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 * couple the period to the dirty_ratio:
135 * period/2 ~ roundup_pow_of_two(dirty limit)
137 static int calc_period_shift(void)
139 unsigned long dirty_total;
142 dirty_total = vm_dirty_bytes / PAGE_SIZE;
144 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
146 return 2 + ilog2(dirty_total - 1);
150 * update the period when the dirty threshold changes.
152 static void update_completion_period(void)
154 int shift = calc_period_shift();
155 prop_change_shift(&vm_completions, shift);
157 writeback_set_ratelimit();
160 int dirty_background_ratio_handler(struct ctl_table *table, int write,
161 void __user *buffer, size_t *lenp,
166 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
167 if (ret == 0 && write)
168 dirty_background_bytes = 0;
172 int dirty_background_bytes_handler(struct ctl_table *table, int write,
173 void __user *buffer, size_t *lenp,
178 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
179 if (ret == 0 && write)
180 dirty_background_ratio = 0;
184 int dirty_ratio_handler(struct ctl_table *table, int write,
185 void __user *buffer, size_t *lenp,
188 int old_ratio = vm_dirty_ratio;
191 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
192 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
193 update_completion_period();
200 int dirty_bytes_handler(struct ctl_table *table, int write,
201 void __user *buffer, size_t *lenp,
204 unsigned long old_bytes = vm_dirty_bytes;
207 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
208 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
209 update_completion_period();
216 * Increment the BDI's writeout completion count and the global writeout
217 * completion count. Called from test_clear_page_writeback().
219 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
221 __inc_bdi_stat(bdi, BDI_WRITTEN);
222 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
226 void bdi_writeout_inc(struct backing_dev_info *bdi)
230 local_irq_save(flags);
231 __bdi_writeout_inc(bdi);
232 local_irq_restore(flags);
234 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
237 * Obtain an accurate fraction of the BDI's portion.
239 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
240 long *numerator, long *denominator)
242 prop_fraction_percpu(&vm_completions, &bdi->completions,
243 numerator, denominator);
247 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
248 * registered backing devices, which, for obvious reasons, can not
251 static unsigned int bdi_min_ratio;
253 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
257 spin_lock_bh(&bdi_lock);
258 if (min_ratio > bdi->max_ratio) {
261 min_ratio -= bdi->min_ratio;
262 if (bdi_min_ratio + min_ratio < 100) {
263 bdi_min_ratio += min_ratio;
264 bdi->min_ratio += min_ratio;
269 spin_unlock_bh(&bdi_lock);
274 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
281 spin_lock_bh(&bdi_lock);
282 if (bdi->min_ratio > max_ratio) {
285 bdi->max_ratio = max_ratio;
286 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
288 spin_unlock_bh(&bdi_lock);
292 EXPORT_SYMBOL(bdi_set_max_ratio);
295 * Work out the current dirty-memory clamping and background writeout
298 * The main aim here is to lower them aggressively if there is a lot of mapped
299 * memory around. To avoid stressing page reclaim with lots of unreclaimable
300 * pages. It is better to clamp down on writers than to start swapping, and
301 * performing lots of scanning.
303 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
305 * We don't permit the clamping level to fall below 5% - that is getting rather
308 * We make sure that the background writeout level is below the adjusted
312 static unsigned long highmem_dirtyable_memory(unsigned long total)
314 #ifdef CONFIG_HIGHMEM
318 for_each_node_state(node, N_HIGH_MEMORY) {
320 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
322 x += zone_page_state(z, NR_FREE_PAGES) +
323 zone_reclaimable_pages(z);
326 * Make sure that the number of highmem pages is never larger
327 * than the number of the total dirtyable memory. This can only
328 * occur in very strange VM situations but we want to make sure
329 * that this does not occur.
331 return min(x, total);
338 * determine_dirtyable_memory - amount of memory that may be used
340 * Returns the numebr of pages that can currently be freed and used
341 * by the kernel for direct mappings.
343 unsigned long determine_dirtyable_memory(void)
347 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
349 if (!vm_highmem_is_dirtyable)
350 x -= highmem_dirtyable_memory(x);
352 return x + 1; /* Ensure that we never return 0 */
355 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
356 unsigned long bg_thresh)
358 return (thresh + bg_thresh) / 2;
361 static unsigned long hard_dirty_limit(unsigned long thresh)
363 return max(thresh, global_dirty_limit);
367 * global_dirty_limits - background-writeback and dirty-throttling thresholds
369 * Calculate the dirty thresholds based on sysctl parameters
370 * - vm.dirty_background_ratio or vm.dirty_background_bytes
371 * - vm.dirty_ratio or vm.dirty_bytes
372 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
375 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
377 unsigned long background;
379 unsigned long uninitialized_var(available_memory);
380 struct task_struct *tsk;
382 if (!vm_dirty_bytes || !dirty_background_bytes)
383 available_memory = determine_dirtyable_memory();
386 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
388 dirty = (vm_dirty_ratio * available_memory) / 100;
390 if (dirty_background_bytes)
391 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
393 background = (dirty_background_ratio * available_memory) / 100;
395 if (background >= dirty)
396 background = dirty / 2;
398 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
399 background += background / 4;
402 *pbackground = background;
404 trace_global_dirty_state(background, dirty);
408 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
409 * @bdi: the backing_dev_info to query
410 * @dirty: global dirty limit in pages
412 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
413 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
415 * Note that balance_dirty_pages() will only seriously take it as a hard limit
416 * when sleeping max_pause per page is not enough to keep the dirty pages under
417 * control. For example, when the device is completely stalled due to some error
418 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
419 * In the other normal situations, it acts more gently by throttling the tasks
420 * more (rather than completely block them) when the bdi dirty pages go high.
422 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
423 * - starving fast devices
424 * - piling up dirty pages (that will take long time to sync) on slow devices
426 * The bdi's share of dirty limit will be adapting to its throughput and
427 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
429 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
432 long numerator, denominator;
435 * Calculate this BDI's share of the dirty ratio.
