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 = div64_s64(((s64)setpoint - (s64)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 = div64_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 * @written may have decreased due to account_page_redirty().
665 * Avoid underflowing @bw calculation.
667 bw = written - min(written, bdi->written_stamp);
669 if (unlikely(elapsed > period)) {
674 bw += (u64)bdi->write_bandwidth * (period - elapsed);
675 bw >>= ilog2(period);
678 * one more level of smoothing, for filtering out sudden spikes
680 if (avg > old && old >= (unsigned long)bw)
681 avg -= (avg - old) >> 3;
683 if (avg < old && old <= (unsigned long)bw)
684 avg += (old - avg) >> 3;
687 bdi->write_bandwidth = bw;
688 bdi->avg_write_bandwidth = avg;
692 * The global dirtyable memory and dirty threshold could be suddenly knocked
693 * down by a large amount (eg. on the startup of KVM in a swapless system).
694 * This may throw the system into deep dirty exceeded state and throttle
695 * heavy/light dirtiers alike. To retain good responsiveness, maintain
696 * global_dirty_limit for tracking slowly down to the knocked down dirty
699 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
701 unsigned long limit = global_dirty_limit;
704 * Follow up in one step.
706 if (limit < thresh) {
712 * Follow down slowly. Use the higher one as the target, because thresh
713 * may drop below dirty. This is exactly the reason to introduce
714 * global_dirty_limit which is guaranteed to lie above the dirty pages.
716 thresh = max(thresh, dirty);
717 if (limit > thresh) {
718 limit -= (limit - thresh) >> 5;
723 global_dirty_limit = limit;
726 static void global_update_bandwidth(unsigned long thresh,
730 static DEFINE_SPINLOCK(dirty_lock);
731 static unsigned long update_time = INITIAL_JIFFIES;
734 * check locklessly first to optimize away locking for the most time
736 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
739 spin_lock(&dirty_lock);
740 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
741 update_dirty_limit(thresh, dirty);
744 spin_unlock(&dirty_lock);
748 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
750 * Normal bdi tasks will be curbed at or below it in long term.
751 * Obviously it should be around (write_bw / N) when there are N dd tasks.
753 static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
754 unsigned long thresh,
755 unsigned long bg_thresh,
757 unsigned long bdi_thresh,
758 unsigned long bdi_dirty,
759 unsigned long dirtied,
760 unsigned long elapsed)
762 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
763 unsigned long limit = hard_dirty_limit(thresh);
764 unsigned long setpoint = (freerun + limit) / 2;
765 unsigned long write_bw = bdi->avg_write_bandwidth;
766 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
767 unsigned long dirty_rate;
768 unsigned long task_ratelimit;
769 unsigned long balanced_dirty_ratelimit;
770 unsigned long pos_ratio;
775 * The dirty rate will match the writeout rate in long term, except
776 * when dirty pages are truncated by userspace or re-dirtied by FS.
778 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
780 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
781 bdi_thresh, bdi_dirty);
783 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
785 task_ratelimit = (u64)dirty_ratelimit *
786 pos_ratio >> RATELIMIT_CALC_SHIFT;
787 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
790 * A linear estimation of the "balanced" throttle rate. The theory is,
791 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
792 * dirty_rate will be measured to be (N * task_ratelimit). So the below
793 * formula will yield the balanced rate limit (write_bw / N).
795 * Note that the expanded form is not a pure rate feedback:
796 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
797 * but also takes pos_ratio into account:
798 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
800 * (1) is not realistic because pos_ratio also takes part in balancing
801 * the dirty rate. Consider the state
802 * pos_ratio = 0.5 (3)
803 * rate = 2 * (write_bw / N) (4)
804 * If (1) is used, it will stuck in that state! Because each dd will
806 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
808 * dirty_rate = N * task_ratelimit = write_bw (6)
809 * put (6) into (1) we get
810 * rate_(i+1) = rate_(i) (7)
812 * So we end up using (2) to always keep
813 * rate_(i+1) ~= (write_bw / N) (8)
814 * regardless of the value of pos_ratio. As long as (8) is satisfied,
815 * pos_ratio is able to drive itself to 1.0, which is not only where
816 * the dirty count meet the setpoint, but also where the slope of
817 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
819 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
823 * We could safely do this and return immediately:
825 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
827 * However to get a more stable dirty_ratelimit, the below elaborated
828 * code makes use of task_ratelimit to filter out sigular points and
829 * limit the step size.
