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
149 static unsigned long highmem_dirtyable_memory(unsigned long total)
151 #ifdef CONFIG_HIGHMEM
155 for_each_node_state(node, N_HIGH_MEMORY) {
157 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
159 x += zone_page_state(z, NR_FREE_PAGES) +
160 zone_reclaimable_pages(z) - z->dirty_balance_reserve;
163 * Make sure that the number of highmem pages is never larger
164 * than the number of the total dirtyable memory. This can only
165 * occur in very strange VM situations but we want to make sure
166 * that this does not occur.
168 return min(x, total);
175 * determine_dirtyable_memory - amount of memory that may be used
177 * Returns the numebr of pages that can currently be freed and used
178 * by the kernel for direct mappings.
180 static unsigned long determine_dirtyable_memory(void)
184 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages() -
185 dirty_balance_reserve;
187 if (!vm_highmem_is_dirtyable)
188 x -= highmem_dirtyable_memory(x);
190 return x + 1; /* Ensure that we never return 0 */
194 * couple the period to the dirty_ratio:
196 * period/2 ~ roundup_pow_of_two(dirty limit)
198 static int calc_period_shift(void)
200 unsigned long dirty_total;
203 dirty_total = vm_dirty_bytes / PAGE_SIZE;
205 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
207 return 2 + ilog2(dirty_total - 1);
211 * update the period when the dirty threshold changes.
213 static void update_completion_period(void)
215 int shift = calc_period_shift();
216 prop_change_shift(&vm_completions, shift);
218 writeback_set_ratelimit();
221 int dirty_background_ratio_handler(struct ctl_table *table, int write,
222 void __user *buffer, size_t *lenp,
227 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
228 if (ret == 0 && write)
229 dirty_background_bytes = 0;
233 int dirty_background_bytes_handler(struct ctl_table *table, int write,
234 void __user *buffer, size_t *lenp,
239 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
240 if (ret == 0 && write)
241 dirty_background_ratio = 0;
245 int dirty_ratio_handler(struct ctl_table *table, int write,
246 void __user *buffer, size_t *lenp,
249 int old_ratio = vm_dirty_ratio;
252 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
253 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
254 update_completion_period();
260 int dirty_bytes_handler(struct ctl_table *table, int write,
261 void __user *buffer, size_t *lenp,
264 unsigned long old_bytes = vm_dirty_bytes;
267 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
268 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
269 update_completion_period();
276 * Increment the BDI's writeout completion count and the global writeout
277 * completion count. Called from test_clear_page_writeback().
279 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
281 __inc_bdi_stat(bdi, BDI_WRITTEN);
282 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
286 void bdi_writeout_inc(struct backing_dev_info *bdi)
290 local_irq_save(flags);
291 __bdi_writeout_inc(bdi);
292 local_irq_restore(flags);
294 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
297 * Obtain an accurate fraction of the BDI's portion.
299 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
300 long *numerator, long *denominator)
302 prop_fraction_percpu(&vm_completions, &bdi->completions,
303 numerator, denominator);
307 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
308 * registered backing devices, which, for obvious reasons, can not
311 static unsigned int bdi_min_ratio;
313 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
317 spin_lock_bh(&bdi_lock);
318 if (min_ratio > bdi->max_ratio) {
321 min_ratio -= bdi->min_ratio;
322 if (bdi_min_ratio + min_ratio < 100) {
323 bdi_min_ratio += min_ratio;
324 bdi->min_ratio += min_ratio;
329 spin_unlock_bh(&bdi_lock);
334 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
341 spin_lock_bh(&bdi_lock);
342 if (bdi->min_ratio > max_ratio) {
345 bdi->max_ratio = max_ratio;
346 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
348 spin_unlock_bh(&bdi_lock);
352 EXPORT_SYMBOL(bdi_set_max_ratio);
354 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
355 unsigned long bg_thresh)
357 return (thresh + bg_thresh) / 2;
360 static unsigned long hard_dirty_limit(unsigned long thresh)
362 return max(thresh, global_dirty_limit);
366 * global_dirty_limits - background-writeback and dirty-throttling thresholds
368 * Calculate the dirty thresholds based on sysctl parameters
369 * - vm.dirty_background_ratio or vm.dirty_background_bytes
370 * - vm.dirty_ratio or vm.dirty_bytes
371 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
374 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
376 unsigned long background;
378 unsigned long uninitialized_var(available_memory);
379 struct task_struct *tsk;
381 if (!vm_dirty_bytes || !dirty_background_bytes)
382 available_memory = determine_dirtyable_memory();
385 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
387 dirty = (vm_dirty_ratio * available_memory) / 100;
389 if (dirty_background_bytes)
390 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
392 background = (dirty_background_ratio * available_memory) / 100;
394 if (background >= dirty)
395 background = dirty / 2;
397 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
398 background += background / 4;
401 *pbackground = background;
403 trace_global_dirty_state(background, dirty);
407 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
408 * @bdi: the backing_dev_info to query
409 * @dirty: global dirty limit in pages
411 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
412 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
414 * Note that balance_dirty_pages() will only seriously take it as a hard limit
415 * when sleeping max_pause per page is not enough to keep the dirty pages under
416 * control. For example, when the device is completely stalled due to some error
417 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
418 * In the other normal situations, it acts more gently by throttling the tasks
419 * more (rather than completely block them) when the bdi dirty pages go high.
421 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
422 * - starving fast devices
423 * - piling up dirty pages (that will take long time to sync) on slow devices
425 * The bdi's share of dirty limit will be adapting to its throughput and
426 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
428 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
431 long numerator, denominator;
434 * Calculate this BDI's share of the dirty ratio.
436 bdi_writeout_fraction(bdi, &numerator, &denominator);
438 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
439 bdi_dirty *= numerator;
440 do_div(bdi_dirty, denominator);
442 bdi_dirty += (dirty * bdi->min_ratio) / 100;
443 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
444 bdi_dirty = dirty * bdi->max_ratio / 100;
450 * Dirty position control.
