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> /* __set_page_dirty_buffers */
36 #include <linux/pagevec.h>
37 #include <linux/timer.h>
38 #include <linux/sched/rt.h>
39 #include <linux/mm_inline.h>
40 #include <trace/events/writeback.h>
45 * Sleep at most 200ms at a time in balance_dirty_pages().
47 #define MAX_PAUSE max(HZ/5, 1)
50 * Try to keep balance_dirty_pages() call intervals higher than this many pages
51 * by raising pause time to max_pause when falls below it.
53 #define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
56 * Estimate write bandwidth at 200ms intervals.
58 #define BANDWIDTH_INTERVAL max(HZ/5, 1)
60 #define RATELIMIT_CALC_SHIFT 10
63 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
64 * will look to see if it needs to force writeback or throttling.
66 static long ratelimit_pages = 32;
68 /* The following parameters are exported via /proc/sys/vm */
71 * Start background writeback (via writeback threads) at this percentage
73 int dirty_background_ratio = 10;
76 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
77 * dirty_background_ratio * the amount of dirtyable memory
79 unsigned long dirty_background_bytes;
82 * free highmem will not be subtracted from the total free memory
83 * for calculating free ratios if vm_highmem_is_dirtyable is true
85 int vm_highmem_is_dirtyable;
88 * The generator of dirty data starts writeback at this percentage
90 int vm_dirty_ratio = 20;
93 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
94 * vm_dirty_ratio * the amount of dirtyable memory
96 unsigned long vm_dirty_bytes;
99 * The interval between `kupdate'-style writebacks
101 unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
103 EXPORT_SYMBOL_GPL(dirty_writeback_interval);
106 * The longest time for which data is allowed to remain dirty
108 unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
111 * Flag that makes the machine dump writes/reads and block dirtyings.
116 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
117 * a full sync is triggered after this time elapses without any disk activity.
121 EXPORT_SYMBOL(laptop_mode);
123 /* End of sysctl-exported parameters */
125 struct wb_domain global_wb_domain;
128 * Length of period for aging writeout fractions of bdis. This is an
129 * arbitrarily chosen number. The longer the period, the slower fractions will
130 * reflect changes in current writeout rate.
132 #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
134 #ifdef CONFIG_CGROUP_WRITEBACK
136 static void wb_min_max_ratio(struct bdi_writeback *wb,
137 unsigned long *minp, unsigned long *maxp)
139 unsigned long this_bw = wb->avg_write_bandwidth;
140 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
141 unsigned long long min = wb->bdi->min_ratio;
142 unsigned long long max = wb->bdi->max_ratio;
145 * @wb may already be clean by the time control reaches here and
146 * the total may not include its bw.
148 if (this_bw < tot_bw) {
163 #else /* CONFIG_CGROUP_WRITEBACK */
165 static void wb_min_max_ratio(struct bdi_writeback *wb,
166 unsigned long *minp, unsigned long *maxp)
168 *minp = wb->bdi->min_ratio;
169 *maxp = wb->bdi->max_ratio;
172 #endif /* CONFIG_CGROUP_WRITEBACK */
175 * In a memory zone, there is a certain amount of pages we consider
176 * available for the page cache, which is essentially the number of
177 * free and reclaimable pages, minus some zone reserves to protect
178 * lowmem and the ability to uphold the zone's watermarks without
179 * requiring writeback.
181 * This number of dirtyable pages is the base value of which the
182 * user-configurable dirty ratio is the effictive number of pages that
183 * are allowed to be actually dirtied. Per individual zone, or
184 * globally by using the sum of dirtyable pages over all zones.
186 * Because the user is allowed to specify the dirty limit globally as
187 * absolute number of bytes, calculating the per-zone dirty limit can
188 * require translating the configured limit into a percentage of
189 * global dirtyable memory first.
193 * zone_dirtyable_memory - number of dirtyable pages in a zone
196 * Returns the zone's number of pages potentially available for dirty
197 * page cache. This is the base value for the per-zone dirty limits.
199 static unsigned long zone_dirtyable_memory(struct zone *zone)
201 unsigned long nr_pages;
203 nr_pages = zone_page_state(zone, NR_FREE_PAGES);
204 nr_pages -= min(nr_pages, zone->dirty_balance_reserve);
206 nr_pages += zone_page_state(zone, NR_INACTIVE_FILE);
207 nr_pages += zone_page_state(zone, NR_ACTIVE_FILE);
212 static unsigned long highmem_dirtyable_memory(unsigned long total)
214 #ifdef CONFIG_HIGHMEM
218 for_each_node_state(node, N_HIGH_MEMORY) {
219 struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
221 x += zone_dirtyable_memory(z);
224 * Unreclaimable memory (kernel memory or anonymous memory
225 * without swap) can bring down the dirtyable pages below
226 * the zone's dirty balance reserve and the above calculation
227 * will underflow. However we still want to add in nodes
228 * which are below threshold (negative values) to get a more
229 * accurate calculation but make sure that the total never
236 * Make sure that the number of highmem pages is never larger
237 * than the number of the total dirtyable memory. This can only
238 * occur in very strange VM situations but we want to make sure
239 * that this does not occur.
241 return min(x, total);
248 * global_dirtyable_memory - number of globally dirtyable pages
250 * Returns the global number of pages potentially available for dirty
251 * page cache. This is the base value for the global dirty limits.
253 static unsigned long global_dirtyable_memory(void)
257 x = global_page_state(NR_FREE_PAGES);
258 x -= min(x, dirty_balance_reserve);
260 x += global_page_state(NR_INACTIVE_FILE);
261 x += global_page_state(NR_ACTIVE_FILE);
263 if (!vm_highmem_is_dirtyable)
264 x -= highmem_dirtyable_memory(x);
266 return x + 1; /* Ensure that we never return 0 */
270 * global_dirty_limits - background-writeback and dirty-throttling thresholds
272 * Calculate the dirty thresholds based on sysctl parameters
273 * - vm.dirty_background_ratio or vm.dirty_background_bytes
274 * - vm.dirty_ratio or vm.dirty_bytes
275 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
278 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
280 const unsigned long available_memory = global_dirtyable_memory();
281 unsigned long background;
283 struct task_struct *tsk;
286 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
288 dirty = (vm_dirty_ratio * available_memory) / 100;
290 if (dirty_background_bytes)
291 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
293 background = (dirty_background_ratio * available_memory) / 100;
295 if (background >= dirty)
296 background = dirty / 2;
298 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
299 background += background / 4;
302 *pbackground = background;
304 trace_global_dirty_state(background, dirty);
308 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
311 * Returns the maximum number of dirty pages allowed in a zone, based
312 * on the zone's dirtyable memory.
314 static unsigned long zone_dirty_limit(struct zone *zone)
316 unsigned long zone_memory = zone_dirtyable_memory(zone);
317 struct task_struct *tsk = current;
321 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
322 zone_memory / global_dirtyable_memory();
324 dirty = vm_dirty_ratio * zone_memory / 100;
326 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
333 * zone_dirty_ok - tells whether a zone is within its dirty limits
334 * @zone: the zone to check
336 * Returns %true when the dirty pages in @zone are within the zone's
337 * dirty limit, %false if the limit is exceeded.
339 bool zone_dirty_ok(struct zone *zone)
341 unsigned long limit = zone_dirty_limit(zone);
343 return zone_page_state(zone, NR_FILE_DIRTY) +
344 zone_page_state(zone, NR_UNSTABLE_NFS) +
345 zone_page_state(zone, NR_WRITEBACK) <= limit;
348 int dirty_background_ratio_handler(struct ctl_table *table, int write,
349 void __user *buffer, size_t *lenp,
354 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
355 if (ret == 0 && write)
356 dirty_background_bytes = 0;
360 int dirty_background_bytes_handler(struct ctl_table *table, int write,
361 void __user *buffer, size_t *lenp,
366 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
367 if (ret == 0 && write)
368 dirty_background_ratio = 0;
372 int dirty_ratio_handler(struct ctl_table *table, int write,
373 void __user *buffer, size_t *lenp,
376 int old_ratio = vm_dirty_ratio;
379 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
380 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
381 writeback_set_ratelimit();
387 int dirty_bytes_handler(struct ctl_table *table, int write,
388 void __user *buffer, size_t *lenp,
391 unsigned long old_bytes = vm_dirty_bytes;
394 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
395 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
396 writeback_set_ratelimit();
402 static unsigned long wp_next_time(unsigned long cur_time)
404 cur_time += VM_COMPLETIONS_PERIOD_LEN;
405 /* 0 has a special meaning... */
412 * Increment the wb's writeout completion count and the global writeout
413 * completion count. Called from test_clear_page_writeback().
