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 * 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 */
99 * The longest time for which data is allowed to remain dirty
101 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
104 * Flag that makes the machine dump writes/reads and block dirtyings.
109 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
110 * a full sync is triggered after this time elapses without any disk activity.
114 EXPORT_SYMBOL(laptop_mode);
116 /* End of sysctl-exported parameters */
118 unsigned long global_dirty_limit;
121 * Scale the writeback cache size proportional to the relative writeout speeds.
123 * We do this by keeping a floating proportion between BDIs, based on page
124 * writeback completions [end_page_writeback()]. Those devices that write out
125 * pages fastest will get the larger share, while the slower will get a smaller
128 * We use page writeout completions because we are interested in getting rid of
129 * dirty pages. Having them written out is the primary goal.
131 * We introduce a concept of time, a period over which we measure these events,
132 * because demand can/will vary over time. The length of this period itself is
133 * measured in page writeback completions.
136 static struct prop_descriptor vm_completions;
139 * couple the period to the dirty_ratio:
141 * period/2 ~ roundup_pow_of_two(dirty limit)
143 static int calc_period_shift(void)
145 unsigned long dirty_total;
148 dirty_total = vm_dirty_bytes / PAGE_SIZE;
150 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
152 return 2 + ilog2(dirty_total - 1);
156 * update the period when the dirty threshold changes.
158 static void update_completion_period(void)
160 int shift = calc_period_shift();
161 prop_change_shift(&vm_completions, shift);
163 writeback_set_ratelimit();
166 int dirty_background_ratio_handler(struct ctl_table *table, int write,
167 void __user *buffer, size_t *lenp,
172 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
173 if (ret == 0 && write)
174 dirty_background_bytes = 0;
178 int dirty_background_bytes_handler(struct ctl_table *table, int write,
179 void __user *buffer, size_t *lenp,
184 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
185 if (ret == 0 && write)
186 dirty_background_ratio = 0;
190 int dirty_ratio_handler(struct ctl_table *table, int write,
191 void __user *buffer, size_t *lenp,
194 int old_ratio = vm_dirty_ratio;
197 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
198 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
199 update_completion_period();
206 int dirty_bytes_handler(struct ctl_table *table, int write,
207 void __user *buffer, size_t *lenp,
210 unsigned long old_bytes = vm_dirty_bytes;
213 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
214 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
215 update_completion_period();
222 * Increment the BDI's writeout completion count and the global writeout
223 * completion count. Called from test_clear_page_writeback().
225 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
227 __inc_bdi_stat(bdi, BDI_WRITTEN);
228 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
232 void bdi_writeout_inc(struct backing_dev_info *bdi)
236 local_irq_save(flags);
237 __bdi_writeout_inc(bdi);
238 local_irq_restore(flags);
240 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
243 * Obtain an accurate fraction of the BDI's portion.
245 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
246 long *numerator, long *denominator)
248 prop_fraction_percpu(&vm_completions, &bdi->completions,
249 numerator, denominator);
253 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
254 * registered backing devices, which, for obvious reasons, can not
257 static unsigned int bdi_min_ratio;
259 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
263 spin_lock_bh(&bdi_lock);
264 if (min_ratio > bdi->max_ratio) {
267 min_ratio -= bdi->min_ratio;
268 if (bdi_min_ratio + min_ratio < 100) {
269 bdi_min_ratio += min_ratio;
270 bdi->min_ratio += min_ratio;
275 spin_unlock_bh(&bdi_lock);
280 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
287 spin_lock_bh(&bdi_lock);
288 if (bdi->min_ratio > max_ratio) {
291 bdi->max_ratio = max_ratio;
292 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
294 spin_unlock_bh(&bdi_lock);
298 EXPORT_SYMBOL(bdi_set_max_ratio);
301 * Work out the current dirty-memory clamping and background writeout
304 * The main aim here is to lower them aggressively if there is a lot of mapped
305 * memory around. To avoid stressing page reclaim with lots of unreclaimable
306 * pages. It is better to clamp down on writers than to start swapping, and
307 * performing lots of scanning.
309 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
311 * We don't permit the clamping level to fall below 5% - that is getting rather
314 * We make sure that the background writeout level is below the adjusted
318 static unsigned long highmem_dirtyable_memory(unsigned long total)
320 #ifdef CONFIG_HIGHMEM
324 for_each_node_state(node, N_HIGH_MEMORY) {
326 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
328 x += zone_page_state(z, NR_FREE_PAGES) +
329 zone_reclaimable_pages(z);
332 * Make sure that the number of highmem pages is never larger
333 * than the number of the total dirtyable memory. This can only
334 * occur in very strange VM situations but we want to make sure
335 * that this does not occur.
337 return min(x, total);
344 * determine_dirtyable_memory - amount of memory that may be used
346 * Returns the numebr of pages that can currently be freed and used
347 * by the kernel for direct mappings.
349 unsigned long determine_dirtyable_memory(void)
353 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
355 if (!vm_highmem_is_dirtyable)
356 x -= highmem_dirtyable_memory(x);
358 return x + 1; /* Ensure that we never return 0 */
361 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
362 unsigned long bg_thresh)
364 return (thresh + bg_thresh) / 2;
367 static unsigned long hard_dirty_limit(unsigned long thresh)
369 return max(thresh, global_dirty_limit);
373 * global_dirty_limits - background-writeback and dirty-throttling thresholds
375 * Calculate the dirty thresholds based on sysctl parameters
376 * - vm.dirty_background_ratio or vm.dirty_background_bytes
377 * - vm.dirty_ratio or vm.dirty_bytes
378 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
381 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
383 unsigned long background;
385 unsigned long uninitialized_var(available_memory);
386 struct task_struct *tsk;
388 if (!vm_dirty_bytes || !dirty_background_bytes)
389 available_memory = determine_dirtyable_memory();
392 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
394 dirty = (vm_dirty_ratio * available_memory) / 100;
396 if (dirty_background_bytes)
397 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
399 background = (dirty_background_ratio * available_memory) / 100;
401 if (background >= dirty)
402 background = dirty / 2;
404 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
405 background += background / 4;
408 *pbackground = background;
410 trace_global_dirty_state(background, dirty);
414 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
415 * @bdi: the backing_dev_info to query
416 * @dirty: global dirty limit in pages
418 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
419 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
421 * Note that balance_dirty_pages() will only seriously take it as a hard limit
422 * when sleeping max_pause per page is not enough to keep the dirty pages under
423 * control. For example, when the device is completely stalled due to some error
424 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
425 * In the other normal situations, it acts more gently by throttling the tasks
426 * more (rather than completely block them) when the bdi dirty pages go high.