452 * (o) global/bdi setpoints
454 * We want the dirty pages be balanced around the global/bdi setpoints.
455 * When the number of dirty pages is higher/lower than the setpoint, the
456 * dirty position control ratio (and hence task dirty ratelimit) will be
457 * decreased/increased to bring the dirty pages back to the setpoint.
459 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
461 * if (dirty < setpoint) scale up pos_ratio
462 * if (dirty > setpoint) scale down pos_ratio
464 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
465 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
467 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
469 * (o) global control line
473 * | |<===== global dirty control scope ======>|
481 * 1.0 ................................*
487 * 0 +------------.------------------.----------------------*------------->
488 * freerun^ setpoint^ limit^ dirty pages
490 * (o) bdi control line
498 * | * |<=========== span ============>|
499 * 1.0 .......................*
511 * 1/4 ...............................................* * * * * * * * * * * *
515 * 0 +----------------------.-------------------------------.------------->
516 * bdi_setpoint^ x_intercept^
518 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
519 * be smoothly throttled down to normal if it starts high in situations like
520 * - start writing to a slow SD card and a fast disk at the same time. The SD
521 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
522 * - the bdi dirty thresh drops quickly due to change of JBOD workload
524 static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
525 unsigned long thresh,
526 unsigned long bg_thresh,
528 unsigned long bdi_thresh,
529 unsigned long bdi_dirty)
531 unsigned long write_bw = bdi->avg_write_bandwidth;
532 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
533 unsigned long limit = hard_dirty_limit(thresh);
534 unsigned long x_intercept;
535 unsigned long setpoint; /* dirty pages' target balance point */
536 unsigned long bdi_setpoint;
538 long long pos_ratio; /* for scaling up/down the rate limit */
541 if (unlikely(dirty >= limit))
548 * f(dirty) := 1.0 + (----------------)
551 * it's a 3rd order polynomial that subjects to
553 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
554 * (2) f(setpoint) = 1.0 => the balance point
555 * (3) f(limit) = 0 => the hard limit
556 * (4) df/dx <= 0 => negative feedback control
557 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
558 * => fast response on large errors; small oscillation near setpoint
560 setpoint = (freerun + limit) / 2;
561 x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
562 limit - setpoint + 1);
564 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
565 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
566 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
569 * We have computed basic pos_ratio above based on global situation. If
570 * the bdi is over/under its share of dirty pages, we want to scale
571 * pos_ratio further down/up. That is done by the following mechanism.
577 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
579 * x_intercept - bdi_dirty
580 * := --------------------------
581 * x_intercept - bdi_setpoint
583 * The main bdi control line is a linear function that subjects to
585 * (1) f(bdi_setpoint) = 1.0
586 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
587 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
589 * For single bdi case, the dirty pages are observed to fluctuate
590 * regularly within range
591 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
592 * for various filesystems, where (2) can yield in a reasonable 12.5%
593 * fluctuation range for pos_ratio.
595 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
596 * own size, so move the slope over accordingly and choose a slope that
597 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
599 if (unlikely(bdi_thresh > thresh))
602 * It's very possible that bdi_thresh is close to 0 not because the
603 * device is slow, but that it has remained inactive for long time.
604 * Honour such devices a reasonable good (hopefully IO efficient)
605 * threshold, so that the occasional writes won't be blocked and active
606 * writes can rampup the threshold quickly.
608 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
610 * scale global setpoint to bdi's:
611 * bdi_setpoint = setpoint * bdi_thresh / thresh
613 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
614 bdi_setpoint = setpoint * (u64)x >> 16;
616 * Use span=(8*write_bw) in single bdi case as indicated by
617 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
619 * bdi_thresh thresh - bdi_thresh
620 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
623 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
624 x_intercept = bdi_setpoint + span;
626 if (bdi_dirty < x_intercept - span / 4) {
627 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
628 x_intercept - bdi_setpoint + 1);
633 * bdi reserve area, safeguard against dirty pool underrun and disk idle
634 * It may push the desired control point of global dirty pages higher
637 x_intercept = bdi_thresh / 2;
638 if (bdi_dirty < x_intercept) {
639 if (bdi_dirty > x_intercept / 8)
640 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
648 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
649 unsigned long elapsed,
650 unsigned long written)
652 const unsigned long period = roundup_pow_of_two(3 * HZ);
653 unsigned long avg = bdi->avg_write_bandwidth;
654 unsigned long old = bdi->write_bandwidth;
658 * bw = written * HZ / elapsed
660 * bw * elapsed + write_bandwidth * (period - elapsed)
661 * write_bandwidth = ---------------------------------------------------
664 bw = written - bdi->written_stamp;
666 if (unlikely(elapsed > period)) {
671 bw += (u64)bdi->write_bandwidth * (period - elapsed);
672 bw >>= ilog2(period);
675 * one more level of smoothing, for filtering out sudden spikes
677 if (avg > old && old >= (unsigned long)bw)
678 avg -= (avg - old) >> 3;
680 if (avg < old && old <= (unsigned long)bw)
681 avg += (old - avg) >> 3;
684 bdi->write_bandwidth = bw;
685 bdi->avg_write_bandwidth = avg;
689 * The global dirtyable memory and dirty threshold could be suddenly knocked
690 * down by a large amount (eg. on the startup of KVM in a swapless system).
691 * This may throw the system into deep dirty exceeded state and throttle
692 * heavy/light dirtiers alike. To retain good responsiveness, maintain
693 * global_dirty_limit for tracking slowly down to the knocked down dirty
696 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
698 unsigned long limit = global_dirty_limit;
701 * Follow up in one step.
703 if (limit < thresh) {
709 * Follow down slowly. Use the higher one as the target, because thresh
710 * may drop below dirty. This is exactly the reason to introduce
711 * global_dirty_limit which is guaranteed to lie above the dirty pages.
