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/sched/signal.h>
40 #include <linux/mm_inline.h>
41 #include <trace/events/writeback.h>
46 * Sleep at most 200ms at a time in balance_dirty_pages().
48 #define MAX_PAUSE max(HZ/5, 1)
51 * Try to keep balance_dirty_pages() call intervals higher than this many pages
52 * by raising pause time to max_pause when falls below it.
54 #define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
57 * Estimate write bandwidth at 200ms intervals.
59 #define BANDWIDTH_INTERVAL max(HZ/5, 1)
61 #define RATELIMIT_CALC_SHIFT 10
64 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
65 * will look to see if it needs to force writeback or throttling.
67 static long ratelimit_pages
= 32;
69 /* The following parameters are exported via /proc/sys/vm */
72 * Start background writeback (via writeback threads) at this percentage
74 int dirty_background_ratio
= 10;
77 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
78 * dirty_background_ratio * the amount of dirtyable memory
80 unsigned long dirty_background_bytes
;
83 * free highmem will not be subtracted from the total free memory
84 * for calculating free ratios if vm_highmem_is_dirtyable is true
86 int vm_highmem_is_dirtyable
;
89 * The generator of dirty data starts writeback at this percentage
91 int vm_dirty_ratio
= 20;
94 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
95 * vm_dirty_ratio * the amount of dirtyable memory
97 unsigned long vm_dirty_bytes
;
100 * The interval between `kupdate'-style writebacks
102 unsigned int dirty_writeback_interval
= 5 * 100; /* centiseconds */
104 EXPORT_SYMBOL_GPL(dirty_writeback_interval
);
107 * The longest time for which data is allowed to remain dirty
109 unsigned int dirty_expire_interval
= 30 * 100; /* centiseconds */
112 * Flag that makes the machine dump writes/reads and block dirtyings.
117 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
118 * a full sync is triggered after this time elapses without any disk activity.
122 EXPORT_SYMBOL(laptop_mode
);
124 /* End of sysctl-exported parameters */
126 struct wb_domain global_wb_domain
;
128 /* consolidated parameters for balance_dirty_pages() and its subroutines */
129 struct dirty_throttle_control
{
130 #ifdef CONFIG_CGROUP_WRITEBACK
131 struct wb_domain
*dom
;
132 struct dirty_throttle_control
*gdtc
; /* only set in memcg dtc's */
134 struct bdi_writeback
*wb
;
135 struct fprop_local_percpu
*wb_completions
;
137 unsigned long avail
; /* dirtyable */
138 unsigned long dirty
; /* file_dirty + write + nfs */
139 unsigned long thresh
; /* dirty threshold */
140 unsigned long bg_thresh
; /* dirty background threshold */
142 unsigned long wb_dirty
; /* per-wb counterparts */
143 unsigned long wb_thresh
;
144 unsigned long wb_bg_thresh
;
146 unsigned long pos_ratio
;
150 * Length of period for aging writeout fractions of bdis. This is an
151 * arbitrarily chosen number. The longer the period, the slower fractions will
152 * reflect changes in current writeout rate.
154 #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
156 #ifdef CONFIG_CGROUP_WRITEBACK
158 #define GDTC_INIT(__wb) .wb = (__wb), \
159 .dom = &global_wb_domain, \
160 .wb_completions = &(__wb)->completions
162 #define GDTC_INIT_NO_WB .dom = &global_wb_domain
164 #define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
165 .dom = mem_cgroup_wb_domain(__wb), \
166 .wb_completions = &(__wb)->memcg_completions, \
169 static bool mdtc_valid(struct dirty_throttle_control
*dtc
)
174 static struct wb_domain
*dtc_dom(struct dirty_throttle_control
*dtc
)
179 static struct dirty_throttle_control
*mdtc_gdtc(struct dirty_throttle_control
*mdtc
)
184 static struct fprop_local_percpu
*wb_memcg_completions(struct bdi_writeback
*wb
)
186 return &wb
->memcg_completions
;
189 static void wb_min_max_ratio(struct bdi_writeback
*wb
,
190 unsigned long *minp
, unsigned long *maxp
)
192 unsigned long this_bw
= wb
->avg_write_bandwidth
;
193 unsigned long tot_bw
= atomic_long_read(&wb
->bdi
->tot_write_bandwidth
);
194 unsigned long long min
= wb
->bdi
->min_ratio
;
195 unsigned long long max
= wb
->bdi
->max_ratio
;
198 * @wb may already be clean by the time control reaches here and
199 * the total may not include its bw.
201 if (this_bw
< tot_bw
) {
216 #else /* CONFIG_CGROUP_WRITEBACK */
218 #define GDTC_INIT(__wb) .wb = (__wb), \
219 .wb_completions = &(__wb)->completions
220 #define GDTC_INIT_NO_WB
221 #define MDTC_INIT(__wb, __gdtc)
223 static bool mdtc_valid(struct dirty_throttle_control
*dtc
)
228 static struct wb_domain
*dtc_dom(struct dirty_throttle_control
*dtc
)
230 return &global_wb_domain
;
233 static struct dirty_throttle_control
*mdtc_gdtc(struct dirty_throttle_control
*mdtc
)
238 static struct fprop_local_percpu
*wb_memcg_completions(struct bdi_writeback
*wb
)
243 static void wb_min_max_ratio(struct bdi_writeback
*wb
,
244 unsigned long *minp
, unsigned long *maxp
)
246 *minp
= wb
->bdi
->min_ratio
;
247 *maxp
= wb
->bdi
->max_ratio
;
250 #endif /* CONFIG_CGROUP_WRITEBACK */
253 * In a memory zone, there is a certain amount of pages we consider
254 * available for the page cache, which is essentially the number of
255 * free and reclaimable pages, minus some zone reserves to protect
256 * lowmem and the ability to uphold the zone's watermarks without
257 * requiring writeback.
259 * This number of dirtyable pages is the base value of which the
260 * user-configurable dirty ratio is the effictive number of pages that
261 * are allowed to be actually dirtied. Per individual zone, or
262 * globally by using the sum of dirtyable pages over all zones.
264 * Because the user is allowed to specify the dirty limit globally as
265 * absolute number of bytes, calculating the per-zone dirty limit can
266 * require translating the configured limit into a percentage of
267 * global dirtyable memory first.
271 * node_dirtyable_memory - number of dirtyable pages in a node
274 * Returns the node's number of pages potentially available for dirty
275 * page cache. This is the base value for the per-node dirty limits.
277 static unsigned long node_dirtyable_memory(struct pglist_data
*pgdat
)
279 unsigned long nr_pages
= 0;
282 for (z
= 0; z
< MAX_NR_ZONES
; z
++) {
283 struct zone
*zone
= pgdat
->node_zones
+ z
;
285 if (!populated_zone(zone
))
288 nr_pages
+= zone_page_state(zone
, NR_FREE_PAGES
);
292 * Pages reserved for the kernel should not be considered
293 * dirtyable, to prevent a situation where reclaim has to
294 * clean pages in order to balance the zones.
296 nr_pages
-= min(nr_pages
, pgdat
->totalreserve_pages
);
298 nr_pages
+= node_page_state(pgdat
, NR_INACTIVE_FILE
);
299 nr_pages
+= node_page_state(pgdat
, NR_ACTIVE_FILE
);
304 static unsigned long highmem_dirtyable_memory(unsigned long total
)
306 #ifdef CONFIG_HIGHMEM
311 for_each_node_state(node
, N_HIGH_MEMORY
) {
312 for (i
= ZONE_NORMAL
+ 1; i
< MAX_NR_ZONES
; i
++) {
314 unsigned long nr_pages
;
316 if (!is_highmem_idx(i
))
319 z
= &NODE_DATA(node
)->node_zones
[i
];
320 if (!populated_zone(z
))
323 nr_pages
= zone_page_state(z
, NR_FREE_PAGES
);
324 /* watch for underflows */
325 nr_pages
-= min(nr_pages
, high_wmark_pages(z
));
326 nr_pages
+= zone_page_state(z
, NR_ZONE_INACTIVE_FILE
);
327 nr_pages
+= zone_page_state(z
, NR_ZONE_ACTIVE_FILE
);
333 * Unreclaimable memory (kernel memory or anonymous memory
334 * without swap) can bring down the dirtyable pages below
335 * the zone's dirty balance reserve and the above calculation
336 * will underflow. However we still want to add in nodes
337 * which are below threshold (negative values) to get a more
338 * accurate calculation but make sure that the total never
345 * Make sure that the number of highmem pages is never larger
346 * than the number of the total dirtyable memory. This can only
347 * occur in very strange VM situations but we want to make sure
348 * that this does not occur.
350 return min(x
, total
);
357 * global_dirtyable_memory - number of globally dirtyable pages
359 * Returns the global number of pages potentially available for dirty
360 * page cache. This is the base value for the global dirty limits.
362 static unsigned long global_dirtyable_memory(void)
366 x
= global_zone_page_state(NR_FREE_PAGES
);
368 * Pages reserved for the kernel should not be considered
369 * dirtyable, to prevent a situation where reclaim has to
370 * clean pages in order to balance the zones.