428 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
429 * - starving fast devices
430 * - piling up dirty pages (that will take long time to sync) on slow devices
432 * The bdi's share of dirty limit will be adapting to its throughput and
433 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
435 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
438 long numerator, denominator;
441 * Calculate this BDI's share of the dirty ratio.
443 bdi_writeout_fraction(bdi, &numerator, &denominator);
445 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
446 bdi_dirty *= numerator;
447 do_div(bdi_dirty, denominator);
449 bdi_dirty += (dirty * bdi->min_ratio) / 100;
450 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
451 bdi_dirty = dirty * bdi->max_ratio / 100;
457 * Dirty position control.
459 * (o) global/bdi setpoints
461 * We want the dirty pages be balanced around the global/bdi setpoints.
462 * When the number of dirty pages is higher/lower than the setpoint, the
463 * dirty position control ratio (and hence task dirty ratelimit) will be
464 * decreased/increased to bring the dirty pages back to the setpoint.
466 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
468 * if (dirty < setpoint) scale up pos_ratio
469 * if (dirty > setpoint) scale down pos_ratio
471 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
472 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
474 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
476 * (o) global control line
480 * | |<===== global dirty control scope ======>|
488 * 1.0 ................................*
494 * 0 +------------.------------------.----------------------*------------->
495 * freerun^ setpoint^ limit^ dirty pages
497 * (o) bdi control line
505 * | * |<=========== span ============>|
506 * 1.0 .......................*
518 * 1/4 ...............................................* * * * * * * * * * * *
522 * 0 +----------------------.-------------------------------.------------->
523 * bdi_setpoint^ x_intercept^
525 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
526 * be smoothly throttled down to normal if it starts high in situations like
527 * - start writing to a slow SD card and a fast disk at the same time. The SD
528 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
529 * - the bdi dirty thresh drops quickly due to change of JBOD workload
531 static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
532 unsigned long thresh,
533 unsigned long bg_thresh,
535 unsigned long bdi_thresh,
536 unsigned long bdi_dirty)
538 unsigned long write_bw = bdi->avg_write_bandwidth;
539 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
540 unsigned long limit = hard_dirty_limit(thresh);
541 unsigned long x_intercept;
542 unsigned long setpoint; /* dirty pages' target balance point */
543 unsigned long bdi_setpoint;
545 long long pos_ratio; /* for scaling up/down the rate limit */
548 if (unlikely(dirty >= limit))
555 * f(dirty) := 1.0 + (----------------)
558 * it's a 3rd order polynomial that subjects to
560 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
561 * (2) f(setpoint) = 1.0 => the balance point
562 * (3) f(limit) = 0 => the hard limit
563 * (4) df/dx <= 0 => negative feedback control
564 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
565 * => fast response on large errors; small oscillation near setpoint
567 setpoint = (freerun + limit) / 2;
568 x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
569 limit - setpoint + 1);
571 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
572 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
573 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
576 * We have computed basic pos_ratio above based on global situation. If
577 * the bdi is over/under its share of dirty pages, we want to scale
578 * pos_ratio further down/up. That is done by the following mechanism.
584 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
586 * x_intercept - bdi_dirty
587 * := --------------------------
588 * x_intercept - bdi_setpoint
590 * The main bdi control line is a linear function that subjects to
592 * (1) f(bdi_setpoint) = 1.0
593 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
594 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
596 * For single bdi case, the dirty pages are observed to fluctuate
597 * regularly within range
598 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
599 * for various filesystems, where (2) can yield in a reasonable 12.5%
600 * fluctuation range for pos_ratio.
602 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
603 * own size, so move the slope over accordingly and choose a slope that
604 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
606 if (unlikely(bdi_thresh > thresh))
609 * It's very possible that bdi_thresh is close to 0 not because the
610 * device is slow, but that it has remained inactive for long time.
611 * Honour such devices a reasonable good (hopefully IO efficient)
612 * threshold, so that the occasional writes won't be blocked and active
613 * writes can rampup the threshold quickly.
615 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
617 * scale global setpoint to bdi's:
618 * bdi_setpoint = setpoint * bdi_thresh / thresh
620 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
621 bdi_setpoint = setpoint * (u64)x >> 16;
623 * Use span=(8*write_bw) in single bdi case as indicated by
624 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
626 * bdi_thresh thresh - bdi_thresh
627 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
630 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
631 x_intercept = bdi_setpoint + span;
633 if (bdi_dirty < x_intercept - span / 4) {
634 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
635 x_intercept - bdi_setpoint + 1);
640 * bdi reserve area, safeguard against dirty pool underrun and disk idle
641 * It may push the desired control point of global dirty pages higher
644 x_intercept = bdi_thresh / 2;
645 if (bdi_dirty < x_intercept) {
646 if (bdi_dirty > x_intercept / 8)
647 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
655 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
656 unsigned long elapsed,
657 unsigned long written)
659 const unsigned long period = roundup_pow_of_two(3 * HZ);
660 unsigned long avg = bdi->avg_write_bandwidth;
661 unsigned long old = bdi->write_bandwidth;
665 * bw = written * HZ / elapsed
667 * bw * elapsed + write_bandwidth * (period - elapsed)
668 * write_bandwidth = ---------------------------------------------------
671 bw = written - bdi->written_stamp;
673 if (unlikely(elapsed > period)) {
678 bw += (u64)bdi->write_bandwidth * (period - elapsed);
679 bw >>= ilog2(period);
682 * one more level of smoothing, for filtering out sudden spikes
684 if (avg > old && old >= (unsigned long)bw)
685 avg -= (avg - old) >> 3;
687 if (avg < old && old <= (unsigned long)bw)
688 avg += (old - avg) >> 3;
691 bdi->write_bandwidth = bw;
692 bdi->avg_write_bandwidth = avg;
696 * The global dirtyable memory and dirty threshold could be suddenly knocked
697 * down by a large amount (eg. on the startup of KVM in a swapless system).
698 * This may throw the system into deep dirty exceeded state and throttle
699 * heavy/light dirtiers alike. To retain good responsiveness, maintain
700 * global_dirty_limit for tracking slowly down to the knocked down dirty
703 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
705 unsigned long limit = global_dirty_limit;
708 * Follow up in one step.
