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);
302 #ifdef CONFIG_HIGHMEM
303 atomic_t highmem_file_pages;
306 static unsigned long highmem_dirtyable_memory(unsigned long total)
308 #ifdef CONFIG_HIGHMEM
312 unsigned long dirtyable = 0;
314 for_each_node_state(node, N_HIGH_MEMORY) {
315 for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
317 unsigned long nr_pages;
319 if (!is_highmem_idx(i))
322 z = &NODE_DATA(node)->node_zones[i];
323 if (!populated_zone(z))
326 nr_pages = zone_page_state(z, NR_FREE_PAGES);
327 /* watch for underflows */
328 nr_pages -= min(nr_pages, high_wmark_pages(z));
329 dirtyable += nr_pages;
333 x = dirtyable + atomic_read(&highmem_file_pages);
336 * Unreclaimable memory (kernel memory or anonymous memory
337 * without swap) can bring down the dirtyable pages below
338 * the zone's dirty balance reserve and the above calculation
339 * will underflow. However we still want to add in nodes
340 * which are below threshold (negative values) to get a more
341 * accurate calculation but make sure that the total never
348 * Make sure that the number of highmem pages is never larger
349 * than the number of the total dirtyable memory. This can only
350 * occur in very strange VM situations but we want to make sure
351 * that this does not occur.
353 return min(x, total);
360 * global_dirtyable_memory - number of globally dirtyable pages
362 * Returns the global number of pages potentially available for dirty
363 * page cache. This is the base value for the global dirty limits.
365 static unsigned long global_dirtyable_memory(void)
369 x = global_page_state(NR_FREE_PAGES);
371 * Pages reserved for the kernel should not be considered
372 * dirtyable, to prevent a situation where reclaim has to
373 * clean pages in order to balance the zones.
375 x -= min(x, totalreserve_pages);
377 x += global_node_page_state(NR_INACTIVE_FILE);
378 x += global_node_page_state(NR_ACTIVE_FILE);
380 if (!vm_highmem_is_dirtyable)
381 x -= highmem_dirtyable_memory(x);
383 return x + 1; /* Ensure that we never return 0 */
387 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
388 * @dtc: dirty_throttle_control of interest
390 * Calculate @dtc->thresh and ->bg_thresh considering
391 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
392 * must ensure that @dtc->avail is set before calling this function. The
393 * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
396 static void domain_dirty_limits(struct dirty_throttle_control *dtc)
398 const unsigned long available_memory = dtc->avail;
399 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
400 unsigned long bytes = vm_dirty_bytes;
401 unsigned long bg_bytes = dirty_background_bytes;
402 /* convert ratios to per-PAGE_SIZE for higher precision */
403 unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
404 unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
405 unsigned long thresh;
406 unsigned long bg_thresh;
407 struct task_struct *tsk;
409 /* gdtc is !NULL iff @dtc is for memcg domain */
411 unsigned long global_avail = gdtc->avail;
414 * The byte settings can't be applied directly to memcg
415 * domains. Convert them to ratios by scaling against
416 * globally available memory. As the ratios are in
417 * per-PAGE_SIZE, they can be obtained by dividing bytes by
421 ratio = min(DIV_ROUND_UP(bytes, global_avail),
424 bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
426 bytes = bg_bytes = 0;
430 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
432 thresh = (ratio * available_memory) / PAGE_SIZE;
435 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
437 bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
439 if (bg_thresh >= thresh)
440 bg_thresh = thresh / 2;
442 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
443 bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
444 thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
446 dtc->thresh = thresh;
447 dtc->bg_thresh = bg_thresh;
449 /* we should eventually report the domain in the TP */
451 trace_global_dirty_state(bg_thresh, thresh);
455 * global_dirty_limits - background-writeback and dirty-throttling thresholds
456 * @pbackground: out parameter for bg_thresh
457 * @pdirty: out parameter for thresh
459 * Calculate bg_thresh and thresh for global_wb_domain. See
460 * domain_dirty_limits() for details.
462 void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
464 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
466 gdtc.avail = global_dirtyable_memory();
467 domain_dirty_limits(&gdtc);
469 *pbackground = gdtc.bg_thresh;
470 *pdirty = gdtc.thresh;
474 * node_dirty_limit - maximum number of dirty pages allowed in a node
477 * Returns the maximum number of dirty pages allowed in a node, based
478 * on the node's dirtyable memory.
480 static unsigned long node_dirty_limit(struct pglist_data *pgdat)
482 unsigned long node_memory = node_dirtyable_memory(pgdat);
483 struct task_struct *tsk = current;
487 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
488 node_memory / global_dirtyable_memory();
490 dirty = vm_dirty_ratio * node_memory / 100;
492 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
499 * node_dirty_ok - tells whether a node is within its dirty limits
500 * @pgdat: the node to check
502 * Returns %true when the dirty pages in @pgdat are within the node's
503 * dirty limit, %false if the limit is exceeded.
505 bool node_dirty_ok(struct pglist_data *pgdat)
507 unsigned long limit = node_dirty_limit(pgdat);
508 unsigned long nr_pages = 0;
510 nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
511 nr_pages += node_page_state(pgdat, NR_UNSTABLE_NFS);
512 nr_pages += node_page_state(pgdat, NR_WRITEBACK);
514 return nr_pages <= limit;
517 int dirty_background_ratio_handler(struct ctl_table *table, int write,
518 void __user *buffer, size_t *lenp,
523 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
524 if (ret == 0 && write)
525 dirty_background_bytes = 0;
529 int dirty_background_bytes_handler(struct ctl_table *table, int write,
530 void __user *buffer, size_t *lenp,
535 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
536 if (ret == 0 && write)
537 dirty_background_ratio = 0;
541 int dirty_ratio_handler(struct ctl_table *table, int write,
542 void __user *buffer, size_t *lenp,
545 int old_ratio = vm_dirty_ratio;
548 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
549 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
550 writeback_set_ratelimit();
556 int dirty_bytes_handler(struct ctl_table *table, int write,
557 void __user *buffer, size_t *lenp,
560 unsigned long old_bytes = vm_dirty_bytes;
563 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
564 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
565 writeback_set_ratelimit();
571 static unsigned long wp_next_time(unsigned long cur_time)
573 cur_time += VM_COMPLETIONS_PERIOD_LEN;
574 /* 0 has a special meaning... */
580 static void wb_domain_writeout_inc(struct wb_domain *dom,
581 struct fprop_local_percpu *completions,
582 unsigned int max_prop_frac)
584 __fprop_inc_percpu_max(&dom->completions, completions,
586 /* First event after period switching was turned off? */
587 if (!unlikely(dom->period_time)) {
589 * We can race with other __bdi_writeout_inc calls here but
590 * it does not cause any harm since the resulting time when
591 * timer will fire and what is in writeout_period_time will be
594 dom->period_time = wp_next_time(jiffies);
595 mod_timer(&dom->period_timer, dom->period_time);
600 * Increment @wb's writeout completion count and the global writeout
601 * completion count. Called from test_clear_page_writeback().
603 static inline void __wb_writeout_inc(struct bdi_writeback *wb)
605 struct wb_domain *cgdom;
607 __inc_wb_stat(wb, WB_WRITTEN);
608 wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
609 wb->bdi->max_prop_frac);
611 cgdom = mem_cgroup_wb_domain(wb);
613 wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
614 wb->bdi->max_prop_frac);
617 void wb_writeout_inc(struct bdi_writeback *wb)
621 local_irq_save(flags);
622 __wb_writeout_inc(wb);
623 local_irq_restore(flags);
625 EXPORT_SYMBOL_GPL(wb_writeout_inc);
628 * On idle system, we can be called long after we scheduled because we use
629 * deferred timers so count with missed periods.
