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/module.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);
256 static unsigned int bdi_min_ratio;
258 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
262 spin_lock_bh(&bdi_lock);
263 if (min_ratio > bdi->max_ratio) {
266 min_ratio -= bdi->min_ratio;
267 if (bdi_min_ratio + min_ratio < 100) {
268 bdi_min_ratio += min_ratio;
269 bdi->min_ratio += min_ratio;
274 spin_unlock_bh(&bdi_lock);
279 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
286 spin_lock_bh(&bdi_lock);
287 if (bdi->min_ratio > max_ratio) {
290 bdi->max_ratio = max_ratio;
291 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
293 spin_unlock_bh(&bdi_lock);
297 EXPORT_SYMBOL(bdi_set_max_ratio);
300 * Work out the current dirty-memory clamping and background writeout
303 * The main aim here is to lower them aggressively if there is a lot of mapped
304 * memory around. To avoid stressing page reclaim with lots of unreclaimable
305 * pages. It is better to clamp down on writers than to start swapping, and
306 * performing lots of scanning.
308 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
310 * We don't permit the clamping level to fall below 5% - that is getting rather
313 * We make sure that the background writeout level is below the adjusted
317 static unsigned long highmem_dirtyable_memory(unsigned long total)
319 #ifdef CONFIG_HIGHMEM
323 for_each_node_state(node, N_HIGH_MEMORY) {
325 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
327 x += zone_page_state(z, NR_FREE_PAGES) +
328 zone_reclaimable_pages(z);
331 * Make sure that the number of highmem pages is never larger
332 * than the number of the total dirtyable memory. This can only
333 * occur in very strange VM situations but we want to make sure
334 * that this does not occur.
336 return min(x, total);
343 * determine_dirtyable_memory - amount of memory that may be used
345 * Returns the numebr of pages that can currently be freed and used
346 * by the kernel for direct mappings.
348 unsigned long determine_dirtyable_memory(void)
352 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
354 if (!vm_highmem_is_dirtyable)
355 x -= highmem_dirtyable_memory(x);
357 return x + 1; /* Ensure that we never return 0 */
360 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
361 unsigned long bg_thresh)
363 return (thresh + bg_thresh) / 2;
366 static unsigned long hard_dirty_limit(unsigned long thresh)
368 return max(thresh, global_dirty_limit);
372 * global_dirty_limits - background-writeback and dirty-throttling thresholds
374 * Calculate the dirty thresholds based on sysctl parameters
375 * - vm.dirty_background_ratio or vm.dirty_background_bytes
376 * - vm.dirty_ratio or vm.dirty_bytes
377 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
380 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
382 unsigned long background;
384 unsigned long uninitialized_var(available_memory);
385 struct task_struct *tsk;
387 if (!vm_dirty_bytes || !dirty_background_bytes)
388 available_memory = determine_dirtyable_memory();
391 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
393 dirty = (vm_dirty_ratio * available_memory) / 100;
395 if (dirty_background_bytes)
396 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
398 background = (dirty_background_ratio * available_memory) / 100;
400 if (background >= dirty)
401 background = dirty / 2;
403 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
404 background += background / 4;
407 *pbackground = background;
409 trace_global_dirty_state(background, dirty);
413 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
414 * @bdi: the backing_dev_info to query
415 * @dirty: global dirty limit in pages
417 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
418 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
419 * And the "limit" in the name is not seriously taken as hard limit in
420 * balance_dirty_pages().
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 = div_s64((setpoint - 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))
602 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
604 * scale global setpoint to bdi's:
605 * bdi_setpoint = setpoint * bdi_thresh / thresh
607 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
608 bdi_setpoint = setpoint * (u64)x >> 16;
610 * Use span=(8*write_bw) in single bdi case as indicated by
611 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
613 * bdi_thresh thresh - bdi_thresh
614 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
617 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
618 x_intercept = bdi_setpoint + span;
620 if (bdi_dirty < x_intercept - span / 4) {
621 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
622 x_intercept - bdi_setpoint + 1);
627 * bdi reserve area, safeguard against dirty pool underrun and disk idle
628 * It may push the desired control point of global dirty pages higher
631 x_intercept = bdi_thresh / 2;
632 if (bdi_dirty < x_intercept) {
633 if (bdi_dirty > x_intercept / 8)
634 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
642 static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
643 unsigned long elapsed,
644 unsigned long written)
646 const unsigned long period = roundup_pow_of_two(3 * HZ);
647 unsigned long avg = bdi->avg_write_bandwidth;
648 unsigned long old = bdi->write_bandwidth;
652 * bw = written * HZ / elapsed
654 * bw * elapsed + write_bandwidth * (period - elapsed)
655 * write_bandwidth = ---------------------------------------------------
658 bw = written - bdi->written_stamp;
660 if (unlikely(elapsed > period)) {
665 bw += (u64)bdi->write_bandwidth * (period - elapsed);
666 bw >>= ilog2(period);
669 * one more level of smoothing, for filtering out sudden spikes
671 if (avg > old && old >= (unsigned long)bw)
672 avg -= (avg - old) >> 3;
674 if (avg < old && old <= (unsigned long)bw)
675 avg += (old - avg) >> 3;
678 bdi->write_bandwidth = bw;
679 bdi->avg_write_bandwidth = avg;
683 * The global dirtyable memory and dirty threshold could be suddenly knocked
684 * down by a large amount (eg. on the startup of KVM in a swapless system).
