4 * Copyright (C) 2002, Linus Torvalds.
5 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
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 * node_dirtyable_memory - number of dirtyable pages in a node
273 * Returns the node's number of pages potentially available for dirty
274 * page cache. This is the base value for the per-node dirty limits.
276 static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
278 unsigned long nr_pages = 0;
281 for (z = 0; z < MAX_NR_ZONES; z++) {
282 struct zone *zone = pgdat->node_zones + z;
284 if (!populated_zone(zone))
287 nr_pages += zone_page_state(zone, NR_FREE_PAGES);
291 * Pages reserved for the kernel should not be considered
292 * dirtyable, to prevent a situation where reclaim has to
293 * clean pages in order to balance the zones.
295 nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
297 nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
298 nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
303 static unsigned long highmem_dirtyable_memory(unsigned long total)
305 #ifdef CONFIG_HIGHMEM
310 for_each_node_state(node, N_HIGH_MEMORY) {
311 for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
313 unsigned long dirtyable;
315 if (!is_highmem_idx(i))
318 z = &NODE_DATA(node)->node_zones[i];
319 dirtyable = zone_page_state(z, NR_FREE_PAGES) +
320 zone_page_state(z, NR_ZONE_LRU_FILE);
322 /* watch for underflows */
323 dirtyable -= min(dirtyable, high_wmark_pages(z));
330 * Unreclaimable memory (kernel memory or anonymous memory
331 * without swap) can bring down the dirtyable pages below
332 * the zone's dirty balance reserve and the above calculation
333 * will underflow. However we still want to add in nodes
334 * which are below threshold (negative values) to get a more
335 * accurate calculation but make sure that the total never
342 * Make sure that the number of highmem pages is never larger
343 * than the number of the total dirtyable memory. This can only
344 * occur in very strange VM situations but we want to make sure
345 * that this does not occur.
347 return min(x, total);
354 * global_dirtyable_memory - number of globally dirtyable pages
356 * Returns the global number of pages potentially available for dirty
357 * page cache. This is the base value for the global dirty limits.
359 static unsigned long global_dirtyable_memory(void)
363 x = global_page_state(NR_FREE_PAGES);
365 * Pages reserved for the kernel should not be considered
366 * dirtyable, to prevent a situation where reclaim has to
367 * clean pages in order to balance the zones.
369 x -= min(x, totalreserve_pages);
371 x += global_node_page_state(NR_INACTIVE_FILE);
372 x += global_node_page_state(NR_ACTIVE_FILE);
374 if (!vm_highmem_is_dirtyable)
375 x -= highmem_dirtyable_memory(x);
377 return x + 1; /* Ensure that we never return 0 */
381 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
382 * @dtc: dirty_throttle_control of interest
384 * Calculate @dtc->thresh and ->bg_thresh considering
385 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
386 * must ensure that @dtc->avail is set before calling this function. The
387 * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
390 static void domain_dirty_limits(struct dirty_throttle_control *dtc)
392 const unsigned long available_memory = dtc->avail;
393 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
394 unsigned long bytes = vm_dirty_bytes;
395 unsigned long bg_bytes = dirty_background_bytes;
396 /* convert ratios to per-PAGE_SIZE for higher precision */
397 unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
398 unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
399 unsigned long thresh;
400 unsigned long bg_thresh;
401 struct task_struct *tsk;
403 /* gdtc is !NULL iff @dtc is for memcg domain */
405 unsigned long global_avail = gdtc->avail;
408 * The byte settings can't be applied directly to memcg
409 * domains. Convert them to ratios by scaling against
410 * globally available memory. As the ratios are in
411 * per-PAGE_SIZE, they can be obtained by dividing bytes by
415 ratio = min(DIV_ROUND_UP(bytes, global_avail),
418 bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
420 bytes = bg_bytes = 0;
424 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
426 thresh = (ratio * available_memory) / PAGE_SIZE;
429 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
431 bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
433 if (bg_thresh >= thresh)
434 bg_thresh = thresh / 2;
436 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
437 bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
438 thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
440 dtc->thresh = thresh;
441 dtc->bg_thresh = bg_thresh;
443 /* we should eventually report the domain in the TP */
445 trace_global_dirty_state(bg_thresh, thresh);
449 * global_dirty_limits - background-writeback and dirty-throttling thresholds
450 * @pbackground: out parameter for bg_thresh
451 * @pdirty: out parameter for thresh
453 * Calculate bg_thresh and thresh for global_wb_domain. See
454 * domain_dirty_limits() for details.
456 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
458 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
460 gdtc.avail = global_dirtyable_memory();
461 domain_dirty_limits(&gdtc);
463 *pbackground = gdtc.bg_thresh;
464 *pdirty = gdtc.thresh;
468 * node_dirty_limit - maximum number of dirty pages allowed in a node
471 * Returns the maximum number of dirty pages allowed in a node, based
472 * on the node's dirtyable memory.
474 static unsigned long node_dirty_limit(struct pglist_data *pgdat)
476 unsigned long node_memory = node_dirtyable_memory(pgdat);
477 struct task_struct *tsk = current;
481 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
482 node_memory / global_dirtyable_memory();
484 dirty = vm_dirty_ratio * node_memory / 100;
486 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
493 * node_dirty_ok - tells whether a node is within its dirty limits
494 * @pgdat: the node to check
496 * Returns %true when the dirty pages in @pgdat are within the node's
497 * dirty limit, %false if the limit is exceeded.
499 bool node_dirty_ok(struct pglist_data *pgdat)
502 unsigned long limit = node_dirty_limit(pgdat);
503 unsigned long nr_pages = 0;
505 for (z = 0; z < MAX_NR_ZONES; z++) {
506 struct zone *zone = pgdat->node_zones + z;
508 if (!populated_zone(zone))
511 nr_pages += zone_page_state(zone, NR_FILE_DIRTY);
512 nr_pages += zone_page_state(zone, NR_UNSTABLE_NFS);
513 nr_pages += zone_page_state(zone, NR_WRITEBACK);
516 return nr_pages <= limit;
519 int dirty_background_ratio_handler(struct ctl_table *table, int write,
520 void __user *buffer, size_t *lenp,
525 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
526 if (ret == 0 && write)
527 dirty_background_bytes = 0;
531 int dirty_background_bytes_handler(struct ctl_table *table, int write,
532 void __user *buffer, size_t *lenp,
537 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
538 if (ret == 0 && write)
539 dirty_background_ratio = 0;
543 int dirty_ratio_handler(struct ctl_table *table, int write,
544 void __user *buffer, size_t *lenp,
547 int old_ratio = vm_dirty_ratio;
550 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
551 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
552 writeback_set_ratelimit();
558 int dirty_bytes_handler(struct ctl_table *table, int write,
559 void __user *buffer, size_t *lenp,
562 unsigned long old_bytes = vm_dirty_bytes;
565 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
566 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
567 writeback_set_ratelimit();
573 static unsigned long wp_next_time(unsigned long cur_time)
575 cur_time += VM_COMPLETIONS_PERIOD_LEN;
576 /* 0 has a special meaning... */
582 static void wb_domain_writeout_inc(struct wb_domain *dom,
583 struct fprop_local_percpu *completions,
584 unsigned int max_prop_frac)
586 __fprop_inc_percpu_max(&dom->completions, completions,
588 /* First event after period switching was turned off? */
589 if (!unlikely(dom->period_time)) {
591 * We can race with other __bdi_writeout_inc calls here but
592 * it does not cause any harm since the resulting time when
593 * timer will fire and what is in writeout_period_time will be
596 dom->period_time = wp_next_time(jiffies);
597 mod_timer(&dom->period_timer, dom->period_time);
602 * Increment @wb's writeout completion count and the global writeout
603 * completion count. Called from test_clear_page_writeback().
605 static inline void __wb_writeout_inc(struct bdi_writeback *wb)
607 struct wb_domain *cgdom;
609 __inc_wb_stat(wb, WB_WRITTEN);
610 wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
611 wb->bdi->max_prop_frac);
613 cgdom = mem_cgroup_wb_domain(wb);
615 wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
616 wb->bdi->max_prop_frac);
619 void wb_writeout_inc(struct bdi_writeback *wb)
623 local_irq_save(flags);
624 __wb_writeout_inc(wb);
625 local_irq_restore(flags);
627 EXPORT_SYMBOL_GPL(wb_writeout_inc);
630 * On idle system, we can be called long after we scheduled because we use
631 * deferred timers so count with missed periods.