372 x
-= min(x
, totalreserve_pages
);
374 x
+= global_node_page_state(NR_INACTIVE_FILE
);
375 x
+= global_node_page_state(NR_ACTIVE_FILE
);
377 if (!vm_highmem_is_dirtyable
)
378 x
-= highmem_dirtyable_memory(x
);
380 return x
+ 1; /* Ensure that we never return 0 */
384 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
385 * @dtc: dirty_throttle_control of interest
387 * Calculate @dtc->thresh and ->bg_thresh considering
388 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
389 * must ensure that @dtc->avail is set before calling this function. The
390 * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
393 static void domain_dirty_limits(struct dirty_throttle_control
*dtc
)
395 const unsigned long available_memory
= dtc
->avail
;
396 struct dirty_throttle_control
*gdtc
= mdtc_gdtc(dtc
);
397 unsigned long bytes
= vm_dirty_bytes
;
398 unsigned long bg_bytes
= dirty_background_bytes
;
399 /* convert ratios to per-PAGE_SIZE for higher precision */
400 unsigned long ratio
= (vm_dirty_ratio
* PAGE_SIZE
) / 100;
401 unsigned long bg_ratio
= (dirty_background_ratio
* PAGE_SIZE
) / 100;
402 unsigned long thresh
;
403 unsigned long bg_thresh
;
404 struct task_struct
*tsk
;
406 /* gdtc is !NULL iff @dtc is for memcg domain */
408 unsigned long global_avail
= gdtc
->avail
;
411 * The byte settings can't be applied directly to memcg
412 * domains. Convert them to ratios by scaling against
413 * globally available memory. As the ratios are in
414 * per-PAGE_SIZE, they can be obtained by dividing bytes by
418 ratio
= min(DIV_ROUND_UP(bytes
, global_avail
),
421 bg_ratio
= min(DIV_ROUND_UP(bg_bytes
, global_avail
),
423 bytes
= bg_bytes
= 0;
427 thresh
= DIV_ROUND_UP(bytes
, PAGE_SIZE
);
429 thresh
= (ratio
* available_memory
) / PAGE_SIZE
;
432 bg_thresh
= DIV_ROUND_UP(bg_bytes
, PAGE_SIZE
);
434 bg_thresh
= (bg_ratio
* available_memory
) / PAGE_SIZE
;
436 if (unlikely(bg_thresh
>= thresh
)) {
437 pr_warn("vm direct limit must be set greater than background limit.\n");
438 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(struct timer_list
*t
)
633 struct wb_domain
*dom
= from_timer(dom
, t
, period_timer
);
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 timer_setup(&dom
->period_timer
, writeout_period
, TIMER_DEFERRABLE
);
658 dom
->dirty_limit_tstamp
= jiffies
;
660 return fprop_global_init(&dom
->completions
, gfp
);
663 #ifdef CONFIG_CGROUP_WRITEBACK
664 void wb_domain_exit(struct wb_domain
*dom
)
666 del_timer_sync(&dom
->period_timer
);
667 fprop_global_destroy(&dom
->completions
);
672 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
673 * registered backing devices, which, for obvious reasons, can not
676 static unsigned int bdi_min_ratio
;
678 int bdi_set_min_ratio(struct backing_dev_info
*bdi
, unsigned int min_ratio
)
682 spin_lock_bh(&bdi_lock
);
683 if (min_ratio
> bdi
->max_ratio
) {
686 min_ratio
-= bdi
->min_ratio
;
687 if (bdi_min_ratio
+ min_ratio
< 100) {
688 bdi_min_ratio
+= min_ratio
;
689 bdi
->min_ratio
+= min_ratio
;
694 spin_unlock_bh(&bdi_lock
);
699 int bdi_set_max_ratio(struct backing_dev_info
*bdi
, unsigned max_ratio
)
706 spin_lock_bh(&bdi_lock
);
707 if (bdi
->min_ratio
> max_ratio
) {
710 bdi
->max_ratio
= max_ratio
;
711 bdi
->max_prop_frac
= (FPROP_FRAC_BASE
* max_ratio
) / 100;
713 spin_unlock_bh(&bdi_lock
);
717 EXPORT_SYMBOL(bdi_set_max_ratio
);
719 static unsigned long dirty_freerun_ceiling(unsigned long thresh
,
720 unsigned long bg_thresh
)
722 return (thresh
+ bg_thresh
) / 2;
725 static unsigned long hard_dirty_limit(struct wb_domain
*dom
,
726 unsigned long thresh
)
728 return max(thresh
, dom
->dirty_limit
);
732 * Memory which can be further allocated to a memcg domain is capped by
733 * system-wide clean memory excluding the amount being used in the domain.
735 static void mdtc_calc_avail(struct dirty_throttle_control
*mdtc
,
736 unsigned long filepages
, unsigned long headroom
)
738 struct dirty_throttle_control
*gdtc
= mdtc_gdtc(mdtc
);
739 unsigned long clean
= filepages
- min(filepages
, mdtc
->dirty
);
740 unsigned long global_clean
= gdtc
->avail
- min(gdtc
->avail
, gdtc
->dirty
);
741 unsigned long other_clean
= global_clean
- min(global_clean
, clean
);
743 mdtc
->avail
= filepages
+ min(headroom
, other_clean
);
747 * __wb_calc_thresh - @wb's share of dirty throttling threshold
748 * @dtc: dirty_throttle_context of interest
750 * Returns @wb's dirty limit in pages. The term "dirty" in the context of
751 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
753 * Note that balance_dirty_pages() will only seriously take it as a hard limit
754 * when sleeping max_pause per page is not enough to keep the dirty pages under
755 * control. For example, when the device is completely stalled due to some error
756 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
757 * In the other normal situations, it acts more gently by throttling the tasks
758 * more (rather than completely block them) when the wb dirty pages go high.
760 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
761 * - starving fast devices
762 * - piling up dirty pages (that will take long time to sync) on slow devices
764 * The wb's share of dirty limit will be adapting to its throughput and
765 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
767 static unsigned long __wb_calc_thresh(struct dirty_throttle_control
*dtc
)
769 struct wb_domain
*dom
= dtc_dom(dtc
);
770 unsigned long thresh
= dtc
->thresh
;
772 long numerator
, denominator
;
773 unsigned long wb_min_ratio
, wb_max_ratio
;
776 * Calculate this BDI's share of the thresh ratio.
778 fprop_fraction_percpu(&dom
->completions
, dtc
->wb_completions
,
779 &numerator
, &denominator
);
781 wb_thresh
= (thresh
* (100 - bdi_min_ratio
)) / 100;
782 wb_thresh
*= numerator
;
783 do_div(wb_thresh
, denominator
);
785 wb_min_max_ratio(dtc
->wb
, &wb_min_ratio
, &wb_max_ratio
);
787 wb_thresh
+= (thresh
* wb_min_ratio
) / 100;
788 if (wb_thresh
> (thresh
* wb_max_ratio
) / 100)
789 wb_thresh
= thresh
* wb_max_ratio
/ 100;
794 unsigned long wb_calc_thresh(struct bdi_writeback
*wb
, unsigned long thresh
)
796 struct dirty_throttle_control gdtc
= { GDTC_INIT(wb
),
798 return __wb_calc_thresh(&gdtc
);
803 * f(dirty) := 1.0 + (----------------)
806 * it's a 3rd order polynomial that subjects to
808 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
809 * (2) f(setpoint) = 1.0 => the balance point
810 * (3) f(limit) = 0 => the hard limit
811 * (4) df/dx <= 0 => negative feedback control
812 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
813 * => fast response on large errors; small oscillation near setpoint
815 static long long pos_ratio_polynom(unsigned long setpoint
,
822 x
= div64_s64(((s64
)setpoint
- (s64
)dirty
) << RATELIMIT_CALC_SHIFT
,
823 (limit
- setpoint
) | 1);
825 pos_ratio
= pos_ratio
* x
>> RATELIMIT_CALC_SHIFT
;
826 pos_ratio
= pos_ratio
* x
>> RATELIMIT_CALC_SHIFT
;
827 pos_ratio
+= 1 << RATELIMIT_CALC_SHIFT
;
829 return clamp(pos_ratio
, 0LL, 2LL << RATELIMIT_CALC_SHIFT
);
833 * Dirty position control.
835 * (o) global/bdi setpoints
837 * We want the dirty pages be balanced around the global/wb setpoints.
838 * When the number of dirty pages is higher/lower than the setpoint, the
839 * dirty position control ratio (and hence task dirty ratelimit) will be
840 * decreased/increased to bring the dirty pages back to the setpoint.
842 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
844 * if (dirty < setpoint) scale up pos_ratio
845 * if (dirty > setpoint) scale down pos_ratio
847 * if (wb_dirty < wb_setpoint) scale up pos_ratio
848 * if (wb_dirty > wb_setpoint) scale down pos_ratio
850 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
852 * (o) global control line
856 * | |<===== global dirty control scope ======>|
864 * 1.0 ................................*
870 * 0 +------------.------------------.----------------------*------------->
871 * freerun^ setpoint^ limit^ dirty pages
873 * (o) wb control line
881 * | * |<=========== span ============>|
882 * 1.0 .......................*
894 * 1/4 ...............................................* * * * * * * * * * * *
898 * 0 +----------------------.-------------------------------.------------->
899 * wb_setpoint^ x_intercept^
901 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
902 * be smoothly throttled down to normal if it starts high in situations like
903 * - start writing to a slow SD card and a fast disk at the same time. The SD
904 * card's wb_dirty may rush to many times higher than wb_setpoint.
905 * - the wb dirty thresh drops quickly due to change of JBOD workload
907 static void wb_position_ratio(struct dirty_throttle_control
*dtc
)
909 struct bdi_writeback
*wb
= dtc
->wb
;
910 unsigned long write_bw
= wb
->avg_write_bandwidth
;
911 unsigned long freerun
= dirty_freerun_ceiling(dtc
->thresh
, dtc
->bg_thresh
);
912 unsigned long limit
= hard_dirty_limit(dtc_dom(dtc
), dtc
->thresh
);
913 unsigned long wb_thresh
= dtc
->wb_thresh
;
914 unsigned long x_intercept
;
915 unsigned long setpoint
; /* dirty pages' target balance point */
916 unsigned long wb_setpoint
;
918 long long pos_ratio
; /* for scaling up/down the rate limit */
923 if (unlikely(dtc
->dirty
>= limit
))
929 * See comment for pos_ratio_polynom().
931 setpoint
= (freerun
+ limit
) / 2;
932 pos_ratio
= pos_ratio_polynom(setpoint
, dtc
->dirty
, limit
);
935 * The strictlimit feature is a tool preventing mistrusted filesystems
936 * from growing a large number of dirty pages before throttling. For
937 * such filesystems balance_dirty_pages always checks wb counters
938 * against wb limits. Even if global "nr_dirty" is under "freerun".
939 * This is especially important for fuse which sets bdi->max_ratio to
940 * 1% by default. Without strictlimit feature, fuse writeback may
941 * consume arbitrary amount of RAM because it is accounted in
942 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
944 * Here, in wb_position_ratio(), we calculate pos_ratio based on
945 * two values: wb_dirty and wb_thresh. Let's consider an example:
946 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
947 * limits are set by default to 10% and 20% (background and throttle).
