp;
986 unsigned long dirtied;
987 unsigned long written;
990 * rate-limit, only update once every 200ms.
992 if (elapsed < BANDWIDTH_INTERVAL)
995 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
996 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
999 * Skip quiet periods when disk bandwidth is under-utilized.
1000 * (at least 1s idle time between two flusher runs)
1002 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
1006 global_update_bandwidth(thresh, dirty, now);
1007 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1008 bdi_thresh, bdi_dirty,
1011 bdi_update_write_bandwidth(bdi, elapsed, written);
1014 bdi->dirtied_stamp = dirtied;
1015 bdi->written_stamp = written;
1016 bdi->bw_time_stamp = now;
1019 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
1020 unsigned long thresh,
1021 unsigned long bg_thresh,
1022 unsigned long dirty,
1023 unsigned long bdi_thresh,
1024 unsigned long bdi_dirty,
1025 unsigned long start_time)
1027 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1029 spin_lock(&bdi->wb.list_lock);
1030 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1031 bdi_thresh, bdi_dirty, start_time);
1032 spin_unlock(&bdi->wb.list_lock);
1036 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
1037 * will look to see if it needs to start dirty throttling.
1039 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1040 * global_page_state() too often. So scale it near-sqrt to the safety margin
1041 * (the number of pages we may dirty without exceeding the dirty limits).
1043 static unsigned long dirty_poll_interval(unsigned long dirty,
1044 unsigned long thresh)
1047 return 1UL << (ilog2(thresh - dirty) >> 1);
1052 static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
1053 unsigned long bdi_dirty)
1055 unsigned long bw = bdi->avg_write_bandwidth;
1056 unsigned long hi = ilog2(bw);
1057 unsigned long lo = ilog2(bdi->dirty_ratelimit);
1060 /* target for 20ms max pause on 1-dd case */
1064 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1067 * (N * 20ms) on 2^N concurrent tasks.
1070 t += (hi - lo) * (20 * HZ) / 1024;
1073 * Limit pause time for small memory systems. If sleeping for too long
1074 * time, a small pool of dirty/writeback pages may go empty and disk go
1077 * 8 serves as the safety ratio.
1079 t = min(t, bdi_dirty * HZ / (8 * bw + 1));
1082 * The pause time will be settled within range (max_pause/4, max_pause).
1083 * Apply a minimal value of 4 to get a non-zero max_pause/4.
1085 return clamp_val(t, 4, MAX_PAUSE);
1089 * balance_dirty_pages() must be called by processes which are generating dirty
1090 * data. It looks at the number of dirty pages in the machine and will force
1091 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1092 * If we're over `background_thresh' then the writeback threads are woken to
1093 * perform some writeout.
1095 static void balance_dirty_pages(struct address_space *mapping,
1096 unsigned long pages_dirtied)
1098 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1099 unsigned long bdi_reclaimable;
1100 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1101 unsigned long bdi_dirty;
1102 unsigned long freerun;
1103 unsigned long background_thresh;
1104 unsigned long dirty_thresh;
1105 unsigned long bdi_thresh;
1108 long uninitialized_var(max_pause);
1109 bool dirty_exceeded = false;
1110 unsigned long task_ratelimit;
1111 unsigned long uninitialized_var(dirty_ratelimit);
1112 unsigned long pos_ratio;
1113 struct backing_dev_info *bdi = mapping->backing_dev_info;
1114 unsigned long start_time = jiffies;
1117 unsigned long now = jiffies;
1120 * Unstable writes are a feature of certain networked
1121 * filesystems (i.e. NFS) in which data may have been
1122 * written to the server's write cache, but has not yet
1123 * been flushed to permanent storage.
1125 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1126 global_page_state(NR_UNSTABLE_NFS);
1127 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1129 global_dirty_limits(&background_thresh, &dirty_thresh);
1132 * Throttle it only when the background writeback cannot
1133 * catch-up. This avoids (excessively) small writeouts
1134 * when the bdi limits are ramping up.