713 thresh = max(thresh, dirty);
714 if (limit > thresh) {
715 limit -= (limit - thresh) >> 5;
720 global_dirty_limit = limit;
723 static void global_update_bandwidth(unsigned long thresh,
727 static DEFINE_SPINLOCK(dirty_lock);
728 static unsigned long update_time;
731 * check locklessly first to optimize away locking for the most time
733 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
736 spin_lock(&dirty_lock);
737 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
738 update_dirty_limit(thresh, dirty);
741 spin_unlock(&dirty_lock);
745 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
747 * Normal bdi tasks will be curbed at or below it in long term.
748 * Obviously it should be around (write_bw / N) when there are N dd tasks.
750 static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
751 unsigned long thresh,
752 unsigned long bg_thresh,
754 unsigned long bdi_thresh,
755 unsigned long bdi_dirty,
756 unsigned long dirtied,
757 unsigned long elapsed)
759 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
760 unsigned long limit = hard_dirty_limit(thresh);
761 unsigned long setpoint = (freerun + limit) / 2;
762 unsigned long write_bw = bdi->avg_write_bandwidth;
763 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
764 unsigned long dirty_rate;
765 unsigned long task_ratelimit;
766 unsigned long balanced_dirty_ratelimit;
767 unsigned long pos_ratio;
772 * The dirty rate will match the writeout rate in long term, except
773 * when dirty pages are truncated by userspace or re-dirtied by FS.
775 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
777 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
778 bdi_thresh, bdi_dirty);
780 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
782 task_ratelimit = (u64)dirty_ratelimit *
783 pos_ratio >> RATELIMIT_CALC_SHIFT;
784 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
787 * A linear estimation of the "balanced" throttle rate. The theory is,
788 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
789 * dirty_rate will be measured to be (N * task_ratelimit). So the below
790 * formula will yield the balanced rate limit (write_bw / N).
792 * Note that the expanded form is not a pure rate feedback:
793 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
794 * but also takes pos_ratio into account:
795 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
797 * (1) is not realistic because pos_ratio also takes part in balancing
798 * the dirty rate. Consider the state
799 * pos_ratio = 0.5 (3)
800 * rate = 2 * (write_bw / N) (4)
801 * If (1) is used, it will stuck in that state! Because each dd will
803 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
805 * dirty_rate = N * task_ratelimit = write_bw (6)
806 * put (6) into (1) we get
807 * rate_(i+1) = rate_(i) (7)
809 * So we end up using (2) to always keep
810 * rate_(i+1) ~= (write_bw / N) (8)
811 * regardless of the value of pos_ratio. As long as (8) is satisfied,
812 * pos_ratio is able to drive itself to 1.0, which is not only where
813 * the dirty count meet the setpoint, but also where the slope of
814 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
816 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
820 * We could safely do this and return immediately:
822 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
824 * However to get a more stable dirty_ratelimit, the below elaborated
825 * code makes use of task_ratelimit to filter out sigular points and
826 * limit the step size.
828 * The below code essentially only uses the relative value of
830 * task_ratelimit - dirty_ratelimit
831 * = (pos_ratio - 1) * dirty_ratelimit
833 * which reflects the direction and size of dirty position error.
837 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
838 * task_ratelimit is on the same side of dirty_ratelimit, too.
840 * - dirty_ratelimit > balanced_dirty_ratelimit
841 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
842 * lowering dirty_ratelimit will help meet both the position and rate
843 * control targets. Otherwise, don't update dirty_ratelimit if it will
844 * only help meet the rate target. After all, what the users ultimately
845 * feel and care are stable dirty rate and small position error.
847 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
848 * and filter out the sigular points of balanced_dirty_ratelimit. Which
849 * keeps jumping around randomly and can even leap far away at times
850 * due to the small 200ms estimation period of dirty_rate (we want to
851 * keep that period small to reduce time lags).
854 if (dirty < setpoint) {
855 x = min(bdi->balanced_dirty_ratelimit,
856 min(balanced_dirty_ratelimit, task_ratelimit));
857 if (dirty_ratelimit < x)
858 step = x - dirty_ratelimit;
860 x = max(bdi->balanced_dirty_ratelimit,
861 max(balanced_dirty_ratelimit, task_ratelimit));
862 if (dirty_ratelimit > x)
863 step = dirty_ratelimit - x;
867 * Don't pursue 100% rate matching. It's impossible since the balanced
868 * rate itself is constantly fluctuating. So decrease the track speed
869 * when it gets close to the target. Helps eliminate pointless tremors.
871 step >>= dirty_ratelimit / (2 * step + 1);
873 * Limit the tracking speed to avoid overshooting.
875 step = (step + 7) / 8;
877 if (dirty_ratelimit < balanced_dirty_ratelimit)
878 dirty_ratelimit += step;
880 dirty_ratelimit -= step;
882 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
883 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
885 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
888 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
889 unsigned long thresh,
890 unsigned long bg_thresh,
892 unsigned long bdi_thresh,
893 unsigned long bdi_dirty,
894 unsigned long start_time)
896 unsigned long now = jiffies;
897 unsigned long elapsed = now - bdi->bw_time_stamp;
898 unsigned long dirtied;
899 unsigned long written;
902 * rate-limit, only update once every 200ms.
904 if (elapsed < BANDWIDTH_INTERVAL)
907 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
908 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
911 * Skip quiet periods when disk bandwidth is under-utilized.
912 * (at least 1s idle time between two flusher runs)
914 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
918 global_update_bandwidth(thresh, dirty, now);
919 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
920 bdi_thresh, bdi_dirty,
923 bdi_update_write_bandwidth(bdi, elapsed, written);
926 bdi->dirtied_stamp = dirtied;
927 bdi->written_stamp = written;
928 bdi->bw_time_stamp = now;
931 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
932 unsigned long thresh,
933 unsigned long bg_thresh,
935 unsigned long bdi_thresh,
936 unsigned long bdi_dirty,
937 unsigned long start_time)
939 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
941 spin_lock(&bdi->wb.list_lock);
942 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
943 bdi_thresh, bdi_dirty, start_time);
944 spin_unlock(&bdi->wb.list_lock);
948 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
949 * will look to see if it needs to start dirty throttling.