831 * The below code essentially only uses the relative value of
833 * task_ratelimit - dirty_ratelimit
834 * = (pos_ratio - 1) * dirty_ratelimit
836 * which reflects the direction and size of dirty position error.
840 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
841 * task_ratelimit is on the same side of dirty_ratelimit, too.
843 * - dirty_ratelimit > balanced_dirty_ratelimit
844 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
845 * lowering dirty_ratelimit will help meet both the position and rate
846 * control targets. Otherwise, don't update dirty_ratelimit if it will
847 * only help meet the rate target. After all, what the users ultimately
848 * feel and care are stable dirty rate and small position error.
850 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
851 * and filter out the sigular points of balanced_dirty_ratelimit. Which
852 * keeps jumping around randomly and can even leap far away at times
853 * due to the small 200ms estimation period of dirty_rate (we want to
854 * keep that period small to reduce time lags).
857 if (dirty < setpoint) {
858 x = min(bdi->balanced_dirty_ratelimit,
859 min(balanced_dirty_ratelimit, task_ratelimit));
860 if (dirty_ratelimit < x)
861 step = x - dirty_ratelimit;
863 x = max(bdi->balanced_dirty_ratelimit,
864 max(balanced_dirty_ratelimit, task_ratelimit));
865 if (dirty_ratelimit > x)
866 step = dirty_ratelimit - x;
870 * Don't pursue 100% rate matching. It's impossible since the balanced
871 * rate itself is constantly fluctuating. So decrease the track speed
872 * when it gets close to the target. Helps eliminate pointless tremors.
874 step >>= dirty_ratelimit / (2 * step + 1);
876 * Limit the tracking speed to avoid overshooting.
878 step = (step + 7) / 8;
880 if (dirty_ratelimit < balanced_dirty_ratelimit)
881 dirty_ratelimit += step;
883 dirty_ratelimit -= step;
885 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
886 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
888 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
891 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
892 unsigned long thresh,
893 unsigned long bg_thresh,
895 unsigned long bdi_thresh,
896 unsigned long bdi_dirty,
897 unsigned long start_time)
899 unsigned long now = jiffies;
900 unsigned long elapsed = now - bdi->bw_time_stamp;
901 unsigned long dirtied;
902 unsigned long written;
905 * rate-limit, only update once every 200ms.
907 if (elapsed < BANDWIDTH_INTERVAL)
910 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
911 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
914 * Skip quiet periods when disk bandwidth is under-utilized.
915 * (at least 1s idle time between two flusher runs)
917 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
921 global_update_bandwidth(thresh, dirty, now);
922 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
923 bdi_thresh, bdi_dirty,
926 bdi_update_write_bandwidth(bdi, elapsed, written);
929 bdi->dirtied_stamp = dirtied;
930 bdi->written_stamp = written;
931 bdi->bw_time_stamp = now;
934 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
935 unsigned long thresh,
936 unsigned long bg_thresh,
938 unsigned long bdi_thresh,
939 unsigned long bdi_dirty,
940 unsigned long start_time)
942 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
944 spin_lock(&bdi->wb.list_lock);
945 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
946 bdi_thresh, bdi_dirty, start_time);
947 spin_unlock(&bdi->wb.list_lock);
951 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
952 * will look to see if it needs to start dirty throttling.
954 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
955 * global_page_state() too often. So scale it near-sqrt to the safety margin
956 * (the number of pages we may dirty without exceeding the dirty limits).
958 static unsigned long dirty_poll_interval(unsigned long dirty,
959 unsigned long thresh)
962 return 1UL << (ilog2(thresh - dirty) >> 1);
967 static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
968 unsigned long bdi_dirty)
970 unsigned long bw = bdi->avg_write_bandwidth;
971 unsigned long hi = ilog2(bw);
972 unsigned long lo = ilog2(bdi->dirty_ratelimit);
975 /* target for 20ms max pause on 1-dd case */
979 * Scale up pause time for concurrent dirtiers in order to reduce CPU
982 * (N * 20ms) on 2^N concurrent tasks.
985 t += (hi - lo) * (20 * HZ) / 1024;
988 * Limit pause time for small memory systems. If sleeping for too long
989 * time, a small pool of dirty/writeback pages may go empty and disk go
992 * 8 serves as the safety ratio.
994 t = min(t, bdi_dirty * HZ / (8 * bw + 1));
997 * The pause time will be settled within range (max_pause/4, max_pause).
998 * Apply a minimal value of 4 to get a non-zero max_pause/4.