1136 freerun = dirty_freerun_ceiling(dirty_thresh,
1138 if (nr_dirty <= freerun) {
1139 current->dirty_paused_when = now;
1140 current->nr_dirtied = 0;
1144 if (unlikely(!writeback_in_progress(bdi)))
1145 bdi_start_background_writeback(bdi);
1148 * bdi_thresh is not treated as some limiting factor as
1149 * dirty_thresh, due to reasons
1150 * - in JBOD setup, bdi_thresh can fluctuate a lot
1151 * - in a system with HDD and USB key, the USB key may somehow
1152 * go into state (bdi_dirty >> bdi_thresh) either because
1153 * bdi_dirty starts high, or because bdi_thresh drops low.
1154 * In this case we don't want to hard throttle the USB key
1155 * dirtiers for 100 seconds until bdi_dirty drops under
1156 * bdi_thresh. Instead the auxiliary bdi control line in
1157 * bdi_position_ratio() will let the dirtier task progress
1158 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1160 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1163 * In order to avoid the stacked BDI deadlock we need
1164 * to ensure we accurately count the 'dirty' pages when
1165 * the threshold is low.
1167 * Otherwise it would be possible to get thresh+n pages
1168 * reported dirty, even though there are thresh-m pages
1169 * actually dirty; with m+n sitting in the percpu
1172 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1173 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1174 bdi_dirty = bdi_reclaimable +
1175 bdi_stat_sum(bdi, BDI_WRITEBACK);
1177 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1178 bdi_dirty = bdi_reclaimable +
1179 bdi_stat(bdi, BDI_WRITEBACK);
1182 dirty_exceeded = (bdi_dirty > bdi_thresh) ||
1183 (nr_dirty > dirty_thresh);
1184 if (dirty_exceeded && !bdi->dirty_exceeded)
1185 bdi->dirty_exceeded = 1;
1187 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1188 nr_dirty, bdi_thresh, bdi_dirty,
1191 max_pause = bdi_max_pause(bdi, bdi_dirty);
1193 dirty_ratelimit = bdi->dirty_ratelimit;
1194 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1195 background_thresh, nr_dirty,
1196 bdi_thresh, bdi_dirty);
1197 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1198 RATELIMIT_CALC_SHIFT;
1199 if (unlikely(task_ratelimit == 0)) {
1204 period = HZ * pages_dirtied / task_ratelimit;
1206 if (current->dirty_paused_when)
1207 pause -= now - current->dirty_paused_when;
1209 * For less than 1s think time (ext3/4 may block the dirtier
1210 * for up to 800ms from time to time on 1-HDD; so does xfs,
1211 * however at much less frequency), try to compensate it in
1212 * future periods by updating the virtual time; otherwise just
1213 * do a reset, as it may be a light dirtier.
1215 if (unlikely(pause <= 0)) {
1216 trace_balance_dirty_pages(bdi,
1229 current->dirty_paused_when = now;
1230 current->nr_dirtied = 0;
1231 } else if (period) {
1232 current->dirty_paused_when += period;
1233 current->nr_dirtied = 0;
1235 pause = 1; /* avoid resetting nr_dirtied_pause below */
1238 pause = min(pause, max_pause);
1241 trace_balance_dirty_pages(bdi,
1253 __set_current_state(TASK_KILLABLE);
1254 io_schedule_timeout(pause);
1256 current->dirty_paused_when = now + pause;
1257 current->nr_dirtied = 0;
1260 * This is typically equal to (nr_dirty < dirty_thresh) and can
1261 * also keep "1000+ dd on a slow USB stick" under control.
1267 * In the case of an unresponding NFS server and the NFS dirty
1268 * pages exceeds dirty_thresh, give the other good bdi's a pipe
1269 * to go through, so that tasks on them still remain responsive.
1271 * In theory 1 page is enough to keep the comsumer-producer
1272 * pipe going: the flusher cleans 1 page => the task dirties 1
1273 * more page. However bdi_dirty has accounting errors. So use
1274 * the larger and more IO friendly bdi_stat_error.