948 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
949 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
950 * about ~6K pages (as the average of background and throttle wb
951 * limits). The 3rd order polynomial will provide positive feedback if
952 * wb_dirty is under wb_setpoint and vice versa.
954 * Note, that we cannot use global counters in these calculations
955 * because we want to throttle process writing to a strictlimit wb
956 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
957 * in the example above).
959 if (unlikely(wb
->bdi
->capabilities
& BDI_CAP_STRICTLIMIT
)) {
960 long long wb_pos_ratio
;
962 if (dtc
->wb_dirty
< 8) {
963 dtc
->pos_ratio
= min_t(long long, pos_ratio
* 2,
964 2 << RATELIMIT_CALC_SHIFT
);
968 if (dtc
->wb_dirty
>= wb_thresh
)
971 wb_setpoint
= dirty_freerun_ceiling(wb_thresh
,
974 if (wb_setpoint
== 0 || wb_setpoint
== wb_thresh
)
977 wb_pos_ratio
= pos_ratio_polynom(wb_setpoint
, dtc
->wb_dirty
,
981 * Typically, for strictlimit case, wb_setpoint << setpoint
982 * and pos_ratio >> wb_pos_ratio. In the other words global
983 * state ("dirty") is not limiting factor and we have to
984 * make decision based on wb counters. But there is an
985 * important case when global pos_ratio should get precedence:
986 * global limits are exceeded (e.g. due to activities on other
987 * wb's) while given strictlimit wb is below limit.
989 * "pos_ratio * wb_pos_ratio" would work for the case above,
990 * but it would look too non-natural for the case of all
991 * activity in the system coming from a single strictlimit wb
992 * with bdi->max_ratio == 100%.
994 * Note that min() below somewhat changes the dynamics of the
995 * control system. Normally, pos_ratio value can be well over 3
996 * (when globally we are at freerun and wb is well below wb
997 * setpoint). Now the maximum pos_ratio in the same situation
998 * is 2. We might want to tweak this if we observe the control
999 * system is too slow to adapt.
1001 dtc
->pos_ratio
= min(pos_ratio
, wb_pos_ratio
);
1006 * We have computed basic pos_ratio above based on global situation. If
1007 * the wb is over/under its share of dirty pages, we want to scale
1008 * pos_ratio further down/up. That is done by the following mechanism.
1014 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1016 * x_intercept - wb_dirty
1017 * := --------------------------
1018 * x_intercept - wb_setpoint
1020 * The main wb control line is a linear function that subjects to
1022 * (1) f(wb_setpoint) = 1.0
1023 * (2) k = - 1 / (8 * write_bw) (in single wb case)
1024 * or equally: x_intercept = wb_setpoint + 8 * write_bw
1026 * For single wb case, the dirty pages are observed to fluctuate
1027 * regularly within range
1028 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1029 * for various filesystems, where (2) can yield in a reasonable 12.5%
1030 * fluctuation range for pos_ratio.
1032 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1033 * own size, so move the slope over accordingly and choose a slope that
1034 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1036 if (unlikely(wb_thresh
> dtc
->thresh
))
1037 wb_thresh
= dtc
->thresh
;
1039 * It's very possible that wb_thresh is close to 0 not because the
1040 * device is slow, but that it has remained inactive for long time.
1041 * Honour such devices a reasonable good (hopefully IO efficient)
1042 * threshold, so that the occasional writes won't be blocked and active
1043 * writes can rampup the threshold quickly.
1045 wb_thresh
= max(wb_thresh
, (limit
- dtc
->dirty
) / 8);
1047 * scale global setpoint to wb's:
1048 * wb_setpoint = setpoint * wb_thresh / thresh
1050 x
= div_u64((u64
)wb_thresh
<< 16, dtc
->thresh
| 1);
1051 wb_setpoint
= setpoint
* (u64
)x
>> 16;
1053 * Use span=(8*write_bw) in single wb case as indicated by
1054 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1056 * wb_thresh thresh - wb_thresh
1057 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1060 span
= (dtc
->thresh
- wb_thresh
+ 8 * write_bw
) * (u64
)x
>> 16;
1061 x_intercept
= wb_setpoint
+ span
;
1063 if (dtc
->wb_dirty
< x_intercept
- span
/ 4) {
1064 pos_ratio
= div64_u64(pos_ratio
* (x_intercept
- dtc
->wb_dirty
),
1065 (x_intercept
- wb_setpoint
) | 1);
1070 * wb reserve area, safeguard against dirty pool underrun and disk idle
1071 * It may push the desired control point of global dirty pages higher
1074 x_intercept
= wb_thresh
/ 2;
1075 if (dtc
->wb_dirty
< x_intercept
) {
1076 if (dtc
->wb_dirty
> x_intercept
/ 8)
1077 pos_ratio
= div_u64(pos_ratio
* x_intercept
,
1083 dtc
->pos_ratio
= pos_ratio
;
1086 static void wb_update_write_bandwidth(struct bdi_writeback
*wb
,
1087 unsigned long elapsed
,
1088 unsigned long written
)
1090 const unsigned long period
= roundup_pow_of_two(3 * HZ
);
1091 unsigned long avg
= wb
->avg_write_bandwidth
;
1092 unsigned long old
= wb
->write_bandwidth
;
1096 * bw = written * HZ / elapsed
1098 * bw * elapsed + write_bandwidth * (period - elapsed)
1099 * write_bandwidth = ---------------------------------------------------
1102 * @written may have decreased due to account_page_redirty().
1103 * Avoid underflowing @bw calculation.
1105 bw
= written
- min(written
, wb
->written_stamp
);
1107 if (unlikely(elapsed
> period
)) {
1108 do_div(bw
, elapsed
);
1112 bw
+= (u64
)wb
->write_bandwidth
* (period
- elapsed
);
1113 bw
>>= ilog2(period
);
1116 * one more level of smoothing, for filtering out sudden spikes
1118 if (avg
> old
&& old
>= (unsigned long)bw
)
1119 avg
-= (avg
- old
) >> 3;
1121 if (avg
< old
&& old
<= (unsigned long)bw
)
1122 avg
+= (old
- avg
) >> 3;
1125 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1126 avg
= max(avg
, 1LU);
1127 if (wb_has_dirty_io(wb
)) {
1128 long delta
= avg
- wb
->avg_write_bandwidth
;
1129 WARN_ON_ONCE(atomic_long_add_return(delta
,
1130 &wb
->bdi
->tot_write_bandwidth
) <= 0);
1132 wb
->write_bandwidth
= bw
;
1133 wb
->avg_write_bandwidth
= avg
;
1136 static void update_dirty_limit(struct dirty_throttle_control
*dtc
)
1138 struct wb_domain
*dom
= dtc_dom(dtc
);
1139 unsigned long thresh
= dtc
->thresh
;
1140 unsigned long limit
= dom
->dirty_limit
;
1143 * Follow up in one step.
1145 if (limit
< thresh
) {
1151 * Follow down slowly. Use the higher one as the target, because thresh
1152 * may drop below dirty. This is exactly the reason to introduce
1153 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1155 thresh
= max(thresh
, dtc
->dirty
);
1156 if (limit
> thresh
) {
1157 limit
-= (limit
- thresh
) >> 5;
1162 dom
->dirty_limit
= limit
;
1165 static void domain_update_bandwidth(struct dirty_throttle_control
*dtc
,
1168 struct wb_domain
*dom
= dtc_dom(dtc
);
1171 * check locklessly first to optimize away locking for the most time
1173 if (time_before(now
, dom
->dirty_limit_tstamp
+ BANDWIDTH_INTERVAL
))
1176 spin_lock(&dom
->lock
);
1177 if (time_after_eq(now
, dom
->dirty_limit_tstamp
+ BANDWIDTH_INTERVAL
)) {
1178 update_dirty_limit(dtc
);
1179 dom
->dirty_limit_tstamp
= now
;
1181 spin_unlock(&dom
->lock
);
1185 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1187 * Normal wb tasks will be curbed at or below it in long term.
1188 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1190 static void wb_update_dirty_ratelimit(struct dirty_throttle_control
*dtc
,
1191 unsigned long dirtied
,
1192 unsigned long elapsed
)
1194 struct bdi_writeback
*wb
= dtc
->wb
;
1195 unsigned long dirty
= dtc
->dirty
;
1196 unsigned long freerun
= dirty_freerun_ceiling(dtc
->thresh
, dtc
->bg_thresh
);
1197 unsigned long limit
= hard_dirty_limit(dtc_dom(dtc
), dtc
->thresh
);
1198 unsigned long setpoint
= (freerun
+ limit
) / 2;
1199 unsigned long write_bw
= wb
->avg_write_bandwidth
;
1200 unsigned long dirty_ratelimit
= wb
->dirty_ratelimit
;
1201 unsigned long dirty_rate
;
1202 unsigned long task_ratelimit
;
1203 unsigned long balanced_dirty_ratelimit
;
1206 unsigned long shift
;
1209 * The dirty rate will match the writeout rate in long term, except
1210 * when dirty pages are truncated by userspace or re-dirtied by FS.
1212 dirty_rate
= (dirtied
- wb
->dirtied_stamp
) * HZ
/ elapsed
;
1215 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1217 task_ratelimit
= (u64
)dirty_ratelimit
*
1218 dtc
->pos_ratio
>> RATELIMIT_CALC_SHIFT
;
1219 task_ratelimit
++; /* it helps rampup dirty_ratelimit from tiny values */
1222 * A linear estimation of the "balanced" throttle rate. The theory is,
1223 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1224 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1225 * formula will yield the balanced rate limit (write_bw / N).
1227 * Note that the expanded form is not a pure rate feedback:
1228 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1229 * but also takes pos_ratio into account:
1230 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1232 * (1) is not realistic because pos_ratio also takes part in balancing
1233 * the dirty rate. Consider the state
1234 * pos_ratio = 0.5 (3)
1235 * rate = 2 * (write_bw / N) (4)
1236 * If (1) is used, it will stuck in that state! Because each dd will
1238 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1240 * dirty_rate = N * task_ratelimit = write_bw (6)
1241 * put (6) into (1) we get
1242 * rate_(i+1) = rate_(i) (7)
1244 * So we end up using (2) to always keep
1245 * rate_(i+1) ~= (write_bw / N) (8)
1246 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1247 * pos_ratio is able to drive itself to 1.0, which is not only where
1248 * the dirty count meet the setpoint, but also where the slope of
1249 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1251 balanced_dirty_ratelimit
= div_u64((u64
)task_ratelimit
* write_bw
,
1254 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1256 if (unlikely(balanced_dirty_ratelimit
> write_bw
))
1257 balanced_dirty_ratelimit
= write_bw
;
1260 * We could safely do this and return immediately:
1262 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1264 * However to get a more stable dirty_ratelimit, the below elaborated
1265 * code makes use of task_ratelimit to filter out singular points and
1266 * limit the step size.