951 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
952 * global_page_state() too often. So scale it near-sqrt to the safety margin
953 * (the number of pages we may dirty without exceeding the dirty limits).
955 static unsigned long dirty_poll_interval(unsigned long dirty,
956 unsigned long thresh)
959 return 1UL << (ilog2(thresh - dirty) >> 1);
964 static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
965 unsigned long bdi_dirty)
967 unsigned long bw = bdi->avg_write_bandwidth;
968 unsigned long hi = ilog2(bw);
969 unsigned long lo = ilog2(bdi->dirty_ratelimit);
972 /* target for 20ms max pause on 1-dd case */
976 * Scale up pause time for concurrent dirtiers in order to reduce CPU
979 * (N * 20ms) on 2^N concurrent tasks.
982 t += (hi - lo) * (20 * HZ) / 1024;
985 * Limit pause time for small memory systems. If sleeping for too long
986 * time, a small pool of dirty/writeback pages may go empty and disk go
989 * 8 serves as the safety ratio.
991 t = min(t, bdi_dirty * HZ / (8 * bw + 1));
994 * The pause time will be settled within range (max_pause/4, max_pause).
995 * Apply a minimal value of 4 to get a non-zero max_pause/4.
997 return clamp_val(t, 4, MAX_PAUSE);
1001 * balance_dirty_pages() must be called by processes which are generating dirty
1002 * data. It looks at the number of dirty pages in the machine and will force
1003 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1004 * If we're over `background_thresh' then the writeback threads are woken to
1005 * perform some writeout.
1007 static void balance_dirty_pages(struct address_space *mapping,
1008 unsigned long pages_dirtied)
1010 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1011 unsigned long bdi_reclaimable;
1012 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1013 unsigned long bdi_dirty;
1014 unsigned long freerun;
1015 unsigned long background_thresh;
1016 unsigned long dirty_thresh;
1017 unsigned long bdi_thresh;
1019 long uninitialized_var(max_pause);
1020 bool dirty_exceeded = false;
1021 unsigned long task_ratelimit;
1022 unsigned long uninitialized_var(dirty_ratelimit);
1023 unsigned long pos_ratio;
1024 struct backing_dev_info *bdi = mapping->backing_dev_info;
1025 unsigned long start_time = jiffies;
1029 * Unstable writes are a feature of certain networked
1030 * filesystems (i.e. NFS) in which data may have been
1031 * written to the server's write cache, but has not yet
1032 * been flushed to permanent storage.
1034 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1035 global_page_state(NR_UNSTABLE_NFS);
1036 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1038 global_dirty_limits(&background_thresh, &dirty_thresh);
1041 * Throttle it only when the background writeback cannot
1042 * catch-up. This avoids (excessively) small writeouts
1043 * when the bdi limits are ramping up.
1045 freerun = dirty_freerun_ceiling(dirty_thresh,
1047 if (nr_dirty <= freerun)
1050 if (unlikely(!writeback_in_progress(bdi)))
1051 bdi_start_background_writeback(bdi);
1054 * bdi_thresh is not treated as some limiting factor as
1055 * dirty_thresh, due to reasons
1056 * - in JBOD setup, bdi_thresh can fluctuate a lot
1057 * - in a system with HDD and USB key, the USB key may somehow
1058 * go into state (bdi_dirty >> bdi_thresh) either because
1059 * bdi_dirty starts high, or because bdi_thresh drops low.
1060 * In this case we don't want to hard throttle the USB key
1061 * dirtiers for 100 seconds until bdi_dirty drops under
1062 * bdi_thresh. Instead the auxiliary bdi control line in
1063 * bdi_position_ratio() will let the dirtier task progress
1064 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1066 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1069 * In order to avoid the stacked BDI deadlock we need
1070 * to ensure we accurately count the 'dirty' pages when
1071 * the threshold is low.
1073 * Otherwise it would be possible to get thresh+n pages
1074 * reported dirty, even though there are thresh-m pages
1075 * actually dirty; with m+n sitting in the percpu
1078 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1079 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1080 bdi_dirty = bdi_reclaimable +
1081 bdi_stat_sum(bdi, BDI_WRITEBACK);
1083 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1084 bdi_dirty = bdi_reclaimable +
1085 bdi_stat(bdi, BDI_WRITEBACK);
1088 dirty_exceeded = (bdi_dirty > bdi_thresh) ||
1089 (nr_dirty > dirty_thresh);
1090 if (dirty_exceeded && !bdi->dirty_exceeded)
1091 bdi->dirty_exceeded = 1;
1093 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1094 nr_dirty, bdi_thresh, bdi_dirty,
1097 max_pause = bdi_max_pause(bdi, bdi_dirty);
1099 dirty_ratelimit = bdi->dirty_ratelimit;
1100 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1101 background_thresh, nr_dirty,
1102 bdi_thresh, bdi_dirty);
1103 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1104 RATELIMIT_CALC_SHIFT;
1105 if (unlikely(task_ratelimit == 0)) {
1109 pause = HZ * pages_dirtied / task_ratelimit;
1110 if (unlikely(pause <= 0)) {
1111 trace_balance_dirty_pages(bdi,
1122 pause = 1; /* avoid resetting nr_dirtied_pause below */
1125 pause = min(pause, max_pause);
1128 trace_balance_dirty_pages(bdi,
1139 __set_current_state(TASK_KILLABLE);
1140 io_schedule_timeout(pause);
1143 * This is typically equal to (nr_dirty < dirty_thresh) and can
1144 * also keep "1000+ dd on a slow USB stick" under control.