1276 if (bdi_dirty <= bdi_stat_error(bdi))
1279 if (fatal_signal_pending(current))
1283 if (!dirty_exceeded && bdi->dirty_exceeded)
1284 bdi->dirty_exceeded = 0;
1286 if (pause == 0) { /* in freerun area */
1287 current->nr_dirtied_pause =
1288 dirty_poll_interval(nr_dirty, dirty_thresh);
1289 } else if (period <= max_pause / 4 &&
1290 pages_dirtied >= current->nr_dirtied_pause) {
1291 current->nr_dirtied_pause = clamp_val(
1292 dirty_ratelimit * (max_pause / 2) / HZ,
1293 pages_dirtied + pages_dirtied / 8,
1295 } else if (pause >= max_pause) {
1296 current->nr_dirtied_pause = 1 | clamp_val(
1297 dirty_ratelimit * (max_pause / 2) / HZ,
1299 pages_dirtied - pages_dirtied / 8);
1302 if (writeback_in_progress(bdi))
1306 * In laptop mode, we wait until hitting the higher threshold before
1307 * starting background writeout, and then write out all the way down
1308 * to the lower threshold. So slow writers cause minimal disk activity.
1310 * In normal mode, we start background writeout at the lower
1311 * background_thresh, to keep the amount of dirty memory low.
1316 if (nr_reclaimable > background_thresh)
1317 bdi_start_background_writeback(bdi);
1320 void set_page_dirty_balance(struct page *page, int page_mkwrite)
1322 if (set_page_dirty(page) || page_mkwrite) {
1323 struct address_space *mapping = page_mapping(page);
1326 balance_dirty_pages_ratelimited(mapping);
1330 static DEFINE_PER_CPU(int, bdp_ratelimits);
1333 * Normal tasks are throttled by
1335 * dirty tsk->nr_dirtied_pause pages;
1336 * take a snap in balance_dirty_pages();
1338 * However there is a worst case. If every task exit immediately when dirtied
1339 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1340 * called to throttle the page dirties. The solution is to save the not yet
1341 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1342 * randomly into the running tasks. This works well for the above worst case,
1343 * as the new task will pick up and accumulate the old task's leaked dirty
1344 * count and eventually get throttled.
1346 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1349 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
1350 * @mapping: address_space which was dirtied
1351 * @nr_pages_dirtied: number of pages which the caller has just dirtied
1353 * Processes which are dirtying memory should call in here once for each page
1354 * which was newly dirtied. The function will periodically check the system's
1355 * dirty state and will initiate writeback if needed.
1357 * On really big machines, get_writeback_state is expensive, so try to avoid
1358 * calling it too often (ratelimiting). But once we're over the dirty memory
1359 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1360 * from overshooting the limit by (ratelimit_pages) each.
1362 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1363 unsigned long nr_pages_dirtied)
1365 struct backing_dev_info *bdi = mapping->backing_dev_info;
1369 if (!bdi_cap_account_dirty(bdi))
1372 ratelimit = current->nr_dirtied_pause;
1373 if (bdi->dirty_exceeded)
1374 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1378 * This prevents one CPU to accumulate too many dirtied pages without
1379 * calling into balance_dirty_pages(), which can happen when there are
1380 * 1000+ tasks, all of them start dirtying pages at exactly the same
1381 * time, hence all honoured too large initial task->nr_dirtied_pause.
1383 p = &__get_cpu_var(bdp_ratelimits);
1384 if (unlikely(current->nr_dirtied >= ratelimit))
1386 else if (unlikely(*p >= ratelimit_pages)) {
1391 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1392 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1393 * the dirty throttling and livelock other long-run dirtiers.
