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;
131 static struct prop_descriptor vm_dirties;
134 * couple the period to the dirty_ratio:
136 * period/2 ~ roundup_pow_of_two(dirty limit)
138 static int calc_period_shift(void)
140 unsigned long dirty_total;
143 dirty_total = vm_dirty_bytes / PAGE_SIZE;
145 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
147 return 2 + ilog2(dirty_total - 1);
151 * update the period when the dirty threshold changes.
153 static void update_completion_period(void)
155 int shift = calc_period_shift();
156 prop_change_shift(&vm_completions, shift);
157 prop_change_shift(&vm_dirties, shift);
159 writeback_set_ratelimit();
162 int dirty_background_ratio_handler(struct ctl_table *table, int write,
163 void __user *buffer, size_t *lenp,
168 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
169 if (ret == 0 && write)
170 dirty_background_bytes = 0;
174 int dirty_background_bytes_handler(struct ctl_table *table, int write,
175 void __user *buffer, size_t *lenp,
180 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
181 if (ret == 0 && write)
182 dirty_background_ratio = 0;
186 int dirty_ratio_handler(struct ctl_table *table, int write,
187 void __user *buffer, size_t *lenp,
190 int old_ratio = vm_dirty_ratio;
193 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
194 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
195 update_completion_period();
202 int dirty_bytes_handler(struct ctl_table *table, int write,
203 void __user *buffer, size_t *lenp,
206 unsigned long old_bytes = vm_dirty_bytes;
209 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
210 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
211 update_completion_period();
218 * Increment the BDI's writeout completion count and the global writeout
219 * completion count. Called from test_clear_page_writeback().
221 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
223 __inc_bdi_stat(bdi, BDI_WRITTEN);
224 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
228 void bdi_writeout_inc(struct backing_dev_info *bdi)
232 local_irq_save(flags);
233 __bdi_writeout_inc(bdi);
234 local_irq_restore(flags);
236 EXPORT_SYMBOL_GPL(bdi_writeout_inc);
238 void task_dirty_inc(struct task_struct *tsk)
240 prop_inc_single(&vm_dirties, &tsk->dirties);
244 * Obtain an accurate fraction of the BDI's portion.
246 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
247 long *numerator, long *denominator)
249 prop_fraction_percpu(&vm_completions, &bdi->completions,
250 numerator, denominator);
254 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
255 * registered backing devices, which, for obvious reasons, can not
258 static unsigned int bdi_min_ratio;
260 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
264 spin_lock_bh(&bdi_lock);
265 if (min_ratio > bdi->max_ratio) {
268 min_ratio -= bdi->min_ratio;
269 if (bdi_min_ratio + min_ratio < 100) {
270 bdi_min_ratio += min_ratio;
271 bdi->min_ratio += min_ratio;
276 spin_unlock_bh(&bdi_lock);
281 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
288 spin_lock_bh(&bdi_lock);
289 if (bdi->min_ratio > max_ratio) {
292 bdi->max_ratio = max_ratio;
293 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
295 spin_unlock_bh(&bdi_lock);
299 EXPORT_SYMBOL(bdi_set_max_ratio);
302 * Work out the current dirty-memory clamping and background writeout
305 * The main aim here is to lower them aggressively if there is a lot of mapped
306 * memory around. To avoid stressing page reclaim with lots of unreclaimable
307 * pages. It is better to clamp down on writers than to start swapping, and
308 * performing lots of scanning.
310 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
312 * We don't permit the clamping level to fall below 5% - that is getting rather
315 * We make sure that the background writeout level is below the adjusted
319 static unsigned long highmem_dirtyable_memory(unsigned long total)
321 #ifdef CONFIG_HIGHMEM
325 for_each_node_state(node, N_HIGH_MEMORY) {
327 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
329 x += zone_page_state(z, NR_FREE_PAGES) +
330 zone_reclaimable_pages(z);
333 * Make sure that the number of highmem pages is never larger
334 * than the number of the total dirtyable memory. This can only
335 * occur in very strange VM situations but we want to make sure
336 * that this does not occur.
338 return min(x, total);
345 * determine_dirtyable_memory - amount of memory that may be used
347 * Returns the numebr of pages that can currently be freed and used
348 * by the kernel for direct mappings.
350 unsigned long determine_dirtyable_memory(void)
354 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
356 if (!vm_highmem_is_dirtyable)
357 x -= highmem_dirtyable_memory(x);
359 return x + 1; /* Ensure that we never return 0 */
362 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
363 unsigned long bg_thresh)
365 return (thresh + bg_thresh) / 2;
368 static unsigned long hard_dirty_limit(unsigned long thresh)
370 return max(thresh, global_dirty_limit);
374 * global_dirty_limits - background-writeback and dirty-throttling thresholds
376 * Calculate the dirty thresholds based on sysctl parameters
377 * - vm.dirty_background_ratio or vm.dirty_background_bytes
378 * - vm.dirty_ratio or vm.dirty_bytes
379 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
382 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
384 unsigned long background;
386 unsigned long uninitialized_var(available_memory);
387 struct task_struct *tsk;
389 if (!vm_dirty_bytes || !dirty_background_bytes)
390 available_memory = determine_dirtyable_memory();
393 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
395 dirty = (vm_dirty_ratio * available_memory) / 100;
397 if (dirty_background_bytes)
398 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
400 background = (dirty_background_ratio * available_memory) / 100;
402 if (background >= dirty)
403 background = dirty / 2;
405 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
406 background += background / 4;
409 *pbackground = background;
411 trace_global_dirty_state(background, dirty);
415 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
416 * @bdi: the backing_dev_info to query
417 * @dirty: global dirty limit in pages
419 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
420 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
421 * And the "limit" in the name is not seriously taken as hard limit in
422 * balance_dirty_pages().