437 bdi_writeout_fraction(bdi, &numerator, &denominator);
439 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
440 bdi_dirty *= numerator;
441 do_div(bdi_dirty, denominator);
443 bdi_dirty += (dirty * bdi->min_ratio) / 100;
444 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
445 bdi_dirty = dirty * bdi->max_ratio / 100;
451 * Dirty position control.
453 * (o) global/bdi setpoints
455 * We want the dirty pages be balanced around the global/bdi setpoints.
456 * When the number of dirty pages is higher/lower than the setpoint, the
457 * dirty position control ratio (and hence task dirty ratelimit) will be
458 * decreased/increased to bring the dirty pages back to the setpoint.
460 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
462 * if (dirty < setpoint) scale up pos_ratio
463 * if (dirty > setpoint) scale down pos_ratio
465 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
466 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
468 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
470 * (o) global control line
474 * | |<===== global dirty control scope ======>|
482 * 1.0 ................................*
488 * 0 +------------.------------------.----------------------*------------->
489 * freerun^ setpoint^ limit^ dirty pages
491 * (o) bdi control line
499 * | * |<=========== span ============>|
500 * 1.0 .......................*
512 * 1/4 ...............................................* * * * * * * * * * * *
516 * 0 +----------------------.-------------------------------.------------->
517 * bdi_setpoint^ x_intercept^
519 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
520 * be smoothly throttled down to normal if it starts high in situations like
521 * - start writing to a slow SD card and a fast disk at the same time. The SD
522 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
523 * - the bdi dirty thresh drops quickly due to change of JBOD workload
525 static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
526 unsigned long thresh,
527 unsigned long bg_thresh,
529 unsigned long bdi_thresh,
530 unsigned long bdi_dirty)
532 unsigned long write_bw = bdi->avg_write_bandwidth;
533 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
534 unsigned long limit = hard_dirty_limit(thresh);
535 unsigned long x_intercept;
536 unsigned long setpoint; /* dirty pages' target balance point */
537 unsigned long bdi_setpoint;
539 long long pos_ratio; /* for scaling up/down the rate limit */
542 if (unlikely(dirty >= limit))
549 * f(dirty) := 1.0 + (----------------)
552 * it's a 3rd order polynomial that subjects to
554 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
555 * (2) f(setpoint) = 1.0 => the balance point
556 * (3) f(limit) = 0 => the hard limit
557 * (4) df/dx <= 0 => negative feedback control
558 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
559 * => fast response on large errors; small oscillation near setpoint
561 setpoint = (freerun + limit) / 2;
562 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
563 (limit - setpoint) | 1);
565 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
566 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
567 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
570 * We have computed basic pos_ratio above based on global situation. If
571 * the bdi is over/under its share of dirty pages, we want to scale
572 * pos_ratio further down/up. That is done by the following mechanism.
578 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
580 * x_intercept - bdi_dirty
581 * := --------------------------
582 * x_intercept - bdi_setpoint
584 * The main bdi control line is a linear function that subjects to
586 * (1) f(bdi_setpoint) = 1.0
587 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
588 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
590 * For single bdi case, the dirty pages are observed to fluctuate
591 * regularly within range
592 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
593 * for various filesystems, where (2) can yield in a reasonable 12.5%
594 * fluctuation range for pos_ratio.
596 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
597 * own size, so move the slope over accordingly and choose a slope that
598 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
600 if (unlikely(bdi_thresh > thresh))
603 * It's very possible that bdi_thresh is close to 0 not because the
604 * device is slow, but that it has remained inactive for long time.
605 * Honour such devices a reasonable good (hopefully IO efficient)
606 * threshold, so that the occasional writes won't be blocked and active
607 * writes can rampup the threshold quickly.
609 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
611 * scale global setpoint to bdi's:
612 * bdi_setpoint = setpoint * bdi_thresh / thresh
614 x = div_u64((u64)bdi_thresh << 16, thresh | 1);
615 bdi_setpoint = setpoint * (u64)x >> 16;
617 * Use span=(8*write_bw) in single bdi case as indicated by
618 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
620 * bdi_thresh thresh - bdi_thresh
621 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
624 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
625 x_intercept = bdi_setpoint + span;
627 if (bdi_dirty < x_intercept - span / 4) {
628 pos_ratio = div64_u64(pos_ratio * (x_intercept - bdi_dirty),
629 (x_intercept - bdi_setpoint) | 1);
634 * bdi reserve area, safeguard against dirty pool underrun and disk idle
635 * It may push the desired control point of global dirty pages higher
638 x_intercept = bdi_thresh / 2;
639 if (bdi_dirty < x_intercept) {
640 if (bdi_dirty > x_intercept / 8)
641 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
649 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
650 unsigned long elapsed,
651 unsigned long written)
653 const unsigned long period = roundup_pow_of_two(3 * HZ);
654 unsigned long avg = bdi->avg_write_bandwidth;
655 unsigned long old = bdi->write_bandwidth;
659 * bw = written * HZ / elapsed
661 * bw * elapsed + write_bandwidth * (period - elapsed)
662 * write_bandwidth = ---------------------------------------------------
665 * @written may have decreased due to account_page_redirty().
666 * Avoid underflowing @bw calculation.
668 bw = written - min(written, bdi->written_stamp);
670 if (unlikely(elapsed > period)) {
675 bw += (u64)bdi->write_bandwidth * (period - elapsed);
676 bw >>= ilog2(period);
679 * one more level of smoothing, for filtering out sudden spikes
681 if (avg > old && old >= (unsigned long)bw)
682 avg -= (avg - old) >> 3;
684 if (avg < old && old <= (unsigned long)bw)
685 avg += (old - avg) >> 3;
688 bdi->write_bandwidth = bw;
689 bdi->avg_write_bandwidth = avg;
693 * The global dirtyable memory and dirty threshold could be suddenly knocked
694 * down by a large amount (eg. on the startup of KVM in a swapless system).