1268 * The below code essentially only uses the relative value of
1270 * task_ratelimit - dirty_ratelimit
1271 * = (pos_ratio - 1) * dirty_ratelimit
1273 * which reflects the direction and size of dirty position error.
1277 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1278 * task_ratelimit is on the same side of dirty_ratelimit, too.
1280 * - dirty_ratelimit > balanced_dirty_ratelimit
1281 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1282 * lowering dirty_ratelimit will help meet both the position and rate
1283 * control targets. Otherwise, don't update dirty_ratelimit if it will
1284 * only help meet the rate target. After all, what the users ultimately
1285 * feel and care are stable dirty rate and small position error.
1287 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1288 * and filter out the singular points of balanced_dirty_ratelimit. Which
1289 * keeps jumping around randomly and can even leap far away at times
1290 * due to the small 200ms estimation period of dirty_rate (we want to
1291 * keep that period small to reduce time lags).
1296 * For strictlimit case, calculations above were based on wb counters
1297 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1298 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1299 * Hence, to calculate "step" properly, we have to use wb_dirty as
1300 * "dirty" and wb_setpoint as "setpoint".
1302 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1303 * it's possible that wb_thresh is close to zero due to inactivity
1304 * of backing device.
1306 if (unlikely(wb
->bdi
->capabilities
& BDI_CAP_STRICTLIMIT
)) {
1307 dirty
= dtc
->wb_dirty
;
1308 if (dtc
->wb_dirty
< 8)
1309 setpoint
= dtc
->wb_dirty
+ 1;
1311 setpoint
= (dtc
->wb_thresh
+ dtc
->wb_bg_thresh
) / 2;
1314 if (dirty
< setpoint
) {
1315 x
= min3(wb
->balanced_dirty_ratelimit
,
1316 balanced_dirty_ratelimit
, task_ratelimit
);
1317 if (dirty_ratelimit
< x
)
1318 step
= x
- dirty_ratelimit
;
1320 x
= max3(wb
->balanced_dirty_ratelimit
,
1321 balanced_dirty_ratelimit
, task_ratelimit
);
1322 if (dirty_ratelimit
> x
)
1323 step
= dirty_ratelimit
- x
;
1327 * Don't pursue 100% rate matching. It's impossible since the balanced
1328 * rate itself is constantly fluctuating. So decrease the track speed
1329 * when it gets close to the target. Helps eliminate pointless tremors.
1331 shift
= dirty_ratelimit
/ (2 * step
+ 1);
1332 if (shift
< BITS_PER_LONG
)
1333 step
= DIV_ROUND_UP(step
>> shift
, 8);
1337 if (dirty_ratelimit
< balanced_dirty_ratelimit
)
1338 dirty_ratelimit
+= step
;
1340 dirty_ratelimit
-= step
;
1342 wb
->dirty_ratelimit
= max(dirty_ratelimit
, 1UL);
1343 wb
->balanced_dirty_ratelimit
= balanced_dirty_ratelimit
;
1345 trace_bdi_dirty_ratelimit(wb
, dirty_rate
, task_ratelimit
);
1348 static void __wb_update_bandwidth(struct dirty_throttle_control
*gdtc
,
1349 struct dirty_throttle_control
*mdtc
,
1350 unsigned long start_time
,
1351 bool update_ratelimit
)
1353 struct bdi_writeback
*wb
= gdtc
->wb
;
1354 unsigned long now
= jiffies
;
1355 unsigned long elapsed
= now
- wb
->bw_time_stamp
;
1356 unsigned long dirtied
;
1357 unsigned long written
;
1359 lockdep_assert_held(&wb
->list_lock
);
1362 * rate-limit, only update once every 200ms.
1364 if (elapsed
< BANDWIDTH_INTERVAL
)
1367 dirtied
= percpu_counter_read(&wb
->stat
[WB_DIRTIED
]);
1368 written
= percpu_counter_read(&wb
->stat
[WB_WRITTEN
]);
1371 * Skip quiet periods when disk bandwidth is under-utilized.
1372 * (at least 1s idle time between two flusher runs)
1374 if (elapsed
> HZ
&& time_before(wb
->bw_time_stamp
, start_time
))
1377 if (update_ratelimit
) {
1378 domain_update_bandwidth(gdtc
, now
);
1379 wb_update_dirty_ratelimit(gdtc
, dirtied
, elapsed
);
1382 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1383 * compiler has no way to figure that out. Help it.
1385 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK
) && mdtc
) {
1386 domain_update_bandwidth(mdtc
, now
);
1387 wb_update_dirty_ratelimit(mdtc
, dirtied
, elapsed
);
1390 wb_update_write_bandwidth(wb
, elapsed
, written
);
1393 wb
->dirtied_stamp
= dirtied
;
1394 wb
->written_stamp
= written
;
1395 wb
->bw_time_stamp
= now
;
1398 void wb_update_bandwidth(struct bdi_writeback
*wb
, unsigned long start_time
)
1400 struct dirty_throttle_control gdtc
= { GDTC_INIT(wb
) };
1402 __wb_update_bandwidth(&gdtc
, NULL
, start_time
, false);
1406 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1407 * will look to see if it needs to start dirty throttling.
1409 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1410 * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1411 * (the number of pages we may dirty without exceeding the dirty limits).
1413 static unsigned long dirty_poll_interval(unsigned long dirty
,
1414 unsigned long thresh
)
1417 return 1UL << (ilog2(thresh
- dirty
) >> 1);
1422 static unsigned long wb_max_pause(struct bdi_writeback
*wb
,
1423 unsigned long wb_dirty
)
1425 unsigned long bw
= wb
->avg_write_bandwidth
;
1429 * Limit pause time for small memory systems. If sleeping for too long
1430 * time, a small pool of dirty/writeback pages may go empty and disk go
1433 * 8 serves as the safety ratio.
1435 t
= wb_dirty
/ (1 + bw
/ roundup_pow_of_two(1 + HZ
/ 8));
1438 return min_t(unsigned long, t
, MAX_PAUSE
);
1441 static long wb_min_pause(struct bdi_writeback
*wb
,
1443 unsigned long task_ratelimit
,
1444 unsigned long dirty_ratelimit
,
1445 int *nr_dirtied_pause
)
1447 long hi
= ilog2(wb
->avg_write_bandwidth
);
1448 long lo
= ilog2(wb
->dirty_ratelimit
);
1449 long t
; /* target pause */
1450 long pause
; /* estimated next pause */
1451 int pages
; /* target nr_dirtied_pause */
1453 /* target for 10ms pause on 1-dd case */
1454 t
= max(1, HZ
/ 100);
1457 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1460 * (N * 10ms) on 2^N concurrent tasks.
1463 t
+= (hi
- lo
) * (10 * HZ
) / 1024;
1466 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1467 * on the much more stable dirty_ratelimit. However the next pause time
1468 * will be computed based on task_ratelimit and the two rate limits may
1469 * depart considerably at some time. Especially if task_ratelimit goes
1470 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1471 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1472 * result task_ratelimit won't be executed faithfully, which could
1473 * eventually bring down dirty_ratelimit.
1475 * We apply two rules to fix it up:
1476 * 1) try to estimate the next pause time and if necessary, use a lower
1477 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1478 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1479 * 2) limit the target pause time to max_pause/2, so that the normal
1480 * small fluctuations of task_ratelimit won't trigger rule (1) and
1481 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1483 t
= min(t
, 1 + max_pause
/ 2);
1484 pages
= dirty_ratelimit
* t
/ roundup_pow_of_two(HZ
);
1487 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1488 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1489 * When the 16 consecutive reads are often interrupted by some dirty
1490 * throttling pause during the async writes, cfq will go into idles
1491 * (deadline is fine). So push nr_dirtied_pause as high as possible
1492 * until reaches DIRTY_POLL_THRESH=32 pages.
1494 if (pages
< DIRTY_POLL_THRESH
) {
1496 pages
= dirty_ratelimit
* t
/ roundup_pow_of_two(HZ
);
1497 if (pages
> DIRTY_POLL_THRESH
) {
1498 pages
= DIRTY_POLL_THRESH
;
1499 t
= HZ
* DIRTY_POLL_THRESH
/ dirty_ratelimit
;
1503 pause
= HZ
* pages
/ (task_ratelimit
+ 1);
1504 if (pause
> max_pause
) {
1506 pages
= task_ratelimit
* t
/ roundup_pow_of_two(HZ
);
1509 *nr_dirtied_pause
= pages
;
1511 * The minimal pause time will normally be half the target pause time.
1513 return pages
>= DIRTY_POLL_THRESH
? 1 + t
/ 2 : t
;
1516 static inline void wb_dirty_limits(struct dirty_throttle_control
*dtc
)
1518 struct bdi_writeback
*wb
= dtc
->wb
;
1519 unsigned long wb_reclaimable
;
1522 * wb_thresh is not treated as some limiting factor as
1523 * dirty_thresh, due to reasons
1524 * - in JBOD setup, wb_thresh can fluctuate a lot
1525 * - in a system with HDD and USB key, the USB key may somehow
1526 * go into state (wb_dirty >> wb_thresh) either because
1527 * wb_dirty starts high, or because wb_thresh drops low.
1528 * In this case we don't want to hard throttle the USB key
1529 * dirtiers for 100 seconds until wb_dirty drops under
1530 * wb_thresh. Instead the auxiliary wb control line in
1531 * wb_position_ratio() will let the dirtier task progress
1532 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1534 dtc
->wb_thresh
= __wb_calc_thresh(dtc
);
1535 dtc
->wb_bg_thresh
= dtc
->thresh
?