424 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
425 * - starving fast devices
426 * - piling up dirty pages (that will take long time to sync) on slow devices
428 * The bdi's share of dirty limit will be adapting to its throughput and
429 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
431 unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
434 long numerator, denominator;
437 * Calculate this BDI's share of the dirty ratio.
439 bdi_writeout_fraction(bdi, &numerator, &denominator);
441 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
442 bdi_dirty *= numerator;
443 do_div(bdi_dirty, denominator);
445 bdi_dirty += (dirty * bdi->min_ratio) / 100;
446 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
447 bdi_dirty = dirty * bdi->max_ratio / 100;
453 * Dirty position control.
455 * (o) global/bdi setpoints
457 * We want the dirty pages be balanced around the global/bdi setpoints.
458 * When the number of dirty pages is higher/lower than the setpoint, the
459 * dirty position control ratio (and hence task dirty ratelimit) will be
460 * decreased/increased to bring the dirty pages back to the setpoint.
462 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
464 * if (dirty < setpoint) scale up pos_ratio
465 * if (dirty > setpoint) scale down pos_ratio
467 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
468 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
470 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
472 * (o) global control line
476 * | |<===== global dirty control scope ======>|
484 * 1.0 ................................*
490 * 0 +------------.------------------.----------------------*------------->
491 * freerun^ setpoint^ limit^ dirty pages
493 * (o) bdi control line
501 * | * |<=========== span ============>|
502 * 1.0 .......................*
514 * 1/4 ...............................................* * * * * * * * * * * *
518 * 0 +----------------------.-------------------------------.------------->
519 * bdi_setpoint^ x_intercept^
521 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
522 * be smoothly throttled down to normal if it starts high in situations like
523 * - start writing to a slow SD card and a fast disk at the same time. The SD
524 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
525 * - the bdi dirty thresh drops quickly due to change of JBOD workload
527 static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
528 unsigned long thresh,
529 unsigned long bg_thresh,
531 unsigned long bdi_thresh,
532 unsigned long bdi_dirty)
534 unsigned long write_bw = bdi->avg_write_bandwidth;
535 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
536 unsigned long limit = hard_dirty_limit(thresh);
537 unsigned long x_intercept;
538 unsigned long setpoint; /* dirty pages' target balance point */
539 unsigned long bdi_setpoint;
541 long long pos_ratio; /* for scaling up/down the rate limit */
544 if (unlikely(dirty >= limit))
551 * f(dirty) := 1.0 + (----------------)
554 * it's a 3rd order polynomial that subjects to
556 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
557 * (2) f(setpoint) = 1.0 => the balance point
558 * (3) f(limit) = 0 => the hard limit
559 * (4) df/dx <= 0 => negative feedback control
560 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
561 * => fast response on large errors; small oscillation near setpoint
563 setpoint = (freerun + limit) / 2;
564 x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
565 limit - setpoint + 1);
567 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
568 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
569 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
572 * We have computed basic pos_ratio above based on global situation. If
573 * the bdi is over/under its share of dirty pages, we want to scale
574 * pos_ratio further down/up. That is done by the following mechanism.
580 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
582 * x_intercept - bdi_dirty
583 * := --------------------------
584 * x_intercept - bdi_setpoint
586 * The main bdi control line is a linear function that subjects to
588 * (1) f(bdi_setpoint) = 1.0
589 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
590 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
592 * For single bdi case, the dirty pages are observed to fluctuate
593 * regularly within range
594 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
595 * for various filesystems, where (2) can yield in a reasonable 12.5%
596 * fluctuation range for pos_ratio.
598 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
599 * own size, so move the slope over accordingly and choose a slope that
600 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
602 if (unlikely(bdi_thresh > thresh))
604 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
606 * scale global setpoint to bdi's:
607 * bdi_setpoint = setpoint * bdi_thresh / thresh
609 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
610 bdi_setpoint = setpoint * (u64)x >> 16;
612 * Use span=(8*write_bw) in single bdi case as indicated by
613 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
615 * bdi_thresh thresh - bdi_thresh
616 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
619 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
620 x_intercept = bdi_setpoint + span;
622 if (bdi_dirty < x_intercept - span / 4) {
623 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
624 x_intercept - bdi_setpoint + 1);
629 * bdi reserve area, safeguard against dirty pool underrun and disk idle
630 * It may push the desired control point of global dirty pages higher
633 x_intercept = bdi_thresh / 2;
634 if (bdi_dirty < x_intercept) {
635 if (bdi_dirty > x_intercept / 8)
636 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
644 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
645 unsigned long elapsed,
646 unsigned long written)
648 const unsigned long period = roundup_pow_of_two(3 * HZ);
649 unsigned long avg = bdi->avg_write_bandwidth;
650 unsigned long old = bdi->write_bandwidth;
654 * bw = written * HZ / elapsed
656 * bw * elapsed + write_bandwidth * (period - elapsed)
657 * write_bandwidth = ---------------------------------------------------
660 bw = written - bdi->written_stamp;
662 if (unlikely(elapsed > period)) {
667 bw += (u64)bdi->write_bandwidth * (period - elapsed);
668 bw >>= ilog2(period);
671 * one more level of smoothing, for filtering out sudden spikes
673 if (avg > old && old >= (unsigned long)bw)
674 avg -= (avg - old) >> 3;
676 if (avg < old && old <= (unsigned long)bw)
677 avg += (old - avg) >> 3;
680 bdi->write_bandwidth = bw;
681 bdi->avg_write_bandwidth = avg;
685 * The global dirtyable memory and dirty threshold could be suddenly knocked
686 * down by a large amount (eg. on the startup of KVM in a swapless system).
687 * This may throw the system into deep dirty exceeded state and throttle
688 * heavy/light dirtiers alike. To retain good responsiveness, maintain
689 * global_dirty_limit for tracking slowly down to the knocked down dirty
692 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
694 unsigned long limit = global_dirty_limit;
697 * Follow up in one step.
