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;
127 /* consolidated parameters for balance_dirty_pages() and its subroutines */
128 struct dirty_throttle_control {
129 #ifdef CONFIG_CGROUP_WRITEBACK
130 struct wb_domain *dom;
131 struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */
133 struct bdi_writeback *wb;
134 struct fprop_local_percpu *wb_completions;
136 unsigned long avail; /* dirtyable */
137 unsigned long dirty; /* file_dirty + write + nfs */
138 unsigned long thresh; /* dirty threshold */
139 unsigned long bg_thresh; /* dirty background threshold */
141 unsigned long wb_dirty; /* per-wb counterparts */
142 unsigned long wb_thresh;
143 unsigned long wb_bg_thresh;
145 unsigned long pos_ratio;
149 * Length of period for aging writeout fractions of bdis. This is an
150 * arbitrarily chosen number. The longer the period, the slower fractions will
151 * reflect changes in current writeout rate.
153 #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
155 #ifdef CONFIG_CGROUP_WRITEBACK
157 #define GDTC_INIT(__wb) .wb = (__wb), \
158 .dom = &global_wb_domain, \
159 .wb_completions = &(__wb)->completions
161 #define GDTC_INIT_NO_WB .dom = &global_wb_domain
163 #define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
164 .dom = mem_cgroup_wb_domain(__wb), \
165 .wb_completions = &(__wb)->memcg_completions, \
168 static bool mdtc_valid(struct dirty_throttle_control *dtc)
173 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
178 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
183 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
185 return &wb->memcg_completions;
188 static void wb_min_max_ratio(struct bdi_writeback *wb,
189 unsigned long *minp, unsigned long *maxp)
191 unsigned long this_bw = wb->avg_write_bandwidth;
192 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
193 unsigned long long min = wb->bdi->min_ratio;
194 unsigned long long max = wb->bdi->max_ratio;
197 * @wb may already be clean by the time control reaches here and
198 * the total may not include its bw.
200 if (this_bw < tot_bw) {
215 #else /* CONFIG_CGROUP_WRITEBACK */
217 #define GDTC_INIT(__wb) .wb = (__wb), \
218 .wb_completions = &(__wb)->completions
219 #define GDTC_INIT_NO_WB
220 #define MDTC_INIT(__wb, __gdtc)
222 static bool mdtc_valid(struct dirty_throttle_control *dtc)
227 static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
229 return &global_wb_domain;
232 static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
237 static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
242 static void wb_min_max_ratio(struct bdi_writeback *wb,
243 unsigned long *minp, unsigned long *maxp)
245 *minp = wb->bdi->min_ratio;
246 *maxp = wb->bdi->max_ratio;
249 #endif /* CONFIG_CGROUP_WRITEBACK */
252 * In a memory zone, there is a certain amount of pages we consider
253 * available for the page cache, which is essentially the number of
254 * free and reclaimable pages, minus some zone reserves to protect
255 * lowmem and the ability to uphold the zone's watermarks without
256 * requiring writeback.
258 * This number of dirtyable pages is the base value of which the
259 * user-configurable dirty ratio is the effictive number of pages that
260 * are allowed to be actually dirtied. Per individual zone, or
261 * globally by using the sum of dirtyable pages over all zones.
263 * Because the user is allowed to specify the dirty limit globally as
264 * absolute number of bytes, calculating the per-zone dirty limit can
265 * require translating the configured limit into a percentage of
266 * global dirtyable memory first.
270 * zone_dirtyable_memory - number of dirtyable pages in a zone
273 * Returns the zone's number of pages potentially available for dirty
274 * page cache. This is the base value for the per-zone dirty limits.
276 static unsigned long zone_dirtyable_memory(struct zone *zone)
278 unsigned long nr_pages;
280 nr_pages = zone_page_state(zone, NR_FREE_PAGES);
281 nr_pages -= min(nr_pages, zone->dirty_balance_reserve);
283 nr_pages += zone_page_state(zone, NR_INACTIVE_FILE);
284 nr_pages += zone_page_state(zone, NR_ACTIVE_FILE);
289 static unsigned long highmem_dirtyable_memory(unsigned long total)
291 #ifdef CONFIG_HIGHMEM
295 for_each_node_state(node, N_HIGH_MEMORY) {
296 struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
298 x += zone_dirtyable_memory(z);
301 * Unreclaimable memory (kernel memory or anonymous memory
302 * without swap) can bring down the dirtyable pages below
303 * the zone's dirty balance reserve and the above calculation
304 * will underflow. However we still want to add in nodes
305 * which are below threshold (negative values) to get a more
306 * accurate calculation but make sure that the total never
313 * Make sure that the number of highmem pages is never larger
314 * than the number of the total dirtyable memory. This can only
315 * occur in very strange VM situations but we want to make sure
316 * that this does not occur.
318 return min(x, total);
325 * global_dirtyable_memory - number of globally dirtyable pages
327 * Returns the global number of pages potentially available for dirty
328 * page cache. This is the base value for the global dirty limits.
330 static unsigned long global_dirtyable_memory(void)
334 x = global_page_state(NR_FREE_PAGES);
335 x -= min(x, dirty_balance_reserve);
337 x += global_page_state(NR_INACTIVE_FILE);
338 x += global_page_state(NR_ACTIVE_FILE);
340 if (!vm_highmem_is_dirtyable)
341 x -= highmem_dirtyable_memory(x);
343 return x + 1; /* Ensure that we never return 0 */
347 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
348 * @dtc: dirty_throttle_control of interest
350 * Calculate @dtc->thresh and ->bg_thresh considering
351 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
352 * must ensure that @dtc->avail is set before calling this function. The
353 * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
356 static void domain_dirty_limits(struct dirty_throttle_control *dtc)
358 const unsigned long available_memory = dtc->avail;
359 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
360 unsigned long bytes = vm_dirty_bytes;
361 unsigned long bg_bytes = dirty_background_bytes;
362 unsigned long ratio = vm_dirty_ratio;
363 unsigned long bg_ratio = dirty_background_ratio;
364 unsigned long thresh;
365 unsigned long bg_thresh;
366 struct task_struct *tsk;
368 /* gdtc is !NULL iff @dtc is for memcg domain */
370 unsigned long global_avail = gdtc->avail;
373 * The byte settings can't be applied directly to memcg
374 * domains. Convert them to ratios by scaling against
375 * globally available memory.
378 ratio = min(DIV_ROUND_UP(bytes, PAGE_SIZE) * 100 /
379 global_avail, 100UL);
381 bg_ratio = min(DIV_ROUND_UP(bg_bytes, PAGE_SIZE) * 100 /
382 global_avail, 100UL);
383 bytes = bg_bytes = 0;
387 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
389 thresh = (ratio * available_memory) / 100;
392 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
394 bg_thresh = (bg_ratio * available_memory) / 100;
396 if (bg_thresh >= thresh)
397 bg_thresh = thresh / 2;
399 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
400 bg_thresh += bg_thresh / 4;
401 thresh += thresh / 4;
403 dtc->thresh = thresh;
404 dtc->bg_thresh = bg_thresh;
406 /* we should eventually report the domain in the TP */
408 trace_global_dirty_state(bg_thresh, thresh);
412 * global_dirty_limits - background-writeback and dirty-throttling thresholds
413 * @pbackground: out parameter for bg_thresh
414 * @pdirty: out parameter for thresh
416 * Calculate bg_thresh and thresh for global_wb_domain. See
417 * domain_dirty_limits() for details.
419 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
421 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
423 gdtc.avail = global_dirtyable_memory();
424 domain_dirty_limits(&gdtc);
426 *pbackground = gdtc.bg_thresh;
427 *pdirty = gdtc.thresh;
431 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
434 * Returns the maximum number of dirty pages allowed in a zone, based
435 * on the zone's dirtyable memory.
437 static unsigned long zone_dirty_limit(struct zone *zone)
439 unsigned long zone_memory = zone_dirtyable_memory(zone);
440 struct task_struct *tsk = current;
444 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
445 zone_memory / global_dirtyable_memory();
447 dirty = vm_dirty_ratio * zone_memory / 100;
449 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
456 * zone_dirty_ok - tells whether a zone is within its dirty limits
457 * @zone: the zone to check
459 * Returns %true when the dirty pages in @zone are within the zone's
460 * dirty limit, %false if the limit is exceeded.