1150 * In the case of an unresponding NFS server and the NFS dirty
1151 * pages exceeds dirty_thresh, give the other good bdi's a pipe
1152 * to go through, so that tasks on them still remain responsive.
1154 * In theory 1 page is enough to keep the comsumer-producer
1155 * pipe going: the flusher cleans 1 page => the task dirties 1
1156 * more page. However bdi_dirty has accounting errors. So use
1157 * the larger and more IO friendly bdi_stat_error.
1159 if (bdi_dirty <= bdi_stat_error(bdi))
1162 if (fatal_signal_pending(current))
1166 if (!dirty_exceeded && bdi->dirty_exceeded)
1167 bdi->dirty_exceeded = 0;
1169 current->nr_dirtied = 0;
1170 if (pause == 0) { /* in freerun area */
1171 current->nr_dirtied_pause =
1172 dirty_poll_interval(nr_dirty, dirty_thresh);
1173 } else if (pause <= max_pause / 4 &&
1174 pages_dirtied >= current->nr_dirtied_pause) {
1175 current->nr_dirtied_pause = clamp_val(
1176 dirty_ratelimit * (max_pause / 2) / HZ,
1177 pages_dirtied + pages_dirtied / 8,
1179 } else if (pause >= max_pause) {
1180 current->nr_dirtied_pause = 1 | clamp_val(
1181 dirty_ratelimit * (max_pause / 2) / HZ,
1183 pages_dirtied - pages_dirtied / 8);
1186 if (writeback_in_progress(bdi))
1190 * In laptop mode, we wait until hitting the higher threshold before
1191 * starting background writeout, and then write out all the way down
1192 * to the lower threshold. So slow writers cause minimal disk activity.
1194 * In normal mode, we start background writeout at the lower
1195 * background_thresh, to keep the amount of dirty memory low.
1200 if (nr_reclaimable > background_thresh)
1201 bdi_start_background_writeback(bdi);
1204 void set_page_dirty_balance(struct page *page, int page_mkwrite)
1206 if (set_page_dirty(page) || page_mkwrite) {
1207 struct address_space *mapping = page_mapping(page);
1210 balance_dirty_pages_ratelimited(mapping);
1214 static DEFINE_PER_CPU(int, bdp_ratelimits);
1217 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
1218 * @mapping: address_space which was dirtied
1219 * @nr_pages_dirtied: number of pages which the caller has just dirtied
1221 * Processes which are dirtying memory should call in here once for each page
1222 * which was newly dirtied. The function will periodically check the system's
1223 * dirty state and will initiate writeback if needed.
1225 * On really big machines, get_writeback_state is expensive, so try to avoid
1226 * calling it too often (ratelimiting). But once we're over the dirty memory
1227 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1228 * from overshooting the limit by (ratelimit_pages) each.
1230 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1231 unsigned long nr_pages_dirtied)
1233 struct backing_dev_info *bdi = mapping->backing_dev_info;
1237 if (!bdi_cap_account_dirty(bdi))
1240 ratelimit = current->nr_dirtied_pause;
1241 if (bdi->dirty_exceeded)
1242 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1244 current->nr_dirtied += nr_pages_dirtied;
1248 * This prevents one CPU to accumulate too many dirtied pages without
1249 * calling into balance_dirty_pages(), which can happen when there are
1250 * 1000+ tasks, all of them start dirtying pages at exactly the same
1251 * time, hence all honoured too large initial task->nr_dirtied_pause.
1253 p = &__get_cpu_var(bdp_ratelimits);
1254 if (unlikely(current->nr_dirtied >= ratelimit))
1257 *p += nr_pages_dirtied;
1258 if (unlikely(*p >= ratelimit_pages)) {
1265 if (unlikely(current->nr_dirtied >= ratelimit))
1266 balance_dirty_pages(mapping, current->nr_dirtied);
1268 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
1270 void throttle_vm_writeout(gfp_t gfp_mask)
1272 unsigned long background_thresh;
1273 unsigned long dirty_thresh;
1276 global_dirty_limits(&background_thresh, &dirty_thresh);
1279 * Boost the allowable dirty threshold a bit for page
1280 * allocators so they don't get DoS'ed by heavy writers
1282 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1284 if (global_page_state(NR_UNSTABLE_NFS) +
1285 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1287 congestion_wait(BLK_RW_ASYNC, HZ/10);
1290 * The caller might hold locks which can prevent IO completion
1291 * or progress in the filesystem. So we cannot just sit here
1292 * waiting for IO to complete.
1294 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1300 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1302 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1303 void __user *buffer, size_t *length, loff_t *ppos)
1305 proc_dointvec(table, write, buffer, length, ppos);
1306 bdi_arm_supers_timer();
1311 void laptop_mode_timer_fn(unsigned long data)
1313 struct request_queue *q = (struct request_queue *)data;
1314 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1315 global_page_state(NR_UNSTABLE_NFS);
1318 * We want to write everything out, not just down to the dirty
1321 if (bdi_has_dirty_io(&q->backing_dev_info))
1322 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1323 WB_REASON_LAPTOP_TIMER);
1327 * We've spun up the disk and we're in laptop mode: schedule writeback
1328 * of all dirty data a few seconds from now. If the flush is already scheduled
1329 * then push it back - the user is still using the disk.
1331 void laptop_io_completion(struct backing_dev_info *info)
1333 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1337 * We're in laptop mode and we've just synced. The sync's writes will have
1338 * caused another writeback to be scheduled by laptop_io_completion.
1339 * Nothing needs to be written back anymore, so we unschedule the writeback.
1341 void laptop_sync_completion(void)
1343 struct backing_dev_info *bdi;
1347 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1348 del_timer(&bdi->laptop_mode_wb_timer);
1355 * If ratelimit_pages is too high then we can get into dirty-data overload
1356 * if a large number of processes all perform writes at the same time.
1357 * If it is too low then SMP machines will call the (expensive)
1358 * get_writeback_state too often.