699 if (limit < thresh) {
705 * Follow down slowly. Use the higher one as the target, because thresh
706 * may drop below dirty. This is exactly the reason to introduce
707 * global_dirty_limit which is guaranteed to lie above the dirty pages.
709 thresh = max(thresh, dirty);
710 if (limit > thresh) {
711 limit -= (limit - thresh) >> 5;
716 global_dirty_limit = limit;
719 static void global_update_bandwidth(unsigned long thresh,
723 static DEFINE_SPINLOCK(dirty_lock);
724 static unsigned long update_time;
727 * check locklessly first to optimize away locking for the most time
729 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
732 spin_lock(&dirty_lock);
733 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
734 update_dirty_limit(thresh, dirty);
737 spin_unlock(&dirty_lock);
741 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
743 * Normal bdi tasks will be curbed at or below it in long term.
744 * Obviously it should be around (write_bw / N) when there are N dd tasks.
746 static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
747 unsigned long thresh,
748 unsigned long bg_thresh,
750 unsigned long bdi_thresh,
751 unsigned long bdi_dirty,
752 unsigned long dirtied,
753 unsigned long elapsed)
755 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
756 unsigned long limit = hard_dirty_limit(thresh);
757 unsigned long setpoint = (freerun + limit) / 2;
758 unsigned long write_bw = bdi->avg_write_bandwidth;
759 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
760 unsigned long dirty_rate;
761 unsigned long task_ratelimit;
762 unsigned long balanced_dirty_ratelimit;
763 unsigned long pos_ratio;
768 * The dirty rate will match the writeout rate in long term, except
769 * when dirty pages are truncated by userspace or re-dirtied by FS.
771 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
773 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
774 bdi_thresh, bdi_dirty);
776 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
778 task_ratelimit = (u64)dirty_ratelimit *
779 pos_ratio >> RATELIMIT_CALC_SHIFT;
780 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
783 * A linear estimation of the "balanced" throttle rate. The theory is,
784 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
785 * dirty_rate will be measured to be (N * task_ratelimit). So the below
786 * formula will yield the balanced rate limit (write_bw / N).
788 * Note that the expanded form is not a pure rate feedback:
789 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
790 * but also takes pos_ratio into account:
791 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
793 * (1) is not realistic because pos_ratio also takes part in balancing
794 * the dirty rate. Consider the state
795 * pos_ratio = 0.5 (3)
796 * rate = 2 * (write_bw / N) (4)
797 * If (1) is used, it will stuck in that state! Because each dd will
799 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
801 * dirty_rate = N * task_ratelimit = write_bw (6)
802 * put (6) into (1) we get
803 * rate_(i+1) = rate_(i) (7)
805 * So we end up using (2) to always keep
806 * rate_(i+1) ~= (write_bw / N) (8)
807 * regardless of the value of pos_ratio. As long as (8) is satisfied,
808 * pos_ratio is able to drive itself to 1.0, which is not only where
809 * the dirty count meet the setpoint, but also where the slope of
810 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
812 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
816 * We could safely do this and return immediately:
818 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
820 * However to get a more stable dirty_ratelimit, the below elaborated
821 * code makes use of task_ratelimit to filter out sigular points and
822 * limit the step size.
824 * The below code essentially only uses the relative value of
826 * task_ratelimit - dirty_ratelimit
827 * = (pos_ratio - 1) * dirty_ratelimit
829 * which reflects the direction and size of dirty position error.
833 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
834 * task_ratelimit is on the same side of dirty_ratelimit, too.
836 * - dirty_ratelimit > balanced_dirty_ratelimit
837 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
838 * lowering dirty_ratelimit will help meet both the position and rate
839 * control targets. Otherwise, don't update dirty_ratelimit if it will
840 * only help meet the rate target. After all, what the users ultimately
841 * feel and care are stable dirty rate and small position error.
843 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
844 * and filter out the sigular points of balanced_dirty_ratelimit. Which
845 * keeps jumping around randomly and can even leap far away at times
846 * due to the small 200ms estimation period of dirty_rate (we want to
847 * keep that period small to reduce time lags).
850 if (dirty < setpoint) {
851 x = min(bdi->balanced_dirty_ratelimit,
852 min(balanced_dirty_ratelimit, task_ratelimit));
853 if (dirty_ratelimit < x)
854 step = x - dirty_ratelimit;
856 x = max(bdi->balanced_dirty_ratelimit,
857 max(balanced_dirty_ratelimit, task_ratelimit));
858 if (dirty_ratelimit > x)
859 step = dirty_ratelimit - x;
863 * Don't pursue 100% rate matching. It's impossible since the balanced
864 * rate itself is constantly fluctuating. So decrease the track speed
865 * when it gets close to the target. Helps eliminate pointless tremors.
867 step >>= dirty_ratelimit / (2 * step + 1);
869 * Limit the tracking speed to avoid overshooting.
871 step = (step + 7) / 8;
873 if (dirty_ratelimit < balanced_dirty_ratelimit)
874 dirty_ratelimit += step;
876 dirty_ratelimit -= step;
878 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
879 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
881 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
884 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
885 unsigned long thresh,
886 unsigned long bg_thresh,
888 unsigned long bdi_thresh,
889 unsigned long bdi_dirty,
890 unsigned long start_time)
892 unsigned long now = jiffies;
893 unsigned long elapsed = now - bdi->bw_time_stamp;
894 unsigned long dirtied;
895 unsigned long written;
898 * rate-limit, only update once every 200ms.
900 if (elapsed < BANDWIDTH_INTERVAL)
903 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
904 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
907 * Skip quiet periods when disk bandwidth is under-utilized.