1000 return clamp_val(t, 4, MAX_PAUSE);
1004 * balance_dirty_pages() must be called by processes which are generating dirty
1005 * data. It looks at the number of dirty pages in the machine and will force
1006 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1007 * If we're over `background_thresh' then the writeback threads are woken to
1008 * perform some writeout.
1010 static void balance_dirty_pages(struct address_space *mapping,
1011 unsigned long pages_dirtied)
1013 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1014 unsigned long bdi_reclaimable;
1015 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1016 unsigned long bdi_dirty;
1017 unsigned long freerun;
1018 unsigned long background_thresh;
1019 unsigned long dirty_thresh;
1020 unsigned long bdi_thresh;
1022 long uninitialized_var(max_pause);
1023 bool dirty_exceeded = false;
1024 unsigned long task_ratelimit;
1025 unsigned long uninitialized_var(dirty_ratelimit);
1026 unsigned long pos_ratio;
1027 struct backing_dev_info *bdi = mapping->backing_dev_info;
1028 unsigned long start_time = jiffies;
1032 * Unstable writes are a feature of certain networked
1033 * filesystems (i.e. NFS) in which data may have been
1034 * written to the server's write cache, but has not yet
1035 * been flushed to permanent storage.
1037 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1038 global_page_state(NR_UNSTABLE_NFS);
1039 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1041 global_dirty_limits(&background_thresh, &dirty_thresh);
1044 * Throttle it only when the background writeback cannot
1045 * catch-up. This avoids (excessively) small writeouts
1046 * when the bdi limits are ramping up.
1048 freerun = dirty_freerun_ceiling(dirty_thresh,
1050 if (nr_dirty <= freerun)
1053 if (unlikely(!writeback_in_progress(bdi)))
1054 bdi_start_background_writeback(bdi);
1057 * bdi_thresh is not treated as some limiting factor as
1058 * dirty_thresh, due to reasons
1059 * - in JBOD setup, bdi_thresh can fluctuate a lot
1060 * - in a system with HDD and USB key, the USB key may somehow
1061 * go into state (bdi_dirty >> bdi_thresh) either because
1062 * bdi_dirty starts high, or because bdi_thresh drops low.
1063 * In this case we don't want to hard throttle the USB key
1064 * dirtiers for 100 seconds until bdi_dirty drops under
1065 * bdi_thresh. Instead the auxiliary bdi control line in
1066 * bdi_position_ratio() will let the dirtier task progress
1067 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1069 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1072 * In order to avoid the stacked BDI deadlock we need
1073 * to ensure we accurately count the 'dirty' pages when
1074 * the threshold is low.
1076 * Otherwise it would be possible to get thresh+n pages
1077 * reported dirty, even though there are thresh-m pages
1078 * actually dirty; with m+n sitting in the percpu
1081 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1082 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1083 bdi_dirty = bdi_reclaimable +
1084 bdi_stat_sum(bdi, BDI_WRITEBACK);
1086 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1087 bdi_dirty = bdi_reclaimable +
1088 bdi_stat(bdi, BDI_WRITEBACK);
1091 dirty_exceeded = (bdi_dirty > bdi_thresh) ||
1092 (nr_dirty > dirty_thresh);
1093 if (dirty_exceeded && !bdi->dirty_exceeded)
1094 bdi->dirty_exceeded = 1;
1096 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1097 nr_dirty, bdi_thresh, bdi_dirty,
1100 max_pause = bdi_max_pause(bdi, bdi_dirty);
1102 dirty_ratelimit = bdi->dirty_ratelimit;
1103 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1104 background_thresh, nr_dirty,
1105 bdi_thresh, bdi_dirty);
1106 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1107 RATELIMIT_CALC_SHIFT;
1108 if (unlikely(task_ratelimit == 0)) {
1112 pause = HZ * pages_dirtied / task_ratelimit;
1113 if (unlikely(pause <= 0)) {
1114 trace_balance_dirty_pages(bdi,
1125 pause = 1; /* avoid resetting nr_dirtied_pause below */
1128 pause = min(pause, max_pause);
1131 trace_balance_dirty_pages(bdi,
1142 __set_current_state(TASK_KILLABLE);
1143 io_schedule_timeout(pause);
1146 * This is typically equal to (nr_dirty < dirty_thresh) and can
1147 * also keep "1000+ dd on a slow USB stick" under control.
1153 * In the case of an unresponding NFS server and the NFS dirty
1154 * pages exceeds dirty_thresh, give the other good bdi's a pipe
1155 * to go through, so that tasks on them still remain responsive.