462 bool zone_dirty_ok(struct zone *zone)
464 unsigned long limit = zone_dirty_limit(zone);
466 return zone_page_state(zone, NR_FILE_DIRTY) +
467 zone_page_state(zone, NR_UNSTABLE_NFS) +
468 zone_page_state(zone, NR_WRITEBACK) <= limit;
471 int dirty_background_ratio_handler(struct ctl_table *table, int write,
472 void __user *buffer, size_t *lenp,
477 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
478 if (ret == 0 && write)
479 dirty_background_bytes = 0;
483 int dirty_background_bytes_handler(struct ctl_table *table, int write,
484 void __user *buffer, size_t *lenp,
489 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
490 if (ret == 0 && write)
491 dirty_background_ratio = 0;
495 int dirty_ratio_handler(struct ctl_table *table, int write,
496 void __user *buffer, size_t *lenp,
499 int old_ratio = vm_dirty_ratio;
502 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
503 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
504 writeback_set_ratelimit();
510 int dirty_bytes_handler(struct ctl_table *table, int write,
511 void __user *buffer, size_t *lenp,
514 unsigned long old_bytes = vm_dirty_bytes;
517 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
518 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
519 writeback_set_ratelimit();
525 static unsigned long wp_next_time(unsigned long cur_time)
527 cur_time += VM_COMPLETIONS_PERIOD_LEN;
528 /* 0 has a special meaning... */
534 static void wb_domain_writeout_inc(struct wb_domain *dom,
535 struct fprop_local_percpu *completions,
536 unsigned int max_prop_frac)
538 __fprop_inc_percpu_max(&dom->completions, completions,
540 /* First event after period switching was turned off? */
541 if (!unlikely(dom->period_time)) {
543 * We can race with other __bdi_writeout_inc calls here but
544 * it does not cause any harm since the resulting time when
545 * timer will fire and what is in writeout_period_time will be
548 dom->period_time = wp_next_time(jiffies);
549 mod_timer(&dom->period_timer, dom->period_time);
554 * Increment @wb's writeout completion count and the global writeout
555 * completion count. Called from test_clear_page_writeback().
557 static inline void __wb_writeout_inc(struct bdi_writeback *wb)
559 struct wb_domain *cgdom;
561 __inc_wb_stat(wb, WB_WRITTEN);
562 wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
563 wb->bdi->max_prop_frac);
565 cgdom = mem_cgroup_wb_domain(wb);
567 wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
568 wb->bdi->max_prop_frac);
571 void wb_writeout_inc(struct bdi_writeback *wb)
575 local_irq_save(flags);
576 __wb_writeout_inc(wb);
577 local_irq_restore(flags);
579 EXPORT_SYMBOL_GPL(wb_writeout_inc);
582 * On idle system, we can be called long after we scheduled because we use
583 * deferred timers so count with missed periods.
585 static void writeout_period(unsigned long t)
587 struct wb_domain *dom = (void *)t;
588 int miss_periods = (jiffies - dom->period_time) /
589 VM_COMPLETIONS_PERIOD_LEN;
591 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
592 dom->period_time = wp_next_time(dom->period_time +
593 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
594 mod_timer(&dom->period_timer, dom->period_time);
597 * Aging has zeroed all fractions. Stop wasting CPU on period
600 dom->period_time = 0;
604 int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
606 memset(dom, 0, sizeof(*dom));
608 spin_lock_init(&dom->lock);
610 init_timer_deferrable(&dom->period_timer);
611 dom->period_timer.function = writeout_period;
612 dom->period_timer.data = (unsigned long)dom;
614 dom->dirty_limit_tstamp = jiffies;
616 return fprop_global_init(&dom->completions, gfp);
619 #ifdef CONFIG_CGROUP_WRITEBACK
620 void wb_domain_exit(struct wb_domain *dom)
622 del_timer_sync(&dom->period_timer);
623 fprop_global_destroy(&dom->completions);
628 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
629 * registered backing devices, which, for obvious reasons, can not
632 static unsigned int bdi_min_ratio;
634 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
638 spin_lock_bh(&bdi_lock);
639 if (min_ratio > bdi->max_ratio) {
642 min_ratio -= bdi->min_ratio;
643 if (bdi_min_ratio + min_ratio < 100) {
644 bdi_min_ratio += min_ratio;
645 bdi->min_ratio += min_ratio;
650 spin_unlock_bh(&bdi_lock);
655 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
662 spin_lock_bh(&bdi_lock);
663 if (bdi->min_ratio > max_ratio) {
666 bdi->max_ratio = max_ratio;
667 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
669 spin_unlock_bh(&bdi_lock);
673 EXPORT_SYMBOL(bdi_set_max_ratio);
675 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
676 unsigned long bg_thresh)
678 return (thresh + bg_thresh) / 2;
681 static unsigned long hard_dirty_limit(struct wb_domain *dom,
682 unsigned long thresh)
684 return max(thresh, dom->dirty_limit);
687 /* memory available to a memcg domain is capped by system-wide clean memory */
688 static void mdtc_cap_avail(struct dirty_throttle_control *mdtc)
690 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
691 unsigned long clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
693 mdtc->avail = min(mdtc->avail, clean);
697 * __wb_calc_thresh - @wb's share of dirty throttling threshold
698 * @dtc: dirty_throttle_context of interest
700 * Returns @wb's dirty limit in pages. The term "dirty" in the context of
701 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
703 * Note that balance_dirty_pages() will only seriously take it as a hard limit
704 * when sleeping max_pause per page is not enough to keep the dirty pages under
705 * control. For example, when the device is completely stalled due to some error
706 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
707 * In the other normal situations, it acts more gently by throttling the tasks
708 * more (rather than completely block them) when the wb dirty pages go high.
710 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
711 * - starving fast devices
712 * - piling up dirty pages (that will take long time to sync) on slow devices
714 * The wb's share of dirty limit will be adapting to its throughput and
715 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
717 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
719 struct wb_domain *dom = dtc_dom(dtc);
720 unsigned long thresh = dtc->thresh;
722 long numerator, denominator;
723 unsigned long wb_min_ratio, wb_max_ratio;
726 * Calculate this BDI's share of the thresh ratio.
728 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
729 &numerator, &denominator);
731 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
732 wb_thresh *= numerator;
733 do_div(wb_thresh, denominator);
735 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
737 wb_thresh += (thresh * wb_min_ratio) / 100;
738 if (wb_thresh > (thresh * wb_max_ratio) / 100)
739 wb_thresh = thresh * wb_max_ratio / 100;
744 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
746 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
748 return __wb_calc_thresh(&gdtc);
753 * f(dirty) := 1.0 + (----------------)
756 * it's a 3rd order polynomial that subjects to
758 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
759 * (2) f(setpoint) = 1.0 => the balance point
760 * (3) f(limit) = 0 => the hard limit
761 * (4) df/dx <= 0 => negative feedback control
762 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
763 * => fast response on large errors; small oscillation near setpoint
765 static long long pos_ratio_polynom(unsigned long setpoint,
772 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
773 (limit - setpoint) | 1);
775 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
776 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
777 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
779 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
783 * Dirty position control.
785 * (o) global/bdi setpoints
787 * We want the dirty pages be balanced around the global/wb setpoints.
788 * When the number of dirty pages is higher/lower than the setpoint, the
789 * dirty position control ratio (and hence task dirty ratelimit) will be
790 * decreased/increased to bring the dirty pages back to the setpoint.
792 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
794 * if (dirty < setpoint) scale up pos_ratio
795 * if (dirty > setpoint) scale down pos_ratio
797 * if (wb_dirty < wb_setpoint) scale up pos_ratio
798 * if (wb_dirty > wb_setpoint) scale down pos_ratio
800 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
802 * (o) global control line
806 * | |<===== global dirty control scope ======>|
814 * 1.0 ................................*
820 * 0 +------------.------------------.----------------------*------------->
821 * freerun^ setpoint^ limit^ dirty pages
823 * (o) wb control line
831 * | * |<=========== span ============>|
832 * 1.0 .......................*
844 * 1/4 ...............................................* * * * * * * * * * * *
848 * 0 +----------------------.-------------------------------.------------->
849 * wb_setpoint^ x_intercept^
851 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
852 * be smoothly throttled down to normal if it starts high in situations like
853 * - start writing to a slow SD card and a fast disk at the same time. The SD
854 * card's wb_dirty may rush to many times higher than wb_setpoint.
855 * - the wb dirty thresh drops quickly due to change of JBOD workload
857 static void wb_position_ratio(struct dirty_throttle_control *dtc)
859 struct bdi_writeback *wb = dtc->wb;
860 unsigned long write_bw = wb->avg_write_bandwidth;
861 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
862 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
863 unsigned long wb_thresh = dtc->wb_thresh;
864 unsigned long x_intercept;
865 unsigned long setpoint; /* dirty pages' target balance point */
866 unsigned long wb_setpoint;
868 long long pos_ratio; /* for scaling up/down the rate limit */
873 if (unlikely(dtc->dirty >= limit))
879 * See comment for pos_ratio_polynom().