710 if (limit < thresh) {
716 * Follow down slowly. Use the higher one as the target, because thresh
717 * may drop below dirty. This is exactly the reason to introduce
718 * global_dirty_limit which is guaranteed to lie above the dirty pages.
720 thresh = max(thresh, dirty);
721 if (limit > thresh) {
722 limit -= (limit - thresh) >> 5;
727 global_dirty_limit = limit;
730 static void global_update_bandwidth(unsigned long thresh,
734 static DEFINE_SPINLOCK(dirty_lock);
735 static unsigned long update_time;
738 * check locklessly first to optimize away locking for the most time
740 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
743 spin_lock(&dirty_lock);
744 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
745 update_dirty_limit(thresh, dirty);
748 spin_unlock(&dirty_lock);
752 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
754 * Normal bdi tasks will be curbed at or below it in long term.
755 * Obviously it should be around (write_bw / N) when there are N dd tasks.
757 static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
758 unsigned long thresh,
759 unsigned long bg_thresh,
761 unsigned long bdi_thresh,
762 unsigned long bdi_dirty,
763 unsigned long dirtied,
764 unsigned long elapsed)
766 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
767 unsigned long limit = hard_dirty_limit(thresh);
768 unsigned long setpoint = (freerun + limit) / 2;
769 unsigned long write_bw = bdi->avg_write_bandwidth;
770 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
771 unsigned long dirty_rate;
772 unsigned long task_ratelimit;
773 unsigned long balanced_dirty_ratelimit;
774 unsigned long pos_ratio;
779 * The dirty rate will match the writeout rate in long term, except
780 * when dirty pages are truncated by userspace or re-dirtied by FS.
782 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
784 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
785 bdi_thresh, bdi_dirty);
787 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
789 task_ratelimit = (u64)dirty_ratelimit *
790 pos_ratio >> RATELIMIT_CALC_SHIFT;
791 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
794 * A linear estimation of the "balanced" throttle rate. The theory is,
795 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
796 * dirty_rate will be measured to be (N * task_ratelimit). So the below
797 * formula will yield the balanced rate limit (write_bw / N).
799 * Note that the expanded form is not a pure rate feedback:
800 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
801 * but also takes pos_ratio into account:
802 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
804 * (1) is not realistic because pos_ratio also takes part in balancing
805 * the dirty rate. Consider the state
806 * pos_ratio = 0.5 (3)
807 * rate = 2 * (write_bw / N) (4)
808 * If (1) is used, it will stuck in that state! Because each dd will
810 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
812 * dirty_rate = N * task_ratelimit = write_bw (6)
813 * put (6) into (1) we get
814 * rate_(i+1) = rate_(i) (7)
816 * So we end up using (2) to always keep
817 * rate_(i+1) ~= (write_bw / N) (8)
818 * regardless of the value of pos_ratio. As long as (8) is satisfied,
819 * pos_ratio is able to drive itself to 1.0, which is not only where
820 * the dirty count meet the setpoint, but also where the slope of
821 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
823 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
827 * We could safely do this and return immediately:
829 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
831 * However to get a more stable dirty_ratelimit, the below elaborated
832 * code makes use of task_ratelimit to filter out sigular points and
833 * limit the step size.
835 * The below code essentially only uses the relative value of
837 * task_ratelimit - dirty_ratelimit
838 * = (pos_ratio - 1) * dirty_ratelimit
840 * which reflects the direction and size of dirty position error.
844 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
845 * task_ratelimit is on the same side of dirty_ratelimit, too.
847 * - dirty_ratelimit > balanced_dirty_ratelimit
848 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
849 * lowering dirty_ratelimit will help meet both the position and rate
850 * control targets. Otherwise, don't update dirty_ratelimit if it will
851 * only help meet the rate target. After all, what the users ultimately
852 * feel and care are stable dirty rate and small position error.
854 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
855 * and filter out the sigular points of balanced_dirty_ratelimit. Which
856 * keeps jumping around randomly and can even leap far away at times
857 * due to the small 200ms estimation period of dirty_rate (we want to
858 * keep that period small to reduce time lags).
861 if (dirty < setpoint) {
862 x = min(bdi->balanced_dirty_ratelimit,
863 min(balanced_dirty_ratelimit, task_ratelimit));
864 if (dirty_ratelimit < x)
865 step = x - dirty_ratelimit;
867 x = max(bdi->balanced_dirty_ratelimit,
868 max(balanced_dirty_ratelimit, task_ratelimit));
869 if (dirty_ratelimit > x)
870 step = dirty_ratelimit - x;
874 * Don't pursue 100% rate matching. It's impossible since the balanced
875 * rate itself is constantly fluctuating. So decrease the track speed
876 * when it gets close to the target. Helps eliminate pointless tremors.
878 step >>= dirty_ratelimit / (2 * step + 1);
880 * Limit the tracking speed to avoid overshooting.
882 step = (step + 7) / 8;
884 if (dirty_ratelimit < balanced_dirty_ratelimit)
885 dirty_ratelimit += step;
887 dirty_ratelimit -= step;
889 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
890 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
892 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
895 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
896 unsigned long thresh,
897 unsigned long bg_thresh,
899 unsigned long bdi_thresh,
900 unsigned long bdi_dirty,
901 unsigned long start_time)
903 unsigned long now = jiffies;
904 unsigned long elapsed = now - bdi->bw_time_stamp;
905 unsigned long dirtied;
906 unsigned long written;
909 * rate-limit, only update once every 200ms.
911 if (elapsed < BANDWIDTH_INTERVAL)
914 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
915 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
918 * Skip quiet periods when disk bandwidth is under-utilized.
919 * (at least 1s idle time between two flusher runs)
921 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
925 global_update_bandwidth(thresh, dirty, now);
926 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
927 bdi_thresh, bdi_dirty,
930 bdi_update_write_bandwidth(bdi, elapsed, written);
933 bdi->dirtied_stamp = dirtied;
934 bdi->written_stamp = written;
935 bdi->bw_time_stamp = now;
938 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
939 unsigned long thresh,
940 unsigned long bg_thresh,
942 unsigned long bdi_thresh,
943 unsigned long bdi_dirty,
944 unsigned long start_time)
946 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
948 spin_lock(&bdi->wb.list_lock);
949 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
950 bdi_thresh, bdi_dirty, start_time);
951 spin_unlock(&bdi->wb.list_lock);
955 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
956 * will look to see if it needs to start dirty throttling.