1157 * In theory 1 page is enough to keep the comsumer-producer
1158 * pipe going: the flusher cleans 1 page => the task dirties 1
1159 * more page. However bdi_dirty has accounting errors. So use
1160 * the larger and more IO friendly bdi_stat_error.
1162 if (bdi_dirty <= bdi_stat_error(bdi))
1165 if (fatal_signal_pending(current))
1169 if (!dirty_exceeded && bdi->dirty_exceeded)
1170 bdi->dirty_exceeded = 0;
1172 current->nr_dirtied = 0;
1173 if (pause == 0) { /* in freerun area */
1174 current->nr_dirtied_pause =
1175 dirty_poll_interval(nr_dirty, dirty_thresh);
1176 } else if (pause <= max_pause / 4 &&
1177 pages_dirtied >= current->nr_dirtied_pause) {
1178 current->nr_dirtied_pause = clamp_val(
1179 dirty_ratelimit * (max_pause / 2) / HZ,
1180 pages_dirtied + pages_dirtied / 8,
1182 } else if (pause >= max_pause) {
1183 current->nr_dirtied_pause = 1 | clamp_val(
1184 dirty_ratelimit * (max_pause / 2) / HZ,
1186 pages_dirtied - pages_dirtied / 8);
1189 if (writeback_in_progress(bdi))
1193 * In laptop mode, we wait until hitting the higher threshold before
1194 * starting background writeout, and then write out all the way down
1195 * to the lower threshold. So slow writers cause minimal disk activity.
1197 * In normal mode, we start background writeout at the lower
1198 * background_thresh, to keep the amount of dirty memory low.
1203 if (nr_reclaimable > background_thresh)
1204 bdi_start_background_writeback(bdi);
1207 static DEFINE_PER_CPU(int, bdp_ratelimits);
1210 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
1211 * @mapping: address_space which was dirtied
1212 * @nr_pages_dirtied: number of pages which the caller has just dirtied
1214 * Processes which are dirtying memory should call in here once for each page
1215 * which was newly dirtied. The function will periodically check the system's
1216 * dirty state and will initiate writeback if needed.
1218 * On really big machines, get_writeback_state is expensive, so try to avoid
1219 * calling it too often (ratelimiting). But once we're over the dirty memory
1220 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1221 * from overshooting the limit by (ratelimit_pages) each.
1223 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1224 unsigned long nr_pages_dirtied)
1226 struct backing_dev_info *bdi = mapping->backing_dev_info;
1230 if (!bdi_cap_account_dirty(bdi))
1233 ratelimit = current->nr_dirtied_pause;
1234 if (bdi->dirty_exceeded)
1235 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1237 current->nr_dirtied += nr_pages_dirtied;
1241 * This prevents one CPU to accumulate too many dirtied pages without
1242 * calling into balance_dirty_pages(), which can happen when there are
1243 * 1000+ tasks, all of them start dirtying pages at exactly the same
1244 * time, hence all honoured too large initial task->nr_dirtied_pause.
1246 p = &__get_cpu_var(bdp_ratelimits);
1247 if (unlikely(current->nr_dirtied >= ratelimit))
1250 *p += nr_pages_dirtied;
1251 if (unlikely(*p >= ratelimit_pages)) {
1258 if (unlikely(current->nr_dirtied >= ratelimit))
1259 balance_dirty_pages(mapping, current->nr_dirtied);
1261 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
1263 void throttle_vm_writeout(gfp_t gfp_mask)
1265 unsigned long background_thresh;
1266 unsigned long dirty_thresh;
1269 global_dirty_limits(&background_thresh, &dirty_thresh);
1272 * Boost the allowable dirty threshold a bit for page
1273 * allocators so they don't get DoS'ed by heavy writers
1275 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1277 if (global_page_state(NR_UNSTABLE_NFS) +
1278 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1280 congestion_wait(BLK_RW_ASYNC, HZ/10);
1283 * The caller might hold locks which can prevent IO completion
1284 * or progress in the filesystem. So we cannot just sit here
1285 * waiting for IO to complete.
1287 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1293 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1295 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1296 void __user *buffer, size_t *length, loff_t *ppos)
1298 proc_dointvec(table, write, buffer, length, ppos);
1299 bdi_arm_supers_timer();
1304 void laptop_mode_timer_fn(unsigned long data)
1306 struct request_queue *q = (struct request_queue *)data;
1307 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1308 global_page_state(NR_UNSTABLE_NFS);
1311 * We want to write everything out, not just down to the dirty
1314 if (bdi_has_dirty_io(&q->backing_dev_info))
1315 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1316 WB_REASON_LAPTOP_TIMER);
1320 * We've spun up the disk and we're in laptop mode: schedule writeback
1321 * of all dirty data a few seconds from now. If the flush is already scheduled
1322 * then push it back - the user is still using the disk.