881 setpoint = (freerun + limit) / 2;
882 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
885 * The strictlimit feature is a tool preventing mistrusted filesystems
886 * from growing a large number of dirty pages before throttling. For
887 * such filesystems balance_dirty_pages always checks wb counters
888 * against wb limits. Even if global "nr_dirty" is under "freerun".
889 * This is especially important for fuse which sets bdi->max_ratio to
890 * 1% by default. Without strictlimit feature, fuse writeback may
891 * consume arbitrary amount of RAM because it is accounted in
892 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
894 * Here, in wb_position_ratio(), we calculate pos_ratio based on
895 * two values: wb_dirty and wb_thresh. Let's consider an example:
896 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
897 * limits are set by default to 10% and 20% (background and throttle).
898 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
899 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
900 * about ~6K pages (as the average of background and throttle wb
901 * limits). The 3rd order polynomial will provide positive feedback if
902 * wb_dirty is under wb_setpoint and vice versa.
904 * Note, that we cannot use global counters in these calculations
905 * because we want to throttle process writing to a strictlimit wb
906 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
907 * in the example above).
909 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
910 long long wb_pos_ratio;
912 if (dtc->wb_dirty < 8) {
913 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
914 2 << RATELIMIT_CALC_SHIFT);
918 if (dtc->wb_dirty >= wb_thresh)
921 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
924 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
927 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
931 * Typically, for strictlimit case, wb_setpoint << setpoint
932 * and pos_ratio >> wb_pos_ratio. In the other words global
933 * state ("dirty") is not limiting factor and we have to
934 * make decision based on wb counters. But there is an
935 * important case when global pos_ratio should get precedence:
936 * global limits are exceeded (e.g. due to activities on other
937 * wb's) while given strictlimit wb is below limit.
939 * "pos_ratio * wb_pos_ratio" would work for the case above,
940 * but it would look too non-natural for the case of all
941 * activity in the system coming from a single strictlimit wb
942 * with bdi->max_ratio == 100%.
944 * Note that min() below somewhat changes the dynamics of the
945 * control system. Normally, pos_ratio value can be well over 3
946 * (when globally we are at freerun and wb is well below wb
947 * setpoint). Now the maximum pos_ratio in the same situation
948 * is 2. We might want to tweak this if we observe the control
949 * system is too slow to adapt.
951 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
956 * We have computed basic pos_ratio above based on global situation. If
957 * the wb is over/under its share of dirty pages, we want to scale
958 * pos_ratio further down/up. That is done by the following mechanism.
964 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
966 * x_intercept - wb_dirty
967 * := --------------------------
968 * x_intercept - wb_setpoint
970 * The main wb control line is a linear function that subjects to
972 * (1) f(wb_setpoint) = 1.0
973 * (2) k = - 1 / (8 * write_bw) (in single wb case)
974 * or equally: x_intercept = wb_setpoint + 8 * write_bw
976 * For single wb case, the dirty pages are observed to fluctuate
977 * regularly within range
978 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
979 * for various filesystems, where (2) can yield in a reasonable 12.5%
980 * fluctuation range for pos_ratio.
982 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
983 * own size, so move the slope over accordingly and choose a slope that
984 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
986 if (unlikely(wb_thresh > dtc->thresh))
987 wb_thresh = dtc->thresh;
989 * It's very possible that wb_thresh is close to 0 not because the
990 * device is slow, but that it has remained inactive for long time.
991 * Honour such devices a reasonable good (hopefully IO efficient)
992 * threshold, so that the occasional writes won't be blocked and active
993 * writes can rampup the threshold quickly.
995 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
997 * scale global setpoint to wb's:
998 * wb_setpoint = setpoint * wb_thresh / thresh
1000 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1001 wb_setpoint = setpoint * (u64)x >> 16;
1003 * Use span=(8*write_bw) in single wb case as indicated by
1004 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1006 * wb_thresh thresh - wb_thresh
1007 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1010 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1011 x_intercept = wb_setpoint + span;
1013 if (dtc->wb_dirty < x_intercept - span / 4) {
1014 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1015 (x_intercept - wb_setpoint) | 1);
1020 * wb reserve area, safeguard against dirty pool underrun and disk idle
1021 * It may push the desired control point of global dirty pages higher
1024 x_intercept = wb_thresh / 2;
1025 if (dtc->wb_dirty < x_intercept) {
1026 if (dtc->wb_dirty > x_intercept / 8)
1027 pos_ratio = div_u64(pos_ratio * x_intercept,
1033 dtc->pos_ratio = pos_ratio;
1036 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1037 unsigned long elapsed,
1038 unsigned long written)
1040 const unsigned long period = roundup_pow_of_two(3 * HZ);
1041 unsigned long avg = wb->avg_write_bandwidth;
1042 unsigned long old = wb->write_bandwidth;
1046 * bw = written * HZ / elapsed
1048 * bw * elapsed + write_bandwidth * (period - elapsed)
1049 * write_bandwidth = ---------------------------------------------------
1052 * @written may have decreased due to account_page_redirty().
1053 * Avoid underflowing @bw calculation.
1055 bw = written - min(written, wb->written_stamp);
1057 if (unlikely(elapsed > period)) {
1058 do_div(bw, elapsed);
1062 bw += (u64)wb->write_bandwidth * (period - elapsed);
1063 bw >>= ilog2(period);
1066 * one more level of smoothing, for filtering out sudden spikes
1068 if (avg > old && old >= (unsigned long)bw)
1069 avg -= (avg - old) >> 3;
1071 if (avg < old && old <= (unsigned long)bw)
1072 avg += (old - avg) >> 3;
1075 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1076 avg = max(avg, 1LU);
1077 if (wb_has_dirty_io(wb)) {
1078 long delta = avg - wb->avg_write_bandwidth;
1079 WARN_ON_ONCE(atomic_long_add_return(delta,
1080 &wb->bdi->tot_write_bandwidth) <= 0);
1082 wb->write_bandwidth = bw;
1083 wb->avg_write_bandwidth = avg;
1086 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1088 struct wb_domain *dom = dtc_dom(dtc);
1089 unsigned long thresh = dtc->thresh;
1090 unsigned long limit = dom->dirty_limit;
1093 * Follow up in one step.
1095 if (limit < thresh) {
1101 * Follow down slowly. Use the higher one as the target, because thresh
1102 * may drop below dirty. This is exactly the reason to introduce
1103 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1105 thresh = max(thresh, dtc->dirty);
1106 if (limit > thresh) {
1107 limit -= (limit - thresh) >> 5;
1112 dom->dirty_limit = limit;
1115 static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
1118 struct wb_domain *dom = dtc_dom(dtc);
1121 * check locklessly first to optimize away locking for the most time
1123 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1126 spin_lock(&dom->lock);
1127 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1128 update_dirty_limit(dtc);
1129 dom->dirty_limit_tstamp = now;
1131 spin_unlock(&dom->lock);
1135 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1137 * Normal wb tasks will be curbed at or below it in long term.
1138 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1140 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1141 unsigned long dirtied,
1142 unsigned long elapsed)
1144 struct bdi_writeback *wb = dtc->wb;
1145 unsigned long dirty = dtc->dirty;
1146 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1147 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1148 unsigned long setpoint = (freerun + limit) / 2;
1149 unsigned long write_bw = wb->avg_write_bandwidth;
1150 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1151 unsigned long dirty_rate;
1152 unsigned long task_ratelimit;
1153 unsigned long balanced_dirty_ratelimit;
1158 * The dirty rate will match the writeout rate in long term, except
1159 * when dirty pages are truncated by userspace or re-dirtied by FS.
1161 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1164 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1166 task_ratelimit = (u64)dirty_ratelimit *
1167 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1168 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1171 * A linear estimation of the "balanced" throttle rate. The theory is,
1172 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1173 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1174 * formula will yield the balanced rate limit (write_bw / N).
1176 * Note that the expanded form is not a pure rate feedback:
1177 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1178 * but also takes pos_ratio into account:
1179 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1181 * (1) is not realistic because pos_ratio also takes part in balancing
1182 * the dirty rate. Consider the state
1183 * pos_ratio = 0.5 (3)
1184 * rate = 2 * (write_bw / N) (4)
1185 * If (1) is used, it will stuck in that state! Because each dd will
1187 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1189 * dirty_rate = N * task_ratelimit = write_bw (6)
1190 * put (6) into (1) we get
1191 * rate_(i+1) = rate_(i) (7)
1193 * So we end up using (2) to always keep
1194 * rate_(i+1) ~= (write_bw / N) (8)
1195 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1196 * pos_ratio is able to drive itself to 1.0, which is not only where
1197 * the dirty count meet the setpoint, but also where the slope of
1198 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1200 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1203 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1205 if (unlikely(balanced_dirty_ratelimit > write_bw))
1206 balanced_dirty_ratelimit = write_bw;
1209 * We could safely do this and return immediately:
1211 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1213 * However to get a more stable dirty_ratelimit, the below elaborated
1214 * code makes use of task_ratelimit to filter out singular points and
1215 * limit the step size.