633 static void writeout_period(unsigned long t)
635 struct wb_domain *dom = (void *)t;
636 int miss_periods = (jiffies - dom->period_time) /
637 VM_COMPLETIONS_PERIOD_LEN;
639 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
640 dom->period_time = wp_next_time(dom->period_time +
641 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
642 mod_timer(&dom->period_timer, dom->period_time);
645 * Aging has zeroed all fractions. Stop wasting CPU on period
648 dom->period_time = 0;
652 int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
654 memset(dom, 0, sizeof(*dom));
656 spin_lock_init(&dom->lock);
658 init_timer_deferrable(&dom->period_timer);
659 dom->period_timer.function = writeout_period;
660 dom->period_timer.data = (unsigned long)dom;
662 dom->dirty_limit_tstamp = jiffies;
664 return fprop_global_init(&dom->completions, gfp);
667 #ifdef CONFIG_CGROUP_WRITEBACK
668 void wb_domain_exit(struct wb_domain *dom)
670 del_timer_sync(&dom->period_timer);
671 fprop_global_destroy(&dom->completions);
676 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
677 * registered backing devices, which, for obvious reasons, can not
680 static unsigned int bdi_min_ratio;
682 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
686 spin_lock_bh(&bdi_lock);
687 if (min_ratio > bdi->max_ratio) {
690 min_ratio -= bdi->min_ratio;
691 if (bdi_min_ratio + min_ratio < 100) {
692 bdi_min_ratio += min_ratio;
693 bdi->min_ratio += min_ratio;
698 spin_unlock_bh(&bdi_lock);
703 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
710 spin_lock_bh(&bdi_lock);
711 if (bdi->min_ratio > max_ratio) {
714 bdi->max_ratio = max_ratio;
715 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
717 spin_unlock_bh(&bdi_lock);
721 EXPORT_SYMBOL(bdi_set_max_ratio);
723 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
724 unsigned long bg_thresh)
726 return (thresh + bg_thresh) / 2;
729 static unsigned long hard_dirty_limit(struct wb_domain *dom,
730 unsigned long thresh)
732 return max(thresh, dom->dirty_limit);
736 * Memory which can be further allocated to a memcg domain is capped by
737 * system-wide clean memory excluding the amount being used in the domain.
739 static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
740 unsigned long filepages, unsigned long headroom)
742 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
743 unsigned long clean = filepages - min(filepages, mdtc->dirty);
744 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
745 unsigned long other_clean = global_clean - min(global_clean, clean);
747 mdtc->avail = filepages + min(headroom, other_clean);
751 * __wb_calc_thresh - @wb's share of dirty throttling threshold
752 * @dtc: dirty_throttle_context of interest
754 * Returns @wb's dirty limit in pages. The term "dirty" in the context of
755 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
757 * Note that balance_dirty_pages() will only seriously take it as a hard limit
758 * when sleeping max_pause per page is not enough to keep the dirty pages under
759 * control. For example, when the device is completely stalled due to some error
760 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
761 * In the other normal situations, it acts more gently by throttling the tasks
762 * more (rather than completely block them) when the wb dirty pages go high.
764 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
765 * - starving fast devices
766 * - piling up dirty pages (that will take long time to sync) on slow devices
768 * The wb's share of dirty limit will be adapting to its throughput and
769 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
771 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
773 struct wb_domain *dom = dtc_dom(dtc);
774 unsigned long thresh = dtc->thresh;
776 long numerator, denominator;
777 unsigned long wb_min_ratio, wb_max_ratio;
780 * Calculate this BDI's share of the thresh ratio.
782 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
783 &numerator, &denominator);
785 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
786 wb_thresh *= numerator;
787 do_div(wb_thresh, denominator);
789 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
791 wb_thresh += (thresh * wb_min_ratio) / 100;
792 if (wb_thresh > (thresh * wb_max_ratio) / 100)
793 wb_thresh = thresh * wb_max_ratio / 100;
798 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
800 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
802 return __wb_calc_thresh(&gdtc);
807 * f(dirty) := 1.0 + (----------------)
810 * it's a 3rd order polynomial that subjects to
812 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
813 * (2) f(setpoint) = 1.0 => the balance point
814 * (3) f(limit) = 0 => the hard limit
815 * (4) df/dx <= 0 => negative feedback control
816 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
817 * => fast response on large errors; small oscillation near setpoint
819 static long long pos_ratio_polynom(unsigned long setpoint,
826 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
827 (limit - setpoint) | 1);
829 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
830 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
831 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
833 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
837 * Dirty position control.
839 * (o) global/bdi setpoints
841 * We want the dirty pages be balanced around the global/wb setpoints.
842 * When the number of dirty pages is higher/lower than the setpoint, the
843 * dirty position control ratio (and hence task dirty ratelimit) will be
844 * decreased/increased to bring the dirty pages back to the setpoint.
846 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
848 * if (dirty < setpoint) scale up pos_ratio
849 * if (dirty > setpoint) scale down pos_ratio
851 * if (wb_dirty < wb_setpoint) scale up pos_ratio
852 * if (wb_dirty > wb_setpoint) scale down pos_ratio
854 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
856 * (o) global control line
860 * | |<===== global dirty control scope ======>|
868 * 1.0 ................................*
874 * 0 +------------.------------------.----------------------*------------->
875 * freerun^ setpoint^ limit^ dirty pages
877 * (o) wb control line
885 * | * |<=========== span ============>|
886 * 1.0 .......................*
898 * 1/4 ...............................................* * * * * * * * * * * *
902 * 0 +----------------------.-------------------------------.------------->
903 * wb_setpoint^ x_intercept^
905 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
906 * be smoothly throttled down to normal if it starts high in situations like
907 * - start writing to a slow SD card and a fast disk at the same time. The SD
908 * card's wb_dirty may rush to many times higher than wb_setpoint.
909 * - the wb dirty thresh drops quickly due to change of JBOD workload
911 static void wb_position_ratio(struct dirty_throttle_control *dtc)
913 struct bdi_writeback *wb = dtc->wb;
914 unsigned long write_bw = wb->avg_write_bandwidth;
915 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
916 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
917 unsigned long wb_thresh = dtc->wb_thresh;
918 unsigned long x_intercept;
919 unsigned long setpoint; /* dirty pages' target balance point */
920 unsigned long wb_setpoint;
922 long long pos_ratio; /* for scaling up/down the rate limit */
927 if (unlikely(dtc->dirty >= limit))
933 * See comment for pos_ratio_polynom().
935 setpoint = (freerun + limit) / 2;
936 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
939 * The strictlimit feature is a tool preventing mistrusted filesystems
940 * from growing a large number of dirty pages before throttling. For
941 * such filesystems balance_dirty_pages always checks wb counters
942 * against wb limits. Even if global "nr_dirty" is under "freerun".
943 * This is especially important for fuse which sets bdi->max_ratio to
944 * 1% by default. Without strictlimit feature, fuse writeback may
945 * consume arbitrary amount of RAM because it is accounted in
946 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
948 * Here, in wb_position_ratio(), we calculate pos_ratio based on
949 * two values: wb_dirty and wb_thresh. Let's consider an example:
950 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
951 * limits are set by default to 10% and 20% (background and throttle).
952 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
953 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
954 * about ~6K pages (as the average of background and throttle wb
955 * limits). The 3rd order polynomial will provide positive feedback if
956 * wb_dirty is under wb_setpoint and vice versa.
958 * Note, that we cannot use global counters in these calculations
959 * because we want to throttle process writing to a strictlimit wb
960 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
961 * in the example above).
963 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
964 long long wb_pos_ratio;
966 if (dtc->wb_dirty < 8) {
967 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
968 2 << RATELIMIT_CALC_SHIFT);
972 if (dtc->wb_dirty >= wb_thresh)
975 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
978 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
981 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
985 * Typically, for strictlimit case, wb_setpoint << setpoint
986 * and pos_ratio >> wb_pos_ratio. In the other words global
987 * state ("dirty") is not limiting factor and we have to
988 * make decision based on wb counters. But there is an
989 * important case when global pos_ratio should get precedence:
990 * global limits are exceeded (e.g. due to activities on other
991 * wb's) while given strictlimit wb is below limit.
993 * "pos_ratio * wb_pos_ratio" would work for the case above,
994 * but it would look too non-natural for the case of all
995 * activity in the system coming from a single strictlimit wb
996 * with bdi->max_ratio == 100%.