908 * (at least 1s idle time between two flusher runs)
910 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
914 global_update_bandwidth(thresh, dirty, now);
915 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
916 bdi_thresh, bdi_dirty,
919 bdi_update_write_bandwidth(bdi, elapsed, written);
922 bdi->dirtied_stamp = dirtied;
923 bdi->written_stamp = written;
924 bdi->bw_time_stamp = now;
927 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
928 unsigned long thresh,
929 unsigned long bg_thresh,
931 unsigned long bdi_thresh,
932 unsigned long bdi_dirty,
933 unsigned long start_time)
935 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
937 spin_lock(&bdi->wb.list_lock);
938 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
939 bdi_thresh, bdi_dirty, start_time);
940 spin_unlock(&bdi->wb.list_lock);
944 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
945 * will look to see if it needs to start dirty throttling.
947 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
948 * global_page_state() too often. So scale it near-sqrt to the safety margin
949 * (the number of pages we may dirty without exceeding the dirty limits).
951 static unsigned long dirty_poll_interval(unsigned long dirty,
952 unsigned long thresh)
955 return 1UL << (ilog2(thresh - dirty) >> 1);
960 static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
961 unsigned long bdi_dirty)
963 unsigned long bw = bdi->avg_write_bandwidth;
964 unsigned long hi = ilog2(bw);
965 unsigned long lo = ilog2(bdi->dirty_ratelimit);
968 /* target for 20ms max pause on 1-dd case */
972 * Scale up pause time for concurrent dirtiers in order to reduce CPU
975 * (N * 20ms) on 2^N concurrent tasks.
978 t += (hi - lo) * (20 * HZ) / 1024;
981 * Limit pause time for small memory systems. If sleeping for too long
982 * time, a small pool of dirty/writeback pages may go empty and disk go
985 * 8 serves as the safety ratio.
988 t = min(t, bdi_dirty * HZ / (8 * bw + 1));
991 * The pause time will be settled within range (max_pause/4, max_pause).
992 * Apply a minimal value of 4 to get a non-zero max_pause/4.
994 return clamp_val(t, 4, MAX_PAUSE);
998 * balance_dirty_pages() must be called by processes which are generating dirty
999 * data. It looks at the number of dirty pages in the machine and will force
1000 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1001 * If we're over `background_thresh' then the writeback threads are woken to
1002 * perform some writeout.
1004 static void balance_dirty_pages(struct address_space *mapping,
1005 unsigned long pages_dirtied)
1007 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1008 unsigned long bdi_reclaimable;
1009 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1010 unsigned long bdi_dirty;
1011 unsigned long freerun;
1012 unsigned long background_thresh;
1013 unsigned long dirty_thresh;
1014 unsigned long bdi_thresh;
1016 long uninitialized_var(max_pause);
1017 bool dirty_exceeded = false;
1018 unsigned long task_ratelimit;
1019 unsigned long uninitialized_var(dirty_ratelimit);
1020 unsigned long pos_ratio;
1021 struct backing_dev_info *bdi = mapping->backing_dev_info;
1022 unsigned long start_time = jiffies;
1026 * Unstable writes are a feature of certain networked
1027 * filesystems (i.e. NFS) in which data may have been
1028 * written to the server's write cache, but has not yet
1029 * been flushed to permanent storage.
1031 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1032 global_page_state(NR_UNSTABLE_NFS);
1033 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1035 global_dirty_limits(&background_thresh, &dirty_thresh);
1038 * Throttle it only when the background writeback cannot
1039 * catch-up. This avoids (excessively) small writeouts
1040 * when the bdi limits are ramping up.
1042 freerun = dirty_freerun_ceiling(dirty_thresh,
1044 if (nr_dirty <= freerun)
1047 if (unlikely(!writeback_in_progress(bdi)))
1048 bdi_start_background_writeback(bdi);
1051 * bdi_thresh is not treated as some limiting factor as
1052 * dirty_thresh, due to reasons
1053 * - in JBOD setup, bdi_thresh can fluctuate a lot
1054 * - in a system with HDD and USB key, the USB key may somehow
1055 * go into state (bdi_dirty >> bdi_thresh) either because
1056 * bdi_dirty starts high, or because bdi_thresh drops low.
1057 * In this case we don't want to hard throttle the USB key
1058 * dirtiers for 100 seconds until bdi_dirty drops under
1059 * bdi_thresh. Instead the auxiliary bdi control line in
1060 * bdi_position_ratio() will let the dirtier task progress
1061 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1063 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1066 * In order to avoid the stacked BDI deadlock we need
1067 * to ensure we accurately count the 'dirty' pages when
1068 * the threshold is low.
1070 * Otherwise it would be possible to get thresh+n pages
1071 * reported dirty, even though there are thresh-m pages
1072 * actually dirty; with m+n sitting in the percpu
1075 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1076 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1077 bdi_dirty = bdi_reclaimable +
1078 bdi_stat_sum(bdi, BDI_WRITEBACK);
1080 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1081 bdi_dirty = bdi_reclaimable +
1082 bdi_stat(bdi, BDI_WRITEBACK);
1085 dirty_exceeded = (bdi_dirty > bdi_thresh) ||
1086 (nr_dirty > dirty_thresh);
1087 if (dirty_exceeded && !bdi->dirty_exceeded)
1088 bdi->dirty_exceeded = 1;
1090 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1091 nr_dirty, bdi_thresh, bdi_dirty,
1094 max_pause = bdi_max_pause(bdi, bdi_dirty);
1096 dirty_ratelimit = bdi->dirty_ratelimit;
1097 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1098 background_thresh, nr_dirty,
1099 bdi_thresh, bdi_dirty);
1100 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1101 RATELIMIT_CALC_SHIFT;
1102 if (unlikely(task_ratelimit == 0)) {
1106 pause = HZ * pages_dirtied / task_ratelimit;
1107 if (unlikely(pause <= 0)) {
1108 trace_balance_dirty_pages(bdi,
1119 pause = 1; /* avoid resetting nr_dirtied_pause below */
1122 pause = min(pause, max_pause);
1125 trace_balance_dirty_pages(bdi,
1136 __set_current_state(TASK_KILLABLE);
1137 io_schedule_timeout(pause);
1139 dirty_thresh = hard_dirty_limit(dirty_thresh);
1141 * max-pause area. If dirty exceeded but still within this
1142 * area, no need to sleep for more than 200ms: (a) 8 pages per
1143 * 200ms is typically more than enough to curb heavy dirtiers;
1144 * (b) the pause time limit makes the dirtiers more responsive.