1324 void laptop_io_completion(struct backing_dev_info *info)
1326 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1330 * We're in laptop mode and we've just synced. The sync's writes will have
1331 * caused another writeback to be scheduled by laptop_io_completion.
1332 * Nothing needs to be written back anymore, so we unschedule the writeback.
1334 void laptop_sync_completion(void)
1336 struct backing_dev_info *bdi;
1340 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1341 del_timer(&bdi->laptop_mode_wb_timer);
1348 * If ratelimit_pages is too high then we can get into dirty-data overload
1349 * if a large number of processes all perform writes at the same time.
1350 * If it is too low then SMP machines will call the (expensive)
1351 * get_writeback_state too often.
1353 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1354 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1358 void writeback_set_ratelimit(void)
1360 unsigned long background_thresh;
1361 unsigned long dirty_thresh;
1362 global_dirty_limits(&background_thresh, &dirty_thresh);
1363 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1364 if (ratelimit_pages < 16)
1365 ratelimit_pages = 16;
1368 static int __cpuinit
1369 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1371 writeback_set_ratelimit();
1375 static struct notifier_block __cpuinitdata ratelimit_nb = {
1376 .notifier_call = ratelimit_handler,
1381 * Called early on to tune the page writeback dirty limits.
1383 * We used to scale dirty pages according to how total memory
1384 * related to pages that could be allocated for buffers (by
1385 * comparing nr_free_buffer_pages() to vm_total_pages.
1387 * However, that was when we used "dirty_ratio" to scale with
1388 * all memory, and we don't do that any more. "dirty_ratio"
1389 * is now applied to total non-HIGHPAGE memory (by subtracting
1390 * totalhigh_pages from vm_total_pages), and as such we can't
1391 * get into the old insane situation any more where we had
1392 * large amounts of dirty pages compared to a small amount of
1393 * non-HIGHMEM memory.
1395 * But we might still want to scale the dirty_ratio by how
1396 * much memory the box has..
1398 void __init page_writeback_init(void)
1402 writeback_set_ratelimit();
1403 register_cpu_notifier(&ratelimit_nb);
1405 shift = calc_period_shift();
1406 prop_descriptor_init(&vm_completions, shift);
1410 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1411 * @mapping: address space structure to write
1412 * @start: starting page index
1413 * @end: ending page index (inclusive)
1415 * This function scans the page range from @start to @end (inclusive) and tags
1416 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1417 * that write_cache_pages (or whoever calls this function) will then use
1418 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1419 * used to avoid livelocking of writeback by a process steadily creating new
1420 * dirty pages in the file (thus it is important for this function to be quick
1421 * so that it can tag pages faster than a dirtying process can create them).
1424 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1426 void tag_pages_for_writeback(struct address_space *mapping,
1427 pgoff_t start, pgoff_t end)
1429 #define WRITEBACK_TAG_BATCH 4096
1430 unsigned long tagged;
1433 spin_lock_irq(&mapping->tree_lock);
1434 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1435 &start, end, WRITEBACK_TAG_BATCH,
1436 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1437 spin_unlock_irq(&mapping->tree_lock);
1438 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1440 /* We check 'start' to handle wrapping when end == ~0UL */
1441 } while (tagged >= WRITEBACK_TAG_BATCH && start);
1443 EXPORT_SYMBOL(tag_pages_for_writeback);
1446 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1447 * @mapping: address space structure to write
1448 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1449 * @writepage: function called for each page
1450 * @data: data passed to writepage function
1452 * If a page is already under I/O, write_cache_pages() skips it, even
1453 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1454 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1455 * and msync() need to guarantee that all the data which was dirty at the time
1456 * the call was made get new I/O started against them. If wbc->sync_mode is
1457 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1458 * existing IO to complete.
1460 * To avoid livelocks (when other process dirties new pages), we first tag
1461 * pages which should be written back with TOWRITE tag and only then start
1462 * writing them. For data-integrity sync we have to be careful so that we do
1463 * not miss some pages (e.g., because some other process has cleared TOWRITE
1464 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1465 * by the process clearing the DIRTY tag (and submitting the page for IO).