631 static void writeout_period(unsigned long t)
633 struct wb_domain *dom = (void *)t;
634 int miss_periods = (jiffies - dom->period_time) /
635 VM_COMPLETIONS_PERIOD_LEN;
637 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
638 dom->period_time = wp_next_time(dom->period_time +
639 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
640 mod_timer(&dom->period_timer, dom->period_time);
643 * Aging has zeroed all fractions. Stop wasting CPU on period
646 dom->period_time = 0;
650 int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
652 memset(dom, 0, sizeof(*dom));
654 spin_lock_init(&dom->lock);
656 init_timer_deferrable(&dom->period_timer);
657 dom->period_timer.function = writeout_period;
658 dom->period_timer.data = (unsigned long)dom;
660 dom->dirty_limit_tstamp = jiffies;
662 return fprop_global_init(&dom->completions, gfp);
665 #ifdef CONFIG_CGROUP_WRITEBACK
666 void wb_domain_exit(struct wb_domain *dom)
668 del_timer_sync(&dom->period_timer);
669 fprop_global_destroy(&dom->completions);
674 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
675 * registered backing devices, which, for obvious reasons, can not
678 static unsigned int bdi_min_ratio;
680 int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
684 spin_lock_bh(&bdi_lock);
685 if (min_ratio > bdi->max_ratio) {
688 min_ratio -= bdi->min_ratio;
689 if (bdi_min_ratio + min_ratio < 100) {
690 bdi_min_ratio += min_ratio;
691 bdi->min_ratio += min_ratio;
696 spin_unlock_bh(&bdi_lock);
701 int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
708 spin_lock_bh(&bdi_lock);
709 if (bdi->min_ratio > max_ratio) {
712 bdi->max_ratio = max_ratio;
713 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
715 spin_unlock_bh(&bdi_lock);
719 EXPORT_SYMBOL(bdi_set_max_ratio);
721 static unsigned long dirty_freerun_ceiling(unsigned long thresh,
722 unsigned long bg_thresh)
724 return (thresh + bg_thresh) / 2;
727 static unsigned long hard_dirty_limit(struct wb_domain *dom,
728 unsigned long thresh)
730 return max(thresh, dom->dirty_limit);
734 * Memory which can be further allocated to a memcg domain is capped by
735 * system-wide clean memory excluding the amount being used in the domain.
737 static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
738 unsigned long filepages, unsigned long headroom)
740 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
741 unsigned long clean = filepages - min(filepages, mdtc->dirty);
742 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
743 unsigned long other_clean = global_clean - min(global_clean, clean);
745 mdtc->avail = filepages + min(headroom, other_clean);
749 * __wb_calc_thresh - @wb's share of dirty throttling threshold
750 * @dtc: dirty_throttle_context of interest
752 * Returns @wb's dirty limit in pages. The term "dirty" in the context of
753 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
755 * Note that balance_dirty_pages() will only seriously take it as a hard limit
756 * when sleeping max_pause per page is not enough to keep the dirty pages under
757 * control. For example, when the device is completely stalled due to some error
758 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
759 * In the other normal situations, it acts more gently by throttling the tasks
760 * more (rather than completely block them) when the wb dirty pages go high.
762 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
763 * - starving fast devices
764 * - piling up dirty pages (that will take long time to sync) on slow devices
766 * The wb's share of dirty limit will be adapting to its throughput and
767 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
769 static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
771 struct wb_domain *dom = dtc_dom(dtc);
772 unsigned long thresh = dtc->thresh;
774 long numerator, denominator;
775 unsigned long wb_min_ratio, wb_max_ratio;
778 * Calculate this BDI's share of the thresh ratio.
780 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
781 &numerator, &denominator);
783 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
784 wb_thresh *= numerator;
785 do_div(wb_thresh, denominator);
787 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
789 wb_thresh += (thresh * wb_min_ratio) / 100;
790 if (wb_thresh > (thresh * wb_max_ratio) / 100)
791 wb_thresh = thresh * wb_max_ratio / 100;
796 unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
798 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
800 return __wb_calc_thresh(&gdtc);
805 * f(dirty) := 1.0 + (----------------)
808 * it's a 3rd order polynomial that subjects to
810 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
811 * (2) f(setpoint) = 1.0 => the balance point
812 * (3) f(limit) = 0 => the hard limit
813 * (4) df/dx <= 0 => negative feedback control
814 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
815 * => fast response on large errors; small oscillation near setpoint
817 static long long pos_ratio_polynom(unsigned long setpoint,
824 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
825 (limit - setpoint) | 1);
827 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
828 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
829 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
831 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
835 * Dirty position control.
837 * (o) global/bdi setpoints
839 * We want the dirty pages be balanced around the global/wb setpoints.
840 * When the number of dirty pages is higher/lower than the setpoint, the
841 * dirty position control ratio (and hence task dirty ratelimit) will be
842 * decreased/increased to bring the dirty pages back to the setpoint.
844 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
846 * if (dirty < setpoint) scale up pos_ratio
847 * if (dirty > setpoint) scale down pos_ratio
849 * if (wb_dirty < wb_setpoint) scale up pos_ratio
850 * if (wb_dirty > wb_setpoint) scale down pos_ratio
852 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
854 * (o) global control line
858 * | |<===== global dirty control scope ======>|
866 * 1.0 ................................*
872 * 0 +------------.------------------.----------------------*------------->
873 * freerun^ setpoint^ limit^ dirty pages
875 * (o) wb control line
883 * | * |<=========== span ============>|
884 * 1.0 .......................*
896 * 1/4 ...............................................* * * * * * * * * * * *
900 * 0 +----------------------.-------------------------------.------------->
901 * wb_setpoint^ x_intercept^
903 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
904 * be smoothly throttled down to normal if it starts high in situations like
905 * - start writing to a slow SD card and a fast disk at the same time. The SD
906 * card's wb_dirty may rush to many times higher than wb_setpoint.
907 * - the wb dirty thresh drops quickly due to change of JBOD workload
909 static void wb_position_ratio(struct dirty_throttle_control *dtc)
911 struct bdi_writeback *wb = dtc->wb;
912 unsigned long write_bw = wb->avg_write_bandwidth;
913 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
914 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
915 unsigned long wb_thresh = dtc->wb_thresh;
916 unsigned long x_intercept;
917 unsigned long setpoint; /* dirty pages' target balance point */
918 unsigned long wb_setpoint;
920 long long pos_ratio; /* for scaling up/down the rate limit */
925 if (unlikely(dtc->dirty >= limit))
931 * See comment for pos_ratio_polynom().
933 setpoint = (freerun + limit) / 2;
934 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
937 * The strictlimit feature is a tool preventing mistrusted filesystems
938 * from growing a large number of dirty pages before throttling. For
939 * such filesystems balance_dirty_pages always checks wb counters
940 * against wb limits. Even if global "nr_dirty" is under "freerun".
941 * This is especially important for fuse which sets bdi->max_ratio to
942 * 1% by default. Without strictlimit feature, fuse writeback may
943 * consume arbitrary amount of RAM because it is accounted in
944 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
946 * Here, in wb_position_ratio(), we calculate pos_ratio based on
947 * two values: wb_dirty and wb_thresh. Let's consider an example:
948 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
949 * limits are set by default to 10% and 20% (background and throttle).
950 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
951 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
952 * about ~6K pages (as the average of background and throttle wb
953 * limits). The 3rd order polynomial will provide positive feedback if
954 * wb_dirty is under wb_setpoint and vice versa.
956 * Note, that we cannot use global counters in these calculations
957 * because we want to throttle process writing to a strictlimit wb
958 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
959 * in the example above).
961 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
962 long long wb_pos_ratio;
964 if (dtc->wb_dirty < 8) {
965 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
966 2 << RATELIMIT_CALC_SHIFT);
970 if (dtc->wb_dirty >= wb_thresh)
973 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
976 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
979 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
983 * Typically, for strictlimit case, wb_setpoint << setpoint
984 * and pos_ratio >> wb_pos_ratio. In the other words global
985 * state ("dirty") is not limiting factor and we have to
986 * make decision based on wb counters. But there is an
987 * important case when global pos_ratio should get precedence:
988 * global limits are exceeded (e.g. due to activities on other
989 * wb's) while given strictlimit wb is below limit.
991 * "pos_ratio * wb_pos_ratio" would work for the case above,
992 * but it would look too non-natural for the case of all
993 * activity in the system coming from a single strictlimit wb
994 * with bdi->max_ratio == 100%.
996 * Note that min() below somewhat changes the dynamics of the
997 * control system. Normally, pos_ratio value can be well over 3
998 * (when globally we are at freerun and wb is well below wb
999 * setpoint). Now the maximum pos_ratio in the same situation
1000 * is 2. We might want to tweak this if we observe the control
1001 * system is too slow to adapt.
1003 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
1008 * We have computed basic pos_ratio above based on global situation. If
1009 * the wb is over/under its share of dirty pages, we want to scale
1010 * pos_ratio further down/up. That is done by the following mechanism.