1146 if (nr_dirty < dirty_thresh)
1149 if (fatal_signal_pending(current))
1153 if (!dirty_exceeded && bdi->dirty_exceeded)
1154 bdi->dirty_exceeded = 0;
1156 current->nr_dirtied = 0;
1157 if (pause == 0) { /* in freerun area */
1158 current->nr_dirtied_pause =
1159 dirty_poll_interval(nr_dirty, dirty_thresh);
1160 } else if (pause <= max_pause / 4 &&
1161 pages_dirtied >= current->nr_dirtied_pause) {
1162 current->nr_dirtied_pause = clamp_val(
1163 dirty_ratelimit * (max_pause / 2) / HZ,
1164 pages_dirtied + pages_dirtied / 8,
1166 } else if (pause >= max_pause) {
1167 current->nr_dirtied_pause = 1 | clamp_val(
1168 dirty_ratelimit * (max_pause / 2) / HZ,
1170 pages_dirtied - pages_dirtied / 8);
1173 if (writeback_in_progress(bdi))
1177 * In laptop mode, we wait until hitting the higher threshold before
1178 * starting background writeout, and then write out all the way down
1179 * to the lower threshold. So slow writers cause minimal disk activity.
1181 * In normal mode, we start background writeout at the lower
1182 * background_thresh, to keep the amount of dirty memory low.
1187 if (nr_reclaimable > background_thresh)
1188 bdi_start_background_writeback(bdi);
1191 void set_page_dirty_balance(struct page *page, int page_mkwrite)
1193 if (set_page_dirty(page) || page_mkwrite) {
1194 struct address_space *mapping = page_mapping(page);
1197 balance_dirty_pages_ratelimited(mapping);
1201 static DEFINE_PER_CPU(int, bdp_ratelimits);
1204 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
1205 * @mapping: address_space which was dirtied
1206 * @nr_pages_dirtied: number of pages which the caller has just dirtied
1208 * Processes which are dirtying memory should call in here once for each page
1209 * which was newly dirtied. The function will periodically check the system's
1210 * dirty state and will initiate writeback if needed.
1212 * On really big machines, get_writeback_state is expensive, so try to avoid
1213 * calling it too often (ratelimiting). But once we're over the dirty memory
1214 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1215 * from overshooting the limit by (ratelimit_pages) each.
1217 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1218 unsigned long nr_pages_dirtied)
1220 struct backing_dev_info *bdi = mapping->backing_dev_info;
1224 if (!bdi_cap_account_dirty(bdi))
1227 ratelimit = current->nr_dirtied_pause;
1228 if (bdi->dirty_exceeded)
1229 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1231 current->nr_dirtied += nr_pages_dirtied;
1235 * This prevents one CPU to accumulate too many dirtied pages without
1236 * calling into balance_dirty_pages(), which can happen when there are
1237 * 1000+ tasks, all of them start dirtying pages at exactly the same
1238 * time, hence all honoured too large initial task->nr_dirtied_pause.
1240 p = &__get_cpu_var(bdp_ratelimits);
1241 if (unlikely(current->nr_dirtied >= ratelimit))
1244 *p += nr_pages_dirtied;
1245 if (unlikely(*p >= ratelimit_pages)) {
1252 if (unlikely(current->nr_dirtied >= ratelimit))
1253 balance_dirty_pages(mapping, current->nr_dirtied);
1255 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
1257 void throttle_vm_writeout(gfp_t gfp_mask)
1259 unsigned long background_thresh;
1260 unsigned long dirty_thresh;
1263 global_dirty_limits(&background_thresh, &dirty_thresh);
1266 * Boost the allowable dirty threshold a bit for page
1267 * allocators so they don't get DoS'ed by heavy writers
1269 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1271 if (global_page_state(NR_UNSTABLE_NFS) +
1272 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1274 congestion_wait(BLK_RW_ASYNC, HZ/10);
1277 * The caller might hold locks which can prevent IO completion
1278 * or progress in the filesystem. So we cannot just sit here
1279 * waiting for IO to complete.
1281 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1287 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1289 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1290 void __user *buffer, size_t *length, loff_t *ppos)
1292 proc_dointvec(table, write, buffer, length, ppos);
1293 bdi_arm_supers_timer();
1298 void laptop_mode_timer_fn(unsigned long data)
1300 struct request_queue *q = (struct request_queue *)data;
1301 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1302 global_page_state(NR_UNSTABLE_NFS);
1305 * We want to write everything out, not just down to the dirty
1308 if (bdi_has_dirty_io(&q->backing_dev_info))
1309 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1310 WB_REASON_LAPTOP_TIMER);
1314 * We've spun up the disk and we're in laptop mode: schedule writeback
1315 * of all dirty data a few seconds from now. If the flush is already scheduled
1316 * then push it back - the user is still using the disk.
1318 void laptop_io_completion(struct backing_dev_info *info)
1320 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1324 * We're in laptop mode and we've just synced. The sync's writes will have
1325 * caused another writeback to be scheduled by laptop_io_completion.
1326 * Nothing needs to be written back anymore, so we unschedule the writeback.
1328 void laptop_sync_completion(void)
1330 struct backing_dev_info *bdi;
1334 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1335 del_timer(&bdi->laptop_mode_wb_timer);
1342 * If ratelimit_pages is too high then we can get into dirty-data overload
1343 * if a large number of processes all perform writes at the same time.