415 static inline void __wb_writeout_inc(struct bdi_writeback *wb)
417 struct wb_domain *dom = &global_wb_domain;
419 __inc_wb_stat(wb, WB_WRITTEN);
420 __fprop_inc_percpu_max(&dom->completions, &wb->completions,
421 wb->bdi->max_prop_frac);
422 /* First event after period switching was turned off? */
423 if (!unlikely(dom->period_time)) {
425 * We can race with other __bdi_writeout_inc calls here but
426 * it does not cause any harm since the resulting time when
427 * timer will fire and what is in writeout_period_time will be
430 dom->period_time = wp_next_time(jiffies);
431 mod_timer(&dom->period_timer, dom->period_time);
435 void wb_writeout_inc(struct bdi_writeback *wb)
439 local_irq_save(flags);
440 __wb_writeout_inc(wb);
441 local_irq_restore(flags);
443 EXPORT_SYMBOL_GPL(wb_writeout_inc);
446 * On idle system, we can be called long after we scheduled because we use
447 * deferred timers so count with missed periods.
449 static void writeout_period(unsigned long t)
451 struct wb_domain *dom = (void *)t;
452 int miss_periods = (jiffies - dom->period_time) /
453 VM_COMPLETIONS_PERIOD_LEN;
455 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
456 dom->period_time = wp_next_time(dom->period_time +
457 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
458 mod_timer(&dom->period_timer, dom->period_time);
461 * Aging has zeroed all fractions. Stop wasting CPU on period
464 dom->period_time = 0;
468 int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
470 memset(dom, 0, sizeof(*dom));
472 spin_lock_init(&dom->lock);
474 init_timer_deferrable(&dom->period_timer);
475 dom->period_timer.function = writeout_period;
476 dom->period_timer.data = (unsigned long)dom;
478 dom->dirty_limit_tstamp = jiffies;
480 return fprop_global_init(&dom->completions, gfp);
484 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
485 * registered backing devices, which, for obvious reasons, can not
488 static unsigned int bdi_min_ratio;
490 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
494 spin_lock_bh(&bdi_lock);
495 if (min_ratio > bdi->max_ratio) {
498 min_ratio -= bdi->min_ratio;
499 if (bdi_min_ratio + min_ratio < 100) {
500 bdi_min_ratio += min_ratio;
501 bdi->min_ratio += min_ratio;
506 spin_unlock_bh(&bdi_lock);
511 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
518 spin_lock_bh(&bdi_lock);
519 if (bdi->min_ratio > max_ratio) {
522 bdi->max_ratio = max_ratio;
523 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
525 spin_unlock_bh(&bdi_lock);
529 EXPORT_SYMBOL(bdi_set_max_ratio);
531 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
532 unsigned long bg_thresh)
534 return (thresh + bg_thresh) / 2;
537 static unsigned long hard_dirty_limit(unsigned long thresh)
539 struct wb_domain *dom = &global_wb_domain;
541 return max(thresh, dom->dirty_limit);
545 * wb_calc_thresh - @wb's share of dirty throttling threshold
546 * @wb: bdi_writeback to query
547 * @dirty: global dirty limit in pages
549 * Returns @wb's dirty limit in pages. The term "dirty" in the context of
550 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
552 * Note that balance_dirty_pages() will only seriously take it as a hard limit
553 * when sleeping max_pause per page is not enough to keep the dirty pages under
554 * control. For example, when the device is completely stalled due to some error
555 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
556 * In the other normal situations, it acts more gently by throttling the tasks
557 * more (rather than completely block them) when the wb dirty pages go high.
559 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
560 * - starving fast devices
561 * - piling up dirty pages (that will take long time to sync) on slow devices
563 * The wb's share of dirty limit will be adapting to its throughput and
564 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
566 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
568 struct wb_domain *dom = &global_wb_domain;
570 long numerator, denominator;
571 unsigned long wb_min_ratio, wb_max_ratio;
574 * Calculate this BDI's share of the thresh ratio.
576 fprop_fraction_percpu(&dom->completions, &wb->completions,
577 &numerator, &denominator);
579 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
580 wb_thresh *= numerator;
581 do_div(wb_thresh, denominator);
583 wb_min_max_ratio(wb, &wb_min_ratio, &wb_max_ratio);
585 wb_thresh += (thresh * wb_min_ratio) / 100;
586 if (wb_thresh > (thresh * wb_max_ratio) / 100)
587 wb_thresh = thresh * wb_max_ratio / 100;
594 * f(dirty) := 1.0 + (----------------)
597 * it's a 3rd order polynomial that subjects to
599 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
600 * (2) f(setpoint) = 1.0 => the balance point
601 * (3) f(limit) = 0 => the hard limit
602 * (4) df/dx <= 0 => negative feedback control
603 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
604 * => fast response on large errors; small oscillation near setpoint
606 static long long pos_ratio_polynom(unsigned long setpoint,
613 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
614 limit - setpoint + 1);
616 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
617 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
618 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
620 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
624 * Dirty position control.
626 * (o) global/bdi setpoints
628 * We want the dirty pages be balanced around the global/wb setpoints.
629 * When the number of dirty pages is higher/lower than the setpoint, the
630 * dirty position control ratio (and hence task dirty ratelimit) will be
631 * decreased/increased to bring the dirty pages back to the setpoint.
633 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
635 * if (dirty < setpoint) scale up pos_ratio
636 * if (dirty > setpoint) scale down pos_ratio
638 * if (wb_dirty < wb_setpoint) scale up pos_ratio
639 * if (wb_dirty > wb_setpoint) scale down pos_ratio
641 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
643 * (o) global control line
647 * | |<===== global dirty control scope ======>|
655 * 1.0 ................................*
661 * 0 +------------.------------------.----------------------*------------->
662 * freerun^ setpoint^ limit^ dirty pages
664 * (o) wb control line
672 * | * |<=========== span ============>|
673 * 1.0 .......................*
685 * 1/4 ...............................................* * * * * * * * * * * *
689 * 0 +----------------------.-------------------------------.------------->
690 * wb_setpoint^ x_intercept^
692 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
693 * be smoothly throttled down to normal if it starts high in situations like
694 * - start writing to a slow SD card and a fast disk at the same time. The SD
695 * card's wb_dirty may rush to many times higher than wb_setpoint.
696 * - the wb dirty thresh drops quickly due to change of JBOD workload
698 static unsigned long wb_position_ratio(struct bdi_writeback *wb,
699 unsigned long thresh,
700 unsigned long bg_thresh,
702 unsigned long wb_thresh,
703 unsigned long wb_dirty)
705 unsigned long write_bw = wb->avg_write_bandwidth;
706 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
707 unsigned long limit = hard_dirty_limit(thresh);
708 unsigned long x_intercept;
709 unsigned long setpoint; /* dirty pages' target balance point */
710 unsigned long wb_setpoint;
712 long long pos_ratio; /* for scaling up/down the rate limit */
715 if (unlikely(dirty >= limit))
721 * See comment for pos_ratio_polynom().
723 setpoint = (freerun + limit) / 2;
724 pos_ratio = pos_ratio_polynom(setpoint, dirty, limit);
727 * The strictlimit feature is a tool preventing mistrusted filesystems
728 * from growing a large number of dirty pages before throttling. For
729 * such filesystems balance_dirty_pages always checks wb counters
730 * against wb limits. Even if global "nr_dirty" is under "freerun".
731 * This is especially important for fuse which sets bdi->max_ratio to
732 * 1% by default. Without strictlimit feature, fuse writeback may
733 * consume arbitrary amount of RAM because it is accounted in
734 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
736 * Here, in wb_position_ratio(), we calculate pos_ratio based on
737 * two values: wb_dirty and wb_thresh. Let's consider an example:
738 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
739 * limits are set by default to 10% and 20% (background and throttle).