1536 div_u64((u64
)dtc
->wb_thresh
* dtc
->bg_thresh
, dtc
->thresh
) : 0;
1539 * In order to avoid the stacked BDI deadlock we need
1540 * to ensure we accurately count the 'dirty' pages when
1541 * the threshold is low.
1543 * Otherwise it would be possible to get thresh+n pages
1544 * reported dirty, even though there are thresh-m pages
1545 * actually dirty; with m+n sitting in the percpu
1548 if (dtc
->wb_thresh
< 2 * wb_stat_error()) {
1549 wb_reclaimable
= wb_stat_sum(wb
, WB_RECLAIMABLE
);
1550 dtc
->wb_dirty
= wb_reclaimable
+ wb_stat_sum(wb
, WB_WRITEBACK
);
1552 wb_reclaimable
= wb_stat(wb
, WB_RECLAIMABLE
);
1553 dtc
->wb_dirty
= wb_reclaimable
+ wb_stat(wb
, WB_WRITEBACK
);
1558 * balance_dirty_pages() must be called by processes which are generating dirty
1559 * data. It looks at the number of dirty pages in the machine and will force
1560 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1561 * If we're over `background_thresh' then the writeback threads are woken to
1562 * perform some writeout.
1564 static void balance_dirty_pages(struct bdi_writeback
*wb
,
1565 unsigned long pages_dirtied
)
1567 struct dirty_throttle_control gdtc_stor
= { GDTC_INIT(wb
) };
1568 struct dirty_throttle_control mdtc_stor
= { MDTC_INIT(wb
, &gdtc_stor
) };
1569 struct dirty_throttle_control
* const gdtc
= &gdtc_stor
;
1570 struct dirty_throttle_control
* const mdtc
= mdtc_valid(&mdtc_stor
) ?
1572 struct dirty_throttle_control
*sdtc
;
1573 unsigned long nr_reclaimable
; /* = file_dirty + unstable_nfs */
1578 int nr_dirtied_pause
;
1579 bool dirty_exceeded
= false;
1580 unsigned long task_ratelimit
;
1581 unsigned long dirty_ratelimit
;
1582 struct backing_dev_info
*bdi
= wb
->bdi
;
1583 bool strictlimit
= bdi
->capabilities
& BDI_CAP_STRICTLIMIT
;
1584 unsigned long start_time
= jiffies
;
1587 unsigned long now
= jiffies
;
1588 unsigned long dirty
, thresh
, bg_thresh
;
1589 unsigned long m_dirty
= 0; /* stop bogus uninit warnings */
1590 unsigned long m_thresh
= 0;
1591 unsigned long m_bg_thresh
= 0;
1594 * Unstable writes are a feature of certain networked
1595 * filesystems (i.e. NFS) in which data may have been
1596 * written to the server's write cache, but has not yet
1597 * been flushed to permanent storage.
1599 nr_reclaimable
= global_node_page_state(NR_FILE_DIRTY
) +
1600 global_node_page_state(NR_UNSTABLE_NFS
);
1601 gdtc
->avail
= global_dirtyable_memory();
1602 gdtc
->dirty
= nr_reclaimable
+ global_node_page_state(NR_WRITEBACK
);
1604 domain_dirty_limits(gdtc
);
1606 if (unlikely(strictlimit
)) {
1607 wb_dirty_limits(gdtc
);
1609 dirty
= gdtc
->wb_dirty
;
1610 thresh
= gdtc
->wb_thresh
;
1611 bg_thresh
= gdtc
->wb_bg_thresh
;
1613 dirty
= gdtc
->dirty
;
1614 thresh
= gdtc
->thresh
;
1615 bg_thresh
= gdtc
->bg_thresh
;
1619 unsigned long filepages
, headroom
, writeback
;
1622 * If @wb belongs to !root memcg, repeat the same
1623 * basic calculations for the memcg domain.
1625 mem_cgroup_wb_stats(wb
, &filepages
, &headroom
,
1626 &mdtc
->dirty
, &writeback
);
1627 mdtc
->dirty
+= writeback
;
1628 mdtc_calc_avail(mdtc
, filepages
, headroom
);
1630 domain_dirty_limits(mdtc
);
1632 if (unlikely(strictlimit
)) {
1633 wb_dirty_limits(mdtc
);
1634 m_dirty
= mdtc
->wb_dirty
;
1635 m_thresh
= mdtc
->wb_thresh
;
1636 m_bg_thresh
= mdtc
->wb_bg_thresh
;
1638 m_dirty
= mdtc
->dirty
;
1639 m_thresh
= mdtc
->thresh
;
1640 m_bg_thresh
= mdtc
->bg_thresh
;
1645 * Throttle it only when the background writeback cannot
1646 * catch-up. This avoids (excessively) small writeouts
1647 * when the wb limits are ramping up in case of !strictlimit.
1649 * In strictlimit case make decision based on the wb counters
1650 * and limits. Small writeouts when the wb limits are ramping
1651 * up are the price we consciously pay for strictlimit-ing.
1653 * If memcg domain is in effect, @dirty should be under
1654 * both global and memcg freerun ceilings.
1656 if (dirty
<= dirty_freerun_ceiling(thresh
, bg_thresh
) &&
1658 m_dirty
<= dirty_freerun_ceiling(m_thresh
, m_bg_thresh
))) {
1659 unsigned long intv
= dirty_poll_interval(dirty
, thresh
);
1660 unsigned long m_intv
= ULONG_MAX
;
1662 current
->dirty_paused_when
= now
;
1663 current
->nr_dirtied
= 0;
1665 m_intv
= dirty_poll_interval(m_dirty
, m_thresh
);
1666 current
->nr_dirtied_pause
= min(intv
, m_intv
);
1670 if (unlikely(!writeback_in_progress(wb
)))
1671 wb_start_background_writeback(wb
);
1674 * Calculate global domain's pos_ratio and select the
1675 * global dtc by default.
1678 wb_dirty_limits(gdtc
);
1680 dirty_exceeded
= (gdtc
->wb_dirty
> gdtc
->wb_thresh
) &&
1681 ((gdtc
->dirty
> gdtc
->thresh
) || strictlimit
);
1683 wb_position_ratio(gdtc
);
1688 * If memcg domain is in effect, calculate its
1689 * pos_ratio. @wb should satisfy constraints from
1690 * both global and memcg domains. Choose the one
1691 * w/ lower pos_ratio.
1694 wb_dirty_limits(mdtc
);
1696 dirty_exceeded
|= (mdtc
->wb_dirty
> mdtc
->wb_thresh
) &&
1697 ((mdtc
->dirty
> mdtc
->thresh
) || strictlimit
);
1699 wb_position_ratio(mdtc
);
1700 if (mdtc
->pos_ratio
< gdtc
->pos_ratio
)
1704 if (dirty_exceeded
&& !wb
->dirty_exceeded
)
1705 wb
->dirty_exceeded
= 1;
1707 if (time_is_before_jiffies(wb
->bw_time_stamp
+
1708 BANDWIDTH_INTERVAL
)) {
1709 spin_lock(&wb
->list_lock
);
1710 __wb_update_bandwidth(gdtc
, mdtc
, start_time
, true);
1711 spin_unlock(&wb
->list_lock
);
1714 /* throttle according to the chosen dtc */
1715 dirty_ratelimit
= wb
->dirty_ratelimit
;
1716 task_ratelimit
= ((u64
)dirty_ratelimit
* sdtc
->pos_ratio
) >>
1717 RATELIMIT_CALC_SHIFT
;
1718 max_pause
= wb_max_pause(wb
, sdtc
->wb_dirty
);
1719 min_pause
= wb_min_pause(wb
, max_pause
,
1720 task_ratelimit
, dirty_ratelimit
,
1723 if (unlikely(task_ratelimit
== 0)) {
1728 period
= HZ
* pages_dirtied
/ task_ratelimit
;
1730 if (current
->dirty_paused_when
)
1731 pause
-= now
- current
->dirty_paused_when
;
1733 * For less than 1s think time (ext3/4 may block the dirtier
1734 * for up to 800ms from time to time on 1-HDD; so does xfs,
1735 * however at much less frequency), try to compensate it in
1736 * future periods by updating the virtual time; otherwise just
1737 * do a reset, as it may be a light dirtier.
1739 if (pause
< min_pause
) {
1740 trace_balance_dirty_pages(wb
,
1753 current
->dirty_paused_when
= now
;
1754 current
->nr_dirtied
= 0;
1755 } else if (period
) {
1756 current
->dirty_paused_when
+= period
;
1757 current
->nr_dirtied
= 0;
1758 } else if (current
->nr_dirtied_pause
<= pages_dirtied
)
1759 current
->nr_dirtied_pause
+= pages_dirtied
;
1762 if (unlikely(pause
> max_pause
)) {
1763 /* for occasional dropped task_ratelimit */
1764 now
+= min(pause
- max_pause
, max_pause
);
1769 trace_balance_dirty_pages(wb
,
1781 __set_current_state(TASK_KILLABLE
);
1782 wb
->dirty_sleep
= now
;
1783 io_schedule_timeout(pause
);
1785 current
->dirty_paused_when
= now
+ pause
;
1786 current
->nr_dirtied
= 0;
1787 current
->nr_dirtied_pause
= nr_dirtied_pause
;
1790 * This is typically equal to (dirty < thresh) and can also
1791 * keep "1000+ dd on a slow USB stick" under control.
1797 * In the case of an unresponding NFS server and the NFS dirty
1798 * pages exceeds dirty_thresh, give the other good wb's a pipe
1799 * to go through, so that tasks on them still remain responsive.
1801 * In theory 1 page is enough to keep the consumer-producer
1802 * pipe going: the flusher cleans 1 page => the task dirties 1
1803 * more page. However wb_dirty has accounting errors. So use
1804 * the larger and more IO friendly wb_stat_error.
1806 if (sdtc
->wb_dirty
<= wb_stat_error())
1809 if (fatal_signal_pending(current
))
1813 if (!dirty_exceeded
&& wb
->dirty_exceeded
)
1814 wb
->dirty_exceeded
= 0;
1816 if (writeback_in_progress(wb
))
1820 * In laptop mode, we wait until hitting the higher threshold before
1821 * starting background writeout, and then write out all the way down
1822 * to the lower threshold. So slow writers cause minimal disk activity.