958 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
959 * global_page_state() too often. So scale it near-sqrt to the safety margin
960 * (the number of pages we may dirty without exceeding the dirty limits).
962 static unsigned long dirty_poll_interval(unsigned long dirty,
963 unsigned long thresh)
966 return 1UL << (ilog2(thresh - dirty) >> 1);
971 static long bdi_max_pause(struct backing_dev_info *bdi,
972 unsigned long bdi_dirty)
974 long bw = bdi->avg_write_bandwidth;
978 * Limit pause time for small memory systems. If sleeping for too long
979 * time, a small pool of dirty/writeback pages may go empty and disk go
982 * 8 serves as the safety ratio.
984 t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
987 return min_t(long, t, MAX_PAUSE);
990 static long bdi_min_pause(struct backing_dev_info *bdi,
992 unsigned long task_ratelimit,
993 unsigned long dirty_ratelimit,
994 int *nr_dirtied_pause)
996 long hi = ilog2(bdi->avg_write_bandwidth);
997 long lo = ilog2(bdi->dirty_ratelimit);
998 long t; /* target pause */
999 long pause; /* estimated next pause */
1000 int pages; /* target nr_dirtied_pause */
1002 /* target for 10ms pause on 1-dd case */
1003 t = max(1, HZ / 100);
1006 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1009 * (N * 10ms) on 2^N concurrent tasks.
1012 t += (hi - lo) * (10 * HZ) / 1024;
1015 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1016 * on the much more stable dirty_ratelimit. However the next pause time
1017 * will be computed based on task_ratelimit and the two rate limits may
1018 * depart considerably at some time. Especially if task_ratelimit goes
1019 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1020 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1021 * result task_ratelimit won't be executed faithfully, which could
1022 * eventually bring down dirty_ratelimit.
1024 * We apply two rules to fix it up:
1025 * 1) try to estimate the next pause time and if necessary, use a lower
1026 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1027 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1028 * 2) limit the target pause time to max_pause/2, so that the normal
1029 * small fluctuations of task_ratelimit won't trigger rule (1) and
1030 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1032 t = min(t, 1 + max_pause / 2);
1033 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1036 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1037 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1038 * When the 16 consecutive reads are often interrupted by some dirty
1039 * throttling pause during the async writes, cfq will go into idles
1040 * (deadline is fine). So push nr_dirtied_pause as high as possible
1041 * until reaches DIRTY_POLL_THRESH=32 pages.
1043 if (pages < DIRTY_POLL_THRESH) {
1045 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1046 if (pages > DIRTY_POLL_THRESH) {
1047 pages = DIRTY_POLL_THRESH;
1048 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1052 pause = HZ * pages / (task_ratelimit + 1);
1053 if (pause > max_pause) {
1055 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1058 *nr_dirtied_pause = pages;
1060 * The minimal pause time will normally be half the target pause time.
1062 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1066 * balance_dirty_pages() must be called by processes which are generating dirty
1067 * data. It looks at the number of dirty pages in the machine and will force
1068 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1069 * If we're over `background_thresh' then the writeback threads are woken to
1070 * perform some writeout.
1072 static void balance_dirty_pages(struct address_space *mapping,
1073 unsigned long pages_dirtied)
1075 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1076 unsigned long bdi_reclaimable;
1077 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1078 unsigned long bdi_dirty;
1079 unsigned long freerun;
1080 unsigned long background_thresh;
1081 unsigned long dirty_thresh;
1082 unsigned long bdi_thresh;
1087 int nr_dirtied_pause;
1088 bool dirty_exceeded = false;
1089 unsigned long task_ratelimit;
1090 unsigned long dirty_ratelimit;
1091 unsigned long pos_ratio;
1092 struct backing_dev_info *bdi = mapping->backing_dev_info;
1093 unsigned long start_time = jiffies;
1096 unsigned long now = jiffies;
1099 * Unstable writes are a feature of certain networked
1100 * filesystems (i.e. NFS) in which data may have been
1101 * written to the server's write cache, but has not yet
1102 * been flushed to permanent storage.
1104 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1105 global_page_state(NR_UNSTABLE_NFS);
1106 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1108 global_dirty_limits(&background_thresh, &dirty_thresh);
1111 * Throttle it only when the background writeback cannot
1112 * catch-up. This avoids (excessively) small writeouts
1113 * when the bdi limits are ramping up.
1115 freerun = dirty_freerun_ceiling(dirty_thresh,
1117 if (nr_dirty <= freerun) {
1118 current->dirty_paused_when = now;
1119 current->nr_dirtied = 0;
1120 current->nr_dirtied_pause =
1121 dirty_poll_interval(nr_dirty, dirty_thresh);
1125 if (unlikely(!writeback_in_progress(bdi)))
1126 bdi_start_background_writeback(bdi);
1129 * bdi_thresh is not treated as some limiting factor as
1130 * dirty_thresh, due to reasons
1131 * - in JBOD setup, bdi_thresh can fluctuate a lot
1132 * - in a system with HDD and USB key, the USB key may somehow
1133 * go into state (bdi_dirty >> bdi_thresh) either because
1134 * bdi_dirty starts high, or because bdi_thresh drops low.
1135 * In this case we don't want to hard throttle the USB key
1136 * dirtiers for 100 seconds until bdi_dirty drops under
1137 * bdi_thresh. Instead the auxiliary bdi control line in
1138 * bdi_position_ratio() will let the dirtier task progress
1139 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1141 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1144 * In order to avoid the stacked BDI deadlock we need
1145 * to ensure we accurately count the 'dirty' pages when
1146 * the threshold is low.