1467 int write_cache_pages(struct address_space *mapping,
1468 struct writeback_control *wbc, writepage_t writepage,
1473 struct pagevec pvec;
1475 pgoff_t uninitialized_var(writeback_index);
1477 pgoff_t end; /* Inclusive */
1480 int range_whole = 0;
1483 pagevec_init(&pvec, 0);
1484 if (wbc->range_cyclic) {
1485 writeback_index = mapping->writeback_index; /* prev offset */
1486 index = writeback_index;
1493 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1494 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1495 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1497 cycled = 1; /* ignore range_cyclic tests */
1499 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1500 tag = PAGECACHE_TAG_TOWRITE;
1502 tag = PAGECACHE_TAG_DIRTY;
1504 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1505 tag_pages_for_writeback(mapping, index, end);
1507 while (!done && (index <= end)) {
1510 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1511 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1515 for (i = 0; i < nr_pages; i++) {
1516 struct page *page = pvec.pages[i];
1519 * At this point, the page may be truncated or
1520 * invalidated (changing page->mapping to NULL), or
1521 * even swizzled back from swapper_space to tmpfs file
1522 * mapping. However, page->index will not change
1523 * because we have a reference on the page.
1525 if (page->index > end) {
1527 * can't be range_cyclic (1st pass) because
1528 * end == -1 in that case.
1534 done_index = page->index;
1539 * Page truncated or invalidated. We can freely skip it
1540 * then, even for data integrity operations: the page
1541 * has disappeared concurrently, so there could be no
1542 * real expectation of this data interity operation
1543 * even if there is now a new, dirty page at the same
1544 * pagecache address.
1546 if (unlikely(page->mapping != mapping)) {
1552 if (!PageDirty(page)) {
1553 /* someone wrote it for us */
1554 goto continue_unlock;
1557 if (PageWriteback(page)) {
1558 if (wbc->sync_mode != WB_SYNC_NONE)
1559 wait_on_page_writeback(page);
1561 goto continue_unlock;
1564 BUG_ON(PageWriteback(page));
1565 if (!clear_page_dirty_for_io(page))
1566 goto continue_unlock;
1568 trace_wbc_writepage(wbc, mapping->backing_dev_info);
1569 ret = (*writepage)(page, wbc, data);
1570 if (unlikely(ret)) {
1571 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1576 * done_index is set past this page,
1577 * so media errors will not choke
1578 * background writeout for the entire
1579 * file. This has consequences for
1580 * range_cyclic semantics (ie. it may
1581 * not be suitable for data integrity
1584 done_index = page->index + 1;
1591 * We stop writing back only if we are not doing
1592 * integrity sync. In case of integrity sync we have to
1593 * keep going until we have written all the pages
1594 * we tagged for writeback prior to entering this loop.
1596 if (--wbc->nr_to_write <= 0 &&
1597 wbc->sync_mode == WB_SYNC_NONE) {
1602 pagevec_release(&pvec);
1605 if (!cycled && !done) {
1608 * We hit the last page and there is more work to be done: wrap
1609 * back to the start of the file
1613 end = writeback_index - 1;
1616 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1617 mapping->writeback_index = done_index;
1621 EXPORT_SYMBOL(write_cache_pages);
1624 * Function used by generic_writepages to call the real writepage
1625 * function and set the mapping flags on error
1627 static int __writepage(struct page *page, struct writeback_control *wbc,
1630 struct address_space *mapping = data;
1631 int ret = mapping->a_ops->writepage(page, wbc);
1632 mapping_set_error(mapping, ret);
1637 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1638 * @mapping: address space structure to write
1639 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1641 * This is a library function, which implements the writepages()
1642 * address_space_operation.
1644 int generic_writepages(struct address_space *mapping,
1645 struct writeback_control *wbc)
1647 struct blk_plug plug;
1650 /* deal with chardevs and other special file */
1651 if (!mapping->a_ops->writepage)
1654 blk_start_plug(&plug);
1655 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1656 blk_finish_plug(&plug);
1660 EXPORT_SYMBOL(generic_writepages);
1662 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1666 if (wbc->nr_to_write <= 0)
1668 if (mapping->a_ops->writepages)
1669 ret = mapping->a_ops->writepages(mapping, wbc);
1671 ret = generic_writepages(mapping, wbc);
1676 * write_one_page - write out a single page and optionally wait on I/O
1677 * @page: the page to write
1678 * @wait: if true, wait on writeout
1680 * The page must be locked by the caller and will be unlocked upon return.
1682 * write_one_page() returns a negative error code if I/O failed.