1016 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1018 * x_intercept - wb_dirty
1019 * := --------------------------
1020 * x_intercept - wb_setpoint
1022 * The main wb control line is a linear function that subjects to
1024 * (1) f(wb_setpoint) = 1.0
1025 * (2) k = - 1 / (8 * write_bw) (in single wb case)
1026 * or equally: x_intercept = wb_setpoint + 8 * write_bw
1028 * For single wb case, the dirty pages are observed to fluctuate
1029 * regularly within range
1030 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1031 * for various filesystems, where (2) can yield in a reasonable 12.5%
1032 * fluctuation range for pos_ratio.
1034 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1035 * own size, so move the slope over accordingly and choose a slope that
1036 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1038 if (unlikely(wb_thresh > dtc->thresh))
1039 wb_thresh = dtc->thresh;
1041 * It's very possible that wb_thresh is close to 0 not because the
1042 * device is slow, but that it has remained inactive for long time.
1043 * Honour such devices a reasonable good (hopefully IO efficient)
1044 * threshold, so that the occasional writes won't be blocked and active
1045 * writes can rampup the threshold quickly.
1047 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1049 * scale global setpoint to wb's:
1050 * wb_setpoint = setpoint * wb_thresh / thresh
1052 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1053 wb_setpoint = setpoint * (u64)x >> 16;
1055 * Use span=(8*write_bw) in single wb case as indicated by
1056 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1058 * wb_thresh thresh - wb_thresh
1059 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1062 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1063 x_intercept = wb_setpoint + span;
1065 if (dtc->wb_dirty < x_intercept - span / 4) {
1066 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1067 (x_intercept - wb_setpoint) | 1);
1072 * wb reserve area, safeguard against dirty pool underrun and disk idle
1073 * It may push the desired control point of global dirty pages higher
1076 x_intercept = wb_thresh / 2;
1077 if (dtc->wb_dirty < x_intercept) {
1078 if (dtc->wb_dirty > x_intercept / 8)
1079 pos_ratio = div_u64(pos_ratio * x_intercept,
1085 dtc->pos_ratio = pos_ratio;
1088 static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1089 unsigned long elapsed,
1090 unsigned long written)
1092 const unsigned long period = roundup_pow_of_two(3 * HZ);
1093 unsigned long avg = wb->avg_write_bandwidth;
1094 unsigned long old = wb->write_bandwidth;
1098 * bw = written * HZ / elapsed
1100 * bw * elapsed + write_bandwidth * (period - elapsed)
1101 * write_bandwidth = ---------------------------------------------------
1104 * @written may have decreased due to account_page_redirty().
1105 * Avoid underflowing @bw calculation.
1107 bw = written - min(written, wb->written_stamp);
1109 if (unlikely(elapsed > period)) {
1110 do_div(bw, elapsed);
1114 bw += (u64)wb->write_bandwidth * (period - elapsed);
1115 bw >>= ilog2(period);
1118 * one more level of smoothing, for filtering out sudden spikes
1120 if (avg > old && old >= (unsigned long)bw)
1121 avg -= (avg - old) >> 3;
1123 if (avg < old && old <= (unsigned long)bw)
1124 avg += (old - avg) >> 3;
1127 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1128 avg = max(avg, 1LU);
1129 if (wb_has_dirty_io(wb)) {
1130 long delta = avg - wb->avg_write_bandwidth;
1131 WARN_ON_ONCE(atomic_long_add_return(delta,
1132 &wb->bdi->tot_write_bandwidth) <= 0);
1134 wb->write_bandwidth = bw;
1135 wb->avg_write_bandwidth = avg;
1138 static void update_dirty_limit(struct dirty_throttle_control *dtc)
1140 struct wb_domain *dom = dtc_dom(dtc);
1141 unsigned long thresh = dtc->thresh;
1142 unsigned long limit = dom->dirty_limit;
1145 * Follow up in one step.
1147 if (limit < thresh) {
1153 * Follow down slowly. Use the higher one as the target, because thresh
1154 * may drop below dirty. This is exactly the reason to introduce
1155 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1157 thresh = max(thresh, dtc->dirty);
1158 if (limit > thresh) {
1159 limit -= (limit - thresh) >> 5;
1164 dom->dirty_limit = limit;
1167 static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
1170 struct wb_domain *dom = dtc_dom(dtc);
1173 * check locklessly first to optimize away locking for the most time
1175 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1178 spin_lock(&dom->lock);
1179 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1180 update_dirty_limit(dtc);
1181 dom->dirty_limit_tstamp = now;
1183 spin_unlock(&dom->lock);
1187 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1189 * Normal wb tasks will be curbed at or below it in long term.
1190 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1192 static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1193 unsigned long dirtied,
1194 unsigned long elapsed)
1196 struct bdi_writeback *wb = dtc->wb;
1197 unsigned long dirty = dtc->dirty;
1198 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1199 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1200 unsigned long setpoint = (freerun + limit) / 2;
1201 unsigned long write_bw = wb->avg_write_bandwidth;
1202 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1203 unsigned long dirty_rate;
1204 unsigned long task_ratelimit;
1205 unsigned long balanced_dirty_ratelimit;
1208 unsigned long shift;
1211 * The dirty rate will match the writeout rate in long term, except
1212 * when dirty pages are truncated by userspace or re-dirtied by FS.
1214 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1217 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1219 task_ratelimit = (u64)dirty_ratelimit *
1220 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1221 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1224 * A linear estimation of the "balanced" throttle rate. The theory is,
1225 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1226 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1227 * formula will yield the balanced rate limit (write_bw / N).
1229 * Note that the expanded form is not a pure rate feedback:
1230 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1231 * but also takes pos_ratio into account:
1232 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1234 * (1) is not realistic because pos_ratio also takes part in balancing
1235 * the dirty rate. Consider the state
1236 * pos_ratio = 0.5 (3)
1237 * rate = 2 * (write_bw / N) (4)
1238 * If (1) is used, it will stuck in that state! Because each dd will
1240 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1242 * dirty_rate = N * task_ratelimit = write_bw (6)
1243 * put (6) into (1) we get
1244 * rate_(i+1) = rate_(i) (7)
1246 * So we end up using (2) to always keep
1247 * rate_(i+1) ~= (write_bw / N) (8)
1248 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1249 * pos_ratio is able to drive itself to 1.0, which is not only where
1250 * the dirty count meet the setpoint, but also where the slope of
1251 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1253 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1256 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1258 if (unlikely(balanced_dirty_ratelimit > write_bw))
1259 balanced_dirty_ratelimit = write_bw;
1262 * We could safely do this and return immediately:
1264 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1266 * However to get a more stable dirty_ratelimit, the below elaborated
1267 * code makes use of task_ratelimit to filter out singular points and
1268 * limit the step size.
1270 * The below code essentially only uses the relative value of
1272 * task_ratelimit - dirty_ratelimit
1273 * = (pos_ratio - 1) * dirty_ratelimit
1275 * which reflects the direction and size of dirty position error.
1279 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1280 * task_ratelimit is on the same side of dirty_ratelimit, too.
1282 * - dirty_ratelimit > balanced_dirty_ratelimit
1283 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1284 * lowering dirty_ratelimit will help meet both the position and rate
1285 * control targets. Otherwise, don't update dirty_ratelimit if it will
1286 * only help meet the rate target. After all, what the users ultimately
1287 * feel and care are stable dirty rate and small position error.
1289 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1290 * and filter out the singular points of balanced_dirty_ratelimit. Which
1291 * keeps jumping around randomly and can even leap far away at times
1292 * due to the small 200ms estimation period of dirty_rate (we want to
1293 * keep that period small to reduce time lags).
1298 * For strictlimit case, calculations above were based on wb counters
1299 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1300 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1301 * Hence, to calculate "step" properly, we have to use wb_dirty as
1302 * "dirty" and wb_setpoint as "setpoint".
1304 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1305 * it's possible that wb_thresh is close to zero due to inactivity
1306 * of backing device.