998 * Note that min() below somewhat changes the dynamics of the
999 * control system. Normally, pos_ratio value can be well over 3
1000 * (when globally we are at freerun and wb is well below wb
1001 * setpoint). Now the maximum pos_ratio in the same situation
1002 * is 2. We might want to tweak this if we observe the control
1003 * system is too slow to adapt.
1005 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
1010 * We have computed basic pos_ratio above based on global situation. If
1011 * the wb is over/under its share of dirty pages, we want to scale
1012 * pos_ratio further down/up. That is done by the following mechanism.
1018 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1020 * x_intercept - wb_dirty
1021 * := --------------------------
1022 * x_intercept - wb_setpoint
1024 * The main wb control line is a linear function that subjects to
1026 * (1) f(wb_setpoint) = 1.0
1027 * (2) k = - 1 / (8 * write_bw) (in single wb case)
1028 * or equally: x_intercept = wb_setpoint + 8 * write_bw
1030 * For single wb case, the dirty pages are observed to fluctuate
1031 * regularly within range
1032 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1033 * for various filesystems, where (2) can yield in a reasonable 12.5%
1034 * fluctuation range for pos_ratio.
1036 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1037 * own size, so move the slope over accordingly and choose a slope that
1038 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1040 if (unlikely(wb_thresh > dtc->thresh))
1041 wb_thresh = dtc->thresh;
1043 * It's very possible that wb_thresh is close to 0 not because the
1044 * device is slow, but that it has remained inactive for long time.
1045 * Honour such devices a reasonable good (hopefully IO efficient)
1046 * threshold, so that the occasional writes won't be blocked and active
1047 * writes can rampup the threshold quickly.
1049 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1051 * scale global setpoint to wb's:
1052 * wb_setpoint = setpoint * wb_thresh / thresh
1054 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1055 wb_setpoint = setpoint * (u64)x >> 16;
1057 * Use span=(8*write_bw) in single wb case as indicated by
1058 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1060 * wb_thresh thresh - wb_thresh
1061 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1064 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1065 x_intercept = wb_setpoint + span;
1067 if (dtc->wb_dirty < x_intercept - span / 4) {
1068 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1069 (x_intercept - wb_setpoint) | 1);
1074 * wb reserve area, safeguard against dirty pool underrun and disk idle
1075 * It may push the desired control point of global dirty pages higher
1078 x_intercept = wb_thresh / 2;
1079 if (dtc->wb_dirty < x_intercept) {
1080 if (dtc->wb_dirty > x_intercept / 8)
1081 pos_ratio = div_u64(pos_ratio * x_intercept,
1087 dtc->pos_ratio = pos_ratio;
1090 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1091 unsigned long elapsed,
1092 unsigned long written)
1094 const unsigned long period = roundup_pow_of_two(3 * HZ);
1095 unsigned long avg = wb->avg_write_bandwidth;
1096 unsigned long old = wb->write_bandwidth;
1100 * bw = written * HZ / elapsed
1102 * bw * elapsed + write_bandwidth * (period - elapsed)
1103 * write_bandwidth = ---------------------------------------------------
1106 * @written may have decreased due to account_page_redirty().
1107 * Avoid underflowing @bw calculation.
1109 bw = written - min(written, wb->written_stamp);
1111 if (unlikely(elapsed > period)) {
1112 do_div(bw, elapsed);
1116 bw += (u64)wb->write_bandwidth * (period - elapsed);
1117 bw >>= ilog2(period);
1120 * one more level of smoothing, for filtering out sudden spikes
1122 if (avg > old && old >= (unsigned long)bw)
1123 avg -= (avg - old) >> 3;
1125 if (avg < old && old <= (unsigned long)bw)
1126 avg += (old - avg) >> 3;
1129 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1130 avg = max(avg, 1LU);
1131 if (wb_has_dirty_io(wb)) {
1132 long delta = avg - wb->avg_write_bandwidth;
1133 WARN_ON_ONCE(atomic_long_add_return(delta,
1134 &wb->bdi->tot_write_bandwidth) <= 0);
1136 wb->write_bandwidth = bw;
1137 wb->avg_write_bandwidth = avg;
1140 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1142 struct wb_domain *dom = dtc_dom(dtc);
1143 unsigned long thresh = dtc->thresh;
1144 unsigned long limit = dom->dirty_limit;
1147 * Follow up in one step.
1149 if (limit < thresh) {
1155 * Follow down slowly. Use the higher one as the target, because thresh
1156 * may drop below dirty. This is exactly the reason to introduce
1157 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1159 thresh = max(thresh, dtc->dirty);
1160 if (limit > thresh) {
1161 limit -= (limit - thresh) >> 5;
1166 dom->dirty_limit = limit;
1169 static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
1172 struct wb_domain *dom = dtc_dom(dtc);
1175 * check locklessly first to optimize away locking for the most time
1177 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1180 spin_lock(&dom->lock);
1181 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1182 update_dirty_limit(dtc);
1183 dom->dirty_limit_tstamp = now;
1185 spin_unlock(&dom->lock);
1189 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1191 * Normal wb tasks will be curbed at or below it in long term.
1192 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1194 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1195 unsigned long dirtied,
1196 unsigned long elapsed)
1198 struct bdi_writeback *wb = dtc->wb;
1199 unsigned long dirty = dtc->dirty;
1200 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1201 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1202 unsigned long setpoint = (freerun + limit) / 2;
1203 unsigned long write_bw = wb->avg_write_bandwidth;
1204 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1205 unsigned long dirty_rate;
1206 unsigned long task_ratelimit;
1207 unsigned long balanced_dirty_ratelimit;
1210 unsigned long shift;
1213 * The dirty rate will match the writeout rate in long term, except
1214 * when dirty pages are truncated by userspace or re-dirtied by FS.
1216 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1219 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1221 task_ratelimit = (u64)dirty_ratelimit *
1222 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1223 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1226 * A linear estimation of the "balanced" throttle rate. The theory is,
1227 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1228 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1229 * formula will yield the balanced rate limit (write_bw / N).
1231 * Note that the expanded form is not a pure rate feedback:
1232 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1233 * but also takes pos_ratio into account:
1234 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1236 * (1) is not realistic because pos_ratio also takes part in balancing
1237 * the dirty rate. Consider the state
1238 * pos_ratio = 0.5 (3)
1239 * rate = 2 * (write_bw / N) (4)
1240 * If (1) is used, it will stuck in that state! Because each dd will
1242 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1244 * dirty_rate = N * task_ratelimit = write_bw (6)
1245 * put (6) into (1) we get
1246 * rate_(i+1) = rate_(i) (7)
1248 * So we end up using (2) to always keep
1249 * rate_(i+1) ~= (write_bw / N) (8)
1250 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1251 * pos_ratio is able to drive itself to 1.0, which is not only where
1252 * the dirty count meet the setpoint, but also where the slope of
1253 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1255 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1258 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1260 if (unlikely(balanced_dirty_ratelimit > write_bw))
1261 balanced_dirty_ratelimit = write_bw;
1264 * We could safely do this and return immediately:
1266 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1268 * However to get a more stable dirty_ratelimit, the below elaborated
1269 * code makes use of task_ratelimit to filter out singular points and
1270 * limit the step size.
1272 * The below code essentially only uses the relative value of
1274 * task_ratelimit - dirty_ratelimit
1275 * = (pos_ratio - 1) * dirty_ratelimit
1277 * which reflects the direction and size of dirty position error.
1281 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1282 * task_ratelimit is on the same side of dirty_ratelimit, too.
1284 * - dirty_ratelimit > balanced_dirty_ratelimit
1285 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1286 * lowering dirty_ratelimit will help meet both the position and rate
1287 * control targets. Otherwise, don't update dirty_ratelimit if it will
1288 * only help meet the rate target. After all, what the users ultimately
1289 * feel and care are stable dirty rate and small position error.
1291 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1292 * and filter out the singular points of balanced_dirty_ratelimit. Which
1293 * keeps jumping around randomly and can even leap far away at times
1294 * due to the small 200ms estimation period of dirty_rate (we want to
1295 * keep that period small to reduce time lags).
1300 * For strictlimit case, calculations above were based on wb counters
1301 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1302 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1303 * Hence, to calculate "step" properly, we have to use wb_dirty as
1304 * "dirty" and wb_setpoint as "setpoint".