1395 p = &__get_cpu_var(dirty_throttle_leaks);
1396 if (*p > 0 && current->nr_dirtied < ratelimit) {
1397 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1398 *p -= nr_pages_dirtied;
1399 current->nr_dirtied += nr_pages_dirtied;
1403 if (unlikely(current->nr_dirtied >= ratelimit))
1404 balance_dirty_pages(mapping, current->nr_dirtied);
1406 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
1408 void throttle_vm_writeout(gfp_t gfp_mask)
1410 unsigned long background_thresh;
1411 unsigned long dirty_thresh;
1414 global_dirty_limits(&background_thresh, &dirty_thresh);
1417 * Boost the allowable dirty threshold a bit for page
1418 * allocators so they don't get DoS'ed by heavy writers
1420 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1422 if (global_page_state(NR_UNSTABLE_NFS) +
1423 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1425 congestion_wait(BLK_RW_ASYNC, HZ/10);
1428 * The caller might hold locks which can prevent IO completion
1429 * or progress in the filesystem. So we cannot just sit here
1430 * waiting for IO to complete.
1432 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1438 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1440 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1441 void __user *buffer, size_t *length, loff_t *ppos)
1443 proc_dointvec(table, write, buffer, length, ppos);
1444 bdi_arm_supers_timer();
1449 void laptop_mode_timer_fn(unsigned long data)
1451 struct request_queue *q = (struct request_queue *)data;
1452 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1453 global_page_state(NR_UNSTABLE_NFS);
1456 * We want to write everything out, not just down to the dirty
1459 if (bdi_has_dirty_io(&q->backing_dev_info))
1460 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1461 WB_REASON_LAPTOP_TIMER);
1465 * We've spun up the disk and we're in laptop mode: schedule writeback
1466 * of all dirty data a few seconds from now. If the flush is already scheduled
1467 * then push it back - the user is still using the disk.
1469 void laptop_io_completion(struct backing_dev_info *info)
1471 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1475 * We're in laptop mode and we've just synced. The sync's writes will have
1476 * caused another writeback to be scheduled by laptop_io_completion.
1477 * Nothing needs to be written back anymore, so we unschedule the writeback.
1479 void laptop_sync_completion(void)
1481 struct backing_dev_info *bdi;
1485 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1486 del_timer(&bdi->laptop_mode_wb_timer);
1493 * If ratelimit_pages is too high then we can get into dirty-data overload
1494 * if a large number of processes all perform writes at the same time.
1495 * If it is too low then SMP machines will call the (expensive)
1496 * get_writeback_state too often.
1498 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1499 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1503 void writeback_set_ratelimit(void)
1505 unsigned long background_thresh;
1506 unsigned long dirty_thresh;
1507 global_dirty_limits(&background_thresh, &dirty_thresh);
1508 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1509 if (ratelimit_pages < 16)
1510 ratelimit_pages = 16;
1513 static int __cpuinit
1514 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1516 writeback_set_ratelimit();
1520 static struct notifier_block __cpuinitdata ratelimit_nb = {
1521 .notifier_call = ratelimit_handler,
1526 * Called early on to tune the page writeback dirty limits.
1528 * We used to scale dirty pages according to how total memory
1529 * related to pages that could be allocated for buffers (by
1530 * comparing nr_free_buffer_pages() to vm_total_pages.
1532 * However, that was when we used "dirty_ratio" to scale with
1533 * all memory, and we don't do that any more. "dirty_ratio"
1534 * is now applied to total non-HIGHPAGE memory (by subtracting
1535 * totalhigh_pages from vm_total_pages), and as such we can't
1536 * get into the old insane situation any more where we had
1537 * large amounts of dirty pages compared to a small amount of
1538 * non-HIGHMEM memory.
1540 * But we might still want to scale the dirty_ratio by how
1541 * much memory the box has..
1543 void __init page_writeback_init(void)
1547 writeback_set_ratelimit();
1548 register_cpu_notifier(&ratelimit_nb);
1550 shift = calc_period_shift();
1551 prop_descriptor_init(&vm_completions, shift);
1555 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1556 * @mapping: address space structure to write
1557 * @start: starting page index
1558 * @end: ending page index (inclusive)
1560 * This function scans the page range from @start to @end (inclusive) and tags
1561 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1562 * that write_cache_pages (or whoever calls this function) will then use
1563 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1564 * used to avoid livelocking of writeback by a process steadily creating new
1565 * dirty pages in the file (thus it is important for this function to be quick
1566 * so that it can tag pages faster than a dirtying process can create them).