1344 * If it is too low then SMP machines will call the (expensive)
1345 * get_writeback_state too often.
1347 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1348 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1352 void writeback_set_ratelimit(void)
1354 unsigned long background_thresh;
1355 unsigned long dirty_thresh;
1356 global_dirty_limits(&background_thresh, &dirty_thresh);
1357 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1358 if (ratelimit_pages < 16)
1359 ratelimit_pages = 16;
1362 static int __cpuinit
1363 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1365 writeback_set_ratelimit();
1369 static struct notifier_block __cpuinitdata ratelimit_nb = {
1370 .notifier_call = ratelimit_handler,
1375 * Called early on to tune the page writeback dirty limits.
1377 * We used to scale dirty pages according to how total memory
1378 * related to pages that could be allocated for buffers (by
1379 * comparing nr_free_buffer_pages() to vm_total_pages.
1381 * However, that was when we used "dirty_ratio" to scale with
1382 * all memory, and we don't do that any more. "dirty_ratio"
1383 * is now applied to total non-HIGHPAGE memory (by subtracting
1384 * totalhigh_pages from vm_total_pages), and as such we can't
1385 * get into the old insane situation any more where we had
1386 * large amounts of dirty pages compared to a small amount of
1387 * non-HIGHMEM memory.
1389 * But we might still want to scale the dirty_ratio by how
1390 * much memory the box has..
1392 void __init page_writeback_init(void)
1396 writeback_set_ratelimit();
1397 register_cpu_notifier(&ratelimit_nb);
1399 shift = calc_period_shift();
1400 prop_descriptor_init(&vm_completions, shift);
1401 prop_descriptor_init(&vm_dirties, shift);
1405 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1406 * @mapping: address space structure to write
1407 * @start: starting page index
1408 * @end: ending page index (inclusive)
1410 * This function scans the page range from @start to @end (inclusive) and tags
1411 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1412 * that write_cache_pages (or whoever calls this function) will then use
1413 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1414 * used to avoid livelocking of writeback by a process steadily creating new
1415 * dirty pages in the file (thus it is important for this function to be quick
1416 * so that it can tag pages faster than a dirtying process can create them).
1419 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1421 void tag_pages_for_writeback(struct address_space *mapping,
1422 pgoff_t start, pgoff_t end)
1424 #define WRITEBACK_TAG_BATCH 4096
1425 unsigned long tagged;
1428 spin_lock_irq(&mapping->tree_lock);
1429 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1430 &start, end, WRITEBACK_TAG_BATCH,
1431 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1432 spin_unlock_irq(&mapping->tree_lock);
1433 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1435 /* We check 'start' to handle wrapping when end == ~0UL */
1436 } while (tagged >= WRITEBACK_TAG_BATCH && start);
1438 EXPORT_SYMBOL(tag_pages_for_writeback);
1441 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1442 * @mapping: address space structure to write
1443 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1444 * @writepage: function called for each page
1445 * @data: data passed to writepage function
1447 * If a page is already under I/O, write_cache_pages() skips it, even
1448 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1449 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1450 * and msync() need to guarantee that all the data which was dirty at the time
1451 * the call was made get new I/O started against them. If wbc->sync_mode is
1452 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1453 * existing IO to complete.
1455 * To avoid livelocks (when other process dirties new pages), we first tag
1456 * pages which should be written back with TOWRITE tag and only then start
1457 * writing them. For data-integrity sync we have to be careful so that we do
1458 * not miss some pages (e.g., because some other process has cleared TOWRITE
1459 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1460 * by the process clearing the DIRTY tag (and submitting the page for IO).
1462 int write_cache_pages(struct address_space *mapping,
1463 struct writeback_control *wbc, writepage_t writepage,
1468 struct pagevec pvec;
1470 pgoff_t uninitialized_var(writeback_index);
1472 pgoff_t end; /* Inclusive */
1475 int range_whole = 0;
1478 pagevec_init(&pvec, 0);
1479 if (wbc->range_cyclic) {
1480 writeback_index = mapping->writeback_index; /* prev offset */
1481 index = writeback_index;
1488 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1489 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1490 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1492 cycled = 1; /* ignore range_cyclic tests */
1494 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1495 tag = PAGECACHE_TAG_TOWRITE;
1497 tag = PAGECACHE_TAG_DIRTY;
1499 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1500 tag_pages_for_writeback(mapping, index, end);
1502 while (!done && (index <= end)) {
1505 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1506 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1510 for (i = 0; i < nr_pages; i++) {
1511 struct page *page = pvec.pages[i];
1514 * At this point, the page may be truncated or
1515 * invalidated (changing page->mapping to NULL), or
1516 * even swizzled back from swapper_space to tmpfs file
1517 * mapping. However, page->index will not change
1518 * because we have a reference on the page.
1520 if (page->index > end) {
1522 * can't be range_cyclic (1st pass) because
1523 * end == -1 in that case.
1529 done_index = page->index;
1534 * Page truncated or invalidated. We can freely skip it
1535 * then, even for data integrity operations: the page
1536 * has disappeared concurrently, so there could be no
1537 * real expectation of this data interity operation
1538 * even if there is now a new, dirty page at the same
1539 * pagecache address.