1684 int write_one_page(struct page *page, int wait)
1686 struct address_space *mapping = page->mapping;
1688 struct writeback_control wbc = {
1689 .sync_mode = WB_SYNC_ALL,
1693 BUG_ON(!PageLocked(page));
1696 wait_on_page_writeback(page);
1698 if (clear_page_dirty_for_io(page)) {
1699 page_cache_get(page);
1700 ret = mapping->a_ops->writepage(page, &wbc);
1701 if (ret == 0 && wait) {
1702 wait_on_page_writeback(page);
1703 if (PageError(page))
1706 page_cache_release(page);
1712 EXPORT_SYMBOL(write_one_page);
1715 * For address_spaces which do not use buffers nor write back.
1717 int __set_page_dirty_no_writeback(struct page *page)
1719 if (!PageDirty(page))
1720 return !TestSetPageDirty(page);
1725 * Helper function for set_page_dirty family.
1726 * NOTE: This relies on being atomic wrt interrupts.
1728 void account_page_dirtied(struct page *page, struct address_space *mapping)
1730 if (mapping_cap_account_dirty(mapping)) {
1731 __inc_zone_page_state(page, NR_FILE_DIRTY);
1732 __inc_zone_page_state(page, NR_DIRTIED);
1733 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1734 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1735 task_io_account_write(PAGE_CACHE_SIZE);
1738 EXPORT_SYMBOL(account_page_dirtied);
1741 * Helper function for set_page_writeback family.
1742 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1745 void account_page_writeback(struct page *page)
1747 inc_zone_page_state(page, NR_WRITEBACK);
1749 EXPORT_SYMBOL(account_page_writeback);
1752 * For address_spaces which do not use buffers. Just tag the page as dirty in
1755 * This is also used when a single buffer is being dirtied: we want to set the
1756 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1757 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1759 * The caller must ensure this doesn't race with truncation. Most will simply
1760 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
1761 * the pte lock held, which also locks out truncation.
1763 int __set_page_dirty_nobuffers(struct page *page)
1765 if (!TestSetPageDirty(page)) {
1766 struct address_space *mapping = page_mapping(page);
1767 unsigned long flags;
1772 spin_lock_irqsave(&mapping->tree_lock, flags);
1773 BUG_ON(page_mapping(page) != mapping);
1774 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1775 account_page_dirtied(page, mapping);
1776 radix_tree_tag_set(&mapping->page_tree, page_index(page),
1777 PAGECACHE_TAG_DIRTY);
1778 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1779 if (mapping->host) {
1780 /* !PageAnon && !swapper_space */
1781 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1787 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1790 * Call this whenever redirtying a page, to de-account the dirty counters
1791 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
1792 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
1793 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
1796 void account_page_redirty(struct page *page)
1798 struct address_space *mapping = page->mapping;
1799 if (mapping && mapping_cap_account_dirty(mapping)) {
1800 current->nr_dirtied--;
1801 dec_zone_page_state(page, NR_DIRTIED);
1802 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1805 EXPORT_SYMBOL(account_page_redirty);
1808 * When a writepage implementation decides that it doesn't want to write this
1809 * page for some reason, it should redirty the locked page via
1810 * redirty_page_for_writepage() and it should then unlock the page and return 0
1812 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1814 wbc->pages_skipped++;
1815 account_page_redirty(page);
1816 return __set_page_dirty_nobuffers(page);
1818 EXPORT_SYMBOL(redirty_page_for_writepage);
1823 * For pages with a mapping this should be done under the page lock
1824 * for the benefit of asynchronous memory errors who prefer a consistent
1825 * dirty state. This rule can be broken in some special cases,
1826 * but should be better not to.
1828 * If the mapping doesn't provide a set_page_dirty a_op, then
1829 * just fall through and assume that it wants buffer_heads.
1831 int set_page_dirty(struct page *page)
1833 struct address_space *mapping = page_mapping(page);
1835 if (likely(mapping)) {
1836 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1838 * readahead/lru_deactivate_page could remain
1839 * PG_readahead/PG_reclaim due to race with end_page_writeback
1840 * About readahead, if the page is written, the flags would be
1841 * reset. So no problem.
1842 * About lru_deactivate_page, if the page is redirty, the flag
1843 * will be reset. So no problem. but if the page is used by readahead
1844 * it will confuse readahead and make it restart the size rampup
1845 * process. But it's a trivial problem.
1847 ClearPageReclaim(page);
1850 spd = __set_page_dirty_buffers;
1852 return (*spd)(page);
1854 if (!PageDirty(page)) {
1855 if (!TestSetPageDirty(page))
1860 EXPORT_SYMBOL(set_page_dirty);
1863 * set_page_dirty() is racy if the caller has no reference against
1864 * page->mapping->host, and if the page is unlocked. This is because another
1865 * CPU could truncate the page off the mapping and then free the mapping.