1148 * Otherwise it would be possible to get thresh+n pages
1149 * reported dirty, even though there are thresh-m pages
1150 * actually dirty; with m+n sitting in the percpu
1153 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1154 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1155 bdi_dirty = bdi_reclaimable +
1156 bdi_stat_sum(bdi, BDI_WRITEBACK);
1158 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1159 bdi_dirty = bdi_reclaimable +
1160 bdi_stat(bdi, BDI_WRITEBACK);
1163 dirty_exceeded = (bdi_dirty > bdi_thresh) ||
1164 (nr_dirty > dirty_thresh);
1165 if (dirty_exceeded && !bdi->dirty_exceeded)
1166 bdi->dirty_exceeded = 1;
1168 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1169 nr_dirty, bdi_thresh, bdi_dirty,
1172 dirty_ratelimit = bdi->dirty_ratelimit;
1173 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1174 background_thresh, nr_dirty,
1175 bdi_thresh, bdi_dirty);
1176 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1177 RATELIMIT_CALC_SHIFT;
1178 max_pause = bdi_max_pause(bdi, bdi_dirty);
1179 min_pause = bdi_min_pause(bdi, max_pause,
1180 task_ratelimit, dirty_ratelimit,
1183 if (unlikely(task_ratelimit == 0)) {
1188 period = HZ * pages_dirtied / task_ratelimit;
1190 if (current->dirty_paused_when)
1191 pause -= now - current->dirty_paused_when;
1193 * For less than 1s think time (ext3/4 may block the dirtier
1194 * for up to 800ms from time to time on 1-HDD; so does xfs,
1195 * however at much less frequency), try to compensate it in
1196 * future periods by updating the virtual time; otherwise just
1197 * do a reset, as it may be a light dirtier.
1199 if (pause < min_pause) {
1200 trace_balance_dirty_pages(bdi,
1213 current->dirty_paused_when = now;
1214 current->nr_dirtied = 0;
1215 } else if (period) {
1216 current->dirty_paused_when += period;
1217 current->nr_dirtied = 0;
1218 } else if (current->nr_dirtied_pause <= pages_dirtied)
1219 current->nr_dirtied_pause += pages_dirtied;
1222 if (unlikely(pause > max_pause)) {
1223 /* for occasional dropped task_ratelimit */
1224 now += min(pause - max_pause, max_pause);
1229 trace_balance_dirty_pages(bdi,
1241 __set_current_state(TASK_KILLABLE);
1242 io_schedule_timeout(pause);
1244 current->dirty_paused_when = now + pause;
1245 current->nr_dirtied = 0;
1246 current->nr_dirtied_pause = nr_dirtied_pause;
1249 * This is typically equal to (nr_dirty < dirty_thresh) and can
1250 * also keep "1000+ dd on a slow USB stick" under control.
1256 * In the case of an unresponding NFS server and the NFS dirty
1257 * pages exceeds dirty_thresh, give the other good bdi's a pipe
1258 * to go through, so that tasks on them still remain responsive.
1260 * In theory 1 page is enough to keep the comsumer-producer
1261 * pipe going: the flusher cleans 1 page => the task dirties 1
1262 * more page. However bdi_dirty has accounting errors. So use
1263 * the larger and more IO friendly bdi_stat_error.
1265 if (bdi_dirty <= bdi_stat_error(bdi))
1268 if (fatal_signal_pending(current))
1272 if (!dirty_exceeded && bdi->dirty_exceeded)
1273 bdi->dirty_exceeded = 0;
1275 if (writeback_in_progress(bdi))
1279 * In laptop mode, we wait until hitting the higher threshold before
1280 * starting background writeout, and then write out all the way down
1281 * to the lower threshold. So slow writers cause minimal disk activity.
1283 * In normal mode, we start background writeout at the lower
1284 * background_thresh, to keep the amount of dirty memory low.
1289 if (nr_reclaimable > background_thresh)
1290 bdi_start_background_writeback(bdi);
1293 void set_page_dirty_balance(struct page *page, int page_mkwrite)
1295 if (set_page_dirty(page) || page_mkwrite) {
1296 struct address_space *mapping = page_mapping(page);
1299 balance_dirty_pages_ratelimited(mapping);
1303 static DEFINE_PER_CPU(int, bdp_ratelimits);
1306 * Normal tasks are throttled by
1308 * dirty tsk->nr_dirtied_pause pages;
1309 * take a snap in balance_dirty_pages();
1311 * However there is a worst case. If every task exit immediately when dirtied
1312 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1313 * called to throttle the page dirties. The solution is to save the not yet
1314 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1315 * randomly into the running tasks. This works well for the above worst case,
1316 * as the new task will pick up and accumulate the old task's leaked dirty
1317 * count and eventually get throttled.
1319 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1322 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
1323 * @mapping: address_space which was dirtied
1324 * @nr_pages_dirtied: number of pages which the caller has just dirtied
1326 * Processes which are dirtying memory should call in here once for each page
1327 * which was newly dirtied. The function will periodically check the system's
1328 * dirty state and will initiate writeback if needed.
1330 * On really big machines, get_writeback_state is expensive, so try to avoid
1331 * calling it too often (ratelimiting). But once we're over the dirty memory
1332 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1333 * from overshooting the limit by (ratelimit_pages) each.
1335 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1336 unsigned long nr_pages_dirtied)
1338 struct backing_dev_info *bdi = mapping->backing_dev_info;
1342 if (!bdi_cap_account_dirty(bdi))
1345 ratelimit = current->nr_dirtied_pause;
1346 if (bdi->dirty_exceeded)
1347 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1351 * This prevents one CPU to accumulate too many dirtied pages without
1352 * calling into balance_dirty_pages(), which can happen when there are
1353 * 1000+ tasks, all of them start dirtying pages at exactly the same
1354 * time, hence all honoured too large initial task->nr_dirtied_pause.
1356 p = &__get_cpu_var(bdp_ratelimits);
1357 if (unlikely(current->nr_dirtied >= ratelimit))
1359 else if (unlikely(*p >= ratelimit_pages)) {
1364 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1365 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1366 * the dirty throttling and livelock other long-run dirtiers.
1368 p = &__get_cpu_var(dirty_throttle_leaks);
1369 if (*p > 0 && current->nr_dirtied < ratelimit) {
1370 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1371 *p -= nr_pages_dirtied;
1372 current->nr_dirtied += nr_pages_dirtied;
1376 if (unlikely(current->nr_dirtied >= ratelimit))
1377 balance_dirty_pages(mapping, current->nr_dirtied);
1379 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
1381 void throttle_vm_writeout(gfp_t gfp_mask)
1383 unsigned long background_thresh;
1384 unsigned long dirty_thresh;
1387 global_dirty_limits(&background_thresh, &dirty_thresh);
1390 * Boost the allowable dirty threshold a bit for page
1391 * allocators so they don't get DoS'ed by heavy writers
1393 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1395 if (global_page_state(NR_UNSTABLE_NFS) +
1396 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1398 congestion_wait(BLK_RW_ASYNC, HZ/10);
1401 * The caller might hold locks which can prevent IO completion
1402 * or progress in the filesystem. So we cannot just sit here
1403 * waiting for IO to complete.