685 * This may throw the system into deep dirty exceeded state and throttle
686 * heavy/light dirtiers alike. To retain good responsiveness, maintain
687 * global_dirty_limit for tracking slowly down to the knocked down dirty
690 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
692 unsigned long limit = global_dirty_limit;
695 * Follow up in one step.
697 if (limit < thresh) {
703 * Follow down slowly. Use the higher one as the target, because thresh
704 * may drop below dirty. This is exactly the reason to introduce
705 * global_dirty_limit which is guaranteed to lie above the dirty pages.
707 thresh = max(thresh, dirty);
708 if (limit > thresh) {
709 limit -= (limit - thresh) >> 5;
714 global_dirty_limit = limit;
717 static void global_update_bandwidth(unsigned long thresh,
721 static DEFINE_SPINLOCK(dirty_lock);
722 static unsigned long update_time;
725 * check locklessly first to optimize away locking for the most time
727 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
730 spin_lock(&dirty_lock);
731 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
732 update_dirty_limit(thresh, dirty);
735 spin_unlock(&dirty_lock);
739 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
741 * Normal bdi tasks will be curbed at or below it in long term.
742 * Obviously it should be around (write_bw / N) when there are N dd tasks.
744 static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
745 unsigned long thresh,
746 unsigned long bg_thresh,
748 unsigned long bdi_thresh,
749 unsigned long bdi_dirty,
750 unsigned long dirtied,
751 unsigned long elapsed)
753 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
754 unsigned long limit = hard_dirty_limit(thresh);
755 unsigned long setpoint = (freerun + limit) / 2;
756 unsigned long write_bw = bdi->avg_write_bandwidth;
757 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
758 unsigned long dirty_rate;
759 unsigned long task_ratelimit;
760 unsigned long balanced_dirty_ratelimit;
761 unsigned long pos_ratio;
766 * The dirty rate will match the writeout rate in long term, except
767 * when dirty pages are truncated by userspace or re-dirtied by FS.
769 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
771 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
772 bdi_thresh, bdi_dirty);
774 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
776 task_ratelimit = (u64)dirty_ratelimit *
777 pos_ratio >> RATELIMIT_CALC_SHIFT;
778 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
781 * A linear estimation of the "balanced" throttle rate. The theory is,
782 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
783 * dirty_rate will be measured to be (N * task_ratelimit). So the below
784 * formula will yield the balanced rate limit (write_bw / N).
786 * Note that the expanded form is not a pure rate feedback:
787 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
788 * but also takes pos_ratio into account:
789 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
791 * (1) is not realistic because pos_ratio also takes part in balancing
792 * the dirty rate. Consider the state
793 * pos_ratio = 0.5 (3)
794 * rate = 2 * (write_bw / N) (4)
795 * If (1) is used, it will stuck in that state! Because each dd will
797 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
799 * dirty_rate = N * task_ratelimit = write_bw (6)
800 * put (6) into (1) we get
801 * rate_(i+1) = rate_(i) (7)
803 * So we end up using (2) to always keep
804 * rate_(i+1) ~= (write_bw / N) (8)
805 * regardless of the value of pos_ratio. As long as (8) is satisfied,
806 * pos_ratio is able to drive itself to 1.0, which is not only where
807 * the dirty count meet the setpoint, but also where the slope of
808 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
810 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
814 * We could safely do this and return immediately:
816 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
818 * However to get a more stable dirty_ratelimit, the below elaborated
819 * code makes use of task_ratelimit to filter out sigular points and
820 * limit the step size.
822 * The below code essentially only uses the relative value of
824 * task_ratelimit - dirty_ratelimit
825 * = (pos_ratio - 1) * dirty_ratelimit
827 * which reflects the direction and size of dirty position error.
831 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
832 * task_ratelimit is on the same side of dirty_ratelimit, too.
834 * - dirty_ratelimit > balanced_dirty_ratelimit
835 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
836 * lowering dirty_ratelimit will help meet both the position and rate
837 * control targets. Otherwise, don't update dirty_ratelimit if it will
838 * only help meet the rate target. After all, what the users ultimately
839 * feel and care are stable dirty rate and small position error.
841 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
842 * and filter out the sigular points of balanced_dirty_ratelimit. Which
843 * keeps jumping around randomly and can even leap far away at times
844 * due to the small 200ms estimation period of dirty_rate (we want to
845 * keep that period small to reduce time lags).
848 if (dirty < setpoint) {
849 x = min(bdi->balanced_dirty_ratelimit,
850 min(balanced_dirty_ratelimit, task_ratelimit));
851 if (dirty_ratelimit < x)
852 step = x - dirty_ratelimit;
854 x = max(bdi->balanced_dirty_ratelimit,
855 max(balanced_dirty_ratelimit, task_ratelimit));
856 if (dirty_ratelimit > x)
857 step = dirty_ratelimit - x;
861 * Don't pursue 100% rate matching. It's impossible since the balanced
862 * rate itself is constantly fluctuating. So decrease the track speed
863 * when it gets close to the target. Helps eliminate pointless tremors.