1217 * The below code essentially only uses the relative value of
1219 * task_ratelimit - dirty_ratelimit
1220 * = (pos_ratio - 1) * dirty_ratelimit
1222 * which reflects the direction and size of dirty position error.
1226 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1227 * task_ratelimit is on the same side of dirty_ratelimit, too.
1229 * - dirty_ratelimit > balanced_dirty_ratelimit
1230 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1231 * lowering dirty_ratelimit will help meet both the position and rate
1232 * control targets. Otherwise, don't update dirty_ratelimit if it will
1233 * only help meet the rate target. After all, what the users ultimately
1234 * feel and care are stable dirty rate and small position error.
1236 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1237 * and filter out the singular points of balanced_dirty_ratelimit. Which
1238 * keeps jumping around randomly and can even leap far away at times
1239 * due to the small 200ms estimation period of dirty_rate (we want to
1240 * keep that period small to reduce time lags).
1245 * For strictlimit case, calculations above were based on wb counters
1246 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1247 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1248 * Hence, to calculate "step" properly, we have to use wb_dirty as
1249 * "dirty" and wb_setpoint as "setpoint".
1251 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1252 * it's possible that wb_thresh is close to zero due to inactivity
1253 * of backing device.
1255 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1256 dirty = dtc->wb_dirty;
1257 if (dtc->wb_dirty < 8)
1258 setpoint = dtc->wb_dirty + 1;
1260 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1263 if (dirty < setpoint) {
1264 x = min3(wb->balanced_dirty_ratelimit,
1265 balanced_dirty_ratelimit, task_ratelimit);
1266 if (dirty_ratelimit < x)
1267 step = x - dirty_ratelimit;
1269 x = max3(wb->balanced_dirty_ratelimit,
1270 balanced_dirty_ratelimit, task_ratelimit);
1271 if (dirty_ratelimit > x)
1272 step = dirty_ratelimit - x;
1276 * Don't pursue 100% rate matching. It's impossible since the balanced
1277 * rate itself is constantly fluctuating. So decrease the track speed
1278 * when it gets close to the target. Helps eliminate pointless tremors.
1280 step >>= dirty_ratelimit / (2 * step + 1);
1282 * Limit the tracking speed to avoid overshooting.
1284 step = (step + 7) / 8;
1286 if (dirty_ratelimit < balanced_dirty_ratelimit)
1287 dirty_ratelimit += step;
1289 dirty_ratelimit -= step;
1291 wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1292 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1294 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1297 static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1298 struct dirty_throttle_control *mdtc,
1299 unsigned long start_time,
1300 bool update_ratelimit)
1302 struct bdi_writeback *wb = gdtc->wb;
1303 unsigned long now = jiffies;
1304 unsigned long elapsed = now - wb->bw_time_stamp;
1305 unsigned long dirtied;
1306 unsigned long written;
1308 lockdep_assert_held(&wb->list_lock);
1311 * rate-limit, only update once every 200ms.
1313 if (elapsed < BANDWIDTH_INTERVAL)
1316 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1317 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1320 * Skip quiet periods when disk bandwidth is under-utilized.
1321 * (at least 1s idle time between two flusher runs)
1323 if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1326 if (update_ratelimit) {
1327 domain_update_bandwidth(gdtc, now);
1328 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1331 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1332 * compiler has no way to figure that out. Help it.
1334 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1335 domain_update_bandwidth(mdtc, now);
1336 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1339 wb_update_write_bandwidth(wb, elapsed, written);
1342 wb->dirtied_stamp = dirtied;
1343 wb->written_stamp = written;
1344 wb->bw_time_stamp = now;
1347 void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1349 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1351 __wb_update_bandwidth(&gdtc, NULL, start_time, false);
1355 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1356 * will look to see if it needs to start dirty throttling.
1358 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1359 * global_page_state() too often. So scale it near-sqrt to the safety margin
1360 * (the number of pages we may dirty without exceeding the dirty limits).
1362 static unsigned long dirty_poll_interval(unsigned long dirty,
1363 unsigned long thresh)
1366 return 1UL << (ilog2(thresh - dirty) >> 1);
1371 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1372 unsigned long wb_dirty)
1374 unsigned long bw = wb->avg_write_bandwidth;
1378 * Limit pause time for small memory systems. If sleeping for too long
1379 * time, a small pool of dirty/writeback pages may go empty and disk go
1382 * 8 serves as the safety ratio.
1384 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1387 return min_t(unsigned long, t, MAX_PAUSE);
1390 static long wb_min_pause(struct bdi_writeback *wb,
1392 unsigned long task_ratelimit,
1393 unsigned long dirty_ratelimit,
1394 int *nr_dirtied_pause)
1396 long hi = ilog2(wb->avg_write_bandwidth);
1397 long lo = ilog2(wb->dirty_ratelimit);
1398 long t; /* target pause */
1399 long pause; /* estimated next pause */
1400 int pages; /* target nr_dirtied_pause */
1402 /* target for 10ms pause on 1-dd case */
1403 t = max(1, HZ / 100);
1406 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1409 * (N * 10ms) on 2^N concurrent tasks.
1412 t += (hi - lo) * (10 * HZ) / 1024;
1415 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1416 * on the much more stable dirty_ratelimit. However the next pause time
1417 * will be computed based on task_ratelimit and the two rate limits may
1418 * depart considerably at some time. Especially if task_ratelimit goes
1419 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1420 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1421 * result task_ratelimit won't be executed faithfully, which could
1422 * eventually bring down dirty_ratelimit.
1424 * We apply two rules to fix it up:
1425 * 1) try to estimate the next pause time and if necessary, use a lower
1426 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1427 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1428 * 2) limit the target pause time to max_pause/2, so that the normal
1429 * small fluctuations of task_ratelimit won't trigger rule (1) and
1430 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1432 t = min(t, 1 + max_pause / 2);
1433 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1436 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1437 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1438 * When the 16 consecutive reads are often interrupted by some dirty
1439 * throttling pause during the async writes, cfq will go into idles
1440 * (deadline is fine). So push nr_dirtied_pause as high as possible
1441 * until reaches DIRTY_POLL_THRESH=32 pages.
1443 if (pages < DIRTY_POLL_THRESH) {
1445 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1446 if (pages > DIRTY_POLL_THRESH) {
1447 pages = DIRTY_POLL_THRESH;
1448 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1452 pause = HZ * pages / (task_ratelimit + 1);
1453 if (pause > max_pause) {
1455 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1458 *nr_dirtied_pause = pages;
1460 * The minimal pause time will normally be half the target pause time.
1462 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1465 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1467 struct bdi_writeback *wb = dtc->wb;
1468 unsigned long wb_reclaimable;
1471 * wb_thresh is not treated as some limiting factor as
1472 * dirty_thresh, due to reasons
1473 * - in JBOD setup, wb_thresh can fluctuate a lot
1474 * - in a system with HDD and USB key, the USB key may somehow
1475 * go into state (wb_dirty >> wb_thresh) either because
1476 * wb_dirty starts high, or because wb_thresh drops low.
1477 * In this case we don't want to hard throttle the USB key
1478 * dirtiers for 100 seconds until wb_dirty drops under
1479 * wb_thresh. Instead the auxiliary wb control line in
1480 * wb_position_ratio() will let the dirtier task progress
1481 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1483 dtc->wb_thresh = __wb_calc_thresh(dtc);
1484 dtc->wb_bg_thresh = dtc->thresh ?
1485 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1488 * In order to avoid the stacked BDI deadlock we need
1489 * to ensure we accurately count the 'dirty' pages when
1490 * the threshold is low.
1492 * Otherwise it would be possible to get thresh+n pages
1493 * reported dirty, even though there are thresh-m pages
1494 * actually dirty; with m+n sitting in the percpu
1497 if (dtc->wb_thresh < 2 * wb_stat_error(wb)) {
1498 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1499 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1501 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1502 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1507 * balance_dirty_pages() must be called by processes which are generating dirty
1508 * data. It looks at the number of dirty pages in the machine and will force
1509 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1510 * If we're over `background_thresh' then the writeback threads are woken to
1511 * perform some writeout.