1569 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1571 void tag_pages_for_writeback(struct address_space *mapping,
1572 pgoff_t start, pgoff_t end)
1574 #define WRITEBACK_TAG_BATCH 4096
1575 unsigned long tagged;
1578 spin_lock_irq(&mapping->tree_lock);
1579 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1580 &start, end, WRITEBACK_TAG_BATCH,
1581 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1582 spin_unlock_irq(&mapping->tree_lock);
1583 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1585 /* We check 'start' to handle wrapping when end == ~0UL */
1586 } while (tagged >= WRITEBACK_TAG_BATCH && start);
1588 EXPORT_SYMBOL(tag_pages_for_writeback);
1591 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1592 * @mapping: address space structure to write
1593 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1594 * @writepage: function called for each page
1595 * @data: data passed to writepage function
1597 * If a page is already under I/O, write_cache_pages() skips it, even
1598 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1599 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1600 * and msync() need to guarantee that all the data which was dirty at the time
1601 * the call was made get new I/O started against them. If wbc->sync_mode is
1602 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1603 * existing IO to complete.
1605 * To avoid livelocks (when other process dirties new pages), we first tag
1606 * pages which should be written back with TOWRITE tag and only then start
1607 * writing them. For data-integrity sync we have to be careful so that we do
1608 * not miss some pages (e.g., because some other process has cleared TOWRITE
1609 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1610 * by the process clearing the DIRTY tag (and submitting the page for IO).
1612 int write_cache_pages(struct address_space *mapping,
1613 struct writeback_control *wbc, writepage_t writepage,
1618 struct pagevec pvec;
1620 pgoff_t uninitialized_var(writeback_index);
1622 pgoff_t end; /* Inclusive */
1625 int range_whole = 0;
1628 pagevec_init(&pvec, 0);
1629 if (wbc->range_cyclic) {
1630 writeback_index = mapping->writeback_index; /* prev offset */
1631 index = writeback_index;
1638 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1639 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1640 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1642 cycled = 1; /* ignore range_cyclic tests */
1644 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1645 tag = PAGECACHE_TAG_TOWRITE;
1647 tag = PAGECACHE_TAG_DIRTY;
1649 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1650 tag_pages_for_writeback(mapping, index, end);
1652 while (!done && (index <= end)) {
1655 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1656 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1660 for (i = 0; i < nr_pages; i++) {
1661 struct page *page = pvec.pages[i];
1664 * At this point, the page may be truncated or
1665 * invalidated (changing page->mapping to NULL), or
1666 * even swizzled back from swapper_space to tmpfs file
1667 * mapping. However, page->index will not change
1668 * because we have a reference on the page.
1670 if (page->index > end) {
1672 * can't be range_cyclic (1st pass) because
1673 * end == -1 in that case.
1679 done_index = page->index;
1684 * Page truncated or invalidated. We can freely skip it
1685 * then, even for data integrity operations: the page
1686 * has disappeared concurrently, so there could be no
1687 * real expectation of this data interity operation
1688 * even if there is now a new, dirty page at the same
1689 * pagecache address.
1691 if (unlikely(page->mapping != mapping)) {
1697 if (!PageDirty(page)) {
1698 /* someone wrote it for us */
1699 goto continue_unlock;
1702 if (PageWriteback(page)) {
1703 if (wbc->sync_mode != WB_SYNC_NONE)
1704 wait_on_page_writeback(page);
1706 goto continue_unlock;
1709 BUG_ON(PageWriteback(page));
1710 if (!clear_page_dirty_for_io(page))
1711 goto continue_unlock;
1713 trace_wbc_writepage(wbc, mapping->backing_dev_info);
1714 ret = (*writepage)(page, wbc, data);
1715 if (unlikely(ret)) {
1716 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1721 * done_index is set past this page,
1722 * so media errors will not choke
1723 * background writeout for the entire
1724 * file. This has consequences for
1725 * range_cyclic semantics (ie. it may
1726 * not be suitable for data integrity
1729 done_index = page->index + 1;
1736 * We stop writing back only if we are not doing
1737 * integrity sync. In case of integrity sync we have to
1738 * keep going until we have written all the pages
1739 * we tagged for writeback prior to entering this loop.