865 step >>= dirty_ratelimit / (2 * step + 1);
867 * Limit the tracking speed to avoid overshooting.
869 step = (step + 7) / 8;
871 if (dirty_ratelimit < balanced_dirty_ratelimit)
872 dirty_ratelimit += step;
874 dirty_ratelimit -= step;
876 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
877 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
879 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
882 void __bdi_update_bandwidth(struct backing_dev_info *bdi,
883 unsigned long thresh,
884 unsigned long bg_thresh,
886 unsigned long bdi_thresh,
887 unsigned long bdi_dirty,
888 unsigned long start_time)
890 unsigned long now = jiffies;
891 unsigned long elapsed = now - bdi->bw_time_stamp;
892 unsigned long dirtied;
893 unsigned long written;
896 * rate-limit, only update once every 200ms.
898 if (elapsed < BANDWIDTH_INTERVAL)
901 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
902 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
905 * Skip quiet periods when disk bandwidth is under-utilized.
906 * (at least 1s idle time between two flusher runs)
908 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
912 global_update_bandwidth(thresh, dirty, now);
913 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
914 bdi_thresh, bdi_dirty,
917 bdi_update_write_bandwidth(bdi, elapsed, written);
920 bdi->dirtied_stamp = dirtied;
921 bdi->written_stamp = written;
922 bdi->bw_time_stamp = now;
925 static void bdi_update_bandwidth(struct backing_dev_info *bdi,
926 unsigned long thresh,
927 unsigned long bg_thresh,
929 unsigned long bdi_thresh,
930 unsigned long bdi_dirty,
931 unsigned long start_time)
933 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
935 spin_lock(&bdi->wb.list_lock);
936 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
937 bdi_thresh, bdi_dirty, start_time);
938 spin_unlock(&bdi->wb.list_lock);
942 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
943 * will look to see if it needs to start dirty throttling.
945 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
946 * global_page_state() too often. So scale it near-sqrt to the safety margin
947 * (the number of pages we may dirty without exceeding the dirty limits).
949 static unsigned long dirty_poll_interval(unsigned long dirty,
950 unsigned long thresh)
953 return 1UL << (ilog2(thresh - dirty) >> 1);
958 static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
959 unsigned long bdi_dirty)
961 unsigned long bw = bdi->avg_write_bandwidth;
962 unsigned long hi = ilog2(bw);
963 unsigned long lo = ilog2(bdi->dirty_ratelimit);
966 /* target for 20ms max pause on 1-dd case */
970 * Scale up pause time for concurrent dirtiers in order to reduce CPU
973 * (N * 20ms) on 2^N concurrent tasks.
976 t += (hi - lo) * (20 * HZ) / 1024;
979 * Limit pause time for small memory systems. If sleeping for too long
980 * time, a small pool of dirty/writeback pages may go empty and disk go
983 * 8 serves as the safety ratio.
986 t = min(t, bdi_dirty * HZ / (8 * bw + 1));
989 * The pause time will be settled within range (max_pause/4, max_pause).
990 * Apply a minimal value of 4 to get a non-zero max_pause/4.
992 return clamp_val(t, 4, MAX_PAUSE);
996 * balance_dirty_pages() must be called by processes which are generating dirty
997 * data. It looks at the number of dirty pages in the machine and will force
998 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
999 * If we're over `background_thresh' then the writeback threads are woken to
1000 * perform some writeout.
1002 static void balance_dirty_pages(struct address_space *mapping,
1003 unsigned long pages_dirtied)
1005 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1006 unsigned long bdi_reclaimable;
1007 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1008 unsigned long bdi_dirty;
1009 unsigned long freerun;
1010 unsigned long background_thresh;
1011 unsigned long dirty_thresh;
1012 unsigned long bdi_thresh;
1014 long uninitialized_var(max_pause);
1015 bool dirty_exceeded = false;
1016 unsigned long task_ratelimit;
1017 unsigned long uninitialized_var(dirty_ratelimit);
1018 unsigned long pos_ratio;
1019 struct backing_dev_info *bdi = mapping->backing_dev_info;
1020 unsigned long start_time = jiffies;
1024 * Unstable writes are a feature of certain networked
1025 * filesystems (i.e. NFS) in which data may have been
1026 * written to the server's write cache, but has not yet
1027 * been flushed to permanent storage.
1029 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1030 global_page_state(NR_UNSTABLE_NFS);
1031 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1033 global_dirty_limits(&background_thresh, &dirty_thresh);
1036 * Throttle it only when the background writeback cannot
1037 * catch-up. This avoids (excessively) small writeouts
1038 * when the bdi limits are ramping up.