1513 static void balance_dirty_pages(struct address_space *mapping,
1514 struct bdi_writeback *wb,
1515 unsigned long pages_dirtied)
1517 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1518 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1519 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1520 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1522 struct dirty_throttle_control *sdtc;
1523 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1528 int nr_dirtied_pause;
1529 bool dirty_exceeded = false;
1530 unsigned long task_ratelimit;
1531 unsigned long dirty_ratelimit;
1532 struct backing_dev_info *bdi = wb->bdi;
1533 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1534 unsigned long start_time = jiffies;
1537 unsigned long now = jiffies;
1538 unsigned long dirty, thresh, bg_thresh;
1539 unsigned long m_dirty, m_thresh, m_bg_thresh;
1542 * Unstable writes are a feature of certain networked
1543 * filesystems (i.e. NFS) in which data may have been
1544 * written to the server's write cache, but has not yet
1545 * been flushed to permanent storage.
1547 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1548 global_page_state(NR_UNSTABLE_NFS);
1549 gdtc->avail = global_dirtyable_memory();
1550 gdtc->dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1552 domain_dirty_limits(gdtc);
1554 if (unlikely(strictlimit)) {
1555 wb_dirty_limits(gdtc);
1557 dirty = gdtc->wb_dirty;
1558 thresh = gdtc->wb_thresh;
1559 bg_thresh = gdtc->wb_bg_thresh;
1561 dirty = gdtc->dirty;
1562 thresh = gdtc->thresh;
1563 bg_thresh = gdtc->bg_thresh;
1567 unsigned long writeback;
1570 * If @wb belongs to !root memcg, repeat the same
1571 * basic calculations for the memcg domain.
1573 mem_cgroup_wb_stats(wb, &mdtc->avail, &mdtc->dirty,
1575 mdtc_cap_avail(mdtc);
1576 mdtc->dirty += writeback;
1578 domain_dirty_limits(mdtc);
1580 if (unlikely(strictlimit)) {
1581 wb_dirty_limits(mdtc);
1582 m_dirty = mdtc->wb_dirty;
1583 m_thresh = mdtc->wb_thresh;
1584 m_bg_thresh = mdtc->wb_bg_thresh;
1586 m_dirty = mdtc->dirty;
1587 m_thresh = mdtc->thresh;
1588 m_bg_thresh = mdtc->bg_thresh;
1593 * Throttle it only when the background writeback cannot
1594 * catch-up. This avoids (excessively) small writeouts
1595 * when the wb limits are ramping up in case of !strictlimit.
1597 * In strictlimit case make decision based on the wb counters
1598 * and limits. Small writeouts when the wb limits are ramping
1599 * up are the price we consciously pay for strictlimit-ing.
1601 * If memcg domain is in effect, @dirty should be under
1602 * both global and memcg freerun ceilings.
1604 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1606 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1607 unsigned long intv = dirty_poll_interval(dirty, thresh);
1608 unsigned long m_intv = ULONG_MAX;
1610 current->dirty_paused_when = now;
1611 current->nr_dirtied = 0;
1613 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1614 current->nr_dirtied_pause = min(intv, m_intv);
1618 if (unlikely(!writeback_in_progress(wb)))
1619 wb_start_background_writeback(wb);
1622 * Calculate global domain's pos_ratio and select the
1623 * global dtc by default.
1626 wb_dirty_limits(gdtc);
1628 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1629 ((gdtc->dirty > gdtc->thresh) || strictlimit);
1631 wb_position_ratio(gdtc);
1636 * If memcg domain is in effect, calculate its
1637 * pos_ratio. @wb should satisfy constraints from
1638 * both global and memcg domains. Choose the one
1639 * w/ lower pos_ratio.
1642 wb_dirty_limits(mdtc);
1644 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1645 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1647 wb_position_ratio(mdtc);
1648 if (mdtc->pos_ratio < gdtc->pos_ratio)
1652 if (dirty_exceeded && !wb->dirty_exceeded)
1653 wb->dirty_exceeded = 1;
1655 if (time_is_before_jiffies(wb->bw_time_stamp +
1656 BANDWIDTH_INTERVAL)) {
1657 spin_lock(&wb->list_lock);
1658 __wb_update_bandwidth(gdtc, mdtc, start_time, true);
1659 spin_unlock(&wb->list_lock);
1662 /* throttle according to the chosen dtc */
1663 dirty_ratelimit = wb->dirty_ratelimit;
1664 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1665 RATELIMIT_CALC_SHIFT;
1666 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1667 min_pause = wb_min_pause(wb, max_pause,
1668 task_ratelimit, dirty_ratelimit,
1671 if (unlikely(task_ratelimit == 0)) {
1676 period = HZ * pages_dirtied / task_ratelimit;
1678 if (current->dirty_paused_when)
1679 pause -= now - current->dirty_paused_when;
1681 * For less than 1s think time (ext3/4 may block the dirtier
1682 * for up to 800ms from time to time on 1-HDD; so does xfs,
1683 * however at much less frequency), try to compensate it in
1684 * future periods by updating the virtual time; otherwise just
1685 * do a reset, as it may be a light dirtier.
1687 if (pause < min_pause) {
1688 trace_balance_dirty_pages(wb,
1701 current->dirty_paused_when = now;
1702 current->nr_dirtied = 0;
1703 } else if (period) {
1704 current->dirty_paused_when += period;
1705 current->nr_dirtied = 0;
1706 } else if (current->nr_dirtied_pause <= pages_dirtied)
1707 current->nr_dirtied_pause += pages_dirtied;
1710 if (unlikely(pause > max_pause)) {
1711 /* for occasional dropped task_ratelimit */
1712 now += min(pause - max_pause, max_pause);
1717 trace_balance_dirty_pages(wb,
1729 __set_current_state(TASK_KILLABLE);
1730 io_schedule_timeout(pause);
1732 current->dirty_paused_when = now + pause;
1733 current->nr_dirtied = 0;
1734 current->nr_dirtied_pause = nr_dirtied_pause;
1737 * This is typically equal to (dirty < thresh) and can also
1738 * keep "1000+ dd on a slow USB stick" under control.
1744 * In the case of an unresponding NFS server and the NFS dirty
1745 * pages exceeds dirty_thresh, give the other good wb's a pipe
1746 * to go through, so that tasks on them still remain responsive.
1748 * In theory 1 page is enough to keep the comsumer-producer
1749 * pipe going: the flusher cleans 1 page => the task dirties 1
1750 * more page. However wb_dirty has accounting errors. So use
1751 * the larger and more IO friendly wb_stat_error.
1753 if (sdtc->wb_dirty <= wb_stat_error(wb))
1756 if (fatal_signal_pending(current))
1760 if (!dirty_exceeded && wb->dirty_exceeded)
1761 wb->dirty_exceeded = 0;
1763 if (writeback_in_progress(wb))
1767 * In laptop mode, we wait until hitting the higher threshold before
1768 * starting background writeout, and then write out all the way down
1769 * to the lower threshold. So slow writers cause minimal disk activity.
1771 * In normal mode, we start background writeout at the lower
1772 * background_thresh, to keep the amount of dirty memory low.
1777 if (nr_reclaimable > gdtc->bg_thresh)
1778 wb_start_background_writeback(wb);
1781 static DEFINE_PER_CPU(int, bdp_ratelimits);
1784 * Normal tasks are throttled by
1786 * dirty tsk->nr_dirtied_pause pages;
1787 * take a snap in balance_dirty_pages();
1789 * However there is a worst case. If every task exit immediately when dirtied
1790 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1791 * called to throttle the page dirties. The solution is to save the not yet
1792 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1793 * randomly into the running tasks. This works well for the above worst case,
1794 * as the new task will pick up and accumulate the old task's leaked dirty
1795 * count and eventually get throttled.
1797 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1800 * balance_dirty_pages_ratelimited - balance dirty memory state
1801 * @mapping: address_space which was dirtied
1803 * Processes which are dirtying memory should call in here once for each page
1804 * which was newly dirtied. The function will periodically check the system's
1805 * dirty state and will initiate writeback if needed.
1807 * On really big machines, get_writeback_state is expensive, so try to avoid
1808 * calling it too often (ratelimiting). But once we're over the dirty memory
1809 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1810 * from overshooting the limit by (ratelimit_pages) each.
1812 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1814 struct inode *inode = mapping->host;
1815 struct backing_dev_info *bdi = inode_to_bdi(inode);
1816 struct bdi_writeback *wb = NULL;
1820 if (!bdi_cap_account_dirty(bdi))
1823 if (inode_cgwb_enabled(inode))
1824 wb = wb_get_create_current(bdi, GFP_KERNEL);
1828 ratelimit = current->nr_dirtied_pause;
1829 if (wb->dirty_exceeded)
1830 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1834 * This prevents one CPU to accumulate too many dirtied pages without
1835 * calling into balance_dirty_pages(), which can happen when there are
1836 * 1000+ tasks, all of them start dirtying pages at exactly the same
1837 * time, hence all honoured too large initial task->nr_dirtied_pause.