1308 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1309 dirty = dtc->wb_dirty;
1310 if (dtc->wb_dirty < 8)
1311 setpoint = dtc->wb_dirty + 1;
1313 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1316 if (dirty < setpoint) {
1317 x = min3(wb->balanced_dirty_ratelimit,
1318 balanced_dirty_ratelimit, task_ratelimit);
1319 if (dirty_ratelimit < x)
1320 step = x - dirty_ratelimit;
1322 x = max3(wb->balanced_dirty_ratelimit,
1323 balanced_dirty_ratelimit, task_ratelimit);
1324 if (dirty_ratelimit > x)
1325 step = dirty_ratelimit - x;
1329 * Don't pursue 100% rate matching. It's impossible since the balanced
1330 * rate itself is constantly fluctuating. So decrease the track speed
1331 * when it gets close to the target. Helps eliminate pointless tremors.
1333 shift = dirty_ratelimit / (2 * step + 1);
1334 if (shift < BITS_PER_LONG)
1335 step = DIV_ROUND_UP(step >> shift, 8);
1339 if (dirty_ratelimit < balanced_dirty_ratelimit)
1340 dirty_ratelimit += step;
1342 dirty_ratelimit -= step;
1344 wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1345 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1347 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1350 static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1351 struct dirty_throttle_control *mdtc,
1352 unsigned long start_time,
1353 bool update_ratelimit)
1355 struct bdi_writeback *wb = gdtc->wb;
1356 unsigned long now = jiffies;
1357 unsigned long elapsed = now - wb->bw_time_stamp;
1358 unsigned long dirtied;
1359 unsigned long written;
1361 lockdep_assert_held(&wb->list_lock);
1364 * rate-limit, only update once every 200ms.
1366 if (elapsed < BANDWIDTH_INTERVAL)
1369 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1370 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1373 * Skip quiet periods when disk bandwidth is under-utilized.
1374 * (at least 1s idle time between two flusher runs)
1376 if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1379 if (update_ratelimit) {
1380 domain_update_bandwidth(gdtc, now);
1381 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1384 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1385 * compiler has no way to figure that out. Help it.
1387 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1388 domain_update_bandwidth(mdtc, now);
1389 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1392 wb_update_write_bandwidth(wb, elapsed, written);
1395 wb->dirtied_stamp = dirtied;
1396 wb->written_stamp = written;
1397 wb->bw_time_stamp = now;
1400 void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1402 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1404 __wb_update_bandwidth(&gdtc, NULL, start_time, false);
1408 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1409 * will look to see if it needs to start dirty throttling.
1411 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1412 * global_page_state() too often. So scale it near-sqrt to the safety margin
1413 * (the number of pages we may dirty without exceeding the dirty limits).
1415 static unsigned long dirty_poll_interval(unsigned long dirty,
1416 unsigned long thresh)
1419 return 1UL << (ilog2(thresh - dirty) >> 1);
1424 static unsigned long wb_max_pause(struct bdi_writeback *wb,
1425 unsigned long wb_dirty)
1427 unsigned long bw = wb->avg_write_bandwidth;
1431 * Limit pause time for small memory systems. If sleeping for too long
1432 * time, a small pool of dirty/writeback pages may go empty and disk go
1435 * 8 serves as the safety ratio.
1437 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1440 return min_t(unsigned long, t, MAX_PAUSE);
1443 static long wb_min_pause(struct bdi_writeback *wb,
1445 unsigned long task_ratelimit,
1446 unsigned long dirty_ratelimit,
1447 int *nr_dirtied_pause)
1449 long hi = ilog2(wb->avg_write_bandwidth);
1450 long lo = ilog2(wb->dirty_ratelimit);
1451 long t; /* target pause */
1452 long pause; /* estimated next pause */
1453 int pages; /* target nr_dirtied_pause */
1455 /* target for 10ms pause on 1-dd case */
1456 t = max(1, HZ / 100);
1459 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1462 * (N * 10ms) on 2^N concurrent tasks.
1465 t += (hi - lo) * (10 * HZ) / 1024;
1468 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1469 * on the much more stable dirty_ratelimit. However the next pause time
1470 * will be computed based on task_ratelimit and the two rate limits may
1471 * depart considerably at some time. Especially if task_ratelimit goes
1472 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1473 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1474 * result task_ratelimit won't be executed faithfully, which could
1475 * eventually bring down dirty_ratelimit.
1477 * We apply two rules to fix it up:
1478 * 1) try to estimate the next pause time and if necessary, use a lower
1479 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1480 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1481 * 2) limit the target pause time to max_pause/2, so that the normal
1482 * small fluctuations of task_ratelimit won't trigger rule (1) and
1483 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1485 t = min(t, 1 + max_pause / 2);
1486 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1489 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1490 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1491 * When the 16 consecutive reads are often interrupted by some dirty
1492 * throttling pause during the async writes, cfq will go into idles
1493 * (deadline is fine). So push nr_dirtied_pause as high as possible
1494 * until reaches DIRTY_POLL_THRESH=32 pages.
1496 if (pages < DIRTY_POLL_THRESH) {
1498 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1499 if (pages > DIRTY_POLL_THRESH) {
1500 pages = DIRTY_POLL_THRESH;
1501 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1505 pause = HZ * pages / (task_ratelimit + 1);
1506 if (pause > max_pause) {
1508 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1511 *nr_dirtied_pause = pages;
1513 * The minimal pause time will normally be half the target pause time.
1515 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1518 static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1520 struct bdi_writeback *wb = dtc->wb;
1521 unsigned long wb_reclaimable;
1524 * wb_thresh is not treated as some limiting factor as
1525 * dirty_thresh, due to reasons
1526 * - in JBOD setup, wb_thresh can fluctuate a lot
1527 * - in a system with HDD and USB key, the USB key may somehow
1528 * go into state (wb_dirty >> wb_thresh) either because
1529 * wb_dirty starts high, or because wb_thresh drops low.
1530 * In this case we don't want to hard throttle the USB key
1531 * dirtiers for 100 seconds until wb_dirty drops under
1532 * wb_thresh. Instead the auxiliary wb control line in
1533 * wb_position_ratio() will let the dirtier task progress
1534 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1536 dtc->wb_thresh = __wb_calc_thresh(dtc);
1537 dtc->wb_bg_thresh = dtc->thresh ?
1538 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1541 * In order to avoid the stacked BDI deadlock we need
1542 * to ensure we accurately count the 'dirty' pages when
1543 * the threshold is low.
1545 * Otherwise it would be possible to get thresh+n pages
1546 * reported dirty, even though there are thresh-m pages
1547 * actually dirty; with m+n sitting in the percpu
1550 if (dtc->wb_thresh < 2 * wb_stat_error(wb)) {
1551 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1552 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1554 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1555 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1560 * balance_dirty_pages() must be called by processes which are generating dirty
1561 * data. It looks at the number of dirty pages in the machine and will force
1562 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1563 * If we're over `background_thresh' then the writeback threads are woken to
1564 * perform some writeout.
1566 static void balance_dirty_pages(struct address_space *mapping,
1567 struct bdi_writeback *wb,
1568 unsigned long pages_dirtied)
1570 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1571 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1572 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1573 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1575 struct dirty_throttle_control *sdtc;
1576 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1581 int nr_dirtied_pause;
1582 bool dirty_exceeded = false;
1583 unsigned long task_ratelimit;
1584 unsigned long dirty_ratelimit;
1585 struct backing_dev_info *bdi = wb->bdi;
1586 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1587 unsigned long start_time = jiffies;
1590 unsigned long now = jiffies;
1591 unsigned long dirty, thresh, bg_thresh;
1592 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1593 unsigned long m_thresh = 0;
1594 unsigned long m_bg_thresh = 0;
1597 * Unstable writes are a feature of certain networked
1598 * filesystems (i.e. NFS) in which data may have been
1599 * written to the server's write cache, but has not yet
1600 * been flushed to permanent storage.
1602 nr_reclaimable = global_node_page_state(NR_FILE_DIRTY) +
1603 global_node_page_state(NR_UNSTABLE_NFS);
1604 gdtc->avail = global_dirtyable_memory();
1605 gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1607 domain_dirty_limits(gdtc);
1609 if (unlikely(strictlimit)) {
1610 wb_dirty_limits(gdtc);
1612 dirty = gdtc->wb_dirty;
1613 thresh = gdtc->wb_thresh;
1614 bg_thresh = gdtc->wb_bg_thresh;
1616 dirty = gdtc->dirty;
1617 thresh = gdtc->thresh;
1618 bg_thresh = gdtc->bg_thresh;
1622 unsigned long filepages, headroom, writeback;
1625 * If @wb belongs to !root memcg, repeat the same
1626 * basic calculations for the memcg domain.