1824 * In normal mode, we start background writeout at the lower
1825 * background_thresh, to keep the amount of dirty memory low.
1830 if (nr_reclaimable
> gdtc
->bg_thresh
)
1831 wb_start_background_writeback(wb
);
1834 static DEFINE_PER_CPU(int, bdp_ratelimits
);
1837 * Normal tasks are throttled by
1839 * dirty tsk->nr_dirtied_pause pages;
1840 * take a snap in balance_dirty_pages();
1842 * However there is a worst case. If every task exit immediately when dirtied
1843 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1844 * called to throttle the page dirties. The solution is to save the not yet
1845 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1846 * randomly into the running tasks. This works well for the above worst case,
1847 * as the new task will pick up and accumulate the old task's leaked dirty
1848 * count and eventually get throttled.
1850 DEFINE_PER_CPU(int, dirty_throttle_leaks
) = 0;
1853 * balance_dirty_pages_ratelimited - balance dirty memory state
1854 * @mapping: address_space which was dirtied
1856 * Processes which are dirtying memory should call in here once for each page
1857 * which was newly dirtied. The function will periodically check the system's
1858 * dirty state and will initiate writeback if needed.
1860 * On really big machines, get_writeback_state is expensive, so try to avoid
1861 * calling it too often (ratelimiting). But once we're over the dirty memory
1862 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1863 * from overshooting the limit by (ratelimit_pages) each.
1865 void balance_dirty_pages_ratelimited(struct address_space
*mapping
)
1867 struct inode
*inode
= mapping
->host
;
1868 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
1869 struct bdi_writeback
*wb
= NULL
;
1873 if (!bdi_cap_account_dirty(bdi
))
1876 if (inode_cgwb_enabled(inode
))
1877 wb
= wb_get_create_current(bdi
, GFP_KERNEL
);
1881 ratelimit
= current
->nr_dirtied_pause
;
1882 if (wb
->dirty_exceeded
)
1883 ratelimit
= min(ratelimit
, 32 >> (PAGE_SHIFT
- 10));
1887 * This prevents one CPU to accumulate too many dirtied pages without
1888 * calling into balance_dirty_pages(), which can happen when there are
1889 * 1000+ tasks, all of them start dirtying pages at exactly the same
1890 * time, hence all honoured too large initial task->nr_dirtied_pause.
1892 p
= this_cpu_ptr(&bdp_ratelimits
);
1893 if (unlikely(current
->nr_dirtied
>= ratelimit
))
1895 else if (unlikely(*p
>= ratelimit_pages
)) {
1900 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1901 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1902 * the dirty throttling and livelock other long-run dirtiers.
1904 p
= this_cpu_ptr(&dirty_throttle_leaks
);
1905 if (*p
> 0 && current
->nr_dirtied
< ratelimit
) {
1906 unsigned long nr_pages_dirtied
;
1907 nr_pages_dirtied
= min(*p
, ratelimit
- current
->nr_dirtied
);
1908 *p
-= nr_pages_dirtied
;
1909 current
->nr_dirtied
+= nr_pages_dirtied
;
1913 if (unlikely(current
->nr_dirtied
>= ratelimit
))
1914 balance_dirty_pages(wb
, current
->nr_dirtied
);
1918 EXPORT_SYMBOL(balance_dirty_pages_ratelimited
);
1921 * wb_over_bg_thresh - does @wb need to be written back?
1922 * @wb: bdi_writeback of interest
1924 * Determines whether background writeback should keep writing @wb or it's
1925 * clean enough. Returns %true if writeback should continue.
1927 bool wb_over_bg_thresh(struct bdi_writeback
*wb
)
1929 struct dirty_throttle_control gdtc_stor
= { GDTC_INIT(wb
) };
1930 struct dirty_throttle_control mdtc_stor
= { MDTC_INIT(wb
, &gdtc_stor
) };
1931 struct dirty_throttle_control
* const gdtc
= &gdtc_stor
;
1932 struct dirty_throttle_control
* const mdtc
= mdtc_valid(&mdtc_stor
) ?
1936 * Similar to balance_dirty_pages() but ignores pages being written
1937 * as we're trying to decide whether to put more under writeback.
1939 gdtc
->avail
= global_dirtyable_memory();
1940 gdtc
->dirty
= global_node_page_state(NR_FILE_DIRTY
) +
1941 global_node_page_state(NR_UNSTABLE_NFS
);
1942 domain_dirty_limits(gdtc
);
1944 if (gdtc
->dirty
> gdtc
->bg_thresh
)
1947 if (wb_stat(wb
, WB_RECLAIMABLE
) >
1948 wb_calc_thresh(gdtc
->wb
, gdtc
->bg_thresh
))
1952 unsigned long filepages
, headroom
, writeback
;
1954 mem_cgroup_wb_stats(wb
, &filepages
, &headroom
, &mdtc
->dirty
,
1956 mdtc_calc_avail(mdtc
, filepages
, headroom
);
1957 domain_dirty_limits(mdtc
); /* ditto, ignore writeback */
1959 if (mdtc
->dirty
> mdtc
->bg_thresh
)
1962 if (wb_stat(wb
, WB_RECLAIMABLE
) >
1963 wb_calc_thresh(mdtc
->wb
, mdtc
->bg_thresh
))
1971 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1973 int dirty_writeback_centisecs_handler(struct ctl_table
*table
, int write
,
1974 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1976 unsigned int old_interval
= dirty_writeback_interval
;
1979 ret
= proc_dointvec(table
, write
, buffer
, length
, ppos
);
1982 * Writing 0 to dirty_writeback_interval will disable periodic writeback
1983 * and a different non-zero value will wakeup the writeback threads.
1984 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
1985 * iterate over all bdis and wbs.
1986 * The reason we do this is to make the change take effect immediately.
1988 if (!ret
&& write
&& dirty_writeback_interval
&&
1989 dirty_writeback_interval
!= old_interval
)
1990 wakeup_flusher_threads(WB_REASON_PERIODIC
);
1996 void laptop_mode_timer_fn(unsigned long data
)
1998 struct request_queue
*q
= (struct request_queue
*)data
;
2000 wakeup_flusher_threads_bdi(q
->backing_dev_info
, WB_REASON_LAPTOP_TIMER
);
2004 * We've spun up the disk and we're in laptop mode: schedule writeback
2005 * of all dirty data a few seconds from now. If the flush is already scheduled
2006 * then push it back - the user is still using the disk.
2008 void laptop_io_completion(struct backing_dev_info
*info
)
2010 mod_timer(&info
->laptop_mode_wb_timer
, jiffies
+ laptop_mode
);
2014 * We're in laptop mode and we've just synced. The sync's writes will have
2015 * caused another writeback to be scheduled by laptop_io_completion.
2016 * Nothing needs to be written back anymore, so we unschedule the writeback.
2018 void laptop_sync_completion(void)
2020 struct backing_dev_info
*bdi
;
2024 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
)
2025 del_timer(&bdi
->laptop_mode_wb_timer
);
2032 * If ratelimit_pages is too high then we can get into dirty-data overload
2033 * if a large number of processes all perform writes at the same time.
2034 * If it is too low then SMP machines will call the (expensive)
2035 * get_writeback_state too often.
2037 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2038 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2042 void writeback_set_ratelimit(void)
2044 struct wb_domain
*dom
= &global_wb_domain
;
2045 unsigned long background_thresh
;
2046 unsigned long dirty_thresh
;
2048 global_dirty_limits(&background_thresh
, &dirty_thresh
);
2049 dom
->dirty_limit
= dirty_thresh
;
2050 ratelimit_pages
= dirty_thresh
/ (num_online_cpus() * 32);
2051 if (ratelimit_pages
< 16)
2052 ratelimit_pages
= 16;
2055 static int page_writeback_cpu_online(unsigned int cpu
)
2057 writeback_set_ratelimit();
2062 * Called early on to tune the page writeback dirty limits.
2064 * We used to scale dirty pages according to how total memory
2065 * related to pages that could be allocated for buffers (by
2066 * comparing nr_free_buffer_pages() to vm_total_pages.
2068 * However, that was when we used "dirty_ratio" to scale with
2069 * all memory, and we don't do that any more. "dirty_ratio"
2070 * is now applied to total non-HIGHPAGE memory (by subtracting
2071 * totalhigh_pages from vm_total_pages), and as such we can't
2072 * get into the old insane situation any more where we had
2073 * large amounts of dirty pages compared to a small amount of
2074 * non-HIGHMEM memory.
2076 * But we might still want to scale the dirty_ratio by how
2077 * much memory the box has..
2079 void __init
page_writeback_init(void)
2081 BUG_ON(wb_domain_init(&global_wb_domain
, GFP_KERNEL
));
2083 cpuhp_setup_state(CPUHP_AP_ONLINE_DYN
, "mm/writeback:online",
2084 page_writeback_cpu_online
, NULL
);
2085 cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD
, "mm/writeback:dead", NULL
,
2086 page_writeback_cpu_online
);
2090 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2091 * @mapping: address space structure to write
2092 * @start: starting page index
2093 * @end: ending page index (inclusive)
2095 * This function scans the page range from @start to @end (inclusive) and tags
2096 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2097 * that write_cache_pages (or whoever calls this function) will then use
2098 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2099 * used to avoid livelocking of writeback by a process steadily creating new
2100 * dirty pages in the file (thus it is important for this function to be quick
2101 * so that it can tag pages faster than a dirtying process can create them).