1040 freerun = dirty_freerun_ceiling(dirty_thresh,
1042 if (nr_dirty <= freerun)
1045 if (unlikely(!writeback_in_progress(bdi)))
1046 bdi_start_background_writeback(bdi);
1049 * bdi_thresh is not treated as some limiting factor as
1050 * dirty_thresh, due to reasons
1051 * - in JBOD setup, bdi_thresh can fluctuate a lot
1052 * - in a system with HDD and USB key, the USB key may somehow
1053 * go into state (bdi_dirty >> bdi_thresh) either because
1054 * bdi_dirty starts high, or because bdi_thresh drops low.
1055 * In this case we don't want to hard throttle the USB key
1056 * dirtiers for 100 seconds until bdi_dirty drops under
1057 * bdi_thresh. Instead the auxiliary bdi control line in
1058 * bdi_position_ratio() will let the dirtier task progress
1059 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1061 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1064 * In order to avoid the stacked BDI deadlock we need
1065 * to ensure we accurately count the 'dirty' pages when
1066 * the threshold is low.
1068 * Otherwise it would be possible to get thresh+n pages
1069 * reported dirty, even though there are thresh-m pages
1070 * actually dirty; with m+n sitting in the percpu
1073 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1074 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1075 bdi_dirty = bdi_reclaimable +
1076 bdi_stat_sum(bdi, BDI_WRITEBACK);
1078 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1079 bdi_dirty = bdi_reclaimable +
1080 bdi_stat(bdi, BDI_WRITEBACK);
1083 dirty_exceeded = (bdi_dirty > bdi_thresh) ||
1084 (nr_dirty > dirty_thresh);
1085 if (dirty_exceeded && !bdi->dirty_exceeded)
1086 bdi->dirty_exceeded = 1;
1088 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1089 nr_dirty, bdi_thresh, bdi_dirty,
1092 max_pause = bdi_max_pause(bdi, bdi_dirty);
1094 dirty_ratelimit = bdi->dirty_ratelimit;
1095 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1096 background_thresh, nr_dirty,
1097 bdi_thresh, bdi_dirty);
1098 if (unlikely(pos_ratio == 0)) {
1102 task_ratelimit = (u64)dirty_ratelimit *
1103 pos_ratio >> RATELIMIT_CALC_SHIFT;
1104 pause = (HZ * pages_dirtied) / (task_ratelimit | 1);
1105 if (unlikely(pause <= 0)) {
1106 trace_balance_dirty_pages(bdi,
1117 pause = 1; /* avoid resetting nr_dirtied_pause below */
1120 pause = min(pause, max_pause);
1123 trace_balance_dirty_pages(bdi,
1134 __set_current_state(TASK_UNINTERRUPTIBLE);
1135 io_schedule_timeout(pause);
1137 dirty_thresh = hard_dirty_limit(dirty_thresh);
1139 * max-pause area. If dirty exceeded but still within this
1140 * area, no need to sleep for more than 200ms: (a) 8 pages per
1141 * 200ms is typically more than enough to curb heavy dirtiers;
1142 * (b) the pause time limit makes the dirtiers more responsive.
1144 if (nr_dirty < dirty_thresh)
1148 if (!dirty_exceeded && bdi->dirty_exceeded)
1149 bdi->dirty_exceeded = 0;
1151 current->nr_dirtied = 0;
1152 if (pause == 0) { /* in freerun area */
1153 current->nr_dirtied_pause =
1154 dirty_poll_interval(nr_dirty, dirty_thresh);
1155 } else if (pause <= max_pause / 4 &&
1156 pages_dirtied >= current->nr_dirtied_pause) {
1157 current->nr_dirtied_pause = clamp_val(
1158 dirty_ratelimit * (max_pause / 2) / HZ,
1159 pages_dirtied + pages_dirtied / 8,
1161 } else if (pause >= max_pause) {
1162 current->nr_dirtied_pause = 1 | clamp_val(
1163 dirty_ratelimit * (max_pause / 2) / HZ,
1165 pages_dirtied - pages_dirtied / 8);
1168 if (writeback_in_progress(bdi))
1172 * In laptop mode, we wait until hitting the higher threshold before
1173 * starting background writeout, and then write out all the way down
1174 * to the lower threshold. So slow writers cause minimal disk activity.
1176 * In normal mode, we start background writeout at the lower
1177 * background_thresh, to keep the amount of dirty memory low.
1182 if (nr_reclaimable > background_thresh)
1183 bdi_start_background_writeback(bdi);
1186 void set_page_dirty_balance(struct page *page, int page_mkwrite)
1188 if (set_page_dirty(page) || page_mkwrite) {
1189 struct address_space *mapping = page_mapping(page);
1192 balance_dirty_pages_ratelimited(mapping);
1196 static DEFINE_PER_CPU(int, bdp_ratelimits);
1199 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
1200 * @mapping: address_space which was dirtied
1201 * @nr_pages_dirtied: number of pages which the caller has just dirtied
1203 * Processes which are dirtying memory should call in here once for each page
1204 * which was newly dirtied. The function will periodically check the system's
1205 * dirty state and will initiate writeback if needed.
1207 * On really big machines, get_writeback_state is expensive, so try to avoid
1208 * calling it too often (ratelimiting). But once we're over the dirty memory
1209 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1210 * from overshooting the limit by (ratelimit_pages) each.