740 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
741 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
742 * about ~6K pages (as the average of background and throttle wb
743 * limits). The 3rd order polynomial will provide positive feedback if
744 * wb_dirty is under wb_setpoint and vice versa.
746 * Note, that we cannot use global counters in these calculations
747 * because we want to throttle process writing to a strictlimit wb
748 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
749 * in the example above).
751 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
752 long long wb_pos_ratio;
753 unsigned long wb_bg_thresh;
756 return min_t(long long, pos_ratio * 2,
757 2 << RATELIMIT_CALC_SHIFT);
759 if (wb_dirty >= wb_thresh)
762 wb_bg_thresh = div_u64((u64)wb_thresh * bg_thresh, thresh);
763 wb_setpoint = dirty_freerun_ceiling(wb_thresh, wb_bg_thresh);
765 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
768 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, wb_dirty,
772 * Typically, for strictlimit case, wb_setpoint << setpoint
773 * and pos_ratio >> wb_pos_ratio. In the other words global
774 * state ("dirty") is not limiting factor and we have to
775 * make decision based on wb counters. But there is an
776 * important case when global pos_ratio should get precedence:
777 * global limits are exceeded (e.g. due to activities on other
778 * wb's) while given strictlimit wb is below limit.
780 * "pos_ratio * wb_pos_ratio" would work for the case above,
781 * but it would look too non-natural for the case of all
782 * activity in the system coming from a single strictlimit wb
783 * with bdi->max_ratio == 100%.
785 * Note that min() below somewhat changes the dynamics of the
786 * control system. Normally, pos_ratio value can be well over 3
787 * (when globally we are at freerun and wb is well below wb
788 * setpoint). Now the maximum pos_ratio in the same situation
789 * is 2. We might want to tweak this if we observe the control
790 * system is too slow to adapt.
792 return min(pos_ratio, wb_pos_ratio);
796 * We have computed basic pos_ratio above based on global situation. If
797 * the wb is over/under its share of dirty pages, we want to scale
798 * pos_ratio further down/up. That is done by the following mechanism.
804 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
806 * x_intercept - wb_dirty
807 * := --------------------------
808 * x_intercept - wb_setpoint
810 * The main wb control line is a linear function that subjects to
812 * (1) f(wb_setpoint) = 1.0
813 * (2) k = - 1 / (8 * write_bw) (in single wb case)
814 * or equally: x_intercept = wb_setpoint + 8 * write_bw
816 * For single wb case, the dirty pages are observed to fluctuate
817 * regularly within range
818 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
819 * for various filesystems, where (2) can yield in a reasonable 12.5%
820 * fluctuation range for pos_ratio.
822 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
823 * own size, so move the slope over accordingly and choose a slope that
824 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
826 if (unlikely(wb_thresh > thresh))
829 * It's very possible that wb_thresh is close to 0 not because the
830 * device is slow, but that it has remained inactive for long time.
831 * Honour such devices a reasonable good (hopefully IO efficient)
832 * threshold, so that the occasional writes won't be blocked and active
833 * writes can rampup the threshold quickly.
835 wb_thresh = max(wb_thresh, (limit - dirty) / 8);
837 * scale global setpoint to wb's:
838 * wb_setpoint = setpoint * wb_thresh / thresh
840 x = div_u64((u64)wb_thresh << 16, thresh + 1);
841 wb_setpoint = setpoint * (u64)x >> 16;
843 * Use span=(8*write_bw) in single wb case as indicated by
844 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
846 * wb_thresh thresh - wb_thresh
847 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
850 span = (thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
851 x_intercept = wb_setpoint + span;
853 if (wb_dirty < x_intercept - span / 4) {
854 pos_ratio = div64_u64(pos_ratio * (x_intercept - wb_dirty),
855 x_intercept - wb_setpoint + 1);
860 * wb reserve area, safeguard against dirty pool underrun and disk idle
861 * It may push the desired control point of global dirty pages higher
864 x_intercept = wb_thresh / 2;
865 if (wb_dirty < x_intercept) {
866 if (wb_dirty > x_intercept / 8)
867 pos_ratio = div_u64(pos_ratio * x_intercept, wb_dirty);
875 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
876 unsigned long elapsed,
877 unsigned long written)
879 const unsigned long period = roundup_pow_of_two(3 * HZ);
880 unsigned long avg = wb->avg_write_bandwidth;
881 unsigned long old = wb->write_bandwidth;
885 * bw = written * HZ / elapsed
887 * bw * elapsed + write_bandwidth * (period - elapsed)
888 * write_bandwidth = ---------------------------------------------------
891 * @written may have decreased due to account_page_redirty().
892 * Avoid underflowing @bw calculation.
894 bw = written - min(written, wb->written_stamp);
896 if (unlikely(elapsed > period)) {
901 bw += (u64)wb->write_bandwidth * (period - elapsed);
902 bw >>= ilog2(period);
905 * one more level of smoothing, for filtering out sudden spikes
907 if (avg > old && old >= (unsigned long)bw)
908 avg -= (avg - old) >> 3;
910 if (avg < old && old <= (unsigned long)bw)
911 avg += (old - avg) >> 3;
914 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
916 if (wb_has_dirty_io(wb)) {
917 long delta = avg - wb->avg_write_bandwidth;
918 WARN_ON_ONCE(atomic_long_add_return(delta,
919 &wb->bdi->tot_write_bandwidth) <= 0);
921 wb->write_bandwidth = bw;
922 wb->avg_write_bandwidth = avg;
925 static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
927 struct wb_domain *dom = &global_wb_domain;
928 unsigned long limit = dom->dirty_limit;
931 * Follow up in one step.
933 if (limit < thresh) {
939 * Follow down slowly. Use the higher one as the target, because thresh
940 * may drop below dirty. This is exactly the reason to introduce
941 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
943 thresh = max(thresh, dirty);
944 if (limit > thresh) {
945 limit -= (limit - thresh) >> 5;
950 dom->dirty_limit = limit;
953 static void global_update_bandwidth(unsigned long thresh,
957 struct wb_domain *dom = &global_wb_domain;
960 * check locklessly first to optimize away locking for the most time
962 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
965 spin_lock(&dom->lock);
966 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
967 update_dirty_limit(thresh, dirty);
968 dom->dirty_limit_tstamp = now;
970 spin_unlock(&dom->lock);
974 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
976 * Normal wb tasks will be curbed at or below it in long term.
977 * Obviously it should be around (write_bw / N) when there are N dd tasks.
979 static void wb_update_dirty_ratelimit(struct bdi_writeback *wb,
980 unsigned long thresh,
981 unsigned long bg_thresh,
983 unsigned long wb_thresh,
984 unsigned long wb_dirty,
985 unsigned long dirtied,
986 unsigned long elapsed)
988 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
989 unsigned long limit = hard_dirty_limit(thresh);
990 unsigned long setpoint = (freerun + limit) / 2;
991 unsigned long write_bw = wb->avg_write_bandwidth;
992 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
993 unsigned long dirty_rate;
994 unsigned long task_ratelimit;
995 unsigned long balanced_dirty_ratelimit;
996 unsigned long pos_ratio;
1001 * The dirty rate will match the writeout rate in long term, except
1002 * when dirty pages are truncated by userspace or re-dirtied by FS.
1004 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1006 pos_ratio = wb_position_ratio(wb, thresh, bg_thresh, dirty,
1007 wb_thresh, wb_dirty);
1009 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1011 task_ratelimit = (u64)dirty_ratelimit *
1012 pos_ratio >> RATELIMIT_CALC_SHIFT;
1013 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1016 * A linear estimation of the "balanced" throttle rate. The theory is,
1017 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1018 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1019 * formula will yield the balanced rate limit (write_bw / N).