2104 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
2106 void tag_pages_for_writeback(struct address_space
*mapping
,
2107 pgoff_t start
, pgoff_t end
)
2109 #define WRITEBACK_TAG_BATCH 4096
2110 unsigned long tagged
= 0;
2111 struct radix_tree_iter iter
;
2114 spin_lock_irq(&mapping
->tree_lock
);
2115 radix_tree_for_each_tagged(slot
, &mapping
->page_tree
, &iter
, start
,
2116 PAGECACHE_TAG_DIRTY
) {
2117 if (iter
.index
> end
)
2119 radix_tree_iter_tag_set(&mapping
->page_tree
, &iter
,
2120 PAGECACHE_TAG_TOWRITE
);
2122 if ((tagged
% WRITEBACK_TAG_BATCH
) != 0)
2124 slot
= radix_tree_iter_resume(slot
, &iter
);
2125 spin_unlock_irq(&mapping
->tree_lock
);
2127 spin_lock_irq(&mapping
->tree_lock
);
2129 spin_unlock_irq(&mapping
->tree_lock
);
2131 EXPORT_SYMBOL(tag_pages_for_writeback
);
2134 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2135 * @mapping: address space structure to write
2136 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2137 * @writepage: function called for each page
2138 * @data: data passed to writepage function
2140 * If a page is already under I/O, write_cache_pages() skips it, even
2141 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2142 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2143 * and msync() need to guarantee that all the data which was dirty at the time
2144 * the call was made get new I/O started against them. If wbc->sync_mode is
2145 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2146 * existing IO to complete.
2148 * To avoid livelocks (when other process dirties new pages), we first tag
2149 * pages which should be written back with TOWRITE tag and only then start
2150 * writing them. For data-integrity sync we have to be careful so that we do
2151 * not miss some pages (e.g., because some other process has cleared TOWRITE
2152 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2153 * by the process clearing the DIRTY tag (and submitting the page for IO).
2155 int write_cache_pages(struct address_space
*mapping
,
2156 struct writeback_control
*wbc
, writepage_t writepage
,
2161 struct pagevec pvec
;
2163 pgoff_t
uninitialized_var(writeback_index
);
2165 pgoff_t end
; /* Inclusive */
2168 int range_whole
= 0;
2171 pagevec_init(&pvec
);
2172 if (wbc
->range_cyclic
) {
2173 writeback_index
= mapping
->writeback_index
; /* prev offset */
2174 index
= writeback_index
;
2181 index
= wbc
->range_start
>> PAGE_SHIFT
;
2182 end
= wbc
->range_end
>> PAGE_SHIFT
;
2183 if (wbc
->range_start
== 0 && wbc
->range_end
== LLONG_MAX
)
2185 cycled
= 1; /* ignore range_cyclic tests */
2187 if (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
)
2188 tag
= PAGECACHE_TAG_TOWRITE
;
2190 tag
= PAGECACHE_TAG_DIRTY
;
2192 if (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
)
2193 tag_pages_for_writeback(mapping
, index
, end
);
2195 while (!done
&& (index
<= end
)) {
2198 nr_pages
= pagevec_lookup_range_tag(&pvec
, mapping
, &index
, end
,
2203 for (i
= 0; i
< nr_pages
; i
++) {
2204 struct page
*page
= pvec
.pages
[i
];
2206 done_index
= page
->index
;
2211 * Page truncated or invalidated. We can freely skip it
2212 * then, even for data integrity operations: the page
2213 * has disappeared concurrently, so there could be no
2214 * real expectation of this data interity operation
2215 * even if there is now a new, dirty page at the same
2216 * pagecache address.
2218 if (unlikely(page
->mapping
!= mapping
)) {
2224 if (!PageDirty(page
)) {
2225 /* someone wrote it for us */
2226 goto continue_unlock
;
2229 if (PageWriteback(page
)) {
2230 if (wbc
->sync_mode
!= WB_SYNC_NONE
)
2231 wait_on_page_writeback(page
);
2233 goto continue_unlock
;
2236 BUG_ON(PageWriteback(page
));
2237 if (!clear_page_dirty_for_io(page
))
2238 goto continue_unlock
;
2240 trace_wbc_writepage(wbc
, inode_to_bdi(mapping
->host
));
2241 ret
= (*writepage
)(page
, wbc
, data
);
2242 if (unlikely(ret
)) {
2243 if (ret
== AOP_WRITEPAGE_ACTIVATE
) {
2248 * done_index is set past this page,
2249 * so media errors will not choke
2250 * background writeout for the entire
2251 * file. This has consequences for
2252 * range_cyclic semantics (ie. it may
2253 * not be suitable for data integrity
2256 done_index
= page
->index
+ 1;
2263 * We stop writing back only if we are not doing
2264 * integrity sync. In case of integrity sync we have to
2265 * keep going until we have written all the pages
2266 * we tagged for writeback prior to entering this loop.
2268 if (--wbc
->nr_to_write
<= 0 &&
2269 wbc
->sync_mode
== WB_SYNC_NONE
) {
2274 pagevec_release(&pvec
);
2277 if (!cycled
&& !done
) {
2280 * We hit the last page and there is more work to be done: wrap
2281 * back to the start of the file
2285 end
= writeback_index
- 1;
2288 if (wbc
->range_cyclic
|| (range_whole
&& wbc
->nr_to_write
> 0))
2289 mapping
->writeback_index
= done_index
;
2293 EXPORT_SYMBOL(write_cache_pages
);
2296 * Function used by generic_writepages to call the real writepage
2297 * function and set the mapping flags on error
2299 static int __writepage(struct page
*page
, struct writeback_control
*wbc
,
2302 struct address_space
*mapping
= data
;
2303 int ret
= mapping
->a_ops
->writepage(page
, wbc
);
2304 mapping_set_error(mapping
, ret
);
2309 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2310 * @mapping: address space structure to write
2311 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2313 * This is a library function, which implements the writepages()
2314 * address_space_operation.
2316 int generic_writepages(struct address_space
*mapping
,
2317 struct writeback_control
*wbc
)
2319 struct blk_plug plug
;
2322 /* deal with chardevs and other special file */
2323 if (!mapping
->a_ops
->writepage
)
2326 blk_start_plug(&plug
);
2327 ret
= write_cache_pages(mapping
, wbc
, __writepage
, mapping
);
2328 blk_finish_plug(&plug
);
2332 EXPORT_SYMBOL(generic_writepages
);
2334 int do_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
2338 if (wbc
->nr_to_write
<= 0)
2341 if (mapping
->a_ops
->writepages
)
2342 ret
= mapping
->a_ops
->writepages(mapping
, wbc
);
2344 ret
= generic_writepages(mapping
, wbc
);
2345 if ((ret
!= -ENOMEM
) || (wbc
->sync_mode
!= WB_SYNC_ALL
))
2348 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
2354 * write_one_page - write out a single page and wait on I/O
2355 * @page: the page to write
2357 * The page must be locked by the caller and will be unlocked upon return.
2359 * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2362 int write_one_page(struct page
*page
)
2364 struct address_space
*mapping
= page
->mapping
;
2366 struct writeback_control wbc
= {
2367 .sync_mode
= WB_SYNC_ALL
,
2371 BUG_ON(!PageLocked(page
));
2373 wait_on_page_writeback(page
);
2375 if (clear_page_dirty_for_io(page
)) {
2377 ret
= mapping
->a_ops
->writepage(page
, &wbc
);
2379 wait_on_page_writeback(page
);
2386 ret
= filemap_check_errors(mapping
);
2389 EXPORT_SYMBOL(write_one_page
);
2392 * For address_spaces which do not use buffers nor write back.
2394 int __set_page_dirty_no_writeback(struct page
*page
)
2396 if (!PageDirty(page
))
2397 return !TestSetPageDirty(page
);
2402 * Helper function for set_page_dirty family.
2404 * Caller must hold lock_page_memcg().
2406 * NOTE: This relies on being atomic wrt interrupts.
2408 void account_page_dirtied(struct page
*page
, struct address_space
*mapping
)
2410 struct inode
*inode
= mapping
->host
;
2412 trace_writeback_dirty_page(page
, mapping
);
2414 if (mapping_cap_account_dirty(mapping
)) {
2415 struct bdi_writeback
*wb
;
2417 inode_attach_wb(inode
, page
);
2418 wb
= inode_to_wb(inode
);
2420 __inc_lruvec_page_state(page
, NR_FILE_DIRTY
);
2421 __inc_zone_page_state(page
, NR_ZONE_WRITE_PENDING
);
2422 __inc_node_page_state(page
, NR_DIRTIED
);
2423 inc_wb_stat(wb
, WB_RECLAIMABLE
);
2424 inc_wb_stat(wb
, WB_DIRTIED
);
2425 task_io_account_write(PAGE_SIZE
);
2426 current
->nr_dirtied
++;
2427 this_cpu_inc(bdp_ratelimits
);
2430 EXPORT_SYMBOL(account_page_dirtied
);
2433 * Helper function for deaccounting dirty page without writeback.
2435 * Caller must hold lock_page_memcg().
2437 void account_page_cleaned(struct page
*page
, struct address_space
*mapping
,
2438 struct bdi_writeback
*wb
)
2440 if (mapping_cap_account_dirty(mapping
)) {
2441 dec_lruvec_page_state(page
, NR_FILE_DIRTY
);
2442 dec_zone_page_state(page
, NR_ZONE_WRITE_PENDING
);
2443 dec_wb_stat(wb
, WB_RECLAIMABLE
);
2444 task_io_account_cancelled_write(PAGE_SIZE
);
2449 * For address_spaces which do not use buffers. Just tag the page as dirty in
2452 * This is also used when a single buffer is being dirtied: we want to set the
2453 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2454 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2456 * The caller must ensure this doesn't race with truncation. Most will simply
2457 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2458 * the pte lock held, which also locks out truncation.