1405 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1411 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1413 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1414 void __user *buffer, size_t *length, loff_t *ppos)
1416 proc_dointvec(table, write, buffer, length, ppos);
1417 bdi_arm_supers_timer();
1422 void laptop_mode_timer_fn(unsigned long data)
1424 struct request_queue *q = (struct request_queue *)data;
1425 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1426 global_page_state(NR_UNSTABLE_NFS);
1429 * We want to write everything out, not just down to the dirty
1432 if (bdi_has_dirty_io(&q->backing_dev_info))
1433 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1434 WB_REASON_LAPTOP_TIMER);
1438 * We've spun up the disk and we're in laptop mode: schedule writeback
1439 * of all dirty data a few seconds from now. If the flush is already scheduled
1440 * then push it back - the user is still using the disk.
1442 void laptop_io_completion(struct backing_dev_info *info)
1444 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1448 * We're in laptop mode and we've just synced. The sync's writes will have
1449 * caused another writeback to be scheduled by laptop_io_completion.
1450 * Nothing needs to be written back anymore, so we unschedule the writeback.
1452 void laptop_sync_completion(void)
1454 struct backing_dev_info *bdi;
1458 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1459 del_timer(&bdi->laptop_mode_wb_timer);
1466 * If ratelimit_pages is too high then we can get into dirty-data overload
1467 * if a large number of processes all perform writes at the same time.
1468 * If it is too low then SMP machines will call the (expensive)
1469 * get_writeback_state too often.
1471 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1472 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1476 void writeback_set_ratelimit(void)
1478 unsigned long background_thresh;
1479 unsigned long dirty_thresh;
1480 global_dirty_limits(&background_thresh, &dirty_thresh);
1481 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1482 if (ratelimit_pages < 16)
1483 ratelimit_pages = 16;
1486 static int __cpuinit
1487 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1489 writeback_set_ratelimit();
1493 static struct notifier_block __cpuinitdata ratelimit_nb = {
1494 .notifier_call = ratelimit_handler,
1499 * Called early on to tune the page writeback dirty limits.
1501 * We used to scale dirty pages according to how total memory
1502 * related to pages that could be allocated for buffers (by
1503 * comparing nr_free_buffer_pages() to vm_total_pages.
1505 * However, that was when we used "dirty_ratio" to scale with
1506 * all memory, and we don't do that any more. "dirty_ratio"
1507 * is now applied to total non-HIGHPAGE memory (by subtracting
1508 * totalhigh_pages from vm_total_pages), and as such we can't
1509 * get into the old insane situation any more where we had
1510 * large amounts of dirty pages compared to a small amount of
1511 * non-HIGHMEM memory.
1513 * But we might still want to scale the dirty_ratio by how
1514 * much memory the box has..
1516 void __init page_writeback_init(void)
1520 writeback_set_ratelimit();
1521 register_cpu_notifier(&ratelimit_nb);
1523 shift = calc_period_shift();
1524 prop_descriptor_init(&vm_completions, shift);
1528 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1529 * @mapping: address space structure to write
1530 * @start: starting page index
1531 * @end: ending page index (inclusive)
1533 * This function scans the page range from @start to @end (inclusive) and tags
1534 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1535 * that write_cache_pages (or whoever calls this function) will then use
1536 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1537 * used to avoid livelocking of writeback by a process steadily creating new
1538 * dirty pages in the file (thus it is important for this function to be quick
1539 * so that it can tag pages faster than a dirtying process can create them).
1542 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1544 void tag_pages_for_writeback(struct address_space *mapping,
1545 pgoff_t start, pgoff_t end)
1547 #define WRITEBACK_TAG_BATCH 4096
1548 unsigned long tagged;
1551 spin_lock_irq(&mapping->tree_lock);
1552 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1553 &start, end, WRITEBACK_TAG_BATCH,
1554 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1555 spin_unlock_irq(&mapping->tree_lock);
1556 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1558 /* We check 'start' to handle wrapping when end == ~0UL */
1559 } while (tagged >= WRITEBACK_TAG_BATCH && start);
1561 EXPORT_SYMBOL(tag_pages_for_writeback);
1564 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1565 * @mapping: address space structure to write
1566 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1567 * @writepage: function called for each page
1568 * @data: data passed to writepage function
1570 * If a page is already under I/O, write_cache_pages() skips it, even
1571 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1572 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1573 * and msync() need to guarantee that all the data which was dirty at the time
1574 * the call was made get new I/O started against them. If wbc->sync_mode is
1575 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1576 * existing IO to complete.
1578 * To avoid livelocks (when other process dirties new pages), we first tag
1579 * pages which should be written back with TOWRITE tag and only then start
1580 * writing them. For data-integrity sync we have to be careful so that we do
1581 * not miss some pages (e.g., because some other process has cleared TOWRITE
1582 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1583 * by the process clearing the DIRTY tag (and submitting the page for IO).
1585 int write_cache_pages(struct address_space *mapping,
1586 struct writeback_control *wbc, writepage_t writepage,
1591 struct pagevec pvec;
1593 pgoff_t uninitialized_var(writeback_index);
1595 pgoff_t end; /* Inclusive */
1598 int range_whole = 0;
1601 pagevec_init(&pvec, 0);
1602 if (wbc->range_cyclic) {
1603 writeback_index = mapping->writeback_index; /* prev offset */
1604 index = writeback_index;
1611 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1612 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1613 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1615 cycled = 1; /* ignore range_cyclic tests */
1617 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1618 tag = PAGECACHE_TAG_TOWRITE;
1620 tag = PAGECACHE_TAG_DIRTY;
1622 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1623 tag_pages_for_writeback(mapping, index, end);
1625 while (!done && (index <= end)) {
1628 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1629 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1633 for (i = 0; i < nr_pages; i++) {
1634 struct page *page = pvec.pages[i];
1637 * At this point, the page may be truncated or
1638 * invalidated (changing page->mapping to NULL), or
1639 * even swizzled back from swapper_space to tmpfs file
1640 * mapping. However, page->index will not change
1641 * because we have a reference on the page.
1643 if (page->index > end) {
1645 * can't be range_cyclic (1st pass) because
1646 * end == -1 in that case.
1652 done_index = page->index;
1657 * Page truncated or invalidated. We can freely skip it
1658 * then, even for data integrity operations: the page
1659 * has disappeared concurrently, so there could be no
1660 * real expectation of this data interity operation
1661 * even if there is now a new, dirty page at the same
1662 * pagecache address.