1021 * Note that the expanded form is not a pure rate feedback:
1022 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1023 * but also takes pos_ratio into account:
1024 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1026 * (1) is not realistic because pos_ratio also takes part in balancing
1027 * the dirty rate. Consider the state
1028 * pos_ratio = 0.5 (3)
1029 * rate = 2 * (write_bw / N) (4)
1030 * If (1) is used, it will stuck in that state! Because each dd will
1032 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1034 * dirty_rate = N * task_ratelimit = write_bw (6)
1035 * put (6) into (1) we get
1036 * rate_(i+1) = rate_(i) (7)
1038 * So we end up using (2) to always keep
1039 * rate_(i+1) ~= (write_bw / N) (8)
1040 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1041 * pos_ratio is able to drive itself to 1.0, which is not only where
1042 * the dirty count meet the setpoint, but also where the slope of
1043 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1045 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1048 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1050 if (unlikely(balanced_dirty_ratelimit > write_bw))
1051 balanced_dirty_ratelimit = write_bw;
1054 * We could safely do this and return immediately:
1056 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1058 * However to get a more stable dirty_ratelimit, the below elaborated
1059 * code makes use of task_ratelimit to filter out singular points and
1060 * limit the step size.
1062 * The below code essentially only uses the relative value of
1064 * task_ratelimit - dirty_ratelimit
1065 * = (pos_ratio - 1) * dirty_ratelimit
1067 * which reflects the direction and size of dirty position error.
1071 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1072 * task_ratelimit is on the same side of dirty_ratelimit, too.
1074 * - dirty_ratelimit > balanced_dirty_ratelimit
1075 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1076 * lowering dirty_ratelimit will help meet both the position and rate
1077 * control targets. Otherwise, don't update dirty_ratelimit if it will
1078 * only help meet the rate target. After all, what the users ultimately
1079 * feel and care are stable dirty rate and small position error.
1081 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1082 * and filter out the singular points of balanced_dirty_ratelimit. Which
1083 * keeps jumping around randomly and can even leap far away at times
1084 * due to the small 200ms estimation period of dirty_rate (we want to
1085 * keep that period small to reduce time lags).
1090 * For strictlimit case, calculations above were based on wb counters
1091 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1092 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1093 * Hence, to calculate "step" properly, we have to use wb_dirty as
1094 * "dirty" and wb_setpoint as "setpoint".
1096 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1097 * it's possible that wb_thresh is close to zero due to inactivity
1098 * of backing device (see the implementation of wb_calc_thresh()).
1100 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1103 setpoint = wb_dirty + 1;
1105 setpoint = (wb_thresh +
1106 wb_calc_thresh(wb, bg_thresh)) / 2;
1109 if (dirty < setpoint) {
1110 x = min3(wb->balanced_dirty_ratelimit,
1111 balanced_dirty_ratelimit, task_ratelimit);
1112 if (dirty_ratelimit < x)
1113 step = x - dirty_ratelimit;
1115 x = max3(wb->balanced_dirty_ratelimit,
1116 balanced_dirty_ratelimit, task_ratelimit);
1117 if (dirty_ratelimit > x)
1118 step = dirty_ratelimit - x;
1122 * Don't pursue 100% rate matching. It's impossible since the balanced
1123 * rate itself is constantly fluctuating. So decrease the track speed
1124 * when it gets close to the target. Helps eliminate pointless tremors.
1126 step >>= dirty_ratelimit / (2 * step + 1);
1128 * Limit the tracking speed to avoid overshooting.
1130 step = (step + 7) / 8;
1132 if (dirty_ratelimit < balanced_dirty_ratelimit)
1133 dirty_ratelimit += step;
1135 dirty_ratelimit -= step;
1137 wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1138 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1140 trace_bdi_dirty_ratelimit(wb->bdi, dirty_rate, task_ratelimit);
1143 static void __wb_update_bandwidth(struct bdi_writeback *wb,
1144 unsigned long thresh,
1145 unsigned long bg_thresh,
1146 unsigned long dirty,
1147 unsigned long wb_thresh,
1148 unsigned long wb_dirty,
1149 unsigned long start_time,
1150 bool update_ratelimit)
1152 unsigned long now = jiffies;
1153 unsigned long elapsed = now - wb->bw_time_stamp;
1154 unsigned long dirtied;
1155 unsigned long written;
1157 lockdep_assert_held(&wb->list_lock);
1160 * rate-limit, only update once every 200ms.
1162 if (elapsed < BANDWIDTH_INTERVAL)
1165 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1166 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1169 * Skip quiet periods when disk bandwidth is under-utilized.
1170 * (at least 1s idle time between two flusher runs)
1172 if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1175 if (update_ratelimit) {
1176 global_update_bandwidth(thresh, dirty, now);
1177 wb_update_dirty_ratelimit(wb, thresh, bg_thresh, dirty,
1178 wb_thresh, wb_dirty,
1181 wb_update_write_bandwidth(wb, elapsed, written);
1184 wb->dirtied_stamp = dirtied;
1185 wb->written_stamp = written;
1186 wb->bw_time_stamp = now;
1189 void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1191 __wb_update_bandwidth(wb, 0, 0, 0, 0, 0, start_time, false);
1195 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1196 * will look to see if it needs to start dirty throttling.
1198 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1199 * global_page_state() too often. So scale it near-sqrt to the safety margin
1200 * (the number of pages we may dirty without exceeding the dirty limits).
1202 static unsigned long dirty_poll_interval(unsigned long dirty,
1203 unsigned long thresh)
1206 return 1UL << (ilog2(thresh - dirty) >> 1);
1211 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1212 unsigned long wb_dirty)
1214 unsigned long bw = wb->avg_write_bandwidth;
1218 * Limit pause time for small memory systems. If sleeping for too long
1219 * time, a small pool of dirty/writeback pages may go empty and disk go
1222 * 8 serves as the safety ratio.
1224 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1227 return min_t(unsigned long, t, MAX_PAUSE);
1230 static long wb_min_pause(struct bdi_writeback *wb,
1232 unsigned long task_ratelimit,
1233 unsigned long dirty_ratelimit,
1234 int *nr_dirtied_pause)
1236 long hi = ilog2(wb->avg_write_bandwidth);
1237 long lo = ilog2(wb->dirty_ratelimit);
1238 long t; /* target pause */
1239 long pause; /* estimated next pause */
1240 int pages; /* target nr_dirtied_pause */
1242 /* target for 10ms pause on 1-dd case */
1243 t = max(1, HZ / 100);
1246 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1249 * (N * 10ms) on 2^N concurrent tasks.
1252 t += (hi - lo) * (10 * HZ) / 1024;
1255 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1256 * on the much more stable dirty_ratelimit. However the next pause time
1257 * will be computed based on task_ratelimit and the two rate limits may
1258 * depart considerably at some time. Especially if task_ratelimit goes
1259 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1260 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1261 * result task_ratelimit won't be executed faithfully, which could
1262 * eventually bring down dirty_ratelimit.
1264 * We apply two rules to fix it up:
1265 * 1) try to estimate the next pause time and if necessary, use a lower
1266 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1267 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1268 * 2) limit the target pause time to max_pause/2, so that the normal
1269 * small fluctuations of task_ratelimit won't trigger rule (1) and
1270 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1272 t = min(t, 1 + max_pause / 2);
1273 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1276 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1277 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1278 * When the 16 consecutive reads are often interrupted by some dirty
1279 * throttling pause during the async writes, cfq will go into idles
1280 * (deadline is fine). So push nr_dirtied_pause as high as possible
1281 * until reaches DIRTY_POLL_THRESH=32 pages.
1283 if (pages < DIRTY_POLL_THRESH) {
1285 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1286 if (pages > DIRTY_POLL_THRESH) {
1287 pages = DIRTY_POLL_THRESH;
1288 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1292 pause = HZ * pages / (task_ratelimit + 1);
1293 if (pause > max_pause) {
1295 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1298 *nr_dirtied_pause = pages;
1300 * The minimal pause time will normally be half the target pause time.
1302 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1305 static inline void wb_dirty_limits(struct bdi_writeback *wb,
1306 unsigned long dirty_thresh,
1307 unsigned long background_thresh,
1308 unsigned long *wb_dirty,
1309 unsigned long *wb_thresh,
1310 unsigned long *wb_bg_thresh)
1312 unsigned long wb_reclaimable;
1315 * wb_thresh is not treated as some limiting factor as
1316 * dirty_thresh, due to reasons
1317 * - in JBOD setup, wb_thresh can fluctuate a lot
1318 * - in a system with HDD and USB key, the USB key may somehow
1319 * go into state (wb_dirty >> wb_thresh) either because
1320 * wb_dirty starts high, or because wb_thresh drops low.