1628 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1629 &mdtc->dirty, &writeback);
1630 mdtc->dirty += writeback;
1631 mdtc_calc_avail(mdtc, filepages, headroom);
1633 domain_dirty_limits(mdtc);
1635 if (unlikely(strictlimit)) {
1636 wb_dirty_limits(mdtc);
1637 m_dirty = mdtc->wb_dirty;
1638 m_thresh = mdtc->wb_thresh;
1639 m_bg_thresh = mdtc->wb_bg_thresh;
1641 m_dirty = mdtc->dirty;
1642 m_thresh = mdtc->thresh;
1643 m_bg_thresh = mdtc->bg_thresh;
1648 * Throttle it only when the background writeback cannot
1649 * catch-up. This avoids (excessively) small writeouts
1650 * when the wb limits are ramping up in case of !strictlimit.
1652 * In strictlimit case make decision based on the wb counters
1653 * and limits. Small writeouts when the wb limits are ramping
1654 * up are the price we consciously pay for strictlimit-ing.
1656 * If memcg domain is in effect, @dirty should be under
1657 * both global and memcg freerun ceilings.
1659 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1661 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1662 unsigned long intv = dirty_poll_interval(dirty, thresh);
1663 unsigned long m_intv = ULONG_MAX;
1665 current->dirty_paused_when = now;
1666 current->nr_dirtied = 0;
1668 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1669 current->nr_dirtied_pause = min(intv, m_intv);
1673 if (unlikely(!writeback_in_progress(wb)))
1674 wb_start_background_writeback(wb);
1677 * Calculate global domain's pos_ratio and select the
1678 * global dtc by default.
1681 wb_dirty_limits(gdtc);
1683 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1684 ((gdtc->dirty > gdtc->thresh) || strictlimit);
1686 wb_position_ratio(gdtc);
1691 * If memcg domain is in effect, calculate its
1692 * pos_ratio. @wb should satisfy constraints from
1693 * both global and memcg domains. Choose the one
1694 * w/ lower pos_ratio.
1697 wb_dirty_limits(mdtc);
1699 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1700 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1702 wb_position_ratio(mdtc);
1703 if (mdtc->pos_ratio < gdtc->pos_ratio)
1707 if (dirty_exceeded && !wb->dirty_exceeded)
1708 wb->dirty_exceeded = 1;
1710 if (time_is_before_jiffies(wb->bw_time_stamp +
1711 BANDWIDTH_INTERVAL)) {
1712 spin_lock(&wb->list_lock);
1713 __wb_update_bandwidth(gdtc, mdtc, start_time, true);
1714 spin_unlock(&wb->list_lock);
1717 /* throttle according to the chosen dtc */
1718 dirty_ratelimit = wb->dirty_ratelimit;
1719 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1720 RATELIMIT_CALC_SHIFT;
1721 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1722 min_pause = wb_min_pause(wb, max_pause,
1723 task_ratelimit, dirty_ratelimit,
1726 if (unlikely(task_ratelimit == 0)) {
1731 period = HZ * pages_dirtied / task_ratelimit;
1733 if (current->dirty_paused_when)
1734 pause -= now - current->dirty_paused_when;
1736 * For less than 1s think time (ext3/4 may block the dirtier
1737 * for up to 800ms from time to time on 1-HDD; so does xfs,
1738 * however at much less frequency), try to compensate it in
1739 * future periods by updating the virtual time; otherwise just
1740 * do a reset, as it may be a light dirtier.
1742 if (pause < min_pause) {
1743 trace_balance_dirty_pages(wb,
1756 current->dirty_paused_when = now;
1757 current->nr_dirtied = 0;
1758 } else if (period) {
1759 current->dirty_paused_when += period;
1760 current->nr_dirtied = 0;
1761 } else if (current->nr_dirtied_pause <= pages_dirtied)
1762 current->nr_dirtied_pause += pages_dirtied;
1765 if (unlikely(pause > max_pause)) {
1766 /* for occasional dropped task_ratelimit */
1767 now += min(pause - max_pause, max_pause);
1772 trace_balance_dirty_pages(wb,
1784 __set_current_state(TASK_KILLABLE);
1785 io_schedule_timeout(pause);
1787 current->dirty_paused_when = now + pause;
1788 current->nr_dirtied = 0;
1789 current->nr_dirtied_pause = nr_dirtied_pause;
1792 * This is typically equal to (dirty < thresh) and can also
1793 * keep "1000+ dd on a slow USB stick" under control.
1799 * In the case of an unresponding NFS server and the NFS dirty
1800 * pages exceeds dirty_thresh, give the other good wb's a pipe
1801 * to go through, so that tasks on them still remain responsive.
1803 * In theory 1 page is enough to keep the comsumer-producer
1804 * pipe going: the flusher cleans 1 page => the task dirties 1
1805 * more page. However wb_dirty has accounting errors. So use
1806 * the larger and more IO friendly wb_stat_error.
1808 if (sdtc->wb_dirty <= wb_stat_error(wb))
1811 if (fatal_signal_pending(current))
1815 if (!dirty_exceeded && wb->dirty_exceeded)
1816 wb->dirty_exceeded = 0;
1818 if (writeback_in_progress(wb))
1822 * In laptop mode, we wait until hitting the higher threshold before
1823 * starting background writeout, and then write out all the way down
1824 * to the lower threshold. So slow writers cause minimal disk activity.
1826 * In normal mode, we start background writeout at the lower
1827 * background_thresh, to keep the amount of dirty memory low.
1832 if (nr_reclaimable > gdtc->bg_thresh)
1833 wb_start_background_writeback(wb);
1836 static DEFINE_PER_CPU(int, bdp_ratelimits);
1839 * Normal tasks are throttled by
1841 * dirty tsk->nr_dirtied_pause pages;
1842 * take a snap in balance_dirty_pages();
1844 * However there is a worst case. If every task exit immediately when dirtied
1845 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1846 * called to throttle the page dirties. The solution is to save the not yet
1847 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1848 * randomly into the running tasks. This works well for the above worst case,
1849 * as the new task will pick up and accumulate the old task's leaked dirty
1850 * count and eventually get throttled.
1852 DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1855 * balance_dirty_pages_ratelimited - balance dirty memory state
1856 * @mapping: address_space which was dirtied
1858 * Processes which are dirtying memory should call in here once for each page
1859 * which was newly dirtied. The function will periodically check the system's
1860 * dirty state and will initiate writeback if needed.
1862 * On really big machines, get_writeback_state is expensive, so try to avoid
1863 * calling it too often (ratelimiting). But once we're over the dirty memory
1864 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1865 * from overshooting the limit by (ratelimit_pages) each.
1867 void balance_dirty_pages_ratelimited(struct address_space *mapping)
1869 struct inode *inode = mapping->host;
1870 struct backing_dev_info *bdi = inode_to_bdi(inode);
1871 struct bdi_writeback *wb = NULL;
1875 if (!bdi_cap_account_dirty(bdi))
1878 if (inode_cgwb_enabled(inode))
1879 wb = wb_get_create_current(bdi, GFP_KERNEL);
1883 ratelimit = current->nr_dirtied_pause;
1884 if (wb->dirty_exceeded)
1885 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1889 * This prevents one CPU to accumulate too many dirtied pages without
1890 * calling into balance_dirty_pages(), which can happen when there are
1891 * 1000+ tasks, all of them start dirtying pages at exactly the same
1892 * time, hence all honoured too large initial task->nr_dirtied_pause.
1894 p = this_cpu_ptr(&bdp_ratelimits);
1895 if (unlikely(current->nr_dirtied >= ratelimit))
1897 else if (unlikely(*p >= ratelimit_pages)) {
1902 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1903 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1904 * the dirty throttling and livelock other long-run dirtiers.
1906 p = this_cpu_ptr(&dirty_throttle_leaks);
1907 if (*p > 0 && current->nr_dirtied < ratelimit) {
1908 unsigned long nr_pages_dirtied;
1909 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1910 *p -= nr_pages_dirtied;
1911 current->nr_dirtied += nr_pages_dirtied;
1915 if (unlikely(current->nr_dirtied >= ratelimit))
1916 balance_dirty_pages(mapping, wb, current->nr_dirtied);
1920 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1923 * wb_over_bg_thresh - does @wb need to be written back?
1924 * @wb: bdi_writeback of interest
1926 * Determines whether background writeback should keep writing @wb or it's
1927 * clean enough. Returns %true if writeback should continue.