1306 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1307 * it's possible that wb_thresh is close to zero due to inactivity
1308 * of backing device.
1310 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1311 dirty = dtc->wb_dirty;
1312 if (dtc->wb_dirty < 8)
1313 setpoint = dtc->wb_dirty + 1;
1315 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1318 if (dirty < setpoint) {
1319 x = min3(wb->balanced_dirty_ratelimit,
1320 balanced_dirty_ratelimit, task_ratelimit);
1321 if (dirty_ratelimit < x)
1322 step = x - dirty_ratelimit;
1324 x = max3(wb->balanced_dirty_ratelimit,
1325 balanced_dirty_ratelimit, task_ratelimit);
1326 if (dirty_ratelimit > x)
1327 step = dirty_ratelimit - x;
1331 * Don't pursue 100% rate matching. It's impossible since the balanced
1332 * rate itself is constantly fluctuating. So decrease the track speed
1333 * when it gets close to the target. Helps eliminate pointless tremors.
1335 shift = dirty_ratelimit / (2 * step + 1);
1336 if (shift < BITS_PER_LONG)
1337 step = DIV_ROUND_UP(step >> shift, 8);
1341 if (dirty_ratelimit < balanced_dirty_ratelimit)
1342 dirty_ratelimit += step;
1344 dirty_ratelimit -= step;
1346 wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1347 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1349 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1352 static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1353 struct dirty_throttle_control *mdtc,
1354 unsigned long start_time,
1355 bool update_ratelimit)
1357 struct bdi_writeback *wb = gdtc->wb;
1358 unsigned long now = jiffies;
1359 unsigned long elapsed = now - wb->bw_time_stamp;
1360 unsigned long dirtied;
1361 unsigned long written;
1363 lockdep_assert_held(&wb->list_lock);
1366 * rate-limit, only update once every 200ms.
1368 if (elapsed < BANDWIDTH_INTERVAL)
1371 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1372 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1375 * Skip quiet periods when disk bandwidth is under-utilized.
1376 * (at least 1s idle time between two flusher runs)
1378 if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1381 if (update_ratelimit) {
1382 domain_update_bandwidth(gdtc, now);
1383 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1386 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1387 * compiler has no way to figure that out. Help it.
1389 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1390 domain_update_bandwidth(mdtc, now);
1391 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1394 wb_update_write_bandwidth(wb, elapsed, written);
1397 wb->dirtied_stamp = dirtied;
1398 wb->written_stamp = written;
1399 wb->bw_time_stamp = now;
1402 void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1404 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1406 __wb_update_bandwidth(&gdtc, NULL, start_time, false);
1410 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1411 * will look to see if it needs to start dirty throttling.
1413 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1414 * global_page_state() too often. So scale it near-sqrt to the safety margin
1415 * (the number of pages we may dirty without exceeding the dirty limits).
1417 static unsigned long dirty_poll_interval(unsigned long dirty,
1418 unsigned long thresh)
1421 return 1UL << (ilog2(thresh - dirty) >> 1);
1426 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1427 unsigned long wb_dirty)
1429 unsigned long bw = wb->avg_write_bandwidth;
1433 * Limit pause time for small memory systems. If sleeping for too long
1434 * time, a small pool of dirty/writeback pages may go empty and disk go
1437 * 8 serves as the safety ratio.
1439 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1442 return min_t(unsigned long, t, MAX_PAUSE);
1445 static long wb_min_pause(struct bdi_writeback *wb,
1447 unsigned long task_ratelimit,
1448 unsigned long dirty_ratelimit,
1449 int *nr_dirtied_pause)
1451 long hi = ilog2(wb->avg_write_bandwidth);
1452 long lo = ilog2(wb->dirty_ratelimit);
1453 long t; /* target pause */
1454 long pause; /* estimated next pause */
1455 int pages; /* target nr_dirtied_pause */
1457 /* target for 10ms pause on 1-dd case */
1458 t = max(1, HZ / 100);
1461 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1464 * (N * 10ms) on 2^N concurrent tasks.
1467 t += (hi - lo) * (10 * HZ) / 1024;
1470 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1471 * on the much more stable dirty_ratelimit. However the next pause time
1472 * will be computed based on task_ratelimit and the two rate limits may
1473 * depart considerably at some time. Especially if task_ratelimit goes
1474 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1475 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1476 * result task_ratelimit won't be executed faithfully, which could
1477 * eventually bring down dirty_ratelimit.
1479 * We apply two rules to fix it up:
1480 * 1) try to estimate the next pause time and if necessary, use a lower
1481 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1482 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1483 * 2) limit the target pause time to max_pause/2, so that the normal
1484 * small fluctuations of task_ratelimit won't trigger rule (1) and
1485 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1487 t = min(t, 1 + max_pause / 2);
1488 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1491 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1492 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1493 * When the 16 consecutive reads are often interrupted by some dirty
1494 * throttling pause during the async writes, cfq will go into idles
1495 * (deadline is fine). So push nr_dirtied_pause as high as possible
1496 * until reaches DIRTY_POLL_THRESH=32 pages.
1498 if (pages < DIRTY_POLL_THRESH) {
1500 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1501 if (pages > DIRTY_POLL_THRESH) {
1502 pages = DIRTY_POLL_THRESH;
1503 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1507 pause = HZ * pages / (task_ratelimit + 1);
1508 if (pause > max_pause) {
1510 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1513 *nr_dirtied_pause = pages;
1515 * The minimal pause time will normally be half the target pause time.
1517 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1520 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1522 struct bdi_writeback *wb = dtc->wb;
1523 unsigned long wb_reclaimable;
1526 * wb_thresh is not treated as some limiting factor as
1527 * dirty_thresh, due to reasons
1528 * - in JBOD setup, wb_thresh can fluctuate a lot
1529 * - in a system with HDD and USB key, the USB key may somehow
1530 * go into state (wb_dirty >> wb_thresh) either because
1531 * wb_dirty starts high, or because wb_thresh drops low.
1532 * In this case we don't want to hard throttle the USB key
1533 * dirtiers for 100 seconds until wb_dirty drops under
1534 * wb_thresh. Instead the auxiliary wb control line in
1535 * wb_position_ratio() will let the dirtier task progress
1536 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1538 dtc->wb_thresh = __wb_calc_thresh(dtc);
1539 dtc->wb_bg_thresh = dtc->thresh ?
1540 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1543 * In order to avoid the stacked BDI deadlock we need
1544 * to ensure we accurately count the 'dirty' pages when
1545 * the threshold is low.
1547 * Otherwise it would be possible to get thresh+n pages
1548 * reported dirty, even though there are thresh-m pages
1549 * actually dirty; with m+n sitting in the percpu
1552 if (dtc->wb_thresh < 2 * wb_stat_error(wb)) {
1553 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1554 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1556 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1557 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1562 * balance_dirty_pages() must be called by processes which are generating dirty
1563 * data. It looks at the number of dirty pages in the machine and will force
1564 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1565 * If we're over `background_thresh' then the writeback threads are woken to
1566 * perform some writeout.
1568 static void balance_dirty_pages(struct address_space *mapping,
1569 struct bdi_writeback *wb,
1570 unsigned long pages_dirtied)
1572 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1573 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1574 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1575 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1577 struct dirty_throttle_control *sdtc;
1578 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1583 int nr_dirtied_pause;
1584 bool dirty_exceeded = false;
1585 unsigned long task_ratelimit;
1586 unsigned long dirty_ratelimit;
1587 struct backing_dev_info *bdi = wb->bdi;
1588 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1589 unsigned long start_time = jiffies;
1592 unsigned long now = jiffies;
1593 unsigned long dirty, thresh, bg_thresh;
1594 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1595 unsigned long m_thresh = 0;
1596 unsigned long m_bg_thresh = 0;
1599 * Unstable writes are a feature of certain networked
1600 * filesystems (i.e. NFS) in which data may have been
1601 * written to the server's write cache, but has not yet
1602 * been flushed to permanent storage.
1604 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1605 global_page_state(NR_UNSTABLE_NFS);
1606 gdtc->avail = global_dirtyable_memory();
1607 gdtc->dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1609 domain_dirty_limits(gdtc);
1611 if (unlikely(strictlimit)) {
1612 wb_dirty_limits(gdtc);
1614 dirty = gdtc->wb_dirty;
1615 thresh = gdtc->wb_thresh;
1616 bg_thresh = gdtc->wb_bg_thresh;
1618 dirty = gdtc->dirty;
1619 thresh = gdtc->thresh;
1620 bg_thresh = gdtc->bg_thresh;
1624 unsigned long filepages, headroom, writeback;
1627 * If @wb belongs to !root memcg, repeat the same
1628 * basic calculations for the memcg domain.