1360 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1361 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1365 void writeback_set_ratelimit(void)
1367 unsigned long background_thresh;
1368 unsigned long dirty_thresh;
1369 global_dirty_limits(&background_thresh, &dirty_thresh);
1370 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1371 if (ratelimit_pages < 16)
1372 ratelimit_pages = 16;
1375 static int __cpuinit
1376 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1378 writeback_set_ratelimit();
1382 static struct notifier_block __cpuinitdata ratelimit_nb = {
1383 .notifier_call = ratelimit_handler,
1388 * Called early on to tune the page writeback dirty limits.
1390 * We used to scale dirty pages according to how total memory
1391 * related to pages that could be allocated for buffers (by
1392 * comparing nr_free_buffer_pages() to vm_total_pages.
1394 * However, that was when we used "dirty_ratio" to scale with
1395 * all memory, and we don't do that any more. "dirty_ratio"
1396 * is now applied to total non-HIGHPAGE memory (by subtracting
1397 * totalhigh_pages from vm_total_pages), and as such we can't
1398 * get into the old insane situation any more where we had
1399 * large amounts of dirty pages compared to a small amount of
1400 * non-HIGHMEM memory.
1402 * But we might still want to scale the dirty_ratio by how
1403 * much memory the box has..
1405 void __init page_writeback_init(void)
1409 writeback_set_ratelimit();
1410 register_cpu_notifier(&ratelimit_nb);
1412 shift = calc_period_shift();
1413 prop_descriptor_init(&vm_completions, shift);
1417 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1418 * @mapping: address space structure to write
1419 * @start: starting page index
1420 * @end: ending page index (inclusive)
1422 * This function scans the page range from @start to @end (inclusive) and tags
1423 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1424 * that write_cache_pages (or whoever calls this function) will then use
1425 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1426 * used to avoid livelocking of writeback by a process steadily creating new
1427 * dirty pages in the file (thus it is important for this function to be quick
1428 * so that it can tag pages faster than a dirtying process can create them).
1431 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1433 void tag_pages_for_writeback(struct address_space *mapping,
1434 pgoff_t start, pgoff_t end)
1436 #define WRITEBACK_TAG_BATCH 4096
1437 unsigned long tagged;
1440 spin_lock_irq(&mapping->tree_lock);
1441 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1442 &start, end, WRITEBACK_TAG_BATCH,
1443 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1444 spin_unlock_irq(&mapping->tree_lock);
1445 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1447 /* We check 'start' to handle wrapping when end == ~0UL */
1448 } while (tagged >= WRITEBACK_TAG_BATCH && start);
1450 EXPORT_SYMBOL(tag_pages_for_writeback);
1453 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1454 * @mapping: address space structure to write
1455 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1456 * @writepage: function called for each page
1457 * @data: data passed to writepage function
1459 * If a page is already under I/O, write_cache_pages() skips it, even
1460 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1461 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1462 * and msync() need to guarantee that all the data which was dirty at the time
1463 * the call was made get new I/O started against them. If wbc->sync_mode is
1464 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1465 * existing IO to complete.
1467 * To avoid livelocks (when other process dirties new pages), we first tag
1468 * pages which should be written back with TOWRITE tag and only then start
1469 * writing them. For data-integrity sync we have to be careful so that we do
1470 * not miss some pages (e.g., because some other process has cleared TOWRITE
1471 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1472 * by the process clearing the DIRTY tag (and submitting the page for IO).
1474 int write_cache_pages(struct address_space *mapping,
1475 struct writeback_control *wbc, writepage_t writepage,
1480 struct pagevec pvec;
1482 pgoff_t uninitialized_var(writeback_index);
1484 pgoff_t end; /* Inclusive */
1487 int range_whole = 0;
1490 pagevec_init(&pvec, 0);
1491 if (wbc->range_cyclic) {
1492 writeback_index = mapping->writeback_index; /* prev offset */
1493 index = writeback_index;
1500 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1501 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1502 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1504 cycled = 1; /* ignore range_cyclic tests */
1506 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1507 tag = PAGECACHE_TAG_TOWRITE;
1509 tag = PAGECACHE_TAG_DIRTY;
1511 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1512 tag_pages_for_writeback(mapping, index, end);
1514 while (!done && (index <= end)) {
1517 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1518 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1522 for (i = 0; i < nr_pages; i++) {
1523 struct page *page = pvec.pages[i];
1526 * At this point, the page may be truncated or
1527 * invalidated (changing page->mapping to NULL), or
1528 * even swizzled back from swapper_space to tmpfs file
1529 * mapping. However, page->index will not change
1530 * because we have a reference on the page.
1532 if (page->index > end) {
1534 * can't be range_cyclic (1st pass) because
1535 * end == -1 in that case.
1541 done_index = page->index;
1546 * Page truncated or invalidated. We can freely skip it
1547 * then, even for data integrity operations: the page
1548 * has disappeared concurrently, so there could be no
1549 * real expectation of this data interity operation
1550 * even if there is now a new, dirty page at the same
1551 * pagecache address.
1553 if (unlikely(page->mapping != mapping)) {
1559 if (!PageDirty(page)) {
1560 /* someone wrote it for us */
1561 goto continue_unlock;
1564 if (PageWriteback(page)) {
1565 if (wbc->sync_mode != WB_SYNC_NONE)
1566 wait_on_page_writeback(page);
1568 goto continue_unlock;
1571 BUG_ON(PageWriteback(page));
1572 if (!clear_page_dirty_for_io(page))
1573 goto continue_unlock;
1575 trace_wbc_writepage(wbc, mapping->backing_dev_info);
1576 ret = (*writepage)(page, wbc, data);
1577 if (unlikely(ret)) {
1578 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1583 * done_index is set past this page,
1584 * so media errors will not choke
1585 * background writeout for the entire
1586 * file. This has consequences for
1587 * range_cyclic semantics (ie. it may
1588 * not be suitable for data integrity
1591 done_index = page->index + 1;
1598 * We stop writing back only if we are not doing
1599 * integrity sync. In case of integrity sync we have to
1600 * keep going until we have written all the pages
1601 * we tagged for writeback prior to entering this loop.