1321 * In this case we don't want to hard throttle the USB key
1322 * dirtiers for 100 seconds until wb_dirty drops under
1323 * wb_thresh. Instead the auxiliary wb control line in
1324 * wb_position_ratio() will let the dirtier task progress
1325 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1327 *wb_thresh = wb_calc_thresh(wb, dirty_thresh);
1330 *wb_bg_thresh = dirty_thresh ? div_u64((u64)*wb_thresh *
1335 * In order to avoid the stacked BDI deadlock we need
1336 * to ensure we accurately count the 'dirty' pages when
1337 * the threshold is low.
1339 * Otherwise it would be possible to get thresh+n pages
1340 * reported dirty, even though there are thresh-m pages
1341 * actually dirty; with m+n sitting in the percpu
1344 if (*wb_thresh < 2 * wb_stat_error(wb)) {
1345 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1346 *wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1348 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1349 *wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1354 * balance_dirty_pages() must be called by processes which are generating dirty
1355 * data. It looks at the number of dirty pages in the machine and will force
1356 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1357 * If we're over `background_thresh' then the writeback threads are woken to
1358 * perform some writeout.
1360 static void balance_dirty_pages(struct address_space *mapping,
1361 struct bdi_writeback *wb,
1362 unsigned long pages_dirtied)
1364 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1365 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1366 unsigned long background_thresh;
1367 unsigned long dirty_thresh;
1372 int nr_dirtied_pause;
1373 bool dirty_exceeded = false;
1374 unsigned long task_ratelimit;
1375 unsigned long dirty_ratelimit;
1376 unsigned long pos_ratio;
1377 struct backing_dev_info *bdi = wb->bdi;
1378 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1379 unsigned long start_time = jiffies;
1382 unsigned long now = jiffies;
1383 unsigned long uninitialized_var(wb_thresh);
1384 unsigned long thresh;
1385 unsigned long uninitialized_var(wb_dirty);
1386 unsigned long dirty;
1387 unsigned long bg_thresh;
1390 * Unstable writes are a feature of certain networked
1391 * filesystems (i.e. NFS) in which data may have been
1392 * written to the server's write cache, but has not yet
1393 * been flushed to permanent storage.
1395 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1396 global_page_state(NR_UNSTABLE_NFS);
1397 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1399 global_dirty_limits(&background_thresh, &dirty_thresh);
1401 if (unlikely(strictlimit)) {
1402 wb_dirty_limits(wb, dirty_thresh, background_thresh,
1403 &wb_dirty, &wb_thresh, &bg_thresh);
1409 thresh = dirty_thresh;
1410 bg_thresh = background_thresh;
1414 * Throttle it only when the background writeback cannot
1415 * catch-up. This avoids (excessively) small writeouts
1416 * when the wb limits are ramping up in case of !strictlimit.
1418 * In strictlimit case make decision based on the wb counters
1419 * and limits. Small writeouts when the wb limits are ramping
1420 * up are the price we consciously pay for strictlimit-ing.
1422 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh)) {
1423 current->dirty_paused_when = now;
1424 current->nr_dirtied = 0;
1425 current->nr_dirtied_pause =
1426 dirty_poll_interval(dirty, thresh);
1430 if (unlikely(!writeback_in_progress(wb)))
1431 wb_start_background_writeback(wb);
1434 wb_dirty_limits(wb, dirty_thresh, background_thresh,
1435 &wb_dirty, &wb_thresh, NULL);
1437 dirty_exceeded = (wb_dirty > wb_thresh) &&
1438 ((nr_dirty > dirty_thresh) || strictlimit);
1439 if (dirty_exceeded && !wb->dirty_exceeded)
1440 wb->dirty_exceeded = 1;
1442 if (time_is_before_jiffies(wb->bw_time_stamp +
1443 BANDWIDTH_INTERVAL)) {
1444 spin_lock(&wb->list_lock);
1445 __wb_update_bandwidth(wb, dirty_thresh,
1446 background_thresh, nr_dirty,
1447 wb_thresh, wb_dirty, start_time,
1449 spin_unlock(&wb->list_lock);
1452 dirty_ratelimit = wb->dirty_ratelimit;
1453 pos_ratio = wb_position_ratio(wb, dirty_thresh,
1454 background_thresh, nr_dirty,
1455 wb_thresh, wb_dirty);
1456 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1457 RATELIMIT_CALC_SHIFT;
1458 max_pause = wb_max_pause(wb, wb_dirty);
1459 min_pause = wb_min_pause(wb, max_pause,
1460 task_ratelimit, dirty_ratelimit,
1463 if (unlikely(task_ratelimit == 0)) {
1468 period = HZ * pages_dirtied / task_ratelimit;
1470 if (current->dirty_paused_when)
1471 pause -= now - current->dirty_paused_when;
1473 * For less than 1s think time (ext3/4 may block the dirtier
1474 * for up to 800ms from time to time on 1-HDD; so does xfs,
1475 * however at much less frequency), try to compensate it in
1476 * future periods by updating the virtual time; otherwise just
1477 * do a reset, as it may be a light dirtier.
1479 if (pause < min_pause) {
1480 trace_balance_dirty_pages(bdi,
1493 current->dirty_paused_when = now;
1494 current->nr_dirtied = 0;
1495 } else if (period) {
1496 current->dirty_paused_when += period;
1497 current->nr_dirtied = 0;
1498 } else if (current->nr_dirtied_pause <= pages_dirtied)
1499 current->nr_dirtied_pause += pages_dirtied;
1502 if (unlikely(pause > max_pause)) {
1503 /* for occasional dropped task_ratelimit */
1504 now += min(pause - max_pause, max_pause);
1509 trace_balance_dirty_pages(bdi,
1521 __set_current_state(TASK_KILLABLE);
1522 io_schedule_timeout(pause);
1524 current->dirty_paused_when = now + pause;
1525 current->nr_dirtied = 0;
1526 current->nr_dirtied_pause = nr_dirtied_pause;
1529 * This is typically equal to (nr_dirty < dirty_thresh) and can
1530 * also keep "1000+ dd on a slow USB stick" under control.
1536 * In the case of an unresponding NFS server and the NFS dirty
1537 * pages exceeds dirty_thresh, give the other good wb's a pipe
1538 * to go through, so that tasks on them still remain responsive.
1540 * In theory 1 page is enough to keep the comsumer-producer
1541 * pipe going: the flusher cleans 1 page => the task dirties 1
1542 * more page. However wb_dirty has accounting errors. So use
1543 * the larger and more IO friendly wb_stat_error.
1545 if (wb_dirty <= wb_stat_error(wb))
1548 if (fatal_signal_pending(current))
1552 if (!dirty_exceeded && wb->dirty_exceeded)
1553 wb->dirty_exceeded = 0;
1555 if (writeback_in_progress(wb))
1559 * In laptop mode, we wait until hitting the higher threshold before
1560 * starting background writeout, and then write out all the way down
1561 * to the lower threshold. So slow writers cause minimal disk activity.
1563 * In normal mode, we start background writeout at the lower
1564 * background_thresh, to keep the amount of dirty memory low.
1569 if (nr_reclaimable > background_thresh)
1570 wb_start_background_writeback(wb);
1573 static DEFINE_PER_CPU(int, bdp_ratelimits);
1576 * Normal tasks are throttled by
1578 * dirty tsk->nr_dirtied_pause pages;
1579 * take a snap in balance_dirty_pages();
1581 * However there is a worst case. If every task exit immediately when dirtied
1582 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1583 * called to throttle the page dirties. The solution is to save the not yet
1584 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1585 * randomly into the running tasks. This works well for the above worst case,
1586 * as the new task will pick up and accumulate the old task's leaked dirty
1587 * count and eventually get throttled.
1589 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1592 * balance_dirty_pages_ratelimited - balance dirty memory state
1593 * @mapping: address_space which was dirtied
1595 * Processes which are dirtying memory should call in here once for each page
1596 * which was newly dirtied. The function will periodically check the system's
1597 * dirty state and will initiate writeback if needed.
1599 * On really big machines, get_writeback_state is expensive, so try to avoid
1600 * calling it too often (ratelimiting). But once we're over the dirty memory
1601 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1602 * from overshooting the limit by (ratelimit_pages) each.