1867 * Usually, the page _is_ locked, or the caller is a user-space process which
1868 * holds a reference on the inode by having an open file.
1870 * In other cases, the page should be locked before running set_page_dirty().
1872 int set_page_dirty_lock(struct page *page)
1877 ret = set_page_dirty(page);
1881 EXPORT_SYMBOL(set_page_dirty_lock);
1884 * Clear a page's dirty flag, while caring for dirty memory accounting.
1885 * Returns true if the page was previously dirty.
1887 * This is for preparing to put the page under writeout. We leave the page
1888 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1889 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1890 * implementation will run either set_page_writeback() or set_page_dirty(),
1891 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1894 * This incoherency between the page's dirty flag and radix-tree tag is
1895 * unfortunate, but it only exists while the page is locked.
1897 int clear_page_dirty_for_io(struct page *page)
1899 struct address_space *mapping = page_mapping(page);
1901 BUG_ON(!PageLocked(page));
1903 if (mapping && mapping_cap_account_dirty(mapping)) {
1905 * Yes, Virginia, this is indeed insane.
1907 * We use this sequence to make sure that
1908 * (a) we account for dirty stats properly
1909 * (b) we tell the low-level filesystem to
1910 * mark the whole page dirty if it was
1911 * dirty in a pagetable. Only to then
1912 * (c) clean the page again and return 1 to
1913 * cause the writeback.
1915 * This way we avoid all nasty races with the
1916 * dirty bit in multiple places and clearing
1917 * them concurrently from different threads.
1919 * Note! Normally the "set_page_dirty(page)"
1920 * has no effect on the actual dirty bit - since
1921 * that will already usually be set. But we
1922 * need the side effects, and it can help us
1925 * We basically use the page "master dirty bit"
1926 * as a serialization point for all the different
1927 * threads doing their things.
1929 if (page_mkclean(page))
1930 set_page_dirty(page);
1932 * We carefully synchronise fault handlers against
1933 * installing a dirty pte and marking the page dirty
1934 * at this point. We do this by having them hold the
1935 * page lock while dirtying the page, and pages are
1936 * always locked coming in here, so we get the desired
1939 if (TestClearPageDirty(page)) {
1940 dec_zone_page_state(page, NR_FILE_DIRTY);
1941 dec_bdi_stat(mapping->backing_dev_info,
1947 return TestClearPageDirty(page);
1949 EXPORT_SYMBOL(clear_page_dirty_for_io);
1951 int test_clear_page_writeback(struct page *page)
1953 struct address_space *mapping = page_mapping(page);
1957 struct backing_dev_info *bdi = mapping->backing_dev_info;
1958 unsigned long flags;
1960 spin_lock_irqsave(&mapping->tree_lock, flags);
1961 ret = TestClearPageWriteback(page);
1963 radix_tree_tag_clear(&mapping->page_tree,
1965 PAGECACHE_TAG_WRITEBACK);
1966 if (bdi_cap_account_writeback(bdi)) {
1967 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1968 __bdi_writeout_inc(bdi);
1971 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1973 ret = TestClearPageWriteback(page);
1976 dec_zone_page_state(page, NR_WRITEBACK);
1977 inc_zone_page_state(page, NR_WRITTEN);
1982 int test_set_page_writeback(struct page *page)
1984 struct address_space *mapping = page_mapping(page);
1988 struct backing_dev_info *bdi = mapping->backing_dev_info;
1989 unsigned long flags;
1991 spin_lock_irqsave(&mapping->tree_lock, flags);
1992 ret = TestSetPageWriteback(page);
1994 radix_tree_tag_set(&mapping->page_tree,
1996 PAGECACHE_TAG_WRITEBACK);
1997 if (bdi_cap_account_writeback(bdi))
1998 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2000 if (!PageDirty(page))
2001 radix_tree_tag_clear(&mapping->page_tree,
2003 PAGECACHE_TAG_DIRTY);
2004 radix_tree_tag_clear(&mapping->page_tree,
2006 PAGECACHE_TAG_TOWRITE);
2007 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2009 ret = TestSetPageWriteback(page);
2012 account_page_writeback(page);
2016 EXPORT_SYMBOL(test_set_page_writeback);
2019 * Return true if any of the pages in the mapping are marked with the
2022 int mapping_tagged(struct address_space *mapping, int tag)
2024 return radix_tree_tagged(&mapping->page_tree, tag);
2026 EXPORT_SYMBOL(mapping_tagged);