1839 p = this_cpu_ptr(&bdp_ratelimits);
1840 if (unlikely(current->nr_dirtied >= ratelimit))
1842 else if (unlikely(*p >= ratelimit_pages)) {
1847 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1848 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1849 * the dirty throttling and livelock other long-run dirtiers.
1851 p = this_cpu_ptr(&dirty_throttle_leaks);
1852 if (*p > 0 && current->nr_dirtied < ratelimit) {
1853 unsigned long nr_pages_dirtied;
1854 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1855 *p -= nr_pages_dirtied;
1856 current->nr_dirtied += nr_pages_dirtied;
1860 if (unlikely(current->nr_dirtied >= ratelimit))
1861 balance_dirty_pages(mapping, wb, current->nr_dirtied);
1865 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1868 * wb_over_bg_thresh - does @wb need to be written back?
1869 * @wb: bdi_writeback of interest
1871 * Determines whether background writeback should keep writing @wb or it's
1872 * clean enough. Returns %true if writeback should continue.
1874 bool wb_over_bg_thresh(struct bdi_writeback *wb)
1876 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1877 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1878 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1879 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1883 * Similar to balance_dirty_pages() but ignores pages being written
1884 * as we're trying to decide whether to put more under writeback.
1886 gdtc->avail = global_dirtyable_memory();
1887 gdtc->dirty = global_page_state(NR_FILE_DIRTY) +
1888 global_page_state(NR_UNSTABLE_NFS);
1889 domain_dirty_limits(gdtc);
1891 if (gdtc->dirty > gdtc->bg_thresh)
1894 if (wb_stat(wb, WB_RECLAIMABLE) > __wb_calc_thresh(gdtc))
1898 unsigned long writeback;
1900 mem_cgroup_wb_stats(wb, &mdtc->avail, &mdtc->dirty, &writeback);
1901 mdtc_cap_avail(mdtc);
1902 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1904 if (mdtc->dirty > mdtc->bg_thresh)
1907 if (wb_stat(wb, WB_RECLAIMABLE) > __wb_calc_thresh(mdtc))
1914 void throttle_vm_writeout(gfp_t gfp_mask)
1916 unsigned long background_thresh;
1917 unsigned long dirty_thresh;
1920 global_dirty_limits(&background_thresh, &dirty_thresh);
1921 dirty_thresh = hard_dirty_limit(&global_wb_domain, dirty_thresh);
1924 * Boost the allowable dirty threshold a bit for page
1925 * allocators so they don't get DoS'ed by heavy writers
1927 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1929 if (global_page_state(NR_UNSTABLE_NFS) +
1930 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1932 congestion_wait(BLK_RW_ASYNC, HZ/10);
1935 * The caller might hold locks which can prevent IO completion
1936 * or progress in the filesystem. So we cannot just sit here
1937 * waiting for IO to complete.
1939 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1945 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1947 int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1948 void __user *buffer, size_t *length, loff_t *ppos)
1950 proc_dointvec(table, write, buffer, length, ppos);
1955 void laptop_mode_timer_fn(unsigned long data)
1957 struct request_queue *q = (struct request_queue *)data;
1958 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1959 global_page_state(NR_UNSTABLE_NFS);
1960 struct bdi_writeback *wb;
1963 * We want to write everything out, not just down to the dirty
1966 if (!bdi_has_dirty_io(&q->backing_dev_info))
1970 list_for_each_entry_rcu(wb, &q->backing_dev_info.wb_list, bdi_node)
1971 if (wb_has_dirty_io(wb))
1972 wb_start_writeback(wb, nr_pages, true,
1973 WB_REASON_LAPTOP_TIMER);
1978 * We've spun up the disk and we're in laptop mode: schedule writeback
1979 * of all dirty data a few seconds from now. If the flush is already scheduled
1980 * then push it back - the user is still using the disk.
1982 void laptop_io_completion(struct backing_dev_info *info)
1984 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1988 * We're in laptop mode and we've just synced. The sync's writes will have
1989 * caused another writeback to be scheduled by laptop_io_completion.
1990 * Nothing needs to be written back anymore, so we unschedule the writeback.
1992 void laptop_sync_completion(void)
1994 struct backing_dev_info *bdi;
1998 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1999 del_timer(&bdi->laptop_mode_wb_timer);
2006 * If ratelimit_pages is too high then we can get into dirty-data overload
2007 * if a large number of processes all perform writes at the same time.
2008 * If it is too low then SMP machines will call the (expensive)
2009 * get_writeback_state too often.
2011 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2012 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2016 void writeback_set_ratelimit(void)
2018 struct wb_domain *dom = &global_wb_domain;
2019 unsigned long background_thresh;
2020 unsigned long dirty_thresh;
2022 global_dirty_limits(&background_thresh, &dirty_thresh);
2023 dom->dirty_limit = dirty_thresh;
2024 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2025 if (ratelimit_pages < 16)
2026 ratelimit_pages = 16;
2030 ratelimit_handler(struct notifier_block *self, unsigned long action,
2034 switch (action & ~CPU_TASKS_FROZEN) {
2037 writeback_set_ratelimit();
2044 static struct notifier_block ratelimit_nb = {
2045 .notifier_call = ratelimit_handler,
2050 * Called early on to tune the page writeback dirty limits.
2052 * We used to scale dirty pages according to how total memory
2053 * related to pages that could be allocated for buffers (by
2054 * comparing nr_free_buffer_pages() to vm_total_pages.
2056 * However, that was when we used "dirty_ratio" to scale with
2057 * all memory, and we don't do that any more. "dirty_ratio"
2058 * is now applied to total non-HIGHPAGE memory (by subtracting
2059 * totalhigh_pages from vm_total_pages), and as such we can't
2060 * get into the old insane situation any more where we had
2061 * large amounts of dirty pages compared to a small amount of
2062 * non-HIGHMEM memory.
2064 * But we might still want to scale the dirty_ratio by how
2065 * much memory the box has..
2067 void __init page_writeback_init(void)
2069 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2071 writeback_set_ratelimit();
2072 register_cpu_notifier(&ratelimit_nb);
2076 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2077 * @mapping: address space structure to write
2078 * @start: starting page index
2079 * @end: ending page index (inclusive)
2081 * This function scans the page range from @start to @end (inclusive) and tags
2082 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2083 * that write_cache_pages (or whoever calls this function) will then use
2084 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2085 * used to avoid livelocking of writeback by a process steadily creating new
2086 * dirty pages in the file (thus it is important for this function to be quick
2087 * so that it can tag pages faster than a dirtying process can create them).
2090 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
2092 void tag_pages_for_writeback(struct address_space *mapping,
2093 pgoff_t start, pgoff_t end)
2095 #define WRITEBACK_TAG_BATCH 4096
2096 unsigned long tagged;
2099 spin_lock_irq(&mapping->tree_lock);
2100 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
2101 &start, end, WRITEBACK_TAG_BATCH,
2102 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
2103 spin_unlock_irq(&mapping->tree_lock);
2104 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
2106 /* We check 'start' to handle wrapping when end == ~0UL */
2107 } while (tagged >= WRITEBACK_TAG_BATCH && start);
2109 EXPORT_SYMBOL(tag_pages_for_writeback);
2112 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2113 * @mapping: address space structure to write
2114 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2115 * @writepage: function called for each page
2116 * @data: data passed to writepage function
2118 * If a page is already under I/O, write_cache_pages() skips it, even
2119 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2120 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2121 * and msync() need to guarantee that all the data which was dirty at the time
2122 * the call was made get new I/O started against them. If wbc->sync_mode is
2123 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2124 * existing IO to complete.
2126 * To avoid livelocks (when other process dirties new pages), we first tag
2127 * pages which should be written back with TOWRITE tag and only then start
2128 * writing them. For data-integrity sync we have to be careful so that we do
2129 * not miss some pages (e.g., because some other process has cleared TOWRITE
2130 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2131 * by the process clearing the DIRTY tag (and submitting the page for IO).