695 * This may throw the system into deep dirty exceeded state and throttle
696 * heavy/light dirtiers alike. To retain good responsiveness, maintain
697 * global_dirty_limit for tracking slowly down to the knocked down dirty
700 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
702 unsigned long limit = global_dirty_limit;
705 * Follow up in one step.
707 if (limit < thresh) {
713 * Follow down slowly. Use the higher one as the target, because thresh
714 * may drop below dirty. This is exactly the reason to introduce
715 * global_dirty_limit which is guaranteed to lie above the dirty pages.
717 thresh = max(thresh, dirty);
718 if (limit > thresh) {
719 limit -= (limit - thresh) >> 5;
724 global_dirty_limit = limit;
727 static void global_update_bandwidth(unsigned long thresh,
731 static DEFINE_SPINLOCK(dirty_lock);
732 static unsigned long update_time = INITIAL_JIFFIES;
735 * check locklessly first to optimize away locking for the most time
737 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
740 spin_lock(&dirty_lock);
741 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
742 update_dirty_limit(thresh, dirty);
745 spin_unlock(&dirty_lock);
749 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
751 * Normal bdi tasks will be curbed at or below it in long term.
752 * Obviously it should be around (write_bw / N) when there are N dd tasks.
754 static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
755 unsigned long thresh,
756 unsigned long bg_thresh,
758 unsigned long bdi_thresh,
759 unsigned long bdi_dirty,
760 unsigned long dirtied,
761 unsigned long elapsed)
763 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
764 unsigned long limit = hard_dirty_limit(thresh);
765 unsigned long setpoint = (freerun + limit) / 2;
766 unsigned long write_bw = bdi->avg_write_bandwidth;
767 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
768 unsigned long dirty_rate;
769 unsigned long task_ratelimit;
770 unsigned long balanced_dirty_ratelimit;
771 unsigned long pos_ratio;
776 * The dirty rate will match the writeout rate in long term, except
777 * when dirty pages are truncated by userspace or re-dirtied by FS.
779 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
781 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
782 bdi_thresh, bdi_dirty);
784 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
786 task_ratelimit = (u64)dirty_ratelimit *
787 pos_ratio >> RATELIMIT_CALC_SHIFT;
788 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
791 * A linear estimation of the "balanced" throttle rate. The theory is,
792 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
793 * dirty_rate will be measured to be (N * task_ratelimit). So the below
794 * formula will yield the balanced rate limit (write_bw / N).
796 * Note that the expanded form is not a pure rate feedback:
797 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
798 * but also takes pos_ratio into account:
799 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
801 * (1) is not realistic because pos_ratio also takes part in balancing
802 * the dirty rate. Consider the state
803 * pos_ratio = 0.5 (3)
804 * rate = 2 * (write_bw / N) (4)
805 * If (1) is used, it will stuck in that state! Because each dd will
807 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
809 * dirty_rate = N * task_ratelimit = write_bw (6)
810 * put (6) into (1) we get
811 * rate_(i+1) = rate_(i) (7)
813 * So we end up using (2) to always keep
814 * rate_(i+1) ~= (write_bw / N) (8)
815 * regardless of the value of pos_ratio. As long as (8) is satisfied,
816 * pos_ratio is able to drive itself to 1.0, which is not only where
817 * the dirty count meet the setpoint, but also where the slope of
818 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
820 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
824 * We could safely do this and return immediately:
826 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
828 * However to get a more stable dirty_ratelimit, the below elaborated
829 * code makes use of task_ratelimit to filter out sigular points and
830 * limit the step size.
832 * The below code essentially only uses the relative value of
834 * task_ratelimit - dirty_ratelimit
835 * = (pos_ratio - 1) * dirty_ratelimit
837 * which reflects the direction and size of dirty position error.
841 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
842 * task_ratelimit is on the same side of dirty_ratelimit, too.
844 * - dirty_ratelimit > balanced_dirty_ratelimit
845 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
846 * lowering dirty_ratelimit will help meet both the position and rate
847 * control targets. Otherwise, don't update dirty_ratelimit if it will
848 * only help meet the rate target. After all, what the users ultimately
849 * feel and care are stable dirty rate and small position error.
851 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
852 * and filter out the sigular points of balanced_dirty_ratelimit. Which
853 * keeps jumping around randomly and can even leap far away at times
854 * due to the small 200ms estimation period of dirty_rate (we want to
855 * keep that period small to reduce time lags).
858 if (dirty < setpoint) {
859 x = min(bdi->balanced_dirty_ratelimit,
860 min(balanced_dirty_ratelimit, task_ratelimit));
861 if (dirty_ratelimit < x)
862 step = x - dirty_ratelimit;
864 x = max(bdi->balanced_dirty_ratelimit,
865 max(balanced_dirty_ratelimit, task_ratelimit));
866 if (dirty_ratelimit > x)
867 step = dirty_ratelimit - x;
871 * Don't pursue 100% rate matching. It's impossible since the balanced
872 * rate itself is constantly fluctuating. So decrease the track speed
873 * when it gets close to the target. Helps eliminate pointless tremors.
875 step >>= dirty_ratelimit / (2 * step + 1);
877 * Limit the tracking speed to avoid overshooting.
879 step = (step + 7) / 8;
881 if (dirty_ratelimit < balanced_dirty_ratelimit)
882 dirty_ratelimit += step;
884 dirty_ratelimit -= step;
886 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
887 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
889 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
892 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
893 unsigned long thresh,
894 unsigned long bg_thresh,
896 unsigned long bdi_thresh,
897 unsigned long bdi_dirty,
898 unsigned long start_time)
900 unsigned long now = jiffies;
901 unsigned long elapsed = now - bdi->bw_time_stamp;
902 unsigned long dirtied;
903 unsigned long written;
906 * rate-limit, only update once every 200ms.