1741 if (--wbc->nr_to_write <= 0 &&
1742 wbc->sync_mode == WB_SYNC_NONE) {
1747 pagevec_release(&pvec);
1750 if (!cycled && !done) {
1753 * We hit the last page and there is more work to be done: wrap
1754 * back to the start of the file
1758 end = writeback_index - 1;
1761 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1762 mapping->writeback_index = done_index;
1766 EXPORT_SYMBOL(write_cache_pages);
1769 * Function used by generic_writepages to call the real writepage
1770 * function and set the mapping flags on error
1772 static int __writepage(struct page *page, struct writeback_control *wbc,
1775 struct address_space *mapping = data;
1776 int ret = mapping->a_ops->writepage(page, wbc);
1777 mapping_set_error(mapping, ret);
1782 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1783 * @mapping: address space structure to write
1784 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1786 * This is a library function, which implements the writepages()
1787 * address_space_operation.
1789 int generic_writepages(struct address_space *mapping,
1790 struct writeback_control *wbc)
1792 struct blk_plug plug;
1795 /* deal with chardevs and other special file */
1796 if (!mapping->a_ops->writepage)
1799 blk_start_plug(&plug);
1800 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1801 blk_finish_plug(&plug);
1805 EXPORT_SYMBOL(generic_writepages);
1807 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1811 if (wbc->nr_to_write <= 0)
1813 if (mapping->a_ops->writepages)
1814 ret = mapping->a_ops->writepages(mapping, wbc);
1816 ret = generic_writepages(mapping, wbc);
1821 * write_one_page - write out a single page and optionally wait on I/O
1822 * @page: the page to write
1823 * @wait: if true, wait on writeout
1825 * The page must be locked by the caller and will be unlocked upon return.
1827 * write_one_page() returns a negative error code if I/O failed.
1829 int write_one_page(struct page *page, int wait)
1831 struct address_space *mapping = page->mapping;
1833 struct writeback_control wbc = {
1834 .sync_mode = WB_SYNC_ALL,
1838 BUG_ON(!PageLocked(page));
1841 wait_on_page_writeback(page);
1843 if (clear_page_dirty_for_io(page)) {
1844 page_cache_get(page);
1845 ret = mapping->a_ops->writepage(page, &wbc);
1846 if (ret == 0 && wait) {
1847 wait_on_page_writeback(page);
1848 if (PageError(page))
1851 page_cache_release(page);
1857 EXPORT_SYMBOL(write_one_page);
1860 * For address_spaces which do not use buffers nor write back.
1862 int __set_page_dirty_no_writeback(struct page *page)
1864 if (!PageDirty(page))
1865 return !TestSetPageDirty(page);
1870 * Helper function for set_page_dirty family.
1871 * NOTE: This relies on being atomic wrt interrupts.
1873 void account_page_dirtied(struct page *page, struct address_space *mapping)
1875 if (mapping_cap_account_dirty(mapping)) {
1876 __inc_zone_page_state(page, NR_FILE_DIRTY);
1877 __inc_zone_page_state(page, NR_DIRTIED);
1878 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1879 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1880 task_io_account_write(PAGE_CACHE_SIZE);
1881 current->nr_dirtied++;
1882 this_cpu_inc(bdp_ratelimits);
1885 EXPORT_SYMBOL(account_page_dirtied);
1888 * Helper function for set_page_writeback family.
1889 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1892 void account_page_writeback(struct page *page)
1894 inc_zone_page_state(page, NR_WRITEBACK);
1896 EXPORT_SYMBOL(account_page_writeback);
1899 * For address_spaces which do not use buffers. Just tag the page as dirty in
1902 * This is also used when a single buffer is being dirtied: we want to set the
1903 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1904 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1906 * Most callers have locked the page, which pins the address_space in memory.