1630 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1631 &mdtc->dirty, &writeback);
1632 mdtc->dirty += writeback;
1633 mdtc_calc_avail(mdtc, filepages, headroom);
1635 domain_dirty_limits(mdtc);
1637 if (unlikely(strictlimit)) {
1638 wb_dirty_limits(mdtc);
1639 m_dirty = mdtc->wb_dirty;
1640 m_thresh = mdtc->wb_thresh;
1641 m_bg_thresh = mdtc->wb_bg_thresh;
1643 m_dirty = mdtc->dirty;
1644 m_thresh = mdtc->thresh;
1645 m_bg_thresh = mdtc->bg_thresh;
1650 * Throttle it only when the background writeback cannot
1651 * catch-up. This avoids (excessively) small writeouts
1652 * when the wb limits are ramping up in case of !strictlimit.
1654 * In strictlimit case make decision based on the wb counters
1655 * and limits. Small writeouts when the wb limits are ramping
1656 * up are the price we consciously pay for strictlimit-ing.
1658 * If memcg domain is in effect, @dirty should be under
1659 * both global and memcg freerun ceilings.
1661 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1663 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1664 unsigned long intv = dirty_poll_interval(dirty, thresh);
1665 unsigned long m_intv = ULONG_MAX;
1667 current->dirty_paused_when = now;
1668 current->nr_dirtied = 0;
1670 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1671 current->nr_dirtied_pause = min(intv, m_intv);
1675 if (unlikely(!writeback_in_progress(wb)))
1676 wb_start_background_writeback(wb);
1679 * Calculate global domain's pos_ratio and select the
1680 * global dtc by default.
1683 wb_dirty_limits(gdtc);
1685 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1686 ((gdtc->dirty > gdtc->thresh) || strictlimit);
1688 wb_position_ratio(gdtc);
1693 * If memcg domain is in effect, calculate its
1694 * pos_ratio. @wb should satisfy constraints from
1695 * both global and memcg domains. Choose the one
1696 * w/ lower pos_ratio.
1699 wb_dirty_limits(mdtc);
1701 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1702 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1704 wb_position_ratio(mdtc);
1705 if (mdtc->pos_ratio < gdtc->pos_ratio)
1709 if (dirty_exceeded && !wb->dirty_exceeded)
1710 wb->dirty_exceeded = 1;
1712 if (time_is_before_jiffies(wb->bw_time_stamp +
1713 BANDWIDTH_INTERVAL)) {
1714 spin_lock(&wb->list_lock);
1715 __wb_update_bandwidth(gdtc, mdtc, start_time, true);
1716 spin_unlock(&wb->list_lock);
1719 /* throttle according to the chosen dtc */
1720 dirty_ratelimit = wb->dirty_ratelimit;
1721 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1722 RATELIMIT_CALC_SHIFT;
1723 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1724 min_pause = wb_min_pause(wb, max_pause,
1725 task_ratelimit, dirty_ratelimit,
1728 if (unlikely(task_ratelimit == 0)) {
1733 period = HZ * pages_dirtied / task_ratelimit;
1735 if (current->dirty_paused_when)
1736 pause -= now - current->dirty_paused_when;
1738 * For less than 1s think time (ext3/4 may block the dirtier
1739 * for up to 800ms from time to time on 1-HDD; so does xfs,
1740 * however at much less frequency), try to compensate it in
1741 * future periods by updating the virtual time; otherwise just
1742 * do a reset, as it may be a light dirtier.
1744 if (pause < min_pause) {
1745 trace_balance_dirty_pages(wb,
1758 current->dirty_paused_when = now;
1759 current->nr_dirtied = 0;
1760 } else if (period) {
1761 current->dirty_paused_when += period;
1762 current->nr_dirtied = 0;
1763 } else if (current->nr_dirtied_pause <= pages_dirtied)
1764 current->nr_dirtied_pause += pages_dirtied;
1767 if (unlikely(pause > max_pause)) {
1768 /* for occasional dropped task_ratelimit */
1769 now += min(pause - max_pause, max_pause);
1774 trace_balance_dirty_pages(wb,
1786 __set_current_state(TASK_KILLABLE);
1787 io_schedule_timeout(pause);
1789 current->dirty_paused_when = now + pause;
1790 current->nr_dirtied = 0;
1791 current->nr_dirtied_pause = nr_dirtied_pause;
1794 * This is typically equal to (dirty < thresh) and can also
1795 * keep "1000+ dd on a slow USB stick" under control.
1801 * In the case of an unresponding NFS server and the NFS dirty
1802 * pages exceeds dirty_thresh, give the other good wb's a pipe
1803 * to go through, so that tasks on them still remain responsive.
1805 * In theory 1 page is enough to keep the comsumer-producer
1806 * pipe going: the flusher cleans 1 page => the task dirties 1
1807 * more page. However wb_dirty has accounting errors. So use
1808 * the larger and more IO friendly wb_stat_error.
1810 if (sdtc->wb_dirty <= wb_stat_error(wb))
1813 if (fatal_signal_pending(current))
1817 if (!dirty_exceeded && wb->dirty_exceeded)
1818 wb->dirty_exceeded = 0;
1820 if (writeback_in_progress(wb))
1824 * In laptop mode, we wait until hitting the higher threshold before
1825 * starting background writeout, and then write out all the way down
1826 * to the lower threshold. So slow writers cause minimal disk activity.
1828 * In normal mode, we start background writeout at the lower
1829 * background_thresh, to keep the amount of dirty memory low.
1834 if (nr_reclaimable > gdtc->bg_thresh)
1835 wb_start_background_writeback(wb);
1838 static DEFINE_PER_CPU(int, bdp_ratelimits);
1841 * Normal tasks are throttled by
1843 * dirty tsk->nr_dirtied_pause pages;
1844 * take a snap in balance_dirty_pages();
1846 * However there is a worst case. If every task exit immediately when dirtied
1847 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1848 * called to throttle the page dirties. The solution is to save the not yet
1849 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1850 * randomly into the running tasks. This works well for the above worst case,
1851 * as the new task will pick up and accumulate the old task's leaked dirty
1852 * count and eventually get throttled.
1854 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1857 * balance_dirty_pages_ratelimited - balance dirty memory state
1858 * @mapping: address_space which was dirtied
1860 * Processes which are dirtying memory should call in here once for each page
1861 * which was newly dirtied. The function will periodically check the system's
1862 * dirty state and will initiate writeback if needed.
1864 * On really big machines, get_writeback_state is expensive, so try to avoid
1865 * calling it too often (ratelimiting). But once we're over the dirty memory
1866 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1867 * from overshooting the limit by (ratelimit_pages) each.
1869 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1871 struct inode *inode = mapping->host;
1872 struct backing_dev_info *bdi = inode_to_bdi(inode);
1873 struct bdi_writeback *wb = NULL;
1877 if (!bdi_cap_account_dirty(bdi))
1880 if (inode_cgwb_enabled(inode))
1881 wb = wb_get_create_current(bdi, GFP_KERNEL);
1885 ratelimit = current->nr_dirtied_pause;
1886 if (wb->dirty_exceeded)
1887 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1891 * This prevents one CPU to accumulate too many dirtied pages without
1892 * calling into balance_dirty_pages(), which can happen when there are
1893 * 1000+ tasks, all of them start dirtying pages at exactly the same
1894 * time, hence all honoured too large initial task->nr_dirtied_pause.
1896 p = this_cpu_ptr(&bdp_ratelimits);
1897 if (unlikely(current->nr_dirtied >= ratelimit))
1899 else if (unlikely(*p >= ratelimit_pages)) {
1904 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1905 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1906 * the dirty throttling and livelock other long-run dirtiers.
1908 p = this_cpu_ptr(&dirty_throttle_leaks);
1909 if (*p > 0 && current->nr_dirtied < ratelimit) {
1910 unsigned long nr_pages_dirtied;
1911 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1912 *p -= nr_pages_dirtied;
1913 current->nr_dirtied += nr_pages_dirtied;
1917 if (unlikely(current->nr_dirtied >= ratelimit))
1918 balance_dirty_pages(mapping, wb, current->nr_dirtied);
1922 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1925 * wb_over_bg_thresh - does @wb need to be written back?