1929 bool wb_over_bg_thresh(struct bdi_writeback *wb)
1931 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1932 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1933 struct dirty_throttle_control * const gdtc = &gdtc_stor;
1934 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1938 * Similar to balance_dirty_pages() but ignores pages being written
1939 * as we're trying to decide whether to put more under writeback.
1941 gdtc->avail = global_dirtyable_memory();
1942 gdtc->dirty = global_node_page_state(NR_FILE_DIRTY) +
1943 global_node_page_state(NR_UNSTABLE_NFS);
1944 domain_dirty_limits(gdtc);
1946 if (gdtc->dirty > gdtc->bg_thresh)
1949 if (wb_stat(wb, WB_RECLAIMABLE) >
1950 wb_calc_thresh(gdtc->wb, gdtc->bg_thresh))
1954 unsigned long filepages, headroom, writeback;
1956 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1958 mdtc_calc_avail(mdtc, filepages, headroom);
1959 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1961 if (mdtc->dirty > mdtc->bg_thresh)
1964 if (wb_stat(wb, WB_RECLAIMABLE) >
1965 wb_calc_thresh(mdtc->wb, mdtc->bg_thresh))
1972 void throttle_vm_writeout(gfp_t gfp_mask)
1974 unsigned long background_thresh;
1975 unsigned long dirty_thresh;
1978 global_dirty_limits(&background_thresh, &dirty_thresh);
1979 dirty_thresh = hard_dirty_limit(&global_wb_domain, dirty_thresh);
1982 * Boost the allowable dirty threshold a bit for page
1983 * allocators so they don't get DoS'ed by heavy writers
1985 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1987 if (global_node_page_state(NR_UNSTABLE_NFS) +
1988 global_node_page_state(NR_WRITEBACK) <= dirty_thresh)
1990 congestion_wait(BLK_RW_ASYNC, HZ/10);
1993 * The caller might hold locks which can prevent IO completion
1994 * or progress in the filesystem. So we cannot just sit here
1995 * waiting for IO to complete.
1997 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
2003 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
2005 int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
2006 void __user *buffer, size_t *length, loff_t *ppos)
2008 proc_dointvec(table, write, buffer, length, ppos);
2013 void laptop_mode_timer_fn(unsigned long data)
2015 struct request_queue *q = (struct request_queue *)data;
2016 int nr_pages = global_node_page_state(NR_FILE_DIRTY) +
2017 global_node_page_state(NR_UNSTABLE_NFS);
2018 struct bdi_writeback *wb;
2021 * We want to write everything out, not just down to the dirty
2024 if (!bdi_has_dirty_io(&q->backing_dev_info))
2028 list_for_each_entry_rcu(wb, &q->backing_dev_info.wb_list, bdi_node)
2029 if (wb_has_dirty_io(wb))
2030 wb_start_writeback(wb, nr_pages, true,
2031 WB_REASON_LAPTOP_TIMER);
2036 * We've spun up the disk and we're in laptop mode: schedule writeback
2037 * of all dirty data a few seconds from now. If the flush is already scheduled
2038 * then push it back - the user is still using the disk.
2040 void laptop_io_completion(struct backing_dev_info *info)
2042 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2046 * We're in laptop mode and we've just synced. The sync's writes will have
2047 * caused another writeback to be scheduled by laptop_io_completion.
2048 * Nothing needs to be written back anymore, so we unschedule the writeback.
2050 void laptop_sync_completion(void)
2052 struct backing_dev_info *bdi;
2056 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2057 del_timer(&bdi->laptop_mode_wb_timer);
2064 * If ratelimit_pages is too high then we can get into dirty-data overload
2065 * if a large number of processes all perform writes at the same time.
2066 * If it is too low then SMP machines will call the (expensive)
2067 * get_writeback_state too often.
2069 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2070 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2074 void writeback_set_ratelimit(void)
2076 struct wb_domain *dom = &global_wb_domain;
2077 unsigned long background_thresh;
2078 unsigned long dirty_thresh;
2080 global_dirty_limits(&background_thresh, &dirty_thresh);
2081 dom->dirty_limit = dirty_thresh;
2082 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2083 if (ratelimit_pages < 16)
2084 ratelimit_pages = 16;
2088 ratelimit_handler(struct notifier_block *self, unsigned long action,
2092 switch (action & ~CPU_TASKS_FROZEN) {
2095 writeback_set_ratelimit();
2102 static struct notifier_block ratelimit_nb = {
2103 .notifier_call = ratelimit_handler,
2108 * Called early on to tune the page writeback dirty limits.
2110 * We used to scale dirty pages according to how total memory
2111 * related to pages that could be allocated for buffers (by
2112 * comparing nr_free_buffer_pages() to vm_total_pages.
2114 * However, that was when we used "dirty_ratio" to scale with
2115 * all memory, and we don't do that any more. "dirty_ratio"
2116 * is now applied to total non-HIGHPAGE memory (by subtracting
2117 * totalhigh_pages from vm_total_pages), and as such we can't
2118 * get into the old insane situation any more where we had
2119 * large amounts of dirty pages compared to a small amount of
2120 * non-HIGHMEM memory.
2122 * But we might still want to scale the dirty_ratio by how
2123 * much memory the box has..
2125 void __init page_writeback_init(void)
2127 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2129 writeback_set_ratelimit();
2130 register_cpu_notifier(&ratelimit_nb);
2134 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2135 * @mapping: address space structure to write
2136 * @start: starting page index
2137 * @end: ending page index (inclusive)
2139 * This function scans the page range from @start to @end (inclusive) and tags
2140 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2141 * that write_cache_pages (or whoever calls this function) will then use
2142 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2143 * used to avoid livelocking of writeback by a process steadily creating new
2144 * dirty pages in the file (thus it is important for this function to be quick
2145 * so that it can tag pages faster than a dirtying process can create them).
2148 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
2150 void tag_pages_for_writeback(struct address_space *mapping,
2151 pgoff_t start, pgoff_t end)
2153 #define WRITEBACK_TAG_BATCH 4096
2154 unsigned long tagged;
2157 spin_lock_irq(&mapping->tree_lock);
2158 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
2159 &start, end, WRITEBACK_TAG_BATCH,
2160 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
2161 spin_unlock_irq(&mapping->tree_lock);
2162 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
2164 /* We check 'start' to handle wrapping when end == ~0UL */
2165 } while (tagged >= WRITEBACK_TAG_BATCH && start);
2167 EXPORT_SYMBOL(tag_pages_for_writeback);
2170 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2171 * @mapping: address space structure to write
2172 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2173 * @writepage: function called for each page
2174 * @data: data passed to writepage function
2176 * If a page is already under I/O, write_cache_pages() skips it, even
2177 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2178 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2179 * and msync() need to guarantee that all the data which was dirty at the time
2180 * the call was made get new I/O started against them. If wbc->sync_mode is
2181 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2182 * existing IO to complete.
2184 * To avoid livelocks (when other process dirties new pages), we first tag
2185 * pages which should be written back with TOWRITE tag and only then start
2186 * writing them. For data-integrity sync we have to be careful so that we do
2187 * not miss some pages (e.g., because some other process has cleared TOWRITE
2188 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2189 * by the process clearing the DIRTY tag (and submitting the page for IO).
2191 int write_cache_pages(struct address_space *mapping,
2192 struct writeback_control *wbc, writepage_t writepage,
2197 struct pagevec pvec;
2199 pgoff_t uninitialized_var(writeback_index);
2201 pgoff_t end; /* Inclusive */
2204 int range_whole = 0;
2207 pagevec_init(&pvec, 0);
2208 if (wbc->range_cyclic) {
2209 writeback_index = mapping->writeback_index; /* prev offset */
2210 index = writeback_index;
2217 index = wbc->range_start >> PAGE_SHIFT;
2218 end = wbc->range_end >> PAGE_SHIFT;
2219 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2221 cycled = 1; /* ignore range_cyclic tests */
2223 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2224 tag = PAGECACHE_TAG_TOWRITE;
2226 tag = PAGECACHE_TAG_DIRTY;
2228 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2229 tag_pages_for_writeback(mapping, index, end);
2231 while (!done && (index <= end)) {
2234 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2235 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2239 for (i = 0; i < nr_pages; i++) {
2240 struct page *page = pvec.pages[i];
2243 * At this point, the page may be truncated or
2244 * invalidated (changing page->mapping to NULL), or
2245 * even swizzled back from swapper_space to tmpfs file
2246 * mapping. However, page->index will not change
2247 * because we have a reference on the page.