1212 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1213 unsigned long nr_pages_dirtied)
1215 struct backing_dev_info *bdi = mapping->backing_dev_info;
1219 if (!bdi_cap_account_dirty(bdi))
1222 ratelimit = current->nr_dirtied_pause;
1223 if (bdi->dirty_exceeded)
1224 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1226 current->nr_dirtied += nr_pages_dirtied;
1230 * This prevents one CPU to accumulate too many dirtied pages without
1231 * calling into balance_dirty_pages(), which can happen when there are
1232 * 1000+ tasks, all of them start dirtying pages at exactly the same
1233 * time, hence all honoured too large initial task->nr_dirtied_pause.
1235 p = &__get_cpu_var(bdp_ratelimits);
1236 if (unlikely(current->nr_dirtied >= ratelimit))
1239 *p += nr_pages_dirtied;
1240 if (unlikely(*p >= ratelimit_pages)) {
1247 if (unlikely(current->nr_dirtied >= ratelimit))
1248 balance_dirty_pages(mapping, current->nr_dirtied);
1250 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
1252 void throttle_vm_writeout(gfp_t gfp_mask)
1254 unsigned long background_thresh;
1255 unsigned long dirty_thresh;
1258 global_dirty_limits(&background_thresh, &dirty_thresh);
1261 * Boost the allowable dirty threshold a bit for page
1262 * allocators so they don't get DoS'ed by heavy writers
1264 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1266 if (global_page_state(NR_UNSTABLE_NFS) +
1267 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1269 congestion_wait(BLK_RW_ASYNC, HZ/10);
1272 * The caller might hold locks which can prevent IO completion
1273 * or progress in the filesystem. So we cannot just sit here
1274 * waiting for IO to complete.
1276 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1282 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1284 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1285 void __user *buffer, size_t *length, loff_t *ppos)
1287 proc_dointvec(table, write, buffer, length, ppos);
1288 bdi_arm_supers_timer();
1293 void laptop_mode_timer_fn(unsigned long data)
1295 struct request_queue *q = (struct request_queue *)data;
1296 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1297 global_page_state(NR_UNSTABLE_NFS);
1300 * We want to write everything out, not just down to the dirty
1303 if (bdi_has_dirty_io(&q->backing_dev_info))
1304 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1305 WB_REASON_LAPTOP_TIMER);
1309 * We've spun up the disk and we're in laptop mode: schedule writeback
1310 * of all dirty data a few seconds from now. If the flush is already scheduled
1311 * then push it back - the user is still using the disk.
1313 void laptop_io_completion(struct backing_dev_info *info)
1315 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1319 * We're in laptop mode and we've just synced. The sync's writes will have
1320 * caused another writeback to be scheduled by laptop_io_completion.
1321 * Nothing needs to be written back anymore, so we unschedule the writeback.
1323 void laptop_sync_completion(void)
1325 struct backing_dev_info *bdi;
1329 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1330 del_timer(&bdi->laptop_mode_wb_timer);
1337 * If ratelimit_pages is too high then we can get into dirty-data overload
1338 * if a large number of processes all perform writes at the same time.
1339 * If it is too low then SMP machines will call the (expensive)
1340 * get_writeback_state too often.
1342 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1343 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1347 void writeback_set_ratelimit(void)
1349 unsigned long background_thresh;
1350 unsigned long dirty_thresh;
1351 global_dirty_limits(&background_thresh, &dirty_thresh);
1352 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1353 if (ratelimit_pages < 16)
1354 ratelimit_pages = 16;
1357 static int __cpuinit
1358 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1360 writeback_set_ratelimit();
1364 static struct notifier_block __cpuinitdata ratelimit_nb = {
1365 .notifier_call = ratelimit_handler,
1370 * Called early on to tune the page writeback dirty limits.
1372 * We used to scale dirty pages according to how total memory
1373 * related to pages that could be allocated for buffers (by
1374 * comparing nr_free_buffer_pages() to vm_total_pages.
1376 * However, that was when we used "dirty_ratio" to scale with
1377 * all memory, and we don't do that any more. "dirty_ratio"
1378 * is now applied to total non-HIGHPAGE memory (by subtracting
1379 * totalhigh_pages from vm_total_pages), and as such we can't
1380 * get into the old insane situation any more where we had
1381 * large amounts of dirty pages compared to a small amount of
1382 * non-HIGHMEM memory.
1384 * But we might still want to scale the dirty_ratio by how
1385 * much memory the box has..
1387 void __init page_writeback_init(void)
1391 writeback_set_ratelimit();
1392 register_cpu_notifier(&ratelimit_nb);
1394 shift = calc_period_shift();
1395 prop_descriptor_init(&vm_completions, shift);
1396 prop_descriptor_init(&vm_dirties, shift);
1400 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1401 * @mapping: address space structure to write
1402 * @start: starting page index
1403 * @end: ending page index (inclusive)
1405 * This function scans the page range from @start to @end (inclusive) and tags
1406 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1407 * that write_cache_pages (or whoever calls this function) will then use
1408 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1409 * used to avoid livelocking of writeback by a process steadily creating new
1410 * dirty pages in the file (thus it is important for this function to be quick
1411 * so that it can tag pages faster than a dirtying process can create them).