1926 * @wb: bdi_writeback of interest
1928 * Determines whether background writeback should keep writing @wb or it's
1929 * clean enough. Returns %true if writeback should continue.
1931 bool wb_over_bg_thresh(struct bdi_writeback *wb)
1933 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1934 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1935 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1936 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1940 * Similar to balance_dirty_pages() but ignores pages being written
1941 * as we're trying to decide whether to put more under writeback.
1943 gdtc->avail = global_dirtyable_memory();
1944 gdtc->dirty = global_page_state(NR_FILE_DIRTY) +
1945 global_page_state(NR_UNSTABLE_NFS);
1946 domain_dirty_limits(gdtc);
1948 if (gdtc->dirty > gdtc->bg_thresh)
1951 if (wb_stat(wb, WB_RECLAIMABLE) >
1952 wb_calc_thresh(gdtc->wb, gdtc->bg_thresh))
1956 unsigned long filepages, headroom, writeback;
1958 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1960 mdtc_calc_avail(mdtc, filepages, headroom);
1961 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1963 if (mdtc->dirty > mdtc->bg_thresh)
1966 if (wb_stat(wb, WB_RECLAIMABLE) >
1967 wb_calc_thresh(mdtc->wb, mdtc->bg_thresh))
1974 void throttle_vm_writeout(gfp_t gfp_mask)
1976 unsigned long background_thresh;
1977 unsigned long dirty_thresh;
1980 global_dirty_limits(&background_thresh, &dirty_thresh);
1981 dirty_thresh = hard_dirty_limit(&global_wb_domain, dirty_thresh);
1984 * Boost the allowable dirty threshold a bit for page
1985 * allocators so they don't get DoS'ed by heavy writers
1987 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1989 if (global_page_state(NR_UNSTABLE_NFS) +
1990 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1992 congestion_wait(BLK_RW_ASYNC, HZ/10);
1995 * The caller might hold locks which can prevent IO completion
1996 * or progress in the filesystem. So we cannot just sit here
1997 * waiting for IO to complete.
1999 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
2005 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
2007 int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
2008 void __user *buffer, size_t *length, loff_t *ppos)
2010 proc_dointvec(table, write, buffer, length, ppos);
2015 void laptop_mode_timer_fn(unsigned long data)
2017 struct request_queue *q = (struct request_queue *)data;
2018 int nr_pages = global_page_state(NR_FILE_DIRTY) +
2019 global_page_state(NR_UNSTABLE_NFS);
2020 struct bdi_writeback *wb;
2023 * We want to write everything out, not just down to the dirty
2026 if (!bdi_has_dirty_io(&q->backing_dev_info))
2030 list_for_each_entry_rcu(wb, &q->backing_dev_info.wb_list, bdi_node)
2031 if (wb_has_dirty_io(wb))
2032 wb_start_writeback(wb, nr_pages, true,
2033 WB_REASON_LAPTOP_TIMER);
2038 * We've spun up the disk and we're in laptop mode: schedule writeback
2039 * of all dirty data a few seconds from now. If the flush is already scheduled
2040 * then push it back - the user is still using the disk.
2042 void laptop_io_completion(struct backing_dev_info *info)
2044 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2048 * We're in laptop mode and we've just synced. The sync's writes will have
2049 * caused another writeback to be scheduled by laptop_io_completion.
2050 * Nothing needs to be written back anymore, so we unschedule the writeback.
2052 void laptop_sync_completion(void)
2054 struct backing_dev_info *bdi;
2058 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2059 del_timer(&bdi->laptop_mode_wb_timer);
2066 * If ratelimit_pages is too high then we can get into dirty-data overload
2067 * if a large number of processes all perform writes at the same time.
2068 * If it is too low then SMP machines will call the (expensive)
2069 * get_writeback_state too often.
2071 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2072 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2076 void writeback_set_ratelimit(void)
2078 struct wb_domain *dom = &global_wb_domain;
2079 unsigned long background_thresh;
2080 unsigned long dirty_thresh;
2082 global_dirty_limits(&background_thresh, &dirty_thresh);
2083 dom->dirty_limit = dirty_thresh;
2084 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2085 if (ratelimit_pages < 16)
2086 ratelimit_pages = 16;
2090 ratelimit_handler(struct notifier_block *self, unsigned long action,
2094 switch (action & ~CPU_TASKS_FROZEN) {
2097 writeback_set_ratelimit();
2104 static struct notifier_block ratelimit_nb = {
2105 .notifier_call = ratelimit_handler,
2110 * Called early on to tune the page writeback dirty limits.
2112 * We used to scale dirty pages according to how total memory
2113 * related to pages that could be allocated for buffers (by
2114 * comparing nr_free_buffer_pages() to vm_total_pages.
2116 * However, that was when we used "dirty_ratio" to scale with
2117 * all memory, and we don't do that any more. "dirty_ratio"
2118 * is now applied to total non-HIGHPAGE memory (by subtracting
2119 * totalhigh_pages from vm_total_pages), and as such we can't
2120 * get into the old insane situation any more where we had
2121 * large amounts of dirty pages compared to a small amount of
2122 * non-HIGHMEM memory.
2124 * But we might still want to scale the dirty_ratio by how
2125 * much memory the box has..
2127 void __init page_writeback_init(void)
2129 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2131 writeback_set_ratelimit();
2132 register_cpu_notifier(&ratelimit_nb);
2136 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2137 * @mapping: address space structure to write
2138 * @start: starting page index
2139 * @end: ending page index (inclusive)
2141 * This function scans the page range from @start to @end (inclusive) and tags
2142 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2143 * that write_cache_pages (or whoever calls this function) will then use
2144 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2145 * used to avoid livelocking of writeback by a process steadily creating new
2146 * dirty pages in the file (thus it is important for this function to be quick
2147 * so that it can tag pages faster than a dirtying process can create them).
2150 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
2152 void tag_pages_for_writeback(struct address_space *mapping,
2153 pgoff_t start, pgoff_t end)
2155 #define WRITEBACK_TAG_BATCH 4096
2156 unsigned long tagged;
2159 spin_lock_irq(&mapping->tree_lock);
2160 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
2161 &start, end, WRITEBACK_TAG_BATCH,
2162 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
2163 spin_unlock_irq(&mapping->tree_lock);
2164 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
2166 /* We check 'start' to handle wrapping when end == ~0UL */
2167 } while (tagged >= WRITEBACK_TAG_BATCH && start);
2169 EXPORT_SYMBOL(tag_pages_for_writeback);
2172 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2173 * @mapping: address space structure to write
2174 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2175 * @writepage: function called for each page
2176 * @data: data passed to writepage function
2178 * If a page is already under I/O, write_cache_pages() skips it, even
2179 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2180 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2181 * and msync() need to guarantee that all the data which was dirty at the time
2182 * the call was made get new I/O started against them. If wbc->sync_mode is
2183 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2184 * existing IO to complete.
2186 * To avoid livelocks (when other process dirties new pages), we first tag
2187 * pages which should be written back with TOWRITE tag and only then start
2188 * writing them. For data-integrity sync we have to be careful so that we do
2189 * not miss some pages (e.g., because some other process has cleared TOWRITE
2190 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2191 * by the process clearing the DIRTY tag (and submitting the page for IO).
2193 int write_cache_pages(struct address_space *mapping,
2194 struct writeback_control *wbc, writepage_t writepage,
2199 struct pagevec pvec;
2201 pgoff_t uninitialized_var(writeback_index);
2203 pgoff_t end; /* Inclusive */
2206 int range_whole = 0;
2209 pagevec_init(&pvec, 0);
2210 if (wbc->range_cyclic) {
2211 writeback_index = mapping->writeback_index; /* prev offset */
2212 index = writeback_index;
2219 index = wbc->range_start >> PAGE_SHIFT;
2220 end = wbc->range_end >> PAGE_SHIFT;
2221 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2223 cycled = 1; /* ignore range_cyclic tests */
2225 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2226 tag = PAGECACHE_TAG_TOWRITE;
2228 tag = PAGECACHE_TAG_DIRTY;
2230 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2231 tag_pages_for_writeback(mapping, index, end);
2233 while (!done && (index <= end)) {
2236 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2237 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2241 for (i = 0; i < nr_pages; i++) {
2242 struct page *page = pvec.pages[i];
2245 * At this point, the page may be truncated or
2246 * invalidated (changing page->mapping to NULL), or
2247 * even swizzled back from swapper_space to tmpfs file
2248 * mapping. However, page->index will not change
2249 * because we have a reference on the page.