908 if (elapsed < BANDWIDTH_INTERVAL)
911 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
912 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
915 * Skip quiet periods when disk bandwidth is under-utilized.
916 * (at least 1s idle time between two flusher runs)
918 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
922 global_update_bandwidth(thresh, dirty, now);
923 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
924 bdi_thresh, bdi_dirty,
927 bdi_update_write_bandwidth(bdi, elapsed, written);
930 bdi->dirtied_stamp = dirtied;
931 bdi->written_stamp = written;
932 bdi->bw_time_stamp = now;
935 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
936 unsigned long thresh,
937 unsigned long bg_thresh,
939 unsigned long bdi_thresh,
940 unsigned long bdi_dirty,
941 unsigned long start_time)
943 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
945 spin_lock(&bdi->wb.list_lock);
946 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
947 bdi_thresh, bdi_dirty, start_time);
948 spin_unlock(&bdi->wb.list_lock);
952 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
953 * will look to see if it needs to start dirty throttling.
955 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
956 * global_page_state() too often. So scale it near-sqrt to the safety margin
957 * (the number of pages we may dirty without exceeding the dirty limits).
959 static unsigned long dirty_poll_interval(unsigned long dirty,
960 unsigned long thresh)
963 return 1UL << (ilog2(thresh - dirty) >> 1);
968 static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
969 unsigned long bdi_dirty)
971 unsigned long bw = bdi->avg_write_bandwidth;
972 unsigned long hi = ilog2(bw);
973 unsigned long lo = ilog2(bdi->dirty_ratelimit);
976 /* target for 20ms max pause on 1-dd case */
980 * Scale up pause time for concurrent dirtiers in order to reduce CPU
983 * (N * 20ms) on 2^N concurrent tasks.
986 t += (hi - lo) * (20 * HZ) / 1024;
989 * Limit pause time for small memory systems. If sleeping for too long
990 * time, a small pool of dirty/writeback pages may go empty and disk go
993 * 8 serves as the safety ratio.
995 t = min(t, bdi_dirty * HZ / (8 * bw + 1));
998 * The pause time will be settled within range (max_pause/4, max_pause).
999 * Apply a minimal value of 4 to get a non-zero max_pause/4.
1001 return clamp_val(t, 4, MAX_PAUSE);
1005 * balance_dirty_pages() must be called by processes which are generating dirty
1006 * data. It looks at the number of dirty pages in the machine and will force
1007 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1008 * If we're over `background_thresh' then the writeback threads are woken to
1009 * perform some writeout.
1011 static void balance_dirty_pages(struct address_space *mapping,
1012 unsigned long pages_dirtied)
1014 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1015 unsigned long bdi_reclaimable;
1016 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1017 unsigned long bdi_dirty;
1018 unsigned long freerun;
1019 unsigned long background_thresh;
1020 unsigned long dirty_thresh;
1021 unsigned long bdi_thresh;
1023 long uninitialized_var(max_pause);
1024 bool dirty_exceeded = false;
1025 unsigned long task_ratelimit;
1026 unsigned long uninitialized_var(dirty_ratelimit);
1027 unsigned long pos_ratio;
1028 struct backing_dev_info *bdi = mapping->backing_dev_info;
1029 unsigned long start_time = jiffies;
1033 * Unstable writes are a feature of certain networked
1034 * filesystems (i.e. NFS) in which data may have been
1035 * written to the server's write cache, but has not yet
1036 * been flushed to permanent storage.
1038 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1039 global_page_state(NR_UNSTABLE_NFS);
1040 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1042 global_dirty_limits(&background_thresh, &dirty_thresh);
1045 * Throttle it only when the background writeback cannot
1046 * catch-up. This avoids (excessively) small writeouts
1047 * when the bdi limits are ramping up.
1049 freerun = dirty_freerun_ceiling(dirty_thresh,
1051 if (nr_dirty <= freerun)
1054 if (unlikely(!writeback_in_progress(bdi)))
1055 bdi_start_background_writeback(bdi);
1058 * bdi_thresh is not treated as some limiting factor as
1059 * dirty_thresh, due to reasons
1060 * - in JBOD setup, bdi_thresh can fluctuate a lot
1061 * - in a system with HDD and USB key, the USB key may somehow
1062 * go into state (bdi_dirty >> bdi_thresh) either because
1063 * bdi_dirty starts high, or because bdi_thresh drops low.
1064 * In this case we don't want to hard throttle the USB key
1065 * dirtiers for 100 seconds until bdi_dirty drops under
1066 * bdi_thresh. Instead the auxiliary bdi control line in
1067 * bdi_position_ratio() will let the dirtier task progress
1068 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1070 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1073 * In order to avoid the stacked BDI deadlock we need
1074 * to ensure we accurately count the 'dirty' pages when
1075 * the threshold is low.
1077 * Otherwise it would be possible to get thresh+n pages
1078 * reported dirty, even though there are thresh-m pages
1079 * actually dirty; with m+n sitting in the percpu
1082 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1083 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1084 bdi_dirty = bdi_reclaimable +
1085 bdi_stat_sum(bdi, BDI_WRITEBACK);
1087 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1088 bdi_dirty = bdi_reclaimable +
1089 bdi_stat(bdi, BDI_WRITEBACK);
1092 dirty_exceeded = (bdi_dirty > bdi_thresh) ||
1093 (nr_dirty > dirty_thresh);
1094 if (dirty_exceeded && !bdi->dirty_exceeded)
1095 bdi->dirty_exceeded = 1;
1097 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1098 nr_dirty, bdi_thresh, bdi_dirty,
1101 max_pause = bdi_max_pause(bdi, bdi_dirty);
1103 dirty_ratelimit = bdi->dirty_ratelimit;
1104 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1105 background_thresh, nr_dirty,
1106 bdi_thresh, bdi_dirty);
1107 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1108 RATELIMIT_CALC_SHIFT;
1109 if (unlikely(task_ratelimit == 0)) {
1113 pause = HZ * pages_dirtied / task_ratelimit;
1114 if (unlikely(pause <= 0)) {
1115 trace_balance_dirty_pages(bdi,
1126 pause = 1; /* avoid resetting nr_dirtied_pause below */
1129 pause = min(pause, max_pause);
1132 trace_balance_dirty_pages(bdi,
1143 __set_current_state(TASK_KILLABLE);
1144 io_schedule_timeout(pause);
1147 * This is typically equal to (nr_dirty < dirty_thresh) and can
1148 * also keep "1000+ dd on a slow USB stick" under control.