1603 if (--wbc->nr_to_write <= 0 &&
1604 wbc->sync_mode == WB_SYNC_NONE) {
1609 pagevec_release(&pvec);
1612 if (!cycled && !done) {
1615 * We hit the last page and there is more work to be done: wrap
1616 * back to the start of the file
1620 end = writeback_index - 1;
1623 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1624 mapping->writeback_index = done_index;
1628 EXPORT_SYMBOL(write_cache_pages);
1631 * Function used by generic_writepages to call the real writepage
1632 * function and set the mapping flags on error
1634 static int __writepage(struct page *page, struct writeback_control *wbc,
1637 struct address_space *mapping = data;
1638 int ret = mapping->a_ops->writepage(page, wbc);
1639 mapping_set_error(mapping, ret);
1644 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1645 * @mapping: address space structure to write
1646 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1648 * This is a library function, which implements the writepages()
1649 * address_space_operation.
1651 int generic_writepages(struct address_space *mapping,
1652 struct writeback_control *wbc)
1654 struct blk_plug plug;
1657 /* deal with chardevs and other special file */
1658 if (!mapping->a_ops->writepage)
1661 blk_start_plug(&plug);
1662 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1663 blk_finish_plug(&plug);
1667 EXPORT_SYMBOL(generic_writepages);
1669 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1673 if (wbc->nr_to_write <= 0)
1675 if (mapping->a_ops->writepages)
1676 ret = mapping->a_ops->writepages(mapping, wbc);
1678 ret = generic_writepages(mapping, wbc);
1683 * write_one_page - write out a single page and optionally wait on I/O
1684 * @page: the page to write
1685 * @wait: if true, wait on writeout
1687 * The page must be locked by the caller and will be unlocked upon return.
1689 * write_one_page() returns a negative error code if I/O failed.
1691 int write_one_page(struct page *page, int wait)
1693 struct address_space *mapping = page->mapping;
1695 struct writeback_control wbc = {
1696 .sync_mode = WB_SYNC_ALL,
1700 BUG_ON(!PageLocked(page));
1703 wait_on_page_writeback(page);
1705 if (clear_page_dirty_for_io(page)) {
1706 page_cache_get(page);
1707 ret = mapping->a_ops->writepage(page, &wbc);
1708 if (ret == 0 && wait) {
1709 wait_on_page_writeback(page);
1710 if (PageError(page))
1713 page_cache_release(page);
1719 EXPORT_SYMBOL(write_one_page);
1722 * For address_spaces which do not use buffers nor write back.
1724 int __set_page_dirty_no_writeback(struct page *page)
1726 if (!PageDirty(page))
1727 return !TestSetPageDirty(page);
1732 * Helper function for set_page_dirty family.
1733 * NOTE: This relies on being atomic wrt interrupts.
1735 void account_page_dirtied(struct page *page, struct address_space *mapping)
1737 if (mapping_cap_account_dirty(mapping)) {
1738 __inc_zone_page_state(page, NR_FILE_DIRTY);
1739 __inc_zone_page_state(page, NR_DIRTIED);
1740 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1741 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1742 task_io_account_write(PAGE_CACHE_SIZE);
1745 EXPORT_SYMBOL(account_page_dirtied);
1748 * Helper function for set_page_writeback family.
1749 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1752 void account_page_writeback(struct page *page)
1754 inc_zone_page_state(page, NR_WRITEBACK);
1756 EXPORT_SYMBOL(account_page_writeback);
1759 * For address_spaces which do not use buffers. Just tag the page as dirty in
1762 * This is also used when a single buffer is being dirtied: we want to set the
1763 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1764 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1766 * Most callers have locked the page, which pins the address_space in memory.
1767 * But zap_pte_range() does not lock the page, however in that case the
1768 * mapping is pinned by the vma's ->vm_file reference.
1770 * We take care to handle the case where the page was truncated from the
1771 * mapping by re-checking page_mapping() inside tree_lock.
1773 int __set_page_dirty_nobuffers(struct page *page)
1775 if (!TestSetPageDirty(page)) {
1776 struct address_space *mapping = page_mapping(page);
1777 struct address_space *mapping2;
1778 unsigned long flags;
1783 spin_lock_irqsave(&mapping->tree_lock, flags);
1784 mapping2 = page_mapping(page);
1785 if (mapping2) { /* Race with truncate? */
1786 BUG_ON(mapping2 != mapping);
1787 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1788 account_page_dirtied(page, mapping);
1789 radix_tree_tag_set(&mapping->page_tree,
1790 page_index(page), PAGECACHE_TAG_DIRTY);
1792 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1793 if (mapping->host) {
1794 /* !PageAnon && !swapper_space */
1795 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1801 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1804 * Call this whenever redirtying a page, to de-account the dirty counters
1805 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
1806 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
1807 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
1810 void account_page_redirty(struct page *page)
1812 struct address_space *mapping = page->mapping;
1813 if (mapping && mapping_cap_account_dirty(mapping)) {
1814 current->nr_dirtied--;
1815 dec_zone_page_state(page, NR_DIRTIED);
1816 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1819 EXPORT_SYMBOL(account_page_redirty);
1822 * When a writepage implementation decides that it doesn't want to write this
1823 * page for some reason, it should redirty the locked page via
1824 * redirty_page_for_writepage() and it should then unlock the page and return 0
1826 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1828 wbc->pages_skipped++;
1829 account_page_redirty(page);
1830 return __set_page_dirty_nobuffers(page);
1832 EXPORT_SYMBOL(redirty_page_for_writepage);
1837 * For pages with a mapping this should be done under the page lock
1838 * for the benefit of asynchronous memory errors who prefer a consistent
1839 * dirty state. This rule can be broken in some special cases,
1840 * but should be better not to.