2133 int write_cache_pages(struct address_space *mapping,
2134 struct writeback_control *wbc, writepage_t writepage,
2139 struct pagevec pvec;
2141 pgoff_t uninitialized_var(writeback_index);
2143 pgoff_t end; /* Inclusive */
2146 int range_whole = 0;
2149 pagevec_init(&pvec, 0);
2150 if (wbc->range_cyclic) {
2151 writeback_index = mapping->writeback_index; /* prev offset */
2152 index = writeback_index;
2159 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2160 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2161 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2163 cycled = 1; /* ignore range_cyclic tests */
2165 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2166 tag = PAGECACHE_TAG_TOWRITE;
2168 tag = PAGECACHE_TAG_DIRTY;
2170 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2171 tag_pages_for_writeback(mapping, index, end);
2173 while (!done && (index <= end)) {
2176 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2177 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2181 for (i = 0; i < nr_pages; i++) {
2182 struct page *page = pvec.pages[i];
2185 * At this point, the page may be truncated or
2186 * invalidated (changing page->mapping to NULL), or
2187 * even swizzled back from swapper_space to tmpfs file
2188 * mapping. However, page->index will not change
2189 * because we have a reference on the page.
2191 if (page->index > end) {
2193 * can't be range_cyclic (1st pass) because
2194 * end == -1 in that case.
2200 done_index = page->index;
2205 * Page truncated or invalidated. We can freely skip it
2206 * then, even for data integrity operations: the page
2207 * has disappeared concurrently, so there could be no
2208 * real expectation of this data interity operation
2209 * even if there is now a new, dirty page at the same
2210 * pagecache address.
2212 if (unlikely(page->mapping != mapping)) {
2218 if (!PageDirty(page)) {
2219 /* someone wrote it for us */
2220 goto continue_unlock;
2223 if (PageWriteback(page)) {
2224 if (wbc->sync_mode != WB_SYNC_NONE)
2225 wait_on_page_writeback(page);
2227 goto continue_unlock;
2230 BUG_ON(PageWriteback(page));
2231 if (!clear_page_dirty_for_io(page))
2232 goto continue_unlock;
2234 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2235 ret = (*writepage)(page, wbc, data);
2236 if (unlikely(ret)) {
2237 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2242 * done_index is set past this page,
2243 * so media errors will not choke
2244 * background writeout for the entire
2245 * file. This has consequences for
2246 * range_cyclic semantics (ie. it may
2247 * not be suitable for data integrity
2250 done_index = page->index + 1;
2257 * We stop writing back only if we are not doing
2258 * integrity sync. In case of integrity sync we have to
2259 * keep going until we have written all the pages
2260 * we tagged for writeback prior to entering this loop.
2262 if (--wbc->nr_to_write <= 0 &&
2263 wbc->sync_mode == WB_SYNC_NONE) {
2268 pagevec_release(&pvec);
2271 if (!cycled && !done) {
2274 * We hit the last page and there is more work to be done: wrap
2275 * back to the start of the file
2279 end = writeback_index - 1;
2282 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2283 mapping->writeback_index = done_index;
2287 EXPORT_SYMBOL(write_cache_pages);
2290 * Function used by generic_writepages to call the real writepage
2291 * function and set the mapping flags on error
2293 static int __writepage(struct page *page, struct writeback_control *wbc,
2296 struct address_space *mapping = data;
2297 int ret = mapping->a_ops->writepage(page, wbc);
2298 mapping_set_error(mapping, ret);
2303 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2304 * @mapping: address space structure to write
2305 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2307 * This is a library function, which implements the writepages()
2308 * address_space_operation.
2310 int generic_writepages(struct address_space *mapping,
2311 struct writeback_control *wbc)
2313 struct blk_plug plug;
2316 /* deal with chardevs and other special file */
2317 if (!mapping->a_ops->writepage)
2320 blk_start_plug(&plug);
2321 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2322 blk_finish_plug(&plug);
2326 EXPORT_SYMBOL(generic_writepages);
2328 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2332 if (wbc->nr_to_write <= 0)
2334 if (mapping->a_ops->writepages)
2335 ret = mapping->a_ops->writepages(mapping, wbc);
2337 ret = generic_writepages(mapping, wbc);
2342 * write_one_page - write out a single page and optionally wait on I/O
2343 * @page: the page to write
2344 * @wait: if true, wait on writeout
2346 * The page must be locked by the caller and will be unlocked upon return.
2348 * write_one_page() returns a negative error code if I/O failed.
2350 int write_one_page(struct page *page, int wait)
2352 struct address_space *mapping = page->mapping;
2354 struct writeback_control wbc = {
2355 .sync_mode = WB_SYNC_ALL,
2359 BUG_ON(!PageLocked(page));
2362 wait_on_page_writeback(page);
2364 if (clear_page_dirty_for_io(page)) {
2365 page_cache_get(page);
2366 ret = mapping->a_ops->writepage(page, &wbc);
2367 if (ret == 0 && wait) {
2368 wait_on_page_writeback(page);
2369 if (PageError(page))
2372 page_cache_release(page);
2378 EXPORT_SYMBOL(write_one_page);
2381 * For address_spaces which do not use buffers nor write back.
2383 int __set_page_dirty_no_writeback(struct page *page)
2385 if (!PageDirty(page))
2386 return !TestSetPageDirty(page);
2391 * Helper function for set_page_dirty family.
2393 * Caller must hold mem_cgroup_begin_page_stat().
2395 * NOTE: This relies on being atomic wrt interrupts.
2397 void account_page_dirtied(struct page *page, struct address_space *mapping,
2398 struct mem_cgroup *memcg)
2400 struct inode *inode = mapping->host;
2402 trace_writeback_dirty_page(page, mapping);
2404 if (mapping_cap_account_dirty(mapping)) {
2405 struct bdi_writeback *wb;
2407 inode_attach_wb(inode, page);
2408 wb = inode_to_wb(inode);
2410 mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
2411 __inc_zone_page_state(page, NR_FILE_DIRTY);
2412 __inc_zone_page_state(page, NR_DIRTIED);
2413 __inc_wb_stat(wb, WB_RECLAIMABLE);
2414 __inc_wb_stat(wb, WB_DIRTIED);
2415 task_io_account_write(PAGE_CACHE_SIZE);
2416 current->nr_dirtied++;
2417 this_cpu_inc(bdp_ratelimits);
2420 EXPORT_SYMBOL(account_page_dirtied);
2423 * Helper function for deaccounting dirty page without writeback.
2425 * Caller must hold mem_cgroup_begin_page_stat().
2427 void account_page_cleaned(struct page *page, struct address_space *mapping,
2428 struct mem_cgroup *memcg, struct bdi_writeback *wb)
2430 if (mapping_cap_account_dirty(mapping)) {
2431 mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
2432 dec_zone_page_state(page, NR_FILE_DIRTY);
2433 dec_wb_stat(wb, WB_RECLAIMABLE);
2434 task_io_account_cancelled_write(PAGE_CACHE_SIZE);
2439 * For address_spaces which do not use buffers. Just tag the page as dirty in
2442 * This is also used when a single buffer is being dirtied: we want to set the
2443 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2444 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2446 * The caller must ensure this doesn't race with truncation. Most will simply
2447 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2448 * the pte lock held, which also locks out truncation.
2450 int __set_page_dirty_nobuffers(struct page *page)
2452 struct mem_cgroup *memcg;
2454 memcg = mem_cgroup_begin_page_stat(page);
2455 if (!TestSetPageDirty(page)) {
2456 struct address_space *mapping = page_mapping(page);
2457 unsigned long flags;
2460 mem_cgroup_end_page_stat(memcg);
2464 spin_lock_irqsave(&mapping->tree_lock, flags);
2465 BUG_ON(page_mapping(page) != mapping);
2466 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2467 account_page_dirtied(page, mapping, memcg);
2468 radix_tree_tag_set(&mapping->page_tree, page_index(page),
2469 PAGECACHE_TAG_DIRTY);
2470 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2471 mem_cgroup_end_page_stat(memcg);
2473 if (mapping->host) {
2474 /* !PageAnon && !swapper_space */
2475 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2479 mem_cgroup_end_page_stat(memcg);
2482 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2485 * Call this whenever redirtying a page, to de-account the dirty counters
2486 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2487 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2488 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2491 void account_page_redirty(struct page *page)
2493 struct address_space *mapping = page->mapping;
2495 if (mapping && mapping_cap_account_dirty(mapping)) {
2496 struct inode *inode = mapping->host;
2497 struct bdi_writeback *wb;
2500 wb = unlocked_inode_to_wb_begin(inode, &locked);
2501 current->nr_dirtied--;
2502 dec_zone_page_state(page, NR_DIRTIED);
2503 dec_wb_stat(wb, WB_DIRTIED);
2504 unlocked_inode_to_wb_end(inode, locked);
2507 EXPORT_SYMBOL(account_page_redirty);
2510 * When a writepage implementation decides that it doesn't want to write this
2511 * page for some reason, it should redirty the locked page via
2512 * redirty_page_for_writepage() and it should then unlock the page and return 0
2514 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2518 wbc->pages_skipped++;
2519 ret = __set_page_dirty_nobuffers(page);
2520 account_page_redirty(page);
2523 EXPORT_SYMBOL(redirty_page_for_writepage);
2528 * For pages with a mapping this should be done under the page lock
2529 * for the benefit of asynchronous memory errors who prefer a consistent
2530 * dirty state. This rule can be broken in some special cases,
2531 * but should be better not to.