1154 * In the case of an unresponding NFS server and the NFS dirty
1155 * pages exceeds dirty_thresh, give the other good bdi's a pipe
1156 * to go through, so that tasks on them still remain responsive.
1158 * In theory 1 page is enough to keep the comsumer-producer
1159 * pipe going: the flusher cleans 1 page => the task dirties 1
1160 * more page. However bdi_dirty has accounting errors. So use
1161 * the larger and more IO friendly bdi_stat_error.
1163 if (bdi_dirty <= bdi_stat_error(bdi))
1166 if (fatal_signal_pending(current))
1170 if (!dirty_exceeded && bdi->dirty_exceeded)
1171 bdi->dirty_exceeded = 0;
1173 current->nr_dirtied = 0;
1174 if (pause == 0) { /* in freerun area */
1175 current->nr_dirtied_pause =
1176 dirty_poll_interval(nr_dirty, dirty_thresh);
1177 } else if (pause <= max_pause / 4 &&
1178 pages_dirtied >= current->nr_dirtied_pause) {
1179 current->nr_dirtied_pause = clamp_val(
1180 dirty_ratelimit * (max_pause / 2) / HZ,
1181 pages_dirtied + pages_dirtied / 8,
1183 } else if (pause >= max_pause) {
1184 current->nr_dirtied_pause = 1 | clamp_val(
1185 dirty_ratelimit * (max_pause / 2) / HZ,
1187 pages_dirtied - pages_dirtied / 8);
1190 if (writeback_in_progress(bdi))
1194 * In laptop mode, we wait until hitting the higher threshold before
1195 * starting background writeout, and then write out all the way down
1196 * to the lower threshold. So slow writers cause minimal disk activity.
1198 * In normal mode, we start background writeout at the lower
1199 * background_thresh, to keep the amount of dirty memory low.
1204 if (nr_reclaimable > background_thresh)
1205 bdi_start_background_writeback(bdi);
1208 static DEFINE_PER_CPU(int, bdp_ratelimits);
1211 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
1212 * @mapping: address_space which was dirtied
1213 * @nr_pages_dirtied: number of pages which the caller has just dirtied
1215 * Processes which are dirtying memory should call in here once for each page
1216 * which was newly dirtied. The function will periodically check the system's
1217 * dirty state and will initiate writeback if needed.
1219 * On really big machines, get_writeback_state is expensive, so try to avoid
1220 * calling it too often (ratelimiting). But once we're over the dirty memory
1221 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1222 * from overshooting the limit by (ratelimit_pages) each.
1224 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1225 unsigned long nr_pages_dirtied)
1227 struct backing_dev_info *bdi = mapping->backing_dev_info;
1231 if (!bdi_cap_account_dirty(bdi))
1234 ratelimit = current->nr_dirtied_pause;
1235 if (bdi->dirty_exceeded)
1236 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1238 current->nr_dirtied += nr_pages_dirtied;
1242 * This prevents one CPU to accumulate too many dirtied pages without
1243 * calling into balance_dirty_pages(), which can happen when there are
1244 * 1000+ tasks, all of them start dirtying pages at exactly the same
1245 * time, hence all honoured too large initial task->nr_dirtied_pause.
1247 p = &__get_cpu_var(bdp_ratelimits);
1248 if (unlikely(current->nr_dirtied >= ratelimit))
1251 *p += nr_pages_dirtied;
1252 if (unlikely(*p >= ratelimit_pages)) {
1259 if (unlikely(current->nr_dirtied >= ratelimit))
1260 balance_dirty_pages(mapping, current->nr_dirtied);
1262 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
1264 void throttle_vm_writeout(gfp_t gfp_mask)
1266 unsigned long background_thresh;
1267 unsigned long dirty_thresh;
1270 global_dirty_limits(&background_thresh, &dirty_thresh);
1273 * Boost the allowable dirty threshold a bit for page
1274 * allocators so they don't get DoS'ed by heavy writers
1276 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1278 if (global_page_state(NR_UNSTABLE_NFS) +
1279 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1281 congestion_wait(BLK_RW_ASYNC, HZ/10);
1284 * The caller might hold locks which can prevent IO completion
1285 * or progress in the filesystem. So we cannot just sit here
1286 * waiting for IO to complete.
1288 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1294 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1296 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1297 void __user *buffer, size_t *length, loff_t *ppos)
1299 proc_dointvec(table, write, buffer, length, ppos);
1300 bdi_arm_supers_timer();
1305 void laptop_mode_timer_fn(unsigned long data)
1307 struct request_queue *q = (struct request_queue *)data;
1308 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1309 global_page_state(NR_UNSTABLE_NFS);
1312 * We want to write everything out, not just down to the dirty
1315 if (bdi_has_dirty_io(&q->backing_dev_info))
1316 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1317 WB_REASON_LAPTOP_TIMER);
1321 * We've spun up the disk and we're in laptop mode: schedule writeback
1322 * of all dirty data a few seconds from now. If the flush is already scheduled
1323 * then push it back - the user is still using the disk.
1325 void laptop_io_completion(struct backing_dev_info *info)
1327 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1331 * We're in laptop mode and we've just synced. The sync's writes will have
1332 * caused another writeback to be scheduled by laptop_io_completion.