1604 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1606 struct inode *inode = mapping->host;
1607 struct backing_dev_info *bdi = inode_to_bdi(inode);
1608 struct bdi_writeback *wb = NULL;
1612 if (!bdi_cap_account_dirty(bdi))
1615 if (inode_cgwb_enabled(inode))
1616 wb = wb_get_create_current(bdi, GFP_KERNEL);
1620 ratelimit = current->nr_dirtied_pause;
1621 if (wb->dirty_exceeded)
1622 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1626 * This prevents one CPU to accumulate too many dirtied pages without
1627 * calling into balance_dirty_pages(), which can happen when there are
1628 * 1000+ tasks, all of them start dirtying pages at exactly the same
1629 * time, hence all honoured too large initial task->nr_dirtied_pause.
1631 p = this_cpu_ptr(&bdp_ratelimits);
1632 if (unlikely(current->nr_dirtied >= ratelimit))
1634 else if (unlikely(*p >= ratelimit_pages)) {
1639 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1640 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1641 * the dirty throttling and livelock other long-run dirtiers.
1643 p = this_cpu_ptr(&dirty_throttle_leaks);
1644 if (*p > 0 && current->nr_dirtied < ratelimit) {
1645 unsigned long nr_pages_dirtied;
1646 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1647 *p -= nr_pages_dirtied;
1648 current->nr_dirtied += nr_pages_dirtied;
1652 if (unlikely(current->nr_dirtied >= ratelimit))
1653 balance_dirty_pages(mapping, wb, current->nr_dirtied);
1657 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1659 void throttle_vm_writeout(gfp_t gfp_mask)
1661 unsigned long background_thresh;
1662 unsigned long dirty_thresh;
1665 global_dirty_limits(&background_thresh, &dirty_thresh);
1666 dirty_thresh = hard_dirty_limit(dirty_thresh);
1669 * Boost the allowable dirty threshold a bit for page
1670 * allocators so they don't get DoS'ed by heavy writers
1672 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1674 if (global_page_state(NR_UNSTABLE_NFS) +
1675 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1677 congestion_wait(BLK_RW_ASYNC, HZ/10);
1680 * The caller might hold locks which can prevent IO completion
1681 * or progress in the filesystem. So we cannot just sit here
1682 * waiting for IO to complete.
1684 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1690 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1692 int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1693 void __user *buffer, size_t *length, loff_t *ppos)
1695 proc_dointvec(table, write, buffer, length, ppos);
1700 void laptop_mode_timer_fn(unsigned long data)
1702 struct request_queue *q = (struct request_queue *)data;
1703 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1704 global_page_state(NR_UNSTABLE_NFS);
1705 struct bdi_writeback *wb;
1706 struct wb_iter iter;
1709 * We want to write everything out, not just down to the dirty
1712 if (!bdi_has_dirty_io(&q->backing_dev_info))
1715 bdi_for_each_wb(wb, &q->backing_dev_info, &iter, 0)
1716 if (wb_has_dirty_io(wb))
1717 wb_start_writeback(wb, nr_pages, true,
1718 WB_REASON_LAPTOP_TIMER);
1722 * We've spun up the disk and we're in laptop mode: schedule writeback
1723 * of all dirty data a few seconds from now. If the flush is already scheduled
1724 * then push it back - the user is still using the disk.
1726 void laptop_io_completion(struct backing_dev_info *info)
1728 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1732 * We're in laptop mode and we've just synced. The sync's writes will have
1733 * caused another writeback to be scheduled by laptop_io_completion.
1734 * Nothing needs to be written back anymore, so we unschedule the writeback.
1736 void laptop_sync_completion(void)
1738 struct backing_dev_info *bdi;
1742 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1743 del_timer(&bdi->laptop_mode_wb_timer);
1750 * If ratelimit_pages is too high then we can get into dirty-data overload
1751 * if a large number of processes all perform writes at the same time.
1752 * If it is too low then SMP machines will call the (expensive)
1753 * get_writeback_state too often.
1755 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1756 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1760 void writeback_set_ratelimit(void)
1762 struct wb_domain *dom = &global_wb_domain;
1763 unsigned long background_thresh;
1764 unsigned long dirty_thresh;
1766 global_dirty_limits(&background_thresh, &dirty_thresh);
1767 dom->dirty_limit = dirty_thresh;
1768 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1769 if (ratelimit_pages < 16)
1770 ratelimit_pages = 16;
1774 ratelimit_handler(struct notifier_block *self, unsigned long action,
1778 switch (action & ~CPU_TASKS_FROZEN) {
1781 writeback_set_ratelimit();
1788 static struct notifier_block ratelimit_nb = {
1789 .notifier_call = ratelimit_handler,
1794 * Called early on to tune the page writeback dirty limits.
1796 * We used to scale dirty pages according to how total memory
1797 * related to pages that could be allocated for buffers (by
1798 * comparing nr_free_buffer_pages() to vm_total_pages.
1800 * However, that was when we used "dirty_ratio" to scale with
1801 * all memory, and we don't do that any more. "dirty_ratio"
1802 * is now applied to total non-HIGHPAGE memory (by subtracting
1803 * totalhigh_pages from vm_total_pages), and as such we can't
1804 * get into the old insane situation any more where we had
1805 * large amounts of dirty pages compared to a small amount of
1806 * non-HIGHMEM memory.
1808 * But we might still want to scale the dirty_ratio by how
1809 * much memory the box has..
1811 void __init page_writeback_init(void)
1813 writeback_set_ratelimit();
1814 register_cpu_notifier(&ratelimit_nb);
1816 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
1820 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1821 * @mapping: address space structure to write
1822 * @start: starting page index
1823 * @end: ending page index (inclusive)
1825 * This function scans the page range from @start to @end (inclusive) and tags
1826 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1827 * that write_cache_pages (or whoever calls this function) will then use
1828 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1829 * used to avoid livelocking of writeback by a process steadily creating new
1830 * dirty pages in the file (thus it is important for this function to be quick
1831 * so that it can tag pages faster than a dirtying process can create them).
1834 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1836 void tag_pages_for_writeback(struct address_space *mapping,
1837 pgoff_t start, pgoff_t end)
1839 #define WRITEBACK_TAG_BATCH 4096
1840 unsigned long tagged;
1843 spin_lock_irq(&mapping->tree_lock);
1844 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1845 &start, end, WRITEBACK_TAG_BATCH,
1846 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1847 spin_unlock_irq(&mapping->tree_lock);
1848 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1850 /* We check 'start' to handle wrapping when end == ~0UL */
1851 } while (tagged >= WRITEBACK_TAG_BATCH && start);
1853 EXPORT_SYMBOL(tag_pages_for_writeback);
1856 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1857 * @mapping: address space structure to write
1858 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1859 * @writepage: function called for each page
1860 * @data: data passed to writepage function
1862 * If a page is already under I/O, write_cache_pages() skips it, even
1863 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1864 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1865 * and msync() need to guarantee that all the data which was dirty at the time
1866 * the call was made get new I/O started against them. If wbc->sync_mode is
1867 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1868 * existing IO to complete.
1870 * To avoid livelocks (when other process dirties new pages), we first tag
1871 * pages which should be written back with TOWRITE tag and only then start
1872 * writing them. For data-integrity sync we have to be careful so that we do
1873 * not miss some pages (e.g., because some other process has cleared TOWRITE
1874 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1875 * by the process clearing the DIRTY tag (and submitting the page for IO).
1877 int write_cache_pages(struct address_space *mapping,
1878 struct writeback_control *wbc, writepage_t writepage,
1883 struct pagevec pvec;
1885 pgoff_t uninitialized_var(writeback_index);
1887 pgoff_t end; /* Inclusive */
1890 int range_whole = 0;
1893 pagevec_init(&pvec, 0);
1894 if (wbc->range_cyclic) {
1895 writeback_index = mapping->writeback_index; /* prev offset */
1896 index = writeback_index;
1903 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1904 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1905 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1907 cycled = 1; /* ignore range_cyclic tests */
1909 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1910 tag = PAGECACHE_TAG_TOWRITE;
1912 tag = PAGECACHE_TAG_DIRTY;
1914 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1915 tag_pages_for_writeback(mapping, index, end);
1917 while (!done && (index <= end)) {
1920 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1921 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1925 for (i = 0; i < nr_pages; i++) {
1926 struct page *page = pvec.pages[i];
1929 * At this point, the page may be truncated or
1930 * invalidated (changing page->mapping to NULL), or
1931 * even swizzled back from swapper_space to tmpfs file
1932 * mapping. However, page->index will not change
1933 * because we have a reference on the page.