1664 if (unlikely(page->mapping != mapping)) {
1670 if (!PageDirty(page)) {
1671 /* someone wrote it for us */
1672 goto continue_unlock;
1675 if (PageWriteback(page)) {
1676 if (wbc->sync_mode != WB_SYNC_NONE)
1677 wait_on_page_writeback(page);
1679 goto continue_unlock;
1682 BUG_ON(PageWriteback(page));
1683 if (!clear_page_dirty_for_io(page))
1684 goto continue_unlock;
1686 trace_wbc_writepage(wbc, mapping->backing_dev_info);
1687 ret = (*writepage)(page, wbc, data);
1688 if (unlikely(ret)) {
1689 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1694 * done_index is set past this page,
1695 * so media errors will not choke
1696 * background writeout for the entire
1697 * file. This has consequences for
1698 * range_cyclic semantics (ie. it may
1699 * not be suitable for data integrity
1702 done_index = page->index + 1;
1709 * We stop writing back only if we are not doing
1710 * integrity sync. In case of integrity sync we have to
1711 * keep going until we have written all the pages
1712 * we tagged for writeback prior to entering this loop.
1714 if (--wbc->nr_to_write <= 0 &&
1715 wbc->sync_mode == WB_SYNC_NONE) {
1720 pagevec_release(&pvec);
1723 if (!cycled && !done) {
1726 * We hit the last page and there is more work to be done: wrap
1727 * back to the start of the file
1731 end = writeback_index - 1;
1734 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1735 mapping->writeback_index = done_index;
1739 EXPORT_SYMBOL(write_cache_pages);
1742 * Function used by generic_writepages to call the real writepage
1743 * function and set the mapping flags on error
1745 static int __writepage(struct page *page, struct writeback_control *wbc,
1748 struct address_space *mapping = data;
1749 int ret = mapping->a_ops->writepage(page, wbc);
1750 mapping_set_error(mapping, ret);
1755 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1756 * @mapping: address space structure to write
1757 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1759 * This is a library function, which implements the writepages()
1760 * address_space_operation.
1762 int generic_writepages(struct address_space *mapping,
1763 struct writeback_control *wbc)
1765 struct blk_plug plug;
1768 /* deal with chardevs and other special file */
1769 if (!mapping->a_ops->writepage)
1772 blk_start_plug(&plug);
1773 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1774 blk_finish_plug(&plug);
1778 EXPORT_SYMBOL(generic_writepages);
1780 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1784 if (wbc->nr_to_write <= 0)
1786 if (mapping->a_ops->writepages)
1787 ret = mapping->a_ops->writepages(mapping, wbc);
1789 ret = generic_writepages(mapping, wbc);
1794 * write_one_page - write out a single page and optionally wait on I/O
1795 * @page: the page to write
1796 * @wait: if true, wait on writeout
1798 * The page must be locked by the caller and will be unlocked upon return.
1800 * write_one_page() returns a negative error code if I/O failed.
1802 int write_one_page(struct page *page, int wait)
1804 struct address_space *mapping = page->mapping;
1806 struct writeback_control wbc = {
1807 .sync_mode = WB_SYNC_ALL,
1811 BUG_ON(!PageLocked(page));
1814 wait_on_page_writeback(page);
1816 if (clear_page_dirty_for_io(page)) {
1817 page_cache_get(page);
1818 ret = mapping->a_ops->writepage(page, &wbc);
1819 if (ret == 0 && wait) {
1820 wait_on_page_writeback(page);
1821 if (PageError(page))
1824 page_cache_release(page);
1830 EXPORT_SYMBOL(write_one_page);
1833 * For address_spaces which do not use buffers nor write back.
1835 int __set_page_dirty_no_writeback(struct page *page)
1837 if (!PageDirty(page))
1838 return !TestSetPageDirty(page);
1843 * Helper function for set_page_dirty family.
1844 * NOTE: This relies on being atomic wrt interrupts.
1846 void account_page_dirtied(struct page *page, struct address_space *mapping)
1848 if (mapping_cap_account_dirty(mapping)) {
1849 __inc_zone_page_state(page, NR_FILE_DIRTY);
1850 __inc_zone_page_state(page, NR_DIRTIED);
1851 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1852 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1853 task_io_account_write(PAGE_CACHE_SIZE);
1854 current->nr_dirtied++;
1855 this_cpu_inc(bdp_ratelimits);
1858 EXPORT_SYMBOL(account_page_dirtied);
1861 * Helper function for set_page_writeback family.
1862 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1865 void account_page_writeback(struct page *page)
1867 inc_zone_page_state(page, NR_WRITEBACK);
1869 EXPORT_SYMBOL(account_page_writeback);
1872 * For address_spaces which do not use buffers. Just tag the page as dirty in
1875 * This is also used when a single buffer is being dirtied: we want to set the
1876 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1877 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1879 * Most callers have locked the page, which pins the address_space in memory.
1880 * But zap_pte_range() does not lock the page, however in that case the
1881 * mapping is pinned by the vma's ->vm_file reference.
1883 * We take care to handle the case where the page was truncated from the
1884 * mapping by re-checking page_mapping() inside tree_lock.
1886 int __set_page_dirty_nobuffers(struct page *page)
1888 if (!TestSetPageDirty(page)) {
1889 struct address_space *mapping = page_mapping(page);
1890 struct address_space *mapping2;
1895 spin_lock_irq(&mapping->tree_lock);
1896 mapping2 = page_mapping(page);
1897 if (mapping2) { /* Race with truncate? */
1898 BUG_ON(mapping2 != mapping);
1899 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1900 account_page_dirtied(page, mapping);
1901 radix_tree_tag_set(&mapping->page_tree,
1902 page_index(page), PAGECACHE_TAG_DIRTY);
1904 spin_unlock_irq(&mapping->tree_lock);
1905 if (mapping->host) {
1906 /* !PageAnon && !swapper_space */
1907 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1913 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1916 * Call this whenever redirtying a page, to de-account the dirty counters
1917 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
1918 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
1919 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
1922 void account_page_redirty(struct page *page)
1924 struct address_space *mapping = page->mapping;
1925 if (mapping && mapping_cap_account_dirty(mapping)) {
1926 current->nr_dirtied--;
1927 dec_zone_page_state(page, NR_DIRTIED);
1928 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1931 EXPORT_SYMBOL(account_page_redirty);
1934 * When a writepage implementation decides that it doesn't want to write this
1935 * page for some reason, it should redirty the locked page via
1936 * redirty_page_for_writepage() and it should then unlock the page and return 0
1938 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1940 wbc->pages_skipped++;
1941 account_page_redirty(page);
1942 return __set_page_dirty_nobuffers(page);
1944 EXPORT_SYMBOL(redirty_page_for_writepage);
1949 * For pages with a mapping this should be done under the page lock
1950 * for the benefit of asynchronous memory errors who prefer a consistent
1951 * dirty state. This rule can be broken in some special cases,
1952 * but should be better not to.