1333 * Nothing needs to be written back anymore, so we unschedule the writeback.
1335 void laptop_sync_completion(void)
1337 struct backing_dev_info *bdi;
1341 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1342 del_timer(&bdi->laptop_mode_wb_timer);
1349 * If ratelimit_pages is too high then we can get into dirty-data overload
1350 * if a large number of processes all perform writes at the same time.
1351 * If it is too low then SMP machines will call the (expensive)
1352 * get_writeback_state too often.
1354 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1355 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1359 void writeback_set_ratelimit(void)
1361 unsigned long background_thresh;
1362 unsigned long dirty_thresh;
1363 global_dirty_limits(&background_thresh, &dirty_thresh);
1364 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1365 if (ratelimit_pages < 16)
1366 ratelimit_pages = 16;
1369 static int __cpuinit
1370 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1372 writeback_set_ratelimit();
1376 static struct notifier_block __cpuinitdata ratelimit_nb = {
1377 .notifier_call = ratelimit_handler,
1382 * Called early on to tune the page writeback dirty limits.
1384 * We used to scale dirty pages according to how total memory
1385 * related to pages that could be allocated for buffers (by
1386 * comparing nr_free_buffer_pages() to vm_total_pages.
1388 * However, that was when we used "dirty_ratio" to scale with
1389 * all memory, and we don't do that any more. "dirty_ratio"
1390 * is now applied to total non-HIGHPAGE memory (by subtracting
1391 * totalhigh_pages from vm_total_pages), and as such we can't
1392 * get into the old insane situation any more where we had
1393 * large amounts of dirty pages compared to a small amount of
1394 * non-HIGHMEM memory.
1396 * But we might still want to scale the dirty_ratio by how
1397 * much memory the box has..
1399 void __init page_writeback_init(void)
1403 writeback_set_ratelimit();
1404 register_cpu_notifier(&ratelimit_nb);
1406 shift = calc_period_shift();
1407 prop_descriptor_init(&vm_completions, shift);
1411 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1412 * @mapping: address space structure to write
1413 * @start: starting page index
1414 * @end: ending page index (inclusive)
1416 * This function scans the page range from @start to @end (inclusive) and tags
1417 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1418 * that write_cache_pages (or whoever calls this function) will then use
1419 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1420 * used to avoid livelocking of writeback by a process steadily creating new
1421 * dirty pages in the file (thus it is important for this function to be quick
1422 * so that it can tag pages faster than a dirtying process can create them).
1425 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1427 void tag_pages_for_writeback(struct address_space *mapping,
1428 pgoff_t start, pgoff_t end)
1430 #define WRITEBACK_TAG_BATCH 4096
1431 unsigned long tagged;
1434 spin_lock_irq(&mapping->tree_lock);
1435 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1436 &start, end, WRITEBACK_TAG_BATCH,
1437 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1438 spin_unlock_irq(&mapping->tree_lock);
1439 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1441 /* We check 'start' to handle wrapping when end == ~0UL */
1442 } while (tagged >= WRITEBACK_TAG_BATCH && start);
1444 EXPORT_SYMBOL(tag_pages_for_writeback);
1447 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1448 * @mapping: address space structure to write
1449 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1450 * @writepage: function called for each page
1451 * @data: data passed to writepage function
1453 * If a page is already under I/O, write_cache_pages() skips it, even
1454 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1455 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1456 * and msync() need to guarantee that all the data which was dirty at the time
1457 * the call was made get new I/O started against them. If wbc->sync_mode is
1458 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1459 * existing IO to complete.
1461 * To avoid livelocks (when other process dirties new pages), we first tag
1462 * pages which should be written back with TOWRITE tag and only then start
1463 * writing them. For data-integrity sync we have to be careful so that we do
1464 * not miss some pages (e.g., because some other process has cleared TOWRITE
1465 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1466 * by the process clearing the DIRTY tag (and submitting the page for IO).
1468 int write_cache_pages(struct address_space *mapping,
1469 struct writeback_control *wbc, writepage_t writepage,
1474 struct pagevec pvec;
1476 pgoff_t uninitialized_var(writeback_index);
1478 pgoff_t end; /* Inclusive */
1481 int range_whole = 0;
1484 pagevec_init(&pvec, 0);
1485 if (wbc->range_cyclic) {
1486 writeback_index = mapping->writeback_index; /* prev offset */
1487 index = writeback_index;
1494 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1495 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1496 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1498 cycled = 1; /* ignore range_cyclic tests */
1500 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1501 tag = PAGECACHE_TAG_TOWRITE;
1503 tag = PAGECACHE_TAG_DIRTY;
1505 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1506 tag_pages_for_writeback(mapping, index, end);
1508 while (!done && (index <= end)) {
1511 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1512 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1516 for (i = 0; i < nr_pages; i++) {
1517 struct page *page = pvec.pages[i];
1520 * At this point, the page may be truncated or
1521 * invalidated (changing page->mapping to NULL), or
1522 * even swizzled back from swapper_space to tmpfs file
1523 * mapping. However, page->index will not change
1524 * because we have a reference on the page.
1526 if (page->index > end) {
1528 * can't be range_cyclic (1st pass) because
1529 * end == -1 in that case.
1535 done_index = page->index;
1540 * Page truncated or invalidated. We can freely skip it
1541 * then, even for data integrity operations: the page
1542 * has disappeared concurrently, so there could be no
1543 * real expectation of this data interity operation
1544 * even if there is now a new, dirty page at the same
1545 * pagecache address.