1414 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1416 void tag_pages_for_writeback(struct address_space *mapping,
1417 pgoff_t start, pgoff_t end)
1419 #define WRITEBACK_TAG_BATCH 4096
1420 unsigned long tagged;
1423 spin_lock_irq(&mapping->tree_lock);
1424 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1425 &start, end, WRITEBACK_TAG_BATCH,
1426 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1427 spin_unlock_irq(&mapping->tree_lock);
1428 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1430 /* We check 'start' to handle wrapping when end == ~0UL */
1431 } while (tagged >= WRITEBACK_TAG_BATCH && start);
1433 EXPORT_SYMBOL(tag_pages_for_writeback);
1436 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1437 * @mapping: address space structure to write
1438 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1439 * @writepage: function called for each page
1440 * @data: data passed to writepage function
1442 * If a page is already under I/O, write_cache_pages() skips it, even
1443 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1444 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1445 * and msync() need to guarantee that all the data which was dirty at the time
1446 * the call was made get new I/O started against them. If wbc->sync_mode is
1447 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1448 * existing IO to complete.
1450 * To avoid livelocks (when other process dirties new pages), we first tag
1451 * pages which should be written back with TOWRITE tag and only then start
1452 * writing them. For data-integrity sync we have to be careful so that we do
1453 * not miss some pages (e.g., because some other process has cleared TOWRITE
1454 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1455 * by the process clearing the DIRTY tag (and submitting the page for IO).
1457 int write_cache_pages(struct address_space *mapping,
1458 struct writeback_control *wbc, writepage_t writepage,
1463 struct pagevec pvec;
1465 pgoff_t uninitialized_var(writeback_index);
1467 pgoff_t end; /* Inclusive */
1470 int range_whole = 0;
1473 pagevec_init(&pvec, 0);
1474 if (wbc->range_cyclic) {
1475 writeback_index = mapping->writeback_index; /* prev offset */
1476 index = writeback_index;
1483 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1484 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1485 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1487 cycled = 1; /* ignore range_cyclic tests */
1489 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1490 tag = PAGECACHE_TAG_TOWRITE;
1492 tag = PAGECACHE_TAG_DIRTY;
1494 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1495 tag_pages_for_writeback(mapping, index, end);
1497 while (!done && (index <= end)) {
1500 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1501 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1505 for (i = 0; i < nr_pages; i++) {
1506 struct page *page = pvec.pages[i];
1509 * At this point, the page may be truncated or
1510 * invalidated (changing page->mapping to NULL), or
1511 * even swizzled back from swapper_space to tmpfs file
1512 * mapping. However, page->index will not change
1513 * because we have a reference on the page.
1515 if (page->index > end) {
1517 * can't be range_cyclic (1st pass) because
1518 * end == -1 in that case.
1524 done_index = page->index;
1529 * Page truncated or invalidated. We can freely skip it
1530 * then, even for data integrity operations: the page
1531 * has disappeared concurrently, so there could be no
1532 * real expectation of this data interity operation
1533 * even if there is now a new, dirty page at the same
1534 * pagecache address.
1536 if (unlikely(page->mapping != mapping)) {
1542 if (!PageDirty(page)) {
1543 /* someone wrote it for us */
1544 goto continue_unlock;
1547 if (PageWriteback(page)) {
1548 if (wbc->sync_mode != WB_SYNC_NONE)
1549 wait_on_page_writeback(page);
1551 goto continue_unlock;
1554 BUG_ON(PageWriteback(page));
1555 if (!clear_page_dirty_for_io(page))
1556 goto continue_unlock;
1558 trace_wbc_writepage(wbc, mapping->backing_dev_info);
1559 ret = (*writepage)(page, wbc, data);
1560 if (unlikely(ret)) {
1561 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1566 * done_index is set past this page,
1567 * so media errors will not choke
1568 * background writeout for the entire
1569 * file. This has consequences for
1570 * range_cyclic semantics (ie. it may
1571 * not be suitable for data integrity
1574 done_index = page->index + 1;
1581 * We stop writing back only if we are not doing
1582 * integrity sync. In case of integrity sync we have to
1583 * keep going until we have written all the pages
1584 * we tagged for writeback prior to entering this loop.
1586 if (--wbc->nr_to_write <= 0 &&
1587 wbc->sync_mode == WB_SYNC_NONE) {
1592 pagevec_release(&pvec);
1595 if (!cycled && !done) {
1598 * We hit the last page and there is more work to be done: wrap
1599 * back to the start of the file
1603 end = writeback_index - 1;
1606 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1607 mapping->writeback_index = done_index;
1611 EXPORT_SYMBOL(write_cache_pages);
1614 * Function used by generic_writepages to call the real writepage
1615 * function and set the mapping flags on error
1617 static int __writepage(struct page *page, struct writeback_control *wbc,
1620 struct address_space *mapping = data;
1621 int ret = mapping->a_ops->writepage(page, wbc);
1622 mapping_set_error(mapping, ret);
1627 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1628 * @mapping: address space structure to write
1629 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1631 * This is a library function, which implements the writepages()
1632 * address_space_operation.