1954 * If the mapping doesn't provide a set_page_dirty a_op, then
1955 * just fall through and assume that it wants buffer_heads.
1957 int set_page_dirty(struct page *page)
1959 struct address_space *mapping = page_mapping(page);
1961 if (likely(mapping)) {
1962 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1964 * readahead/lru_deactivate_page could remain
1965 * PG_readahead/PG_reclaim due to race with end_page_writeback
1966 * About readahead, if the page is written, the flags would be
1967 * reset. So no problem.
1968 * About lru_deactivate_page, if the page is redirty, the flag
1969 * will be reset. So no problem. but if the page is used by readahead
1970 * it will confuse readahead and make it restart the size rampup
1971 * process. But it's a trivial problem.
1973 ClearPageReclaim(page);
1976 spd = __set_page_dirty_buffers;
1978 return (*spd)(page);
1980 if (!PageDirty(page)) {
1981 if (!TestSetPageDirty(page))
1986 EXPORT_SYMBOL(set_page_dirty);
1989 * set_page_dirty() is racy if the caller has no reference against
1990 * page->mapping->host, and if the page is unlocked. This is because another
1991 * CPU could truncate the page off the mapping and then free the mapping.
1993 * Usually, the page _is_ locked, or the caller is a user-space process which
1994 * holds a reference on the inode by having an open file.
1996 * In other cases, the page should be locked before running set_page_dirty().
1998 int set_page_dirty_lock(struct page *page)
2003 ret = set_page_dirty(page);
2007 EXPORT_SYMBOL(set_page_dirty_lock);
2010 * Clear a page's dirty flag, while caring for dirty memory accounting.
2011 * Returns true if the page was previously dirty.
2013 * This is for preparing to put the page under writeout. We leave the page
2014 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2015 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2016 * implementation will run either set_page_writeback() or set_page_dirty(),
2017 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2020 * This incoherency between the page's dirty flag and radix-tree tag is
2021 * unfortunate, but it only exists while the page is locked.
2023 int clear_page_dirty_for_io(struct page *page)
2025 struct address_space *mapping = page_mapping(page);
2027 BUG_ON(!PageLocked(page));
2029 if (mapping && mapping_cap_account_dirty(mapping)) {
2031 * Yes, Virginia, this is indeed insane.
2033 * We use this sequence to make sure that
2034 * (a) we account for dirty stats properly
2035 * (b) we tell the low-level filesystem to
2036 * mark the whole page dirty if it was
2037 * dirty in a pagetable. Only to then
2038 * (c) clean the page again and return 1 to
2039 * cause the writeback.
2041 * This way we avoid all nasty races with the
2042 * dirty bit in multiple places and clearing
2043 * them concurrently from different threads.
2045 * Note! Normally the "set_page_dirty(page)"
2046 * has no effect on the actual dirty bit - since
2047 * that will already usually be set. But we
2048 * need the side effects, and it can help us
2051 * We basically use the page "master dirty bit"
2052 * as a serialization point for all the different
2053 * threads doing their things.
2055 if (page_mkclean(page))
2056 set_page_dirty(page);
2058 * We carefully synchronise fault handlers against
2059 * installing a dirty pte and marking the page dirty
2060 * at this point. We do this by having them hold the
2061 * page lock at some point after installing their
2062 * pte, but before marking the page dirty.
2063 * Pages are always locked coming in here, so we get
2064 * the desired exclusion. See mm/memory.c:do_wp_page()
2065 * for more comments.
2067 if (TestClearPageDirty(page)) {
2068 dec_zone_page_state(page, NR_FILE_DIRTY);
2069 dec_bdi_stat(mapping->backing_dev_info,
2075 return TestClearPageDirty(page);
2077 EXPORT_SYMBOL(clear_page_dirty_for_io);
2079 int test_clear_page_writeback(struct page *page)
2081 struct address_space *mapping = page_mapping(page);
2085 struct backing_dev_info *bdi = mapping->backing_dev_info;
2086 unsigned long flags;
2088 spin_lock_irqsave(&mapping->tree_lock, flags);
2089 ret = TestClearPageWriteback(page);
2091 radix_tree_tag_clear(&mapping->page_tree,
2093 PAGECACHE_TAG_WRITEBACK);
2094 if (bdi_cap_account_writeback(bdi)) {
2095 __dec_bdi_stat(bdi, BDI_WRITEBACK);
2096 __bdi_writeout_inc(bdi);
2099 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2101 ret = TestClearPageWriteback(page);
2104 dec_zone_page_state(page, NR_WRITEBACK);
2105 inc_zone_page_state(page, NR_WRITTEN);
2110 int test_set_page_writeback(struct page *page)
2112 struct address_space *mapping = page_mapping(page);
2116 struct backing_dev_info *bdi = mapping->backing_dev_info;
2117 unsigned long flags;
2119 spin_lock_irqsave(&mapping->tree_lock, flags);
2120 ret = TestSetPageWriteback(page);
2122 radix_tree_tag_set(&mapping->page_tree,
2124 PAGECACHE_TAG_WRITEBACK);
2125 if (bdi_cap_account_writeback(bdi))
2126 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2128 if (!PageDirty(page))
2129 radix_tree_tag_clear(&mapping->page_tree,
2131 PAGECACHE_TAG_DIRTY);
2132 radix_tree_tag_clear(&mapping->page_tree,
2134 PAGECACHE_TAG_TOWRITE);
2135 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2137 ret = TestSetPageWriteback(page);
2140 account_page_writeback(page);
2144 EXPORT_SYMBOL(test_set_page_writeback);
2147 * Return true if any of the pages in the mapping are marked with the
2150 int mapping_tagged(struct address_space *mapping, int tag)
2152 return radix_tree_tagged(&mapping->page_tree, tag);
2154 EXPORT_SYMBOL(mapping_tagged);