1547 if (unlikely(page->mapping != mapping)) {
1553 if (!PageDirty(page)) {
1554 /* someone wrote it for us */
1555 goto continue_unlock;
1558 if (PageWriteback(page)) {
1559 if (wbc->sync_mode != WB_SYNC_NONE)
1560 wait_on_page_writeback(page);
1562 goto continue_unlock;
1565 BUG_ON(PageWriteback(page));
1566 if (!clear_page_dirty_for_io(page))
1567 goto continue_unlock;
1569 trace_wbc_writepage(wbc, mapping->backing_dev_info);
1570 ret = (*writepage)(page, wbc, data);
1571 if (unlikely(ret)) {
1572 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1577 * done_index is set past this page,
1578 * so media errors will not choke
1579 * background writeout for the entire
1580 * file. This has consequences for
1581 * range_cyclic semantics (ie. it may
1582 * not be suitable for data integrity
1585 done_index = page->index + 1;
1592 * We stop writing back only if we are not doing
1593 * integrity sync. In case of integrity sync we have to
1594 * keep going until we have written all the pages
1595 * we tagged for writeback prior to entering this loop.
1597 if (--wbc->nr_to_write <= 0 &&
1598 wbc->sync_mode == WB_SYNC_NONE) {
1603 pagevec_release(&pvec);
1606 if (!cycled && !done) {
1609 * We hit the last page and there is more work to be done: wrap
1610 * back to the start of the file
1614 end = writeback_index - 1;
1617 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1618 mapping->writeback_index = done_index;
1622 EXPORT_SYMBOL(write_cache_pages);
1625 * Function used by generic_writepages to call the real writepage
1626 * function and set the mapping flags on error
1628 static int __writepage(struct page *page, struct writeback_control *wbc,
1631 struct address_space *mapping = data;
1632 int ret = mapping->a_ops->writepage(page, wbc);
1633 mapping_set_error(mapping, ret);
1638 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1639 * @mapping: address space structure to write
1640 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1642 * This is a library function, which implements the writepages()
1643 * address_space_operation.
1645 int generic_writepages(struct address_space *mapping,
1646 struct writeback_control *wbc)
1648 struct blk_plug plug;
1651 /* deal with chardevs and other special file */
1652 if (!mapping->a_ops->writepage)
1655 blk_start_plug(&plug);
1656 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1657 blk_finish_plug(&plug);
1661 EXPORT_SYMBOL(generic_writepages);
1663 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1667 if (wbc->nr_to_write <= 0)
1669 if (mapping->a_ops->writepages)
1670 ret = mapping->a_ops->writepages(mapping, wbc);
1672 ret = generic_writepages(mapping, wbc);
1677 * write_one_page - write out a single page and optionally wait on I/O
1678 * @page: the page to write
1679 * @wait: if true, wait on writeout
1681 * The page must be locked by the caller and will be unlocked upon return.
1683 * write_one_page() returns a negative error code if I/O failed.
1685 int write_one_page(struct page *page, int wait)
1687 struct address_space *mapping = page->mapping;
1689 struct writeback_control wbc = {
1690 .sync_mode = WB_SYNC_ALL,
1694 BUG_ON(!PageLocked(page));
1697 wait_on_page_writeback(page);
1699 if (clear_page_dirty_for_io(page)) {
1700 page_cache_get(page);
1701 ret = mapping->a_ops->writepage(page, &wbc);
1702 if (ret == 0 && wait) {
1703 wait_on_page_writeback(page);
1704 if (PageError(page))
1707 page_cache_release(page);
1713 EXPORT_SYMBOL(write_one_page);
1716 * For address_spaces which do not use buffers nor write back.
1718 int __set_page_dirty_no_writeback(struct page *page)
1720 if (!PageDirty(page))
1721 return !TestSetPageDirty(page);
1726 * Helper function for set_page_dirty family.
1727 * NOTE: This relies on being atomic wrt interrupts.
1729 void account_page_dirtied(struct page *page, struct address_space *mapping)
1731 if (mapping_cap_account_dirty(mapping)) {
1732 __inc_zone_page_state(page, NR_FILE_DIRTY);
1733 __inc_zone_page_state(page, NR_DIRTIED);
1734 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1735 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1736 task_io_account_write(PAGE_CACHE_SIZE);
1739 EXPORT_SYMBOL(account_page_dirtied);
1742 * Helper function for set_page_writeback family.
1743 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1746 void account_page_writeback(struct page *page)
1748 inc_zone_page_state(page, NR_WRITEBACK);
1750 EXPORT_SYMBOL(account_page_writeback);
1753 * For address_spaces which do not use buffers. Just tag the page as dirty in
1756 * This is also used when a single buffer is being dirtied: we want to set the
1757 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1758 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1760 * The caller must ensure this doesn't race with truncation. Most will simply
1761 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
1762 * the pte lock held, which also locks out truncation.
1764 int __set_page_dirty_nobuffers(struct page *page)
1766 if (!TestSetPageDirty(page)) {
1767 struct address_space *mapping = page_mapping(page);
1768 unsigned long flags;
1773 spin_lock_irqsave(&mapping->tree_lock, flags);
1774 BUG_ON(page_mapping(page) != mapping);
1775 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1776 account_page_dirtied(page, mapping);
1777 radix_tree_tag_set(&mapping->page_tree, page_index(page),
1778 PAGECACHE_TAG_DIRTY);
1779 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1780 if (mapping->host) {
1781 /* !PageAnon && !swapper_space */
1782 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1788 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1791 * Call this whenever redirtying a page, to de-account the dirty counters
1792 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
1793 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
1794 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
1797 void account_page_redirty(struct page *page)
1799 struct address_space *mapping = page->mapping;
1800 if (mapping && mapping_cap_account_dirty(mapping)) {
1801 current->nr_dirtied--;
1802 dec_zone_page_state(page, NR_DIRTIED);
1803 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1806 EXPORT_SYMBOL(account_page_redirty);
1809 * When a writepage implementation decides that it doesn't want to write this
1810 * page for some reason, it should redirty the locked page via
1811 * redirty_page_for_writepage() and it should then unlock the page and return 0
1813 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1815 wbc->pages_skipped++;
1816 account_page_redirty(page);
1817 return __set_page_dirty_nobuffers(page);
1819 EXPORT_SYMBOL(redirty_page_for_writepage);
1824 * For pages with a mapping this should be done under the page lock
1825 * for the benefit of asynchronous memory errors who prefer a consistent
1826 * dirty state. This rule can be broken in some special cases,
1827 * but should be better not to.