1634 int generic_writepages(struct address_space *mapping,
1635 struct writeback_control *wbc)
1637 struct blk_plug plug;
1640 /* deal with chardevs and other special file */
1641 if (!mapping->a_ops->writepage)
1644 blk_start_plug(&plug);
1645 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1646 blk_finish_plug(&plug);
1650 EXPORT_SYMBOL(generic_writepages);
1652 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1656 if (wbc->nr_to_write <= 0)
1658 if (mapping->a_ops->writepages)
1659 ret = mapping->a_ops->writepages(mapping, wbc);
1661 ret = generic_writepages(mapping, wbc);
1666 * write_one_page - write out a single page and optionally wait on I/O
1667 * @page: the page to write
1668 * @wait: if true, wait on writeout
1670 * The page must be locked by the caller and will be unlocked upon return.
1672 * write_one_page() returns a negative error code if I/O failed.
1674 int write_one_page(struct page *page, int wait)
1676 struct address_space *mapping = page->mapping;
1678 struct writeback_control wbc = {
1679 .sync_mode = WB_SYNC_ALL,
1683 BUG_ON(!PageLocked(page));
1686 wait_on_page_writeback(page);
1688 if (clear_page_dirty_for_io(page)) {
1689 page_cache_get(page);
1690 ret = mapping->a_ops->writepage(page, &wbc);
1691 if (ret == 0 && wait) {
1692 wait_on_page_writeback(page);
1693 if (PageError(page))
1696 page_cache_release(page);
1702 EXPORT_SYMBOL(write_one_page);
1705 * For address_spaces which do not use buffers nor write back.
1707 int __set_page_dirty_no_writeback(struct page *page)
1709 if (!PageDirty(page))
1710 return !TestSetPageDirty(page);
1715 * Helper function for set_page_dirty family.
1716 * NOTE: This relies on being atomic wrt interrupts.
1718 void account_page_dirtied(struct page *page, struct address_space *mapping)
1720 if (mapping_cap_account_dirty(mapping)) {
1721 __inc_zone_page_state(page, NR_FILE_DIRTY);
1722 __inc_zone_page_state(page, NR_DIRTIED);
1723 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1724 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1725 task_dirty_inc(current);
1726 task_io_account_write(PAGE_CACHE_SIZE);
1729 EXPORT_SYMBOL(account_page_dirtied);
1732 * Helper function for set_page_writeback family.
1733 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1736 void account_page_writeback(struct page *page)
1738 inc_zone_page_state(page, NR_WRITEBACK);
1740 EXPORT_SYMBOL(account_page_writeback);
1743 * For address_spaces which do not use buffers. Just tag the page as dirty in
1746 * This is also used when a single buffer is being dirtied: we want to set the
1747 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1748 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1750 * Most callers have locked the page, which pins the address_space in memory.
1751 * But zap_pte_range() does not lock the page, however in that case the
1752 * mapping is pinned by the vma's ->vm_file reference.
1754 * We take care to handle the case where the page was truncated from the
1755 * mapping by re-checking page_mapping() inside tree_lock.
1757 int __set_page_dirty_nobuffers(struct page *page)
1759 if (!TestSetPageDirty(page)) {
1760 struct address_space *mapping = page_mapping(page);
1761 struct address_space *mapping2;
1766 spin_lock_irq(&mapping->tree_lock);
1767 mapping2 = page_mapping(page);
1768 if (mapping2) { /* Race with truncate? */
1769 BUG_ON(mapping2 != mapping);
1770 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1771 account_page_dirtied(page, mapping);
1772 radix_tree_tag_set(&mapping->page_tree,
1773 page_index(page), PAGECACHE_TAG_DIRTY);
1775 spin_unlock_irq(&mapping->tree_lock);
1776 if (mapping->host) {
1777 /* !PageAnon && !swapper_space */
1778 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1784 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1787 * When a writepage implementation decides that it doesn't want to write this
1788 * page for some reason, it should redirty the locked page via
1789 * redirty_page_for_writepage() and it should then unlock the page and return 0
1791 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1793 wbc->pages_skipped++;
1794 return __set_page_dirty_nobuffers(page);
1796 EXPORT_SYMBOL(redirty_page_for_writepage);
1801 * For pages with a mapping this should be done under the page lock
1802 * for the benefit of asynchronous memory errors who prefer a consistent
1803 * dirty state. This rule can be broken in some special cases,
1804 * but should be better not to.
1806 * If the mapping doesn't provide a set_page_dirty a_op, then
1807 * just fall through and assume that it wants buffer_heads.
1809 int set_page_dirty(struct page *page)
1811 struct address_space *mapping = page_mapping(page);
1813 if (likely(mapping)) {
1814 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1816 * readahead/lru_deactivate_page could remain
1817 * PG_readahead/PG_reclaim due to race with end_page_writeback
1818 * About readahead, if the page is written, the flags would be
1819 * reset. So no problem.
1820 * About lru_deactivate_page, if the page is redirty, the flag
1821 * will be reset. So no problem. but if the page is used by readahead
1822 * it will confuse readahead and make it restart the size rampup
1823 * process. But it's a trivial problem.