2251 if (page->index > end) {
2253 * can't be range_cyclic (1st pass) because
2254 * end == -1 in that case.
2260 done_index = page->index;
2265 * Page truncated or invalidated. We can freely skip it
2266 * then, even for data integrity operations: the page
2267 * has disappeared concurrently, so there could be no
2268 * real expectation of this data interity operation
2269 * even if there is now a new, dirty page at the same
2270 * pagecache address.
2272 if (unlikely(page->mapping != mapping)) {
2278 if (!PageDirty(page)) {
2279 /* someone wrote it for us */
2280 goto continue_unlock;
2283 if (PageWriteback(page)) {
2284 if (wbc->sync_mode != WB_SYNC_NONE)
2285 wait_on_page_writeback(page);
2287 goto continue_unlock;
2290 BUG_ON(PageWriteback(page));
2291 if (!clear_page_dirty_for_io(page))
2292 goto continue_unlock;
2294 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2295 ret = (*writepage)(page, wbc, data);
2296 if (unlikely(ret)) {
2297 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2302 * done_index is set past this page,
2303 * so media errors will not choke
2304 * background writeout for the entire
2305 * file. This has consequences for
2306 * range_cyclic semantics (ie. it may
2307 * not be suitable for data integrity
2310 done_index = page->index + 1;
2317 * We stop writing back only if we are not doing
2318 * integrity sync. In case of integrity sync we have to
2319 * keep going until we have written all the pages
2320 * we tagged for writeback prior to entering this loop.
2322 if (--wbc->nr_to_write <= 0 &&
2323 wbc->sync_mode == WB_SYNC_NONE) {
2328 pagevec_release(&pvec);
2331 if (!cycled && !done) {
2334 * We hit the last page and there is more work to be done: wrap
2335 * back to the start of the file
2339 end = writeback_index - 1;
2342 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2343 mapping->writeback_index = done_index;
2347 EXPORT_SYMBOL(write_cache_pages);
2350 * Function used by generic_writepages to call the real writepage
2351 * function and set the mapping flags on error
2353 static int __writepage(struct page *page, struct writeback_control *wbc,
2356 struct address_space *mapping = data;
2357 int ret = mapping->a_ops->writepage(page, wbc);
2358 mapping_set_error(mapping, ret);
2363 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2364 * @mapping: address space structure to write
2365 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2367 * This is a library function, which implements the writepages()
2368 * address_space_operation.
2370 int generic_writepages(struct address_space *mapping,
2371 struct writeback_control *wbc)
2373 struct blk_plug plug;
2376 /* deal with chardevs and other special file */
2377 if (!mapping->a_ops->writepage)
2380 blk_start_plug(&plug);
2381 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2382 blk_finish_plug(&plug);
2386 EXPORT_SYMBOL(generic_writepages);
2388 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2392 if (wbc->nr_to_write <= 0)
2394 if (mapping->a_ops->writepages)
2395 ret = mapping->a_ops->writepages(mapping, wbc);
2397 ret = generic_writepages(mapping, wbc);
2402 * write_one_page - write out a single page and optionally wait on I/O
2403 * @page: the page to write
2404 * @wait: if true, wait on writeout
2406 * The page must be locked by the caller and will be unlocked upon return.
2408 * write_one_page() returns a negative error code if I/O failed.
2410 int write_one_page(struct page *page, int wait)
2412 struct address_space *mapping = page->mapping;
2414 struct writeback_control wbc = {
2415 .sync_mode = WB_SYNC_ALL,
2419 BUG_ON(!PageLocked(page));
2422 wait_on_page_writeback(page);
2424 if (clear_page_dirty_for_io(page)) {
2426 ret = mapping->a_ops->writepage(page, &wbc);
2427 if (ret == 0 && wait) {
2428 wait_on_page_writeback(page);
2429 if (PageError(page))
2438 EXPORT_SYMBOL(write_one_page);
2441 * For address_spaces which do not use buffers nor write back.
2443 int __set_page_dirty_no_writeback(struct page *page)
2445 if (!PageDirty(page))
2446 return !TestSetPageDirty(page);
2451 * Helper function for set_page_dirty family.
2453 * Caller must hold lock_page_memcg().
2455 * NOTE: This relies on being atomic wrt interrupts.
2457 void account_page_dirtied(struct page *page, struct address_space *mapping)
2459 struct inode *inode = mapping->host;
2461 trace_writeback_dirty_page(page, mapping);
2463 if (mapping_cap_account_dirty(mapping)) {
2464 struct bdi_writeback *wb;
2466 inode_attach_wb(inode, page);
2467 wb = inode_to_wb(inode);
2469 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_DIRTY);
2470 __inc_zone_page_state(page, NR_FILE_DIRTY);
2471 __inc_zone_page_state(page, NR_DIRTIED);
2472 __inc_wb_stat(wb, WB_RECLAIMABLE);
2473 __inc_wb_stat(wb, WB_DIRTIED);
2474 task_io_account_write(PAGE_SIZE);
2475 current->nr_dirtied++;
2476 this_cpu_inc(bdp_ratelimits);
2479 EXPORT_SYMBOL(account_page_dirtied);
2482 * Helper function for deaccounting dirty page without writeback.
2484 * Caller must hold lock_page_memcg().
2486 void account_page_cleaned(struct page *page, struct address_space *mapping,
2487 struct bdi_writeback *wb)
2489 if (mapping_cap_account_dirty(mapping)) {
2490 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_DIRTY);
2491 dec_zone_page_state(page, NR_FILE_DIRTY);
2492 dec_wb_stat(wb, WB_RECLAIMABLE);
2493 task_io_account_cancelled_write(PAGE_SIZE);
2498 * For address_spaces which do not use buffers. Just tag the page as dirty in
2501 * This is also used when a single buffer is being dirtied: we want to set the
2502 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2503 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2505 * The caller must ensure this doesn't race with truncation. Most will simply
2506 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2507 * the pte lock held, which also locks out truncation.
2509 int __set_page_dirty_nobuffers(struct page *page)
2511 lock_page_memcg(page);
2512 if (!TestSetPageDirty(page)) {
2513 struct address_space *mapping = page_mapping(page);
2514 unsigned long flags;
2517 unlock_page_memcg(page);
2521 spin_lock_irqsave(&mapping->tree_lock, flags);
2522 BUG_ON(page_mapping(page) != mapping);
2523 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2524 account_page_dirtied(page, mapping);
2525 radix_tree_tag_set(&mapping->page_tree, page_index(page),
2526 PAGECACHE_TAG_DIRTY);
2527 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2528 unlock_page_memcg(page);
2530 if (mapping->host) {
2531 /* !PageAnon && !swapper_space */
2532 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2536 unlock_page_memcg(page);
2539 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2542 * Call this whenever redirtying a page, to de-account the dirty counters
2543 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2544 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2545 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2548 void account_page_redirty(struct page *page)
2550 struct address_space *mapping = page->mapping;
2552 if (mapping && mapping_cap_account_dirty(mapping)) {
2553 struct inode *inode = mapping->host;
2554 struct bdi_writeback *wb;
2557 wb = unlocked_inode_to_wb_begin(inode, &locked);
2558 current->nr_dirtied--;
2559 dec_zone_page_state(page, NR_DIRTIED);
2560 dec_wb_stat(wb, WB_DIRTIED);
2561 unlocked_inode_to_wb_end(inode, locked);
2564 EXPORT_SYMBOL(account_page_redirty);
2567 * When a writepage implementation decides that it doesn't want to write this
2568 * page for some reason, it should redirty the locked page via
2569 * redirty_page_for_writepage() and it should then unlock the page and return 0
2571 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2575 wbc->pages_skipped++;
2576 ret = __set_page_dirty_nobuffers(page);
2577 account_page_redirty(page);
2580 EXPORT_SYMBOL(redirty_page_for_writepage);
2585 * For pages with a mapping this should be done under the page lock
2586 * for the benefit of asynchronous memory errors who prefer a consistent
2587 * dirty state. This rule can be broken in some special cases,
2588 * but should be better not to.