1907 * But zap_pte_range() does not lock the page, however in that case the
1908 * mapping is pinned by the vma's ->vm_file reference.
1910 * We take care to handle the case where the page was truncated from the
1911 * mapping by re-checking page_mapping() inside tree_lock.
1913 int __set_page_dirty_nobuffers(struct page *page)
1915 if (!TestSetPageDirty(page)) {
1916 struct address_space *mapping = page_mapping(page);
1917 struct address_space *mapping2;
1922 spin_lock_irq(&mapping->tree_lock);
1923 mapping2 = page_mapping(page);
1924 if (mapping2) { /* Race with truncate? */
1925 BUG_ON(mapping2 != mapping);
1926 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1927 account_page_dirtied(page, mapping);
1928 radix_tree_tag_set(&mapping->page_tree,
1929 page_index(page), PAGECACHE_TAG_DIRTY);
1931 spin_unlock_irq(&mapping->tree_lock);
1932 if (mapping->host) {
1933 /* !PageAnon && !swapper_space */
1934 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1940 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1943 * Call this whenever redirtying a page, to de-account the dirty counters
1944 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
1945 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
1946 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
1949 void account_page_redirty(struct page *page)
1951 struct address_space *mapping = page->mapping;
1952 if (mapping && mapping_cap_account_dirty(mapping)) {
1953 current->nr_dirtied--;
1954 dec_zone_page_state(page, NR_DIRTIED);
1955 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1958 EXPORT_SYMBOL(account_page_redirty);
1961 * When a writepage implementation decides that it doesn't want to write this
1962 * page for some reason, it should redirty the locked page via
1963 * redirty_page_for_writepage() and it should then unlock the page and return 0
1965 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1967 wbc->pages_skipped++;
1968 account_page_redirty(page);
1969 return __set_page_dirty_nobuffers(page);
1971 EXPORT_SYMBOL(redirty_page_for_writepage);
1976 * For pages with a mapping this should be done under the page lock
1977 * for the benefit of asynchronous memory errors who prefer a consistent
1978 * dirty state. This rule can be broken in some special cases,
1979 * but should be better not to.
1981 * If the mapping doesn't provide a set_page_dirty a_op, then
1982 * just fall through and assume that it wants buffer_heads.
1984 int set_page_dirty(struct page *page)
1986 struct address_space *mapping = page_mapping(page);
1988 if (likely(mapping)) {
1989 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1991 * readahead/lru_deactivate_page could remain
1992 * PG_readahead/PG_reclaim due to race with end_page_writeback
1993 * About readahead, if the page is written, the flags would be
1994 * reset. So no problem.
1995 * About lru_deactivate_page, if the page is redirty, the flag
1996 * will be reset. So no problem. but if the page is used by readahead
1997 * it will confuse readahead and make it restart the size rampup
1998 * process. But it's a trivial problem.
2000 ClearPageReclaim(page);
2003 spd = __set_page_dirty_buffers;
2005 return (*spd)(page);
2007 if (!PageDirty(page)) {
2008 if (!TestSetPageDirty(page))
2013 EXPORT_SYMBOL(set_page_dirty);
2016 * set_page_dirty() is racy if the caller has no reference against
2017 * page->mapping->host, and if the page is unlocked. This is because another
2018 * CPU could truncate the page off the mapping and then free the mapping.
2020 * Usually, the page _is_ locked, or the caller is a user-space process which
2021 * holds a reference on the inode by having an open file.
2023 * In other cases, the page should be locked before running set_page_dirty().
2025 int set_page_dirty_lock(struct page *page)
2030 ret = set_page_dirty(page);
2034 EXPORT_SYMBOL(set_page_dirty_lock);
2037 * Clear a page's dirty flag, while caring for dirty memory accounting.
2038 * Returns true if the page was previously dirty.
2040 * This is for preparing to put the page under writeout. We leave the page
2041 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2042 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2043 * implementation will run either set_page_writeback() or set_page_dirty(),
2044 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2047 * This incoherency between the page's dirty flag and radix-tree tag is
2048 * unfortunate, but it only exists while the page is locked.