2533 * If the mapping doesn't provide a set_page_dirty a_op, then
2534 * just fall through and assume that it wants buffer_heads.
2536 int set_page_dirty(struct page *page)
2538 struct address_space *mapping = page_mapping(page);
2540 if (likely(mapping)) {
2541 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2543 * readahead/lru_deactivate_page could remain
2544 * PG_readahead/PG_reclaim due to race with end_page_writeback
2545 * About readahead, if the page is written, the flags would be
2546 * reset. So no problem.
2547 * About lru_deactivate_page, if the page is redirty, the flag
2548 * will be reset. So no problem. but if the page is used by readahead
2549 * it will confuse readahead and make it restart the size rampup
2550 * process. But it's a trivial problem.
2552 if (PageReclaim(page))
2553 ClearPageReclaim(page);
2556 spd = __set_page_dirty_buffers;
2558 return (*spd)(page);
2560 if (!PageDirty(page)) {
2561 if (!TestSetPageDirty(page))
2566 EXPORT_SYMBOL(set_page_dirty);
2569 * set_page_dirty() is racy if the caller has no reference against
2570 * page->mapping->host, and if the page is unlocked. This is because another
2571 * CPU could truncate the page off the mapping and then free the mapping.
2573 * Usually, the page _is_ locked, or the caller is a user-space process which
2574 * holds a reference on the inode by having an open file.
2576 * In other cases, the page should be locked before running set_page_dirty().
2578 int set_page_dirty_lock(struct page *page)
2583 ret = set_page_dirty(page);
2587 EXPORT_SYMBOL(set_page_dirty_lock);
2590 * This cancels just the dirty bit on the kernel page itself, it does NOT
2591 * actually remove dirty bits on any mmap's that may be around. It also
2592 * leaves the page tagged dirty, so any sync activity will still find it on
2593 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2594 * look at the dirty bits in the VM.
2596 * Doing this should *normally* only ever be done when a page is truncated,
2597 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2598 * this when it notices that somebody has cleaned out all the buffers on a
2599 * page without actually doing it through the VM. Can you say "ext3 is
2600 * horribly ugly"? Thought you could.
2602 void cancel_dirty_page(struct page *page)
2604 struct address_space *mapping = page_mapping(page);
2606 if (mapping_cap_account_dirty(mapping)) {
2607 struct inode *inode = mapping->host;
2608 struct bdi_writeback *wb;
2609 struct mem_cgroup *memcg;
2612 memcg = mem_cgroup_begin_page_stat(page);
2613 wb = unlocked_inode_to_wb_begin(inode, &locked);
2615 if (TestClearPageDirty(page))
2616 account_page_cleaned(page, mapping, memcg, wb);
2618 unlocked_inode_to_wb_end(inode, locked);
2619 mem_cgroup_end_page_stat(memcg);
2621 ClearPageDirty(page);
2624 EXPORT_SYMBOL(cancel_dirty_page);
2627 * Clear a page's dirty flag, while caring for dirty memory accounting.
2628 * Returns true if the page was previously dirty.
2630 * This is for preparing to put the page under writeout. We leave the page
2631 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2632 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2633 * implementation will run either set_page_writeback() or set_page_dirty(),
2634 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2637 * This incoherency between the page's dirty flag and radix-tree tag is
2638 * unfortunate, but it only exists while the page is locked.
2640 int clear_page_dirty_for_io(struct page *page)
2642 struct address_space *mapping = page_mapping(page);
2645 BUG_ON(!PageLocked(page));
2647 if (mapping && mapping_cap_account_dirty(mapping)) {
2648 struct inode *inode = mapping->host;
2649 struct bdi_writeback *wb;
2650 struct mem_cgroup *memcg;
2654 * Yes, Virginia, this is indeed insane.
2656 * We use this sequence to make sure that
2657 * (a) we account for dirty stats properly
2658 * (b) we tell the low-level filesystem to
2659 * mark the whole page dirty if it was
2660 * dirty in a pagetable. Only to then
2661 * (c) clean the page again and return 1 to
2662 * cause the writeback.
2664 * This way we avoid all nasty races with the
2665 * dirty bit in multiple places and clearing
2666 * them concurrently from different threads.
2668 * Note! Normally the "set_page_dirty(page)"
2669 * has no effect on the actual dirty bit - since
2670 * that will already usually be set. But we
2671 * need the side effects, and it can help us
2674 * We basically use the page "master dirty bit"
2675 * as a serialization point for all the different
2676 * threads doing their things.
2678 if (page_mkclean(page))
2679 set_page_dirty(page);
2681 * We carefully synchronise fault handlers against
2682 * installing a dirty pte and marking the page dirty
2683 * at this point. We do this by having them hold the
2684 * page lock while dirtying the page, and pages are
2685 * always locked coming in here, so we get the desired
2688 memcg = mem_cgroup_begin_page_stat(page);
2689 wb = unlocked_inode_to_wb_begin(inode, &locked);
2690 if (TestClearPageDirty(page)) {
2691 mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
2692 dec_zone_page_state(page, NR_FILE_DIRTY);
2693 dec_wb_stat(wb, WB_RECLAIMABLE);
2696 unlocked_inode_to_wb_end(inode, locked);
2697 mem_cgroup_end_page_stat(memcg);
2700 return TestClearPageDirty(page);
2702 EXPORT_SYMBOL(clear_page_dirty_for_io);
2704 int test_clear_page_writeback(struct page *page)
2706 struct address_space *mapping = page_mapping(page);
2707 struct mem_cgroup *memcg;
2710 memcg = mem_cgroup_begin_page_stat(page);
2712 struct inode *inode = mapping->host;
2713 struct backing_dev_info *bdi = inode_to_bdi(inode);
2714 unsigned long flags;
2716 spin_lock_irqsave(&mapping->tree_lock, flags);
2717 ret = TestClearPageWriteback(page);
2719 radix_tree_tag_clear(&mapping->page_tree,
2721 PAGECACHE_TAG_WRITEBACK);
2722 if (bdi_cap_account_writeback(bdi)) {
2723 struct bdi_writeback *wb = inode_to_wb(inode);
2725 __dec_wb_stat(wb, WB_WRITEBACK);
2726 __wb_writeout_inc(wb);
2729 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2731 ret = TestClearPageWriteback(page);
2734 mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
2735 dec_zone_page_state(page, NR_WRITEBACK);
2736 inc_zone_page_state(page, NR_WRITTEN);
2738 mem_cgroup_end_page_stat(memcg);
2742 int __test_set_page_writeback(struct page *page, bool keep_write)
2744 struct address_space *mapping = page_mapping(page);
2745 struct mem_cgroup *memcg;
2748 memcg = mem_cgroup_begin_page_stat(page);
2750 struct inode *inode = mapping->host;
2751 struct backing_dev_info *bdi = inode_to_bdi(inode);
2752 unsigned long flags;
2754 spin_lock_irqsave(&mapping->tree_lock, flags);
2755 ret = TestSetPageWriteback(page);
2757 radix_tree_tag_set(&mapping->page_tree,
2759 PAGECACHE_TAG_WRITEBACK);
2760 if (bdi_cap_account_writeback(bdi))
2761 __inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2763 if (!PageDirty(page))
2764 radix_tree_tag_clear(&mapping->page_tree,
2766 PAGECACHE_TAG_DIRTY);
2768 radix_tree_tag_clear(&mapping->page_tree,
2770 PAGECACHE_TAG_TOWRITE);
2771 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2773 ret = TestSetPageWriteback(page);
2776 mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
2777 inc_zone_page_state(page, NR_WRITEBACK);
2779 mem_cgroup_end_page_stat(memcg);
2783 EXPORT_SYMBOL(__test_set_page_writeback);
2786 * Return true if any of the pages in the mapping are marked with the
2789 int mapping_tagged(struct address_space *mapping, int tag)
2791 return radix_tree_tagged(&mapping->page_tree, tag);
2793 EXPORT_SYMBOL(mapping_tagged);
2796 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2797 * @page: The page to wait on.
2799 * This function determines if the given page is related to a backing device
2800 * that requires page contents to be held stable during writeback. If so, then
2801 * it will wait for any pending writeback to complete.
2803 void wait_for_stable_page(struct page *page)
2805 if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2806 wait_on_page_writeback(page);
2808 EXPORT_SYMBOL_GPL(wait_for_stable_page);