1935 if (page->index > end) {
1937 * can't be range_cyclic (1st pass) because
1938 * end == -1 in that case.
1944 done_index = page->index;
1949 * Page truncated or invalidated. We can freely skip it
1950 * then, even for data integrity operations: the page
1951 * has disappeared concurrently, so there could be no
1952 * real expectation of this data interity operation
1953 * even if there is now a new, dirty page at the same
1954 * pagecache address.
1956 if (unlikely(page->mapping != mapping)) {
1962 if (!PageDirty(page)) {
1963 /* someone wrote it for us */
1964 goto continue_unlock;
1967 if (PageWriteback(page)) {
1968 if (wbc->sync_mode != WB_SYNC_NONE)
1969 wait_on_page_writeback(page);
1971 goto continue_unlock;
1974 BUG_ON(PageWriteback(page));
1975 if (!clear_page_dirty_for_io(page))
1976 goto continue_unlock;
1978 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
1979 ret = (*writepage)(page, wbc, data);
1980 if (unlikely(ret)) {
1981 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1986 * done_index is set past this page,
1987 * so media errors will not choke
1988 * background writeout for the entire
1989 * file. This has consequences for
1990 * range_cyclic semantics (ie. it may
1991 * not be suitable for data integrity
1994 done_index = page->index + 1;
2001 * We stop writing back only if we are not doing
2002 * integrity sync. In case of integrity sync we have to
2003 * keep going until we have written all the pages
2004 * we tagged for writeback prior to entering this loop.
2006 if (--wbc->nr_to_write <= 0 &&
2007 wbc->sync_mode == WB_SYNC_NONE) {
2012 pagevec_release(&pvec);
2015 if (!cycled && !done) {
2018 * We hit the last page and there is more work to be done: wrap
2019 * back to the start of the file
2023 end = writeback_index - 1;
2026 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2027 mapping->writeback_index = done_index;
2031 EXPORT_SYMBOL(write_cache_pages);
2034 * Function used by generic_writepages to call the real writepage
2035 * function and set the mapping flags on error
2037 static int __writepage(struct page *page, struct writeback_control *wbc,
2040 struct address_space *mapping = data;
2041 int ret = mapping->a_ops->writepage(page, wbc);
2042 mapping_set_error(mapping, ret);
2047 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2048 * @mapping: address space structure to write
2049 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2051 * This is a library function, which implements the writepages()
2052 * address_space_operation.
2054 int generic_writepages(struct address_space *mapping,
2055 struct writeback_control *wbc)
2057 struct blk_plug plug;
2060 /* deal with chardevs and other special file */
2061 if (!mapping->a_ops->writepage)
2064 blk_start_plug(&plug);
2065 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2066 blk_finish_plug(&plug);
2070 EXPORT_SYMBOL(generic_writepages);
2072 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2076 if (wbc->nr_to_write <= 0)
2078 if (mapping->a_ops->writepages)
2079 ret = mapping->a_ops->writepages(mapping, wbc);
2081 ret = generic_writepages(mapping, wbc);
2086 * write_one_page - write out a single page and optionally wait on I/O
2087 * @page: the page to write
2088 * @wait: if true, wait on writeout
2090 * The page must be locked by the caller and will be unlocked upon return.
2092 * write_one_page() returns a negative error code if I/O failed.
2094 int write_one_page(struct page *page, int wait)
2096 struct address_space *mapping = page->mapping;
2098 struct writeback_control wbc = {
2099 .sync_mode = WB_SYNC_ALL,
2103 BUG_ON(!PageLocked(page));
2106 wait_on_page_writeback(page);
2108 if (clear_page_dirty_for_io(page)) {
2109 page_cache_get(page);
2110 ret = mapping->a_ops->writepage(page, &wbc);
2111 if (ret == 0 && wait) {
2112 wait_on_page_writeback(page);
2113 if (PageError(page))
2116 page_cache_release(page);
2122 EXPORT_SYMBOL(write_one_page);
2125 * For address_spaces which do not use buffers nor write back.
2127 int __set_page_dirty_no_writeback(struct page *page)
2129 if (!PageDirty(page))
2130 return !TestSetPageDirty(page);
2135 * Helper function for set_page_dirty family.
2137 * Caller must hold mem_cgroup_begin_page_stat().
2139 * NOTE: This relies on being atomic wrt interrupts.
2141 void account_page_dirtied(struct page *page, struct address_space *mapping,
2142 struct mem_cgroup *memcg)
2144 struct inode *inode = mapping->host;
2146 trace_writeback_dirty_page(page, mapping);
2148 if (mapping_cap_account_dirty(mapping)) {
2149 struct bdi_writeback *wb;
2151 inode_attach_wb(inode, page);
2152 wb = inode_to_wb(inode);
2154 mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
2155 __inc_zone_page_state(page, NR_FILE_DIRTY);
2156 __inc_zone_page_state(page, NR_DIRTIED);
2157 __inc_wb_stat(wb, WB_RECLAIMABLE);
2158 __inc_wb_stat(wb, WB_DIRTIED);
2159 task_io_account_write(PAGE_CACHE_SIZE);
2160 current->nr_dirtied++;
2161 this_cpu_inc(bdp_ratelimits);
2164 EXPORT_SYMBOL(account_page_dirtied);
2167 * Helper function for deaccounting dirty page without writeback.
2169 * Caller must hold mem_cgroup_begin_page_stat().
2171 void account_page_cleaned(struct page *page, struct address_space *mapping,
2172 struct mem_cgroup *memcg)
2174 if (mapping_cap_account_dirty(mapping)) {
2175 mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
2176 dec_zone_page_state(page, NR_FILE_DIRTY);
2177 dec_wb_stat(inode_to_wb(mapping->host), WB_RECLAIMABLE);
2178 task_io_account_cancelled_write(PAGE_CACHE_SIZE);
2183 * For address_spaces which do not use buffers. Just tag the page as dirty in
2186 * This is also used when a single buffer is being dirtied: we want to set the
2187 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2188 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2190 * The caller must ensure this doesn't race with truncation. Most will simply
2191 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2192 * the pte lock held, which also locks out truncation.
2194 int __set_page_dirty_nobuffers(struct page *page)
2196 struct mem_cgroup *memcg;
2198 memcg = mem_cgroup_begin_page_stat(page);
2199 if (!TestSetPageDirty(page)) {
2200 struct address_space *mapping = page_mapping(page);
2201 unsigned long flags;
2204 mem_cgroup_end_page_stat(memcg);
2208 spin_lock_irqsave(&mapping->tree_lock, flags);
2209 BUG_ON(page_mapping(page) != mapping);
2210 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2211 account_page_dirtied(page, mapping, memcg);
2212 radix_tree_tag_set(&mapping->page_tree, page_index(page),
2213 PAGECACHE_TAG_DIRTY);
2214 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2215 mem_cgroup_end_page_stat(memcg);
2217 if (mapping->host) {
2218 /* !PageAnon && !swapper_space */
2219 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2223 mem_cgroup_end_page_stat(memcg);
2226 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2229 * Call this whenever redirtying a page, to de-account the dirty counters
2230 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2231 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2232 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2235 void account_page_redirty(struct page *page)
2237 struct address_space *mapping = page->mapping;
2239 if (mapping && mapping_cap_account_dirty(mapping)) {
2240 struct bdi_writeback *wb = inode_to_wb(mapping->host);
2242 current->nr_dirtied--;
2243 dec_zone_page_state(page, NR_DIRTIED);
2244 dec_wb_stat(wb, WB_DIRTIED);
2247 EXPORT_SYMBOL(account_page_redirty);
2250 * When a writepage implementation decides that it doesn't want to write this
2251 * page for some reason, it should redirty the locked page via
2252 * redirty_page_for_writepage() and it should then unlock the page and return 0
2254 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2258 wbc->pages_skipped++;
2259 ret = __set_page_dirty_nobuffers(page);
2260 account_page_redirty(page);
2263 EXPORT_SYMBOL(redirty_page_for_writepage);
2268 * For pages with a mapping this should be done under the page lock
2269 * for the benefit of asynchronous memory errors who prefer a consistent
2270 * dirty state. This rule can be broken in some special cases,
2271 * but should be better not to.