2249 if (page->index > end) {
2251 * can't be range_cyclic (1st pass) because
2252 * end == -1 in that case.
2258 done_index = page->index;
2263 * Page truncated or invalidated. We can freely skip it
2264 * then, even for data integrity operations: the page
2265 * has disappeared concurrently, so there could be no
2266 * real expectation of this data interity operation
2267 * even if there is now a new, dirty page at the same
2268 * pagecache address.
2270 if (unlikely(page->mapping != mapping)) {
2276 if (!PageDirty(page)) {
2277 /* someone wrote it for us */
2278 goto continue_unlock;
2281 if (PageWriteback(page)) {
2282 if (wbc->sync_mode != WB_SYNC_NONE)
2283 wait_on_page_writeback(page);
2285 goto continue_unlock;
2288 BUG_ON(PageWriteback(page));
2289 if (!clear_page_dirty_for_io(page))
2290 goto continue_unlock;
2292 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2293 ret = (*writepage)(page, wbc, data);
2294 if (unlikely(ret)) {
2295 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2300 * done_index is set past this page,
2301 * so media errors will not choke
2302 * background writeout for the entire
2303 * file. This has consequences for
2304 * range_cyclic semantics (ie. it may
2305 * not be suitable for data integrity
2308 done_index = page->index + 1;
2315 * We stop writing back only if we are not doing
2316 * integrity sync. In case of integrity sync we have to
2317 * keep going until we have written all the pages
2318 * we tagged for writeback prior to entering this loop.
2320 if (--wbc->nr_to_write <= 0 &&
2321 wbc->sync_mode == WB_SYNC_NONE) {
2326 pagevec_release(&pvec);
2329 if (!cycled && !done) {
2332 * We hit the last page and there is more work to be done: wrap
2333 * back to the start of the file
2337 end = writeback_index - 1;
2340 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2341 mapping->writeback_index = done_index;
2345 EXPORT_SYMBOL(write_cache_pages);
2348 * Function used by generic_writepages to call the real writepage
2349 * function and set the mapping flags on error
2351 static int __writepage(struct page *page, struct writeback_control *wbc,
2354 struct address_space *mapping = data;
2355 int ret = mapping->a_ops->writepage(page, wbc);
2356 mapping_set_error(mapping, ret);
2361 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2362 * @mapping: address space structure to write
2363 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2365 * This is a library function, which implements the writepages()
2366 * address_space_operation.
2368 int generic_writepages(struct address_space *mapping,
2369 struct writeback_control *wbc)
2371 struct blk_plug plug;
2374 /* deal with chardevs and other special file */
2375 if (!mapping->a_ops->writepage)
2378 blk_start_plug(&plug);
2379 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2380 blk_finish_plug(&plug);
2384 EXPORT_SYMBOL(generic_writepages);
2386 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2390 if (wbc->nr_to_write <= 0)
2392 if (mapping->a_ops->writepages)
2393 ret = mapping->a_ops->writepages(mapping, wbc);
2395 ret = generic_writepages(mapping, wbc);
2400 * write_one_page - write out a single page and optionally wait on I/O
2401 * @page: the page to write
2402 * @wait: if true, wait on writeout
2404 * The page must be locked by the caller and will be unlocked upon return.
2406 * write_one_page() returns a negative error code if I/O failed.
2408 int write_one_page(struct page *page, int wait)
2410 struct address_space *mapping = page->mapping;
2412 struct writeback_control wbc = {
2413 .sync_mode = WB_SYNC_ALL,
2417 BUG_ON(!PageLocked(page));
2420 wait_on_page_writeback(page);
2422 if (clear_page_dirty_for_io(page)) {
2424 ret = mapping->a_ops->writepage(page, &wbc);
2425 if (ret == 0 && wait) {
2426 wait_on_page_writeback(page);
2427 if (PageError(page))
2436 EXPORT_SYMBOL(write_one_page);
2439 * For address_spaces which do not use buffers nor write back.
2441 int __set_page_dirty_no_writeback(struct page *page)
2443 if (!PageDirty(page))
2444 return !TestSetPageDirty(page);
2449 * Helper function for set_page_dirty family.
2451 * Caller must hold lock_page_memcg().
2453 * NOTE: This relies on being atomic wrt interrupts.
2455 void account_page_dirtied(struct page *page, struct address_space *mapping)
2457 struct inode *inode = mapping->host;
2459 trace_writeback_dirty_page(page, mapping);
2461 if (mapping_cap_account_dirty(mapping)) {
2462 struct bdi_writeback *wb;
2464 inode_attach_wb(inode, page);
2465 wb = inode_to_wb(inode);
2467 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_DIRTY);
2468 __inc_node_page_state(page, NR_FILE_DIRTY);
2469 __inc_node_page_state(page, NR_DIRTIED);
2470 __inc_wb_stat(wb, WB_RECLAIMABLE);
2471 __inc_wb_stat(wb, WB_DIRTIED);
2472 task_io_account_write(PAGE_SIZE);
2473 current->nr_dirtied++;
2474 this_cpu_inc(bdp_ratelimits);
2477 EXPORT_SYMBOL(account_page_dirtied);
2480 * Helper function for deaccounting dirty page without writeback.
2482 * Caller must hold lock_page_memcg().
2484 void account_page_cleaned(struct page *page, struct address_space *mapping,
2485 struct bdi_writeback *wb)
2487 if (mapping_cap_account_dirty(mapping)) {
2488 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_DIRTY);
2489 dec_node_page_state(page, NR_FILE_DIRTY);
2490 dec_wb_stat(wb, WB_RECLAIMABLE);
2491 task_io_account_cancelled_write(PAGE_SIZE);
2496 * For address_spaces which do not use buffers. Just tag the page as dirty in
2499 * This is also used when a single buffer is being dirtied: we want to set the
2500 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2501 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2503 * The caller must ensure this doesn't race with truncation. Most will simply
2504 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2505 * the pte lock held, which also locks out truncation.
2507 int __set_page_dirty_nobuffers(struct page *page)
2509 lock_page_memcg(page);
2510 if (!TestSetPageDirty(page)) {
2511 struct address_space *mapping = page_mapping(page);
2512 unsigned long flags;
2515 unlock_page_memcg(page);
2519 spin_lock_irqsave(&mapping->tree_lock, flags);
2520 BUG_ON(page_mapping(page) != mapping);
2521 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2522 account_page_dirtied(page, mapping);
2523 radix_tree_tag_set(&mapping->page_tree, page_index(page),
2524 PAGECACHE_TAG_DIRTY);
2525 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2526 unlock_page_memcg(page);
2528 if (mapping->host) {
2529 /* !PageAnon && !swapper_space */
2530 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2534 unlock_page_memcg(page);
2537 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2540 * Call this whenever redirtying a page, to de-account the dirty counters
2541 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2542 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2543 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2546 void account_page_redirty(struct page *page)
2548 struct address_space *mapping = page->mapping;
2550 if (mapping && mapping_cap_account_dirty(mapping)) {
2551 struct inode *inode = mapping->host;
2552 struct bdi_writeback *wb;
2555 wb = unlocked_inode_to_wb_begin(inode, &locked);
2556 current->nr_dirtied--;
2557 dec_node_page_state(page, NR_DIRTIED);
2558 dec_wb_stat(wb, WB_DIRTIED);
2559 unlocked_inode_to_wb_end(inode, locked);
2562 EXPORT_SYMBOL(account_page_redirty);
2565 * When a writepage implementation decides that it doesn't want to write this
2566 * page for some reason, it should redirty the locked page via
2567 * redirty_page_for_writepage() and it should then unlock the page and return 0
2569 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2573 wbc->pages_skipped++;
2574 ret = __set_page_dirty_nobuffers(page);
2575 account_page_redirty(page);
2578 EXPORT_SYMBOL(redirty_page_for_writepage);
2583 * For pages with a mapping this should be done under the page lock
2584 * for the benefit of asynchronous memory errors who prefer a consistent
2585 * dirty state. This rule can be broken in some special cases,
2586 * but should be better not to.
2588 * If the mapping doesn't provide a set_page_dirty a_op, then
2589 * just fall through and assume that it wants buffer_heads.