1842 * If the mapping doesn't provide a set_page_dirty a_op, then
1843 * just fall through and assume that it wants buffer_heads.
1845 int set_page_dirty(struct page *page)
1847 struct address_space *mapping = page_mapping(page);
1849 if (likely(mapping)) {
1850 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1852 * readahead/lru_deactivate_page could remain
1853 * PG_readahead/PG_reclaim due to race with end_page_writeback
1854 * About readahead, if the page is written, the flags would be
1855 * reset. So no problem.
1856 * About lru_deactivate_page, if the page is redirty, the flag
1857 * will be reset. So no problem. but if the page is used by readahead
1858 * it will confuse readahead and make it restart the size rampup
1859 * process. But it's a trivial problem.
1861 ClearPageReclaim(page);
1864 spd = __set_page_dirty_buffers;
1866 return (*spd)(page);
1868 if (!PageDirty(page)) {
1869 if (!TestSetPageDirty(page))
1874 EXPORT_SYMBOL(set_page_dirty);
1877 * set_page_dirty() is racy if the caller has no reference against
1878 * page->mapping->host, and if the page is unlocked. This is because another
1879 * CPU could truncate the page off the mapping and then free the mapping.
1881 * Usually, the page _is_ locked, or the caller is a user-space process which
1882 * holds a reference on the inode by having an open file.
1884 * In other cases, the page should be locked before running set_page_dirty().
1886 int set_page_dirty_lock(struct page *page)
1891 ret = set_page_dirty(page);
1895 EXPORT_SYMBOL(set_page_dirty_lock);
1898 * Clear a page's dirty flag, while caring for dirty memory accounting.
1899 * Returns true if the page was previously dirty.
1901 * This is for preparing to put the page under writeout. We leave the page
1902 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1903 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1904 * implementation will run either set_page_writeback() or set_page_dirty(),
1905 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1908 * This incoherency between the page's dirty flag and radix-tree tag is
1909 * unfortunate, but it only exists while the page is locked.
1911 int clear_page_dirty_for_io(struct page *page)
1913 struct address_space *mapping = page_mapping(page);
1915 BUG_ON(!PageLocked(page));
1917 if (mapping && mapping_cap_account_dirty(mapping)) {
1919 * Yes, Virginia, this is indeed insane.
1921 * We use this sequence to make sure that
1922 * (a) we account for dirty stats properly
1923 * (b) we tell the low-level filesystem to
1924 * mark the whole page dirty if it was
1925 * dirty in a pagetable. Only to then
1926 * (c) clean the page again and return 1 to
1927 * cause the writeback.
1929 * This way we avoid all nasty races with the
1930 * dirty bit in multiple places and clearing
1931 * them concurrently from different threads.
1933 * Note! Normally the "set_page_dirty(page)"
1934 * has no effect on the actual dirty bit - since
1935 * that will already usually be set. But we
1936 * need the side effects, and it can help us
1939 * We basically use the page "master dirty bit"
1940 * as a serialization point for all the different
1941 * threads doing their things.
1943 if (page_mkclean(page))
1944 set_page_dirty(page);
1946 * We carefully synchronise fault handlers against
1947 * installing a dirty pte and marking the page dirty
1948 * at this point. We do this by having them hold the
1949 * page lock at some point after installing their
1950 * pte, but before marking the page dirty.
1951 * Pages are always locked coming in here, so we get
1952 * the desired exclusion. See mm/memory.c:do_wp_page()
1953 * for more comments.
1955 if (TestClearPageDirty(page)) {
1956 dec_zone_page_state(page, NR_FILE_DIRTY);
1957 dec_bdi_stat(mapping->backing_dev_info,
1963 return TestClearPageDirty(page);
1965 EXPORT_SYMBOL(clear_page_dirty_for_io);
1967 int test_clear_page_writeback(struct page *page)
1969 struct address_space *mapping = page_mapping(page);
1973 struct backing_dev_info *bdi = mapping->backing_dev_info;
1974 unsigned long flags;
1976 spin_lock_irqsave(&mapping->tree_lock, flags);
1977 ret = TestClearPageWriteback(page);
1979 radix_tree_tag_clear(&mapping->page_tree,
1981 PAGECACHE_TAG_WRITEBACK);
1982 if (bdi_cap_account_writeback(bdi)) {
1983 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1984 __bdi_writeout_inc(bdi);
1987 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1989 ret = TestClearPageWriteback(page);
1992 dec_zone_page_state(page, NR_WRITEBACK);
1993 inc_zone_page_state(page, NR_WRITTEN);
1998 int test_set_page_writeback(struct page *page)
2000 struct address_space *mapping = page_mapping(page);
2004 struct backing_dev_info *bdi = mapping->backing_dev_info;
2005 unsigned long flags;
2007 spin_lock_irqsave(&mapping->tree_lock, flags);
2008 ret = TestSetPageWriteback(page);
2010 radix_tree_tag_set(&mapping->page_tree,
2012 PAGECACHE_TAG_WRITEBACK);
2013 if (bdi_cap_account_writeback(bdi))
2014 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2016 if (!PageDirty(page))
2017 radix_tree_tag_clear(&mapping->page_tree,
2019 PAGECACHE_TAG_DIRTY);
2020 radix_tree_tag_clear(&mapping->page_tree,
2022 PAGECACHE_TAG_TOWRITE);
2023 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2025 ret = TestSetPageWriteback(page);
2028 account_page_writeback(page);
2032 EXPORT_SYMBOL(test_set_page_writeback);
2035 * Return true if any of the pages in the mapping are marked with the
2038 int mapping_tagged(struct address_space *mapping, int tag)
2040 return radix_tree_tagged(&mapping->page_tree, tag);
2042 EXPORT_SYMBOL(mapping_tagged);