1541 if (unlikely(page->mapping != mapping)) {
1547 if (!PageDirty(page)) {
1548 /* someone wrote it for us */
1549 goto continue_unlock;
1552 if (PageWriteback(page)) {
1553 if (wbc->sync_mode != WB_SYNC_NONE)
1554 wait_on_page_writeback(page);
1556 goto continue_unlock;
1559 BUG_ON(PageWriteback(page));
1560 if (!clear_page_dirty_for_io(page))
1561 goto continue_unlock;
1563 trace_wbc_writepage(wbc, mapping->backing_dev_info);
1564 ret = (*writepage)(page, wbc, data);
1565 if (unlikely(ret)) {
1566 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1571 * done_index is set past this page,
1572 * so media errors will not choke
1573 * background writeout for the entire
1574 * file. This has consequences for
1575 * range_cyclic semantics (ie. it may
1576 * not be suitable for data integrity
1579 done_index = page->index + 1;
1586 * We stop writing back only if we are not doing
1587 * integrity sync. In case of integrity sync we have to
1588 * keep going until we have written all the pages
1589 * we tagged for writeback prior to entering this loop.
1591 if (--wbc->nr_to_write <= 0 &&
1592 wbc->sync_mode == WB_SYNC_NONE) {
1597 pagevec_release(&pvec);
1600 if (!cycled && !done) {
1603 * We hit the last page and there is more work to be done: wrap
1604 * back to the start of the file
1608 end = writeback_index - 1;
1611 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1612 mapping->writeback_index = done_index;
1616 EXPORT_SYMBOL(write_cache_pages);
1619 * Function used by generic_writepages to call the real writepage
1620 * function and set the mapping flags on error
1622 static int __writepage(struct page *page, struct writeback_control *wbc,
1625 struct address_space *mapping = data;
1626 int ret = mapping->a_ops->writepage(page, wbc);
1627 mapping_set_error(mapping, ret);
1632 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1633 * @mapping: address space structure to write
1634 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1636 * This is a library function, which implements the writepages()
1637 * address_space_operation.
1639 int generic_writepages(struct address_space *mapping,
1640 struct writeback_control *wbc)
1642 struct blk_plug plug;
1645 /* deal with chardevs and other special file */
1646 if (!mapping->a_ops->writepage)
1649 blk_start_plug(&plug);
1650 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1651 blk_finish_plug(&plug);
1655 EXPORT_SYMBOL(generic_writepages);
1657 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1661 if (wbc->nr_to_write <= 0)
1663 if (mapping->a_ops->writepages)
1664 ret = mapping->a_ops->writepages(mapping, wbc);
1666 ret = generic_writepages(mapping, wbc);
1671 * write_one_page - write out a single page and optionally wait on I/O
1672 * @page: the page to write
1673 * @wait: if true, wait on writeout
1675 * The page must be locked by the caller and will be unlocked upon return.
1677 * write_one_page() returns a negative error code if I/O failed.
1679 int write_one_page(struct page *page, int wait)
1681 struct address_space *mapping = page->mapping;
1683 struct writeback_control wbc = {
1684 .sync_mode = WB_SYNC_ALL,
1688 BUG_ON(!PageLocked(page));
1691 wait_on_page_writeback(page);
1693 if (clear_page_dirty_for_io(page)) {
1694 page_cache_get(page);
1695 ret = mapping->a_ops->writepage(page, &wbc);
1696 if (ret == 0 && wait) {
1697 wait_on_page_writeback(page);
1698 if (PageError(page))
1701 page_cache_release(page);
1707 EXPORT_SYMBOL(write_one_page);
1710 * For address_spaces which do not use buffers nor write back.
1712 int __set_page_dirty_no_writeback(struct page *page)
1714 if (!PageDirty(page))
1715 return !TestSetPageDirty(page);
1720 * Helper function for set_page_dirty family.
1721 * NOTE: This relies on being atomic wrt interrupts.
1723 void account_page_dirtied(struct page *page, struct address_space *mapping)
1725 if (mapping_cap_account_dirty(mapping)) {
1726 __inc_zone_page_state(page, NR_FILE_DIRTY);
1727 __inc_zone_page_state(page, NR_DIRTIED);
1728 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1729 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1730 task_dirty_inc(current);
1731 task_io_account_write(PAGE_CACHE_SIZE);
1734 EXPORT_SYMBOL(account_page_dirtied);
1737 * Helper function for set_page_writeback family.
1738 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1741 void account_page_writeback(struct page *page)
1743 inc_zone_page_state(page, NR_WRITEBACK);
1745 EXPORT_SYMBOL(account_page_writeback);
1748 * For address_spaces which do not use buffers. Just tag the page as dirty in
1751 * This is also used when a single buffer is being dirtied: we want to set the
1752 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1753 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1755 * Most callers have locked the page, which pins the address_space in memory.
1756 * But zap_pte_range() does not lock the page, however in that case the
1757 * mapping is pinned by the vma's ->vm_file reference.
1759 * We take care to handle the case where the page was truncated from the
1760 * mapping by re-checking page_mapping() inside tree_lock.
1762 int __set_page_dirty_nobuffers(struct page *page)
1764 if (!TestSetPageDirty(page)) {
1765 struct address_space *mapping = page_mapping(page);
1766 struct address_space *mapping2;
1771 spin_lock_irq(&mapping->tree_lock);
1772 mapping2 = page_mapping(page);
1773 if (mapping2) { /* Race with truncate? */
1774 BUG_ON(mapping2 != mapping);
1775 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1776 account_page_dirtied(page, mapping);
1777 radix_tree_tag_set(&mapping->page_tree,
1778 page_index(page), PAGECACHE_TAG_DIRTY);
1780 spin_unlock_irq(&mapping->tree_lock);
1781 if (mapping->host) {
1782 /* !PageAnon && !swapper_space */
1783 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1789 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1792 * When a writepage implementation decides that it doesn't want to write this
1793 * page for some reason, it should redirty the locked page via
1794 * redirty_page_for_writepage() and it should then unlock the page and return 0
1796 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1798 wbc->pages_skipped++;
1799 return __set_page_dirty_nobuffers(page);
1801 EXPORT_SYMBOL(redirty_page_for_writepage);
1806 * For pages with a mapping this should be done under the page lock
1807 * for the benefit of asynchronous memory errors who prefer a consistent
1808 * dirty state. This rule can be broken in some special cases,
1809 * but should be better not to.