2590 * If the mapping doesn't provide a set_page_dirty a_op, then
2591 * just fall through and assume that it wants buffer_heads.
2593 int set_page_dirty(struct page *page)
2595 struct address_space *mapping = page_mapping(page);
2597 page = compound_head(page);
2598 if (likely(mapping)) {
2599 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2601 * readahead/lru_deactivate_page could remain
2602 * PG_readahead/PG_reclaim due to race with end_page_writeback
2603 * About readahead, if the page is written, the flags would be
2604 * reset. So no problem.
2605 * About lru_deactivate_page, if the page is redirty, the flag
2606 * will be reset. So no problem. but if the page is used by readahead
2607 * it will confuse readahead and make it restart the size rampup
2608 * process. But it's a trivial problem.
2610 if (PageReclaim(page))
2611 ClearPageReclaim(page);
2614 spd = __set_page_dirty_buffers;
2616 return (*spd)(page);
2618 if (!PageDirty(page)) {
2619 if (!TestSetPageDirty(page))
2624 EXPORT_SYMBOL(set_page_dirty);
2627 * set_page_dirty() is racy if the caller has no reference against
2628 * page->mapping->host, and if the page is unlocked. This is because another
2629 * CPU could truncate the page off the mapping and then free the mapping.
2631 * Usually, the page _is_ locked, or the caller is a user-space process which
2632 * holds a reference on the inode by having an open file.
2634 * In other cases, the page should be locked before running set_page_dirty().
2636 int set_page_dirty_lock(struct page *page)
2641 ret = set_page_dirty(page);
2645 EXPORT_SYMBOL(set_page_dirty_lock);
2648 * This cancels just the dirty bit on the kernel page itself, it does NOT
2649 * actually remove dirty bits on any mmap's that may be around. It also
2650 * leaves the page tagged dirty, so any sync activity will still find it on
2651 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2652 * look at the dirty bits in the VM.
2654 * Doing this should *normally* only ever be done when a page is truncated,
2655 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2656 * this when it notices that somebody has cleaned out all the buffers on a
2657 * page without actually doing it through the VM. Can you say "ext3 is
2658 * horribly ugly"? Thought you could.
2660 void cancel_dirty_page(struct page *page)
2662 struct address_space *mapping = page_mapping(page);
2664 if (mapping_cap_account_dirty(mapping)) {
2665 struct inode *inode = mapping->host;
2666 struct bdi_writeback *wb;
2669 lock_page_memcg(page);
2670 wb = unlocked_inode_to_wb_begin(inode, &locked);
2672 if (TestClearPageDirty(page))
2673 account_page_cleaned(page, mapping, wb);
2675 unlocked_inode_to_wb_end(inode, locked);
2676 unlock_page_memcg(page);
2678 ClearPageDirty(page);
2681 EXPORT_SYMBOL(cancel_dirty_page);
2684 * Clear a page's dirty flag, while caring for dirty memory accounting.
2685 * Returns true if the page was previously dirty.
2687 * This is for preparing to put the page under writeout. We leave the page
2688 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2689 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2690 * implementation will run either set_page_writeback() or set_page_dirty(),
2691 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2694 * This incoherency between the page's dirty flag and radix-tree tag is
2695 * unfortunate, but it only exists while the page is locked.
2697 int clear_page_dirty_for_io(struct page *page)
2699 struct address_space *mapping = page_mapping(page);
2702 BUG_ON(!PageLocked(page));
2704 if (mapping && mapping_cap_account_dirty(mapping)) {
2705 struct inode *inode = mapping->host;
2706 struct bdi_writeback *wb;
2710 * Yes, Virginia, this is indeed insane.
2712 * We use this sequence to make sure that
2713 * (a) we account for dirty stats properly
2714 * (b) we tell the low-level filesystem to
2715 * mark the whole page dirty if it was
2716 * dirty in a pagetable. Only to then
2717 * (c) clean the page again and return 1 to
2718 * cause the writeback.
2720 * This way we avoid all nasty races with the
2721 * dirty bit in multiple places and clearing
2722 * them concurrently from different threads.
2724 * Note! Normally the "set_page_dirty(page)"
2725 * has no effect on the actual dirty bit - since
2726 * that will already usually be set. But we
2727 * need the side effects, and it can help us
2730 * We basically use the page "master dirty bit"
2731 * as a serialization point for all the different
2732 * threads doing their things.
2734 if (page_mkclean(page))
2735 set_page_dirty(page);
2737 * We carefully synchronise fault handlers against
2738 * installing a dirty pte and marking the page dirty
2739 * at this point. We do this by having them hold the
2740 * page lock while dirtying the page, and pages are
2741 * always locked coming in here, so we get the desired
2744 wb = unlocked_inode_to_wb_begin(inode, &locked);
2745 if (TestClearPageDirty(page)) {
2746 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_DIRTY);
2747 dec_zone_page_state(page, NR_FILE_DIRTY);
2748 dec_wb_stat(wb, WB_RECLAIMABLE);
2751 unlocked_inode_to_wb_end(inode, locked);
2754 return TestClearPageDirty(page);
2756 EXPORT_SYMBOL(clear_page_dirty_for_io);
2758 int test_clear_page_writeback(struct page *page)
2760 struct address_space *mapping = page_mapping(page);
2763 lock_page_memcg(page);
2765 struct inode *inode = mapping->host;
2766 struct backing_dev_info *bdi = inode_to_bdi(inode);
2767 unsigned long flags;
2769 spin_lock_irqsave(&mapping->tree_lock, flags);
2770 ret = TestClearPageWriteback(page);
2772 radix_tree_tag_clear(&mapping->page_tree,
2774 PAGECACHE_TAG_WRITEBACK);
2775 if (bdi_cap_account_writeback(bdi)) {
2776 struct bdi_writeback *wb = inode_to_wb(inode);
2778 __dec_wb_stat(wb, WB_WRITEBACK);
2779 __wb_writeout_inc(wb);
2783 if (mapping->host && !mapping_tagged(mapping,
2784 PAGECACHE_TAG_WRITEBACK))
2785 sb_clear_inode_writeback(mapping->host);
2787 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2789 ret = TestClearPageWriteback(page);
2792 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
2793 dec_zone_page_state(page, NR_WRITEBACK);
2794 inc_zone_page_state(page, NR_WRITTEN);
2796 unlock_page_memcg(page);
2800 int __test_set_page_writeback(struct page *page, bool keep_write)
2802 struct address_space *mapping = page_mapping(page);
2805 lock_page_memcg(page);
2807 struct inode *inode = mapping->host;
2808 struct backing_dev_info *bdi = inode_to_bdi(inode);
2809 unsigned long flags;
2811 spin_lock_irqsave(&mapping->tree_lock, flags);
2812 ret = TestSetPageWriteback(page);
2816 on_wblist = mapping_tagged(mapping,
2817 PAGECACHE_TAG_WRITEBACK);
2819 radix_tree_tag_set(&mapping->page_tree,
2821 PAGECACHE_TAG_WRITEBACK);
2822 if (bdi_cap_account_writeback(bdi))
2823 __inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2826 * We can come through here when swapping anonymous
2827 * pages, so we don't necessarily have an inode to track
2830 if (mapping->host && !on_wblist)
2831 sb_mark_inode_writeback(mapping->host);
2833 if (!PageDirty(page))
2834 radix_tree_tag_clear(&mapping->page_tree,
2836 PAGECACHE_TAG_DIRTY);
2838 radix_tree_tag_clear(&mapping->page_tree,
2840 PAGECACHE_TAG_TOWRITE);
2841 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2843 ret = TestSetPageWriteback(page);
2846 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
2847 inc_zone_page_state(page, NR_WRITEBACK);
2849 unlock_page_memcg(page);
2853 EXPORT_SYMBOL(__test_set_page_writeback);
2856 * Return true if any of the pages in the mapping are marked with the
2859 int mapping_tagged(struct address_space *mapping, int tag)
2861 return radix_tree_tagged(&mapping->page_tree, tag);
2863 EXPORT_SYMBOL(mapping_tagged);
2866 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2867 * @page: The page to wait on.
2869 * This function determines if the given page is related to a backing device
2870 * that requires page contents to be held stable during writeback. If so, then
2871 * it will wait for any pending writeback to complete.
2873 void wait_for_stable_page(struct page *page)
2875 if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2876 wait_on_page_writeback(page);
2878 EXPORT_SYMBOL_GPL(wait_for_stable_page);