1829 * If the mapping doesn't provide a set_page_dirty a_op, then
1830 * just fall through and assume that it wants buffer_heads.
1832 int set_page_dirty(struct page *page)
1834 struct address_space *mapping = page_mapping(page);
1836 if (likely(mapping)) {
1837 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1839 * readahead/lru_deactivate_page could remain
1840 * PG_readahead/PG_reclaim due to race with end_page_writeback
1841 * About readahead, if the page is written, the flags would be
1842 * reset. So no problem.
1843 * About lru_deactivate_page, if the page is redirty, the flag
1844 * will be reset. So no problem. but if the page is used by readahead
1845 * it will confuse readahead and make it restart the size rampup
1846 * process. But it's a trivial problem.
1848 ClearPageReclaim(page);
1851 spd = __set_page_dirty_buffers;
1853 return (*spd)(page);
1855 if (!PageDirty(page)) {
1856 if (!TestSetPageDirty(page))
1861 EXPORT_SYMBOL(set_page_dirty);
1864 * set_page_dirty() is racy if the caller has no reference against
1865 * page->mapping->host, and if the page is unlocked. This is because another
1866 * CPU could truncate the page off the mapping and then free the mapping.
1868 * Usually, the page _is_ locked, or the caller is a user-space process which
1869 * holds a reference on the inode by having an open file.
1871 * In other cases, the page should be locked before running set_page_dirty().
1873 int set_page_dirty_lock(struct page *page)
1878 ret = set_page_dirty(page);
1882 EXPORT_SYMBOL(set_page_dirty_lock);
1885 * Clear a page's dirty flag, while caring for dirty memory accounting.
1886 * Returns true if the page was previously dirty.
1888 * This is for preparing to put the page under writeout. We leave the page
1889 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1890 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1891 * implementation will run either set_page_writeback() or set_page_dirty(),
1892 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1895 * This incoherency between the page's dirty flag and radix-tree tag is
1896 * unfortunate, but it only exists while the page is locked.
1898 int clear_page_dirty_for_io(struct page *page)
1900 struct address_space *mapping = page_mapping(page);
1902 BUG_ON(!PageLocked(page));
1904 if (mapping && mapping_cap_account_dirty(mapping)) {
1906 * Yes, Virginia, this is indeed insane.
1908 * We use this sequence to make sure that
1909 * (a) we account for dirty stats properly
1910 * (b) we tell the low-level filesystem to
1911 * mark the whole page dirty if it was
1912 * dirty in a pagetable. Only to then
1913 * (c) clean the page again and return 1 to
1914 * cause the writeback.
1916 * This way we avoid all nasty races with the
1917 * dirty bit in multiple places and clearing
1918 * them concurrently from different threads.
1920 * Note! Normally the "set_page_dirty(page)"
1921 * has no effect on the actual dirty bit - since
1922 * that will already usually be set. But we
1923 * need the side effects, and it can help us
1926 * We basically use the page "master dirty bit"
1927 * as a serialization point for all the different
1928 * threads doing their things.
1930 if (page_mkclean(page))
1931 set_page_dirty(page);
1933 * We carefully synchronise fault handlers against
1934 * installing a dirty pte and marking the page dirty
1935 * at this point. We do this by having them hold the
1936 * page lock while dirtying the page, and pages are
1937 * always locked coming in here, so we get the desired
1940 if (TestClearPageDirty(page)) {
1941 dec_zone_page_state(page, NR_FILE_DIRTY);
1942 dec_bdi_stat(mapping->backing_dev_info,
1948 return TestClearPageDirty(page);
1950 EXPORT_SYMBOL(clear_page_dirty_for_io);
1952 int test_clear_page_writeback(struct page *page)
1954 struct address_space *mapping = page_mapping(page);
1958 struct backing_dev_info *bdi = mapping->backing_dev_info;
1959 unsigned long flags;
1961 spin_lock_irqsave(&mapping->tree_lock, flags);
1962 ret = TestClearPageWriteback(page);
1964 radix_tree_tag_clear(&mapping->page_tree,
1966 PAGECACHE_TAG_WRITEBACK);
1967 if (bdi_cap_account_writeback(bdi)) {
1968 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1969 __bdi_writeout_inc(bdi);
1972 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1974 ret = TestClearPageWriteback(page);
1977 dec_zone_page_state(page, NR_WRITEBACK);
1978 inc_zone_page_state(page, NR_WRITTEN);
1983 int test_set_page_writeback(struct page *page)
1985 struct address_space *mapping = page_mapping(page);
1989 struct backing_dev_info *bdi = mapping->backing_dev_info;
1990 unsigned long flags;
1992 spin_lock_irqsave(&mapping->tree_lock, flags);
1993 ret = TestSetPageWriteback(page);
1995 radix_tree_tag_set(&mapping->page_tree,
1997 PAGECACHE_TAG_WRITEBACK);
1998 if (bdi_cap_account_writeback(bdi))
1999 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2001 if (!PageDirty(page))
2002 radix_tree_tag_clear(&mapping->page_tree,
2004 PAGECACHE_TAG_DIRTY);
2005 radix_tree_tag_clear(&mapping->page_tree,
2007 PAGECACHE_TAG_TOWRITE);
2008 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2010 ret = TestSetPageWriteback(page);
2013 account_page_writeback(page);
2017 EXPORT_SYMBOL(test_set_page_writeback);
2020 * Return true if any of the pages in the mapping are marked with the
2023 int mapping_tagged(struct address_space *mapping, int tag)
2025 return radix_tree_tagged(&mapping->page_tree, tag);
2027 EXPORT_SYMBOL(mapping_tagged);