1811 * If the mapping doesn't provide a set_page_dirty a_op, then
1812 * just fall through and assume that it wants buffer_heads.
1814 int set_page_dirty(struct page *page)
1816 struct address_space *mapping = page_mapping(page);
1818 if (likely(mapping)) {
1819 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1821 * readahead/lru_deactivate_page could remain
1822 * PG_readahead/PG_reclaim due to race with end_page_writeback
1823 * About readahead, if the page is written, the flags would be
1824 * reset. So no problem.
1825 * About lru_deactivate_page, if the page is redirty, the flag
1826 * will be reset. So no problem. but if the page is used by readahead
1827 * it will confuse readahead and make it restart the size rampup
1828 * process. But it's a trivial problem.
1830 ClearPageReclaim(page);
1833 spd = __set_page_dirty_buffers;
1835 return (*spd)(page);
1837 if (!PageDirty(page)) {
1838 if (!TestSetPageDirty(page))
1843 EXPORT_SYMBOL(set_page_dirty);
1846 * set_page_dirty() is racy if the caller has no reference against
1847 * page->mapping->host, and if the page is unlocked. This is because another
1848 * CPU could truncate the page off the mapping and then free the mapping.
1850 * Usually, the page _is_ locked, or the caller is a user-space process which
1851 * holds a reference on the inode by having an open file.
1853 * In other cases, the page should be locked before running set_page_dirty().
1855 int set_page_dirty_lock(struct page *page)
1860 ret = set_page_dirty(page);
1864 EXPORT_SYMBOL(set_page_dirty_lock);
1867 * Clear a page's dirty flag, while caring for dirty memory accounting.
1868 * Returns true if the page was previously dirty.
1870 * This is for preparing to put the page under writeout. We leave the page
1871 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1872 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1873 * implementation will run either set_page_writeback() or set_page_dirty(),
1874 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1877 * This incoherency between the page's dirty flag and radix-tree tag is
1878 * unfortunate, but it only exists while the page is locked.
1880 int clear_page_dirty_for_io(struct page *page)
1882 struct address_space *mapping = page_mapping(page);
1884 BUG_ON(!PageLocked(page));
1886 if (mapping && mapping_cap_account_dirty(mapping)) {
1888 * Yes, Virginia, this is indeed insane.
1890 * We use this sequence to make sure that
1891 * (a) we account for dirty stats properly
1892 * (b) we tell the low-level filesystem to
1893 * mark the whole page dirty if it was
1894 * dirty in a pagetable. Only to then
1895 * (c) clean the page again and return 1 to
1896 * cause the writeback.
1898 * This way we avoid all nasty races with the
1899 * dirty bit in multiple places and clearing
1900 * them concurrently from different threads.
1902 * Note! Normally the "set_page_dirty(page)"
1903 * has no effect on the actual dirty bit - since
1904 * that will already usually be set. But we
1905 * need the side effects, and it can help us
1908 * We basically use the page "master dirty bit"
1909 * as a serialization point for all the different
1910 * threads doing their things.
1912 if (page_mkclean(page))
1913 set_page_dirty(page);
1915 * We carefully synchronise fault handlers against
1916 * installing a dirty pte and marking the page dirty
1917 * at this point. We do this by having them hold the
1918 * page lock at some point after installing their
1919 * pte, but before marking the page dirty.
1920 * Pages are always locked coming in here, so we get
1921 * the desired exclusion. See mm/memory.c:do_wp_page()
1922 * for more comments.
1924 if (TestClearPageDirty(page)) {
1925 dec_zone_page_state(page, NR_FILE_DIRTY);
1926 dec_bdi_stat(mapping->backing_dev_info,
1932 return TestClearPageDirty(page);
1934 EXPORT_SYMBOL(clear_page_dirty_for_io);
1936 int test_clear_page_writeback(struct page *page)
1938 struct address_space *mapping = page_mapping(page);
1942 struct backing_dev_info *bdi = mapping->backing_dev_info;
1943 unsigned long flags;
1945 spin_lock_irqsave(&mapping->tree_lock, flags);
1946 ret = TestClearPageWriteback(page);
1948 radix_tree_tag_clear(&mapping->page_tree,
1950 PAGECACHE_TAG_WRITEBACK);
1951 if (bdi_cap_account_writeback(bdi)) {
1952 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1953 __bdi_writeout_inc(bdi);
1956 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1958 ret = TestClearPageWriteback(page);
1961 dec_zone_page_state(page, NR_WRITEBACK);
1962 inc_zone_page_state(page, NR_WRITTEN);
1967 int test_set_page_writeback(struct page *page)
1969 struct address_space *mapping = page_mapping(page);
1973 struct backing_dev_info *bdi = mapping->backing_dev_info;
1974 unsigned long flags;
1976 spin_lock_irqsave(&mapping->tree_lock, flags);
1977 ret = TestSetPageWriteback(page);
1979 radix_tree_tag_set(&mapping->page_tree,
1981 PAGECACHE_TAG_WRITEBACK);
1982 if (bdi_cap_account_writeback(bdi))
1983 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1985 if (!PageDirty(page))
1986 radix_tree_tag_clear(&mapping->page_tree,
1988 PAGECACHE_TAG_DIRTY);
1989 radix_tree_tag_clear(&mapping->page_tree,
1991 PAGECACHE_TAG_TOWRITE);
1992 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1994 ret = TestSetPageWriteback(page);
1997 account_page_writeback(page);
2001 EXPORT_SYMBOL(test_set_page_writeback);
2004 * Return true if any of the pages in the mapping are marked with the
2007 int mapping_tagged(struct address_space *mapping, int tag)
2009 return radix_tree_tagged(&mapping->page_tree, tag);
2011 EXPORT_SYMBOL(mapping_tagged);