2273 * If the mapping doesn't provide a set_page_dirty a_op, then
2274 * just fall through and assume that it wants buffer_heads.
2276 int set_page_dirty(struct page *page)
2278 struct address_space *mapping = page_mapping(page);
2280 if (likely(mapping)) {
2281 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2283 * readahead/lru_deactivate_page could remain
2284 * PG_readahead/PG_reclaim due to race with end_page_writeback
2285 * About readahead, if the page is written, the flags would be
2286 * reset. So no problem.
2287 * About lru_deactivate_page, if the page is redirty, the flag
2288 * will be reset. So no problem. but if the page is used by readahead
2289 * it will confuse readahead and make it restart the size rampup
2290 * process. But it's a trivial problem.
2292 if (PageReclaim(page))
2293 ClearPageReclaim(page);
2296 spd = __set_page_dirty_buffers;
2298 return (*spd)(page);
2300 if (!PageDirty(page)) {
2301 if (!TestSetPageDirty(page))
2306 EXPORT_SYMBOL(set_page_dirty);
2309 * set_page_dirty() is racy if the caller has no reference against
2310 * page->mapping->host, and if the page is unlocked. This is because another
2311 * CPU could truncate the page off the mapping and then free the mapping.
2313 * Usually, the page _is_ locked, or the caller is a user-space process which
2314 * holds a reference on the inode by having an open file.
2316 * In other cases, the page should be locked before running set_page_dirty().
2318 int set_page_dirty_lock(struct page *page)
2323 ret = set_page_dirty(page);
2327 EXPORT_SYMBOL(set_page_dirty_lock);
2330 * This cancels just the dirty bit on the kernel page itself, it does NOT
2331 * actually remove dirty bits on any mmap's that may be around. It also
2332 * leaves the page tagged dirty, so any sync activity will still find it on
2333 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2334 * look at the dirty bits in the VM.
2336 * Doing this should *normally* only ever be done when a page is truncated,
2337 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2338 * this when it notices that somebody has cleaned out all the buffers on a
2339 * page without actually doing it through the VM. Can you say "ext3 is
2340 * horribly ugly"? Thought you could.
2342 void cancel_dirty_page(struct page *page)
2344 struct address_space *mapping = page_mapping(page);
2346 if (mapping_cap_account_dirty(mapping)) {
2347 struct mem_cgroup *memcg;
2349 memcg = mem_cgroup_begin_page_stat(page);
2351 if (TestClearPageDirty(page))
2352 account_page_cleaned(page, mapping, memcg);
2354 mem_cgroup_end_page_stat(memcg);
2356 ClearPageDirty(page);
2359 EXPORT_SYMBOL(cancel_dirty_page);
2362 * Clear a page's dirty flag, while caring for dirty memory accounting.
2363 * Returns true if the page was previously dirty.
2365 * This is for preparing to put the page under writeout. We leave the page
2366 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2367 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2368 * implementation will run either set_page_writeback() or set_page_dirty(),
2369 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2372 * This incoherency between the page's dirty flag and radix-tree tag is
2373 * unfortunate, but it only exists while the page is locked.
2375 int clear_page_dirty_for_io(struct page *page)
2377 struct address_space *mapping = page_mapping(page);
2378 struct mem_cgroup *memcg;
2381 BUG_ON(!PageLocked(page));
2383 if (mapping && mapping_cap_account_dirty(mapping)) {
2385 * Yes, Virginia, this is indeed insane.
2387 * We use this sequence to make sure that
2388 * (a) we account for dirty stats properly
2389 * (b) we tell the low-level filesystem to
2390 * mark the whole page dirty if it was
2391 * dirty in a pagetable. Only to then
2392 * (c) clean the page again and return 1 to
2393 * cause the writeback.
2395 * This way we avoid all nasty races with the
2396 * dirty bit in multiple places and clearing
2397 * them concurrently from different threads.
2399 * Note! Normally the "set_page_dirty(page)"
2400 * has no effect on the actual dirty bit - since
2401 * that will already usually be set. But we
2402 * need the side effects, and it can help us
2405 * We basically use the page "master dirty bit"
2406 * as a serialization point for all the different
2407 * threads doing their things.
2409 if (page_mkclean(page))
2410 set_page_dirty(page);
2412 * We carefully synchronise fault handlers against
2413 * installing a dirty pte and marking the page dirty
2414 * at this point. We do this by having them hold the
2415 * page lock while dirtying the page, and pages are
2416 * always locked coming in here, so we get the desired
2419 memcg = mem_cgroup_begin_page_stat(page);
2420 if (TestClearPageDirty(page)) {
2421 mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
2422 dec_zone_page_state(page, NR_FILE_DIRTY);
2423 dec_wb_stat(inode_to_wb(mapping->host), WB_RECLAIMABLE);
2426 mem_cgroup_end_page_stat(memcg);
2429 return TestClearPageDirty(page);
2431 EXPORT_SYMBOL(clear_page_dirty_for_io);
2433 int test_clear_page_writeback(struct page *page)
2435 struct address_space *mapping = page_mapping(page);
2436 struct mem_cgroup *memcg;
2439 memcg = mem_cgroup_begin_page_stat(page);
2441 struct inode *inode = mapping->host;
2442 struct backing_dev_info *bdi = inode_to_bdi(inode);
2443 unsigned long flags;
2445 spin_lock_irqsave(&mapping->tree_lock, flags);
2446 ret = TestClearPageWriteback(page);
2448 radix_tree_tag_clear(&mapping->page_tree,
2450 PAGECACHE_TAG_WRITEBACK);
2451 if (bdi_cap_account_writeback(bdi)) {
2452 struct bdi_writeback *wb = inode_to_wb(inode);
2454 __dec_wb_stat(wb, WB_WRITEBACK);
2455 __wb_writeout_inc(wb);
2458 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2460 ret = TestClearPageWriteback(page);
2463 mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
2464 dec_zone_page_state(page, NR_WRITEBACK);
2465 inc_zone_page_state(page, NR_WRITTEN);
2467 mem_cgroup_end_page_stat(memcg);
2471 int __test_set_page_writeback(struct page *page, bool keep_write)
2473 struct address_space *mapping = page_mapping(page);
2474 struct mem_cgroup *memcg;
2477 memcg = mem_cgroup_begin_page_stat(page);
2479 struct inode *inode = mapping->host;
2480 struct backing_dev_info *bdi = inode_to_bdi(inode);
2481 unsigned long flags;
2483 spin_lock_irqsave(&mapping->tree_lock, flags);
2484 ret = TestSetPageWriteback(page);
2486 radix_tree_tag_set(&mapping->page_tree,
2488 PAGECACHE_TAG_WRITEBACK);
2489 if (bdi_cap_account_writeback(bdi))
2490 __inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2492 if (!PageDirty(page))
2493 radix_tree_tag_clear(&mapping->page_tree,
2495 PAGECACHE_TAG_DIRTY);
2497 radix_tree_tag_clear(&mapping->page_tree,
2499 PAGECACHE_TAG_TOWRITE);
2500 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2502 ret = TestSetPageWriteback(page);
2505 mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
2506 inc_zone_page_state(page, NR_WRITEBACK);
2508 mem_cgroup_end_page_stat(memcg);
2512 EXPORT_SYMBOL(__test_set_page_writeback);
2515 * Return true if any of the pages in the mapping are marked with the
2518 int mapping_tagged(struct address_space *mapping, int tag)
2520 return radix_tree_tagged(&mapping->page_tree, tag);
2522 EXPORT_SYMBOL(mapping_tagged);
2525 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2526 * @page: The page to wait on.
2528 * This function determines if the given page is related to a backing device
2529 * that requires page contents to be held stable during writeback. If so, then
2530 * it will wait for any pending writeback to complete.
2532 void wait_for_stable_page(struct page *page)
2534 if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2535 wait_on_page_writeback(page);
2537 EXPORT_SYMBOL_GPL(wait_for_stable_page);