1825 ClearPageReclaim(page);
1828 spd = __set_page_dirty_buffers;
1830 return (*spd)(page);
1832 if (!PageDirty(page)) {
1833 if (!TestSetPageDirty(page))
1838 EXPORT_SYMBOL(set_page_dirty);
1841 * set_page_dirty() is racy if the caller has no reference against
1842 * page->mapping->host, and if the page is unlocked. This is because another
1843 * CPU could truncate the page off the mapping and then free the mapping.
1845 * Usually, the page _is_ locked, or the caller is a user-space process which
1846 * holds a reference on the inode by having an open file.
1848 * In other cases, the page should be locked before running set_page_dirty().
1850 int set_page_dirty_lock(struct page *page)
1855 ret = set_page_dirty(page);
1859 EXPORT_SYMBOL(set_page_dirty_lock);
1862 * Clear a page's dirty flag, while caring for dirty memory accounting.
1863 * Returns true if the page was previously dirty.
1865 * This is for preparing to put the page under writeout. We leave the page
1866 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1867 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1868 * implementation will run either set_page_writeback() or set_page_dirty(),
1869 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1872 * This incoherency between the page's dirty flag and radix-tree tag is
1873 * unfortunate, but it only exists while the page is locked.
1875 int clear_page_dirty_for_io(struct page *page)
1877 struct address_space *mapping = page_mapping(page);
1879 BUG_ON(!PageLocked(page));
1881 if (mapping && mapping_cap_account_dirty(mapping)) {
1883 * Yes, Virginia, this is indeed insane.
1885 * We use this sequence to make sure that
1886 * (a) we account for dirty stats properly
1887 * (b) we tell the low-level filesystem to
1888 * mark the whole page dirty if it was
1889 * dirty in a pagetable. Only to then
1890 * (c) clean the page again and return 1 to
1891 * cause the writeback.
1893 * This way we avoid all nasty races with the
1894 * dirty bit in multiple places and clearing
1895 * them concurrently from different threads.
1897 * Note! Normally the "set_page_dirty(page)"
1898 * has no effect on the actual dirty bit - since
1899 * that will already usually be set. But we
1900 * need the side effects, and it can help us
1903 * We basically use the page "master dirty bit"
1904 * as a serialization point for all the different
1905 * threads doing their things.
1907 if (page_mkclean(page))
1908 set_page_dirty(page);
1910 * We carefully synchronise fault handlers against
1911 * installing a dirty pte and marking the page dirty
1912 * at this point. We do this by having them hold the
1913 * page lock at some point after installing their
1914 * pte, but before marking the page dirty.
1915 * Pages are always locked coming in here, so we get
1916 * the desired exclusion. See mm/memory.c:do_wp_page()
1917 * for more comments.
1919 if (TestClearPageDirty(page)) {
1920 dec_zone_page_state(page, NR_FILE_DIRTY);
1921 dec_bdi_stat(mapping->backing_dev_info,
1927 return TestClearPageDirty(page);
1929 EXPORT_SYMBOL(clear_page_dirty_for_io);
1931 int test_clear_page_writeback(struct page *page)
1933 struct address_space *mapping = page_mapping(page);
1937 struct backing_dev_info *bdi = mapping->backing_dev_info;
1938 unsigned long flags;
1940 spin_lock_irqsave(&mapping->tree_lock, flags);
1941 ret = TestClearPageWriteback(page);
1943 radix_tree_tag_clear(&mapping->page_tree,
1945 PAGECACHE_TAG_WRITEBACK);
1946 if (bdi_cap_account_writeback(bdi)) {
1947 __dec_bdi_stat(bdi, BDI_WRITEBACK);
1948 __bdi_writeout_inc(bdi);
1951 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1953 ret = TestClearPageWriteback(page);
1956 dec_zone_page_state(page, NR_WRITEBACK);
1957 inc_zone_page_state(page, NR_WRITTEN);
1962 int test_set_page_writeback(struct page *page)
1964 struct address_space *mapping = page_mapping(page);
1968 struct backing_dev_info *bdi = mapping->backing_dev_info;
1969 unsigned long flags;
1971 spin_lock_irqsave(&mapping->tree_lock, flags);
1972 ret = TestSetPageWriteback(page);
1974 radix_tree_tag_set(&mapping->page_tree,
1976 PAGECACHE_TAG_WRITEBACK);
1977 if (bdi_cap_account_writeback(bdi))
1978 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1980 if (!PageDirty(page))
1981 radix_tree_tag_clear(&mapping->page_tree,
1983 PAGECACHE_TAG_DIRTY);
1984 radix_tree_tag_clear(&mapping->page_tree,
1986 PAGECACHE_TAG_TOWRITE);
1987 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1989 ret = TestSetPageWriteback(page);
1992 account_page_writeback(page);
1996 EXPORT_SYMBOL(test_set_page_writeback);
1999 * Return true if any of the pages in the mapping are marked with the
2002 int mapping_tagged(struct address_space *mapping, int tag)
2004 return radix_tree_tagged(&mapping->page_tree, tag);
2006 EXPORT_SYMBOL(mapping_tagged);