2050 int clear_page_dirty_for_io(struct page *page)
2052 struct address_space *mapping = page_mapping(page);
2054 BUG_ON(!PageLocked(page));
2056 if (mapping && mapping_cap_account_dirty(mapping)) {
2058 * Yes, Virginia, this is indeed insane.
2060 * We use this sequence to make sure that
2061 * (a) we account for dirty stats properly
2062 * (b) we tell the low-level filesystem to
2063 * mark the whole page dirty if it was
2064 * dirty in a pagetable. Only to then
2065 * (c) clean the page again and return 1 to
2066 * cause the writeback.
2068 * This way we avoid all nasty races with the
2069 * dirty bit in multiple places and clearing
2070 * them concurrently from different threads.
2072 * Note! Normally the "set_page_dirty(page)"
2073 * has no effect on the actual dirty bit - since
2074 * that will already usually be set. But we
2075 * need the side effects, and it can help us
2078 * We basically use the page "master dirty bit"
2079 * as a serialization point for all the different
2080 * threads doing their things.
2082 if (page_mkclean(page))
2083 set_page_dirty(page);
2085 * We carefully synchronise fault handlers against
2086 * installing a dirty pte and marking the page dirty
2087 * at this point. We do this by having them hold the
2088 * page lock at some point after installing their
2089 * pte, but before marking the page dirty.
2090 * Pages are always locked coming in here, so we get
2091 * the desired exclusion. See mm/memory.c:do_wp_page()
2092 * for more comments.
2094 if (TestClearPageDirty(page)) {
2095 dec_zone_page_state(page, NR_FILE_DIRTY);
2096 dec_bdi_stat(mapping->backing_dev_info,
2102 return TestClearPageDirty(page);
2104 EXPORT_SYMBOL(clear_page_dirty_for_io);
2106 int test_clear_page_writeback(struct page *page)
2108 struct address_space *mapping = page_mapping(page);
2112 struct backing_dev_info *bdi = mapping->backing_dev_info;
2113 unsigned long flags;
2115 spin_lock_irqsave(&mapping->tree_lock, flags);
2116 ret = TestClearPageWriteback(page);
2118 radix_tree_tag_clear(&mapping->page_tree,
2120 PAGECACHE_TAG_WRITEBACK);
2121 if (bdi_cap_account_writeback(bdi)) {
2122 __dec_bdi_stat(bdi, BDI_WRITEBACK);
2123 __bdi_writeout_inc(bdi);
2126 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2128 ret = TestClearPageWriteback(page);
2131 dec_zone_page_state(page, NR_WRITEBACK);
2132 inc_zone_page_state(page, NR_WRITTEN);
2137 int test_set_page_writeback(struct page *page)
2139 struct address_space *mapping = page_mapping(page);
2143 struct backing_dev_info *bdi = mapping->backing_dev_info;
2144 unsigned long flags;
2146 spin_lock_irqsave(&mapping->tree_lock, flags);
2147 ret = TestSetPageWriteback(page);
2149 radix_tree_tag_set(&mapping->page_tree,
2151 PAGECACHE_TAG_WRITEBACK);
2152 if (bdi_cap_account_writeback(bdi))
2153 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2155 if (!PageDirty(page))
2156 radix_tree_tag_clear(&mapping->page_tree,
2158 PAGECACHE_TAG_DIRTY);
2159 radix_tree_tag_clear(&mapping->page_tree,
2161 PAGECACHE_TAG_TOWRITE);
2162 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2164 ret = TestSetPageWriteback(page);
2167 account_page_writeback(page);
2171 EXPORT_SYMBOL(test_set_page_writeback);
2174 * Return true if any of the pages in the mapping are marked with the
2177 int mapping_tagged(struct address_space *mapping, int tag)
2179 return radix_tree_tagged(&mapping->page_tree, tag);
2181 EXPORT_SYMBOL(mapping_tagged);