2591 int set_page_dirty(struct page *page)
2593 struct address_space *mapping = page_mapping(page);
2595 page = compound_head(page);
2596 if (likely(mapping)) {
2597 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2599 * readahead/lru_deactivate_page could remain
2600 * PG_readahead/PG_reclaim due to race with end_page_writeback
2601 * About readahead, if the page is written, the flags would be
2602 * reset. So no problem.
2603 * About lru_deactivate_page, if the page is redirty, the flag
2604 * will be reset. So no problem. but if the page is used by readahead
2605 * it will confuse readahead and make it restart the size rampup
2606 * process. But it's a trivial problem.
2608 if (PageReclaim(page))
2609 ClearPageReclaim(page);
2612 spd = __set_page_dirty_buffers;
2614 return (*spd)(page);
2616 if (!PageDirty(page)) {
2617 if (!TestSetPageDirty(page))
2622 EXPORT_SYMBOL(set_page_dirty);
2625 * set_page_dirty() is racy if the caller has no reference against
2626 * page->mapping->host, and if the page is unlocked. This is because another
2627 * CPU could truncate the page off the mapping and then free the mapping.
2629 * Usually, the page _is_ locked, or the caller is a user-space process which
2630 * holds a reference on the inode by having an open file.
2632 * In other cases, the page should be locked before running set_page_dirty().
2634 int set_page_dirty_lock(struct page *page)
2639 ret = set_page_dirty(page);
2643 EXPORT_SYMBOL(set_page_dirty_lock);
2646 * This cancels just the dirty bit on the kernel page itself, it does NOT
2647 * actually remove dirty bits on any mmap's that may be around. It also
2648 * leaves the page tagged dirty, so any sync activity will still find it on
2649 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2650 * look at the dirty bits in the VM.
2652 * Doing this should *normally* only ever be done when a page is truncated,
2653 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2654 * this when it notices that somebody has cleaned out all the buffers on a
2655 * page without actually doing it through the VM. Can you say "ext3 is
2656 * horribly ugly"? Thought you could.
2658 void cancel_dirty_page(struct page *page)
2660 struct address_space *mapping = page_mapping(page);
2662 if (mapping_cap_account_dirty(mapping)) {
2663 struct inode *inode = mapping->host;
2664 struct bdi_writeback *wb;
2667 lock_page_memcg(page);
2668 wb = unlocked_inode_to_wb_begin(inode, &locked);
2670 if (TestClearPageDirty(page))
2671 account_page_cleaned(page, mapping, wb);
2673 unlocked_inode_to_wb_end(inode, locked);
2674 unlock_page_memcg(page);
2676 ClearPageDirty(page);
2679 EXPORT_SYMBOL(cancel_dirty_page);
2682 * Clear a page's dirty flag, while caring for dirty memory accounting.
2683 * Returns true if the page was previously dirty.
2685 * This is for preparing to put the page under writeout. We leave the page
2686 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2687 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2688 * implementation will run either set_page_writeback() or set_page_dirty(),
2689 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2692 * This incoherency between the page's dirty flag and radix-tree tag is
2693 * unfortunate, but it only exists while the page is locked.
2695 int clear_page_dirty_for_io(struct page *page)
2697 struct address_space *mapping = page_mapping(page);
2700 BUG_ON(!PageLocked(page));
2702 if (mapping && mapping_cap_account_dirty(mapping)) {
2703 struct inode *inode = mapping->host;
2704 struct bdi_writeback *wb;
2708 * Yes, Virginia, this is indeed insane.
2710 * We use this sequence to make sure that
2711 * (a) we account for dirty stats properly
2712 * (b) we tell the low-level filesystem to
2713 * mark the whole page dirty if it was
2714 * dirty in a pagetable. Only to then
2715 * (c) clean the page again and return 1 to
2716 * cause the writeback.
2718 * This way we avoid all nasty races with the
2719 * dirty bit in multiple places and clearing
2720 * them concurrently from different threads.
2722 * Note! Normally the "set_page_dirty(page)"
2723 * has no effect on the actual dirty bit - since
2724 * that will already usually be set. But we
2725 * need the side effects, and it can help us
2728 * We basically use the page "master dirty bit"
2729 * as a serialization point for all the different
2730 * threads doing their things.
2732 if (page_mkclean(page))
2733 set_page_dirty(page);
2735 * We carefully synchronise fault handlers against
2736 * installing a dirty pte and marking the page dirty
2737 * at this point. We do this by having them hold the
2738 * page lock while dirtying the page, and pages are
2739 * always locked coming in here, so we get the desired
2742 wb = unlocked_inode_to_wb_begin(inode, &locked);
2743 if (TestClearPageDirty(page)) {
2744 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_DIRTY);
2745 dec_node_page_state(page, NR_FILE_DIRTY);
2746 dec_wb_stat(wb, WB_RECLAIMABLE);
2749 unlocked_inode_to_wb_end(inode, locked);
2752 return TestClearPageDirty(page);
2754 EXPORT_SYMBOL(clear_page_dirty_for_io);
2756 int test_clear_page_writeback(struct page *page)
2758 struct address_space *mapping = page_mapping(page);
2761 lock_page_memcg(page);
2763 struct inode *inode = mapping->host;
2764 struct backing_dev_info *bdi = inode_to_bdi(inode);
2765 unsigned long flags;
2767 spin_lock_irqsave(&mapping->tree_lock, flags);
2768 ret = TestClearPageWriteback(page);
2770 radix_tree_tag_clear(&mapping->page_tree,
2772 PAGECACHE_TAG_WRITEBACK);
2773 if (bdi_cap_account_writeback(bdi)) {
2774 struct bdi_writeback *wb = inode_to_wb(inode);
2776 __dec_wb_stat(wb, WB_WRITEBACK);
2777 __wb_writeout_inc(wb);
2781 if (mapping->host && !mapping_tagged(mapping,
2782 PAGECACHE_TAG_WRITEBACK))
2783 sb_clear_inode_writeback(mapping->host);
2785 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2787 ret = TestClearPageWriteback(page);
2790 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
2791 dec_node_page_state(page, NR_WRITEBACK);
2792 inc_node_page_state(page, NR_WRITTEN);
2794 unlock_page_memcg(page);
2798 int __test_set_page_writeback(struct page *page, bool keep_write)
2800 struct address_space *mapping = page_mapping(page);
2803 lock_page_memcg(page);
2805 struct inode *inode = mapping->host;
2806 struct backing_dev_info *bdi = inode_to_bdi(inode);
2807 unsigned long flags;
2809 spin_lock_irqsave(&mapping->tree_lock, flags);
2810 ret = TestSetPageWriteback(page);
2814 on_wblist = mapping_tagged(mapping,
2815 PAGECACHE_TAG_WRITEBACK);
2817 radix_tree_tag_set(&mapping->page_tree,
2819 PAGECACHE_TAG_WRITEBACK);
2820 if (bdi_cap_account_writeback(bdi))
2821 __inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2824 * We can come through here when swapping anonymous
2825 * pages, so we don't necessarily have an inode to track
2828 if (mapping->host && !on_wblist)
2829 sb_mark_inode_writeback(mapping->host);
2831 if (!PageDirty(page))
2832 radix_tree_tag_clear(&mapping->page_tree,
2834 PAGECACHE_TAG_DIRTY);
2836 radix_tree_tag_clear(&mapping->page_tree,
2838 PAGECACHE_TAG_TOWRITE);
2839 spin_unlock_irqrestore(&mapping->tree_lock, flags);
2841 ret = TestSetPageWriteback(page);
2844 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
2845 inc_node_page_state(page, NR_WRITEBACK);
2847 unlock_page_memcg(page);
2851 EXPORT_SYMBOL(__test_set_page_writeback);
2854 * Return true if any of the pages in the mapping are marked with the
2857 int mapping_tagged(struct address_space *mapping, int tag)
2859 return radix_tree_tagged(&mapping->page_tree, tag);
2861 EXPORT_SYMBOL(mapping_tagged);
2864 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2865 * @page: The page to wait on.
2867 * This function determines if the given page is related to a backing device
2868 * that requires page contents to be held stable during writeback. If so, then
2869 * it will wait for any pending writeback to complete.
2871 void wait_for_stable_page(struct page *page)
2873 if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2874 wait_on_page_writeback(page);
2876 EXPORT_SYMBOL_GPL(wait_for_stable_page);