2460 int __set_page_dirty_nobuffers(struct page
*page
)
2462 lock_page_memcg(page
);
2463 if (!TestSetPageDirty(page
)) {
2464 struct address_space
*mapping
= page_mapping(page
);
2465 unsigned long flags
;
2468 unlock_page_memcg(page
);
2472 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
2473 BUG_ON(page_mapping(page
) != mapping
);
2474 WARN_ON_ONCE(!PagePrivate(page
) && !PageUptodate(page
));
2475 account_page_dirtied(page
, mapping
);
2476 radix_tree_tag_set(&mapping
->page_tree
, page_index(page
),
2477 PAGECACHE_TAG_DIRTY
);
2478 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
2479 unlock_page_memcg(page
);
2481 if (mapping
->host
) {
2482 /* !PageAnon && !swapper_space */
2483 __mark_inode_dirty(mapping
->host
, I_DIRTY_PAGES
);
2487 unlock_page_memcg(page
);
2490 EXPORT_SYMBOL(__set_page_dirty_nobuffers
);
2493 * Call this whenever redirtying a page, to de-account the dirty counters
2494 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2495 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2496 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2499 void account_page_redirty(struct page
*page
)
2501 struct address_space
*mapping
= page
->mapping
;
2503 if (mapping
&& mapping_cap_account_dirty(mapping
)) {
2504 struct inode
*inode
= mapping
->host
;
2505 struct bdi_writeback
*wb
;
2508 wb
= unlocked_inode_to_wb_begin(inode
, &locked
);
2509 current
->nr_dirtied
--;
2510 dec_node_page_state(page
, NR_DIRTIED
);
2511 dec_wb_stat(wb
, WB_DIRTIED
);
2512 unlocked_inode_to_wb_end(inode
, locked
);
2515 EXPORT_SYMBOL(account_page_redirty
);
2518 * When a writepage implementation decides that it doesn't want to write this
2519 * page for some reason, it should redirty the locked page via
2520 * redirty_page_for_writepage() and it should then unlock the page and return 0
2522 int redirty_page_for_writepage(struct writeback_control
*wbc
, struct page
*page
)
2526 wbc
->pages_skipped
++;
2527 ret
= __set_page_dirty_nobuffers(page
);
2528 account_page_redirty(page
);
2531 EXPORT_SYMBOL(redirty_page_for_writepage
);
2536 * For pages with a mapping this should be done under the page lock
2537 * for the benefit of asynchronous memory errors who prefer a consistent
2538 * dirty state. This rule can be broken in some special cases,
2539 * but should be better not to.
2541 * If the mapping doesn't provide a set_page_dirty a_op, then
2542 * just fall through and assume that it wants buffer_heads.
2544 int set_page_dirty(struct page
*page
)
2546 struct address_space
*mapping
= page_mapping(page
);
2548 page
= compound_head(page
);
2549 if (likely(mapping
)) {
2550 int (*spd
)(struct page
*) = mapping
->a_ops
->set_page_dirty
;
2552 * readahead/lru_deactivate_page could remain
2553 * PG_readahead/PG_reclaim due to race with end_page_writeback
2554 * About readahead, if the page is written, the flags would be
2555 * reset. So no problem.
2556 * About lru_deactivate_page, if the page is redirty, the flag
2557 * will be reset. So no problem. but if the page is used by readahead
2558 * it will confuse readahead and make it restart the size rampup
2559 * process. But it's a trivial problem.
2561 if (PageReclaim(page
))
2562 ClearPageReclaim(page
);
2565 spd
= __set_page_dirty_buffers
;
2567 return (*spd
)(page
);
2569 if (!PageDirty(page
)) {
2570 if (!TestSetPageDirty(page
))
2575 EXPORT_SYMBOL(set_page_dirty
);
2578 * set_page_dirty() is racy if the caller has no reference against
2579 * page->mapping->host, and if the page is unlocked. This is because another
2580 * CPU could truncate the page off the mapping and then free the mapping.
2582 * Usually, the page _is_ locked, or the caller is a user-space process which
2583 * holds a reference on the inode by having an open file.
2585 * In other cases, the page should be locked before running set_page_dirty().
2587 int set_page_dirty_lock(struct page
*page
)
2592 ret
= set_page_dirty(page
);
2596 EXPORT_SYMBOL(set_page_dirty_lock
);
2599 * This cancels just the dirty bit on the kernel page itself, it does NOT
2600 * actually remove dirty bits on any mmap's that may be around. It also
2601 * leaves the page tagged dirty, so any sync activity will still find it on
2602 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2603 * look at the dirty bits in the VM.
2605 * Doing this should *normally* only ever be done when a page is truncated,
2606 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2607 * this when it notices that somebody has cleaned out all the buffers on a
2608 * page without actually doing it through the VM. Can you say "ext3 is
2609 * horribly ugly"? Thought you could.
2611 void __cancel_dirty_page(struct page
*page
)
2613 struct address_space
*mapping
= page_mapping(page
);
2615 if (mapping_cap_account_dirty(mapping
)) {
2616 struct inode
*inode
= mapping
->host
;
2617 struct bdi_writeback
*wb
;
2620 lock_page_memcg(page
);
2621 wb
= unlocked_inode_to_wb_begin(inode
, &locked
);
2623 if (TestClearPageDirty(page
))
2624 account_page_cleaned(page
, mapping
, wb
);
2626 unlocked_inode_to_wb_end(inode
, locked
);
2627 unlock_page_memcg(page
);
2629 ClearPageDirty(page
);
2632 EXPORT_SYMBOL(__cancel_dirty_page
);
2635 * Clear a page's dirty flag, while caring for dirty memory accounting.
2636 * Returns true if the page was previously dirty.
2638 * This is for preparing to put the page under writeout. We leave the page
2639 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2640 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2641 * implementation will run either set_page_writeback() or set_page_dirty(),
2642 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2645 * This incoherency between the page's dirty flag and radix-tree tag is
2646 * unfortunate, but it only exists while the page is locked.
2648 int clear_page_dirty_for_io(struct page
*page
)
2650 struct address_space
*mapping
= page_mapping(page
);
2653 BUG_ON(!PageLocked(page
));
2655 if (mapping
&& mapping_cap_account_dirty(mapping
)) {
2656 struct inode
*inode
= mapping
->host
;
2657 struct bdi_writeback
*wb
;
2661 * Yes, Virginia, this is indeed insane.
2663 * We use this sequence to make sure that
2664 * (a) we account for dirty stats properly
2665 * (b) we tell the low-level filesystem to
2666 * mark the whole page dirty if it was
2667 * dirty in a pagetable. Only to then
2668 * (c) clean the page again and return 1 to
2669 * cause the writeback.
2671 * This way we avoid all nasty races with the
2672 * dirty bit in multiple places and clearing
2673 * them concurrently from different threads.
2675 * Note! Normally the "set_page_dirty(page)"
2676 * has no effect on the actual dirty bit - since
2677 * that will already usually be set. But we
2678 * need the side effects, and it can help us
2681 * We basically use the page "master dirty bit"
2682 * as a serialization point for all the different
2683 * threads doing their things.
2685 if (page_mkclean(page
))
2686 set_page_dirty(page
);
2688 * We carefully synchronise fault handlers against
2689 * installing a dirty pte and marking the page dirty
2690 * at this point. We do this by having them hold the
2691 * page lock while dirtying the page, and pages are
2692 * always locked coming in here, so we get the desired
2695 wb
= unlocked_inode_to_wb_begin(inode
, &locked
);
2696 if (TestClearPageDirty(page
)) {
2697 dec_lruvec_page_state(page
, NR_FILE_DIRTY
);
2698 dec_zone_page_state(page
, NR_ZONE_WRITE_PENDING
);
2699 dec_wb_stat(wb
, WB_RECLAIMABLE
);
2702 unlocked_inode_to_wb_end(inode
, locked
);
2705 return TestClearPageDirty(page
);
2707 EXPORT_SYMBOL(clear_page_dirty_for_io
);
2709 int test_clear_page_writeback(struct page
*page
)
2711 struct address_space
*mapping
= page_mapping(page
);
2712 struct mem_cgroup
*memcg
;
2713 struct lruvec
*lruvec
;
2716 memcg
= lock_page_memcg(page
);
2717 lruvec
= mem_cgroup_page_lruvec(page
, page_pgdat(page
));
2718 if (mapping
&& mapping_use_writeback_tags(mapping
)) {
2719 struct inode
*inode
= mapping
->host
;
2720 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
2721 unsigned long flags
;
2723 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
2724 ret
= TestClearPageWriteback(page
);
2726 radix_tree_tag_clear(&mapping
->page_tree
,
2728 PAGECACHE_TAG_WRITEBACK
);
2729 if (bdi_cap_account_writeback(bdi
)) {
2730 struct bdi_writeback
*wb
= inode_to_wb(inode
);
2732 dec_wb_stat(wb
, WB_WRITEBACK
);
2733 __wb_writeout_inc(wb
);
2737 if (mapping
->host
&& !mapping_tagged(mapping
,
2738 PAGECACHE_TAG_WRITEBACK
))
2739 sb_clear_inode_writeback(mapping
->host
);
2741 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
2743 ret
= TestClearPageWriteback(page
);
2746 * NOTE: Page might be free now! Writeback doesn't hold a page
2747 * reference on its own, it relies on truncation to wait for
2748 * the clearing of PG_writeback. The below can only access
2749 * page state that is static across allocation cycles.
2752 dec_lruvec_state(lruvec
, NR_WRITEBACK
);
2753 dec_zone_page_state(page
, NR_ZONE_WRITE_PENDING
);
2754 inc_node_page_state(page
, NR_WRITTEN
);
2756 __unlock_page_memcg(memcg
);
2760 int __test_set_page_writeback(struct page
*page
, bool keep_write
)
2762 struct address_space
*mapping
= page_mapping(page
);
2765 lock_page_memcg(page
);
2766 if (mapping
&& mapping_use_writeback_tags(mapping
)) {
2767 struct inode
*inode
= mapping
->host
;
2768 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
2769 unsigned long flags
;
2771 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
2772 ret
= TestSetPageWriteback(page
);
2776 on_wblist
= mapping_tagged(mapping
,
2777 PAGECACHE_TAG_WRITEBACK
);
2779 radix_tree_tag_set(&mapping
->page_tree
,
2781 PAGECACHE_TAG_WRITEBACK
);
2782 if (bdi_cap_account_writeback(bdi
))
2783 inc_wb_stat(inode_to_wb(inode
), WB_WRITEBACK
);
2786 * We can come through here when swapping anonymous
2787 * pages, so we don't necessarily have an inode to track
2790 if (mapping
->host
&& !on_wblist
)
2791 sb_mark_inode_writeback(mapping
->host
);
2793 if (!PageDirty(page
))
2794 radix_tree_tag_clear(&mapping
->page_tree
,
2796 PAGECACHE_TAG_DIRTY
);
2798 radix_tree_tag_clear(&mapping
->page_tree
,
2800 PAGECACHE_TAG_TOWRITE
);
2801 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
2803 ret
= TestSetPageWriteback(page
);
2806 inc_lruvec_page_state(page
, NR_WRITEBACK
);
2807 inc_zone_page_state(page
, NR_ZONE_WRITE_PENDING
);
2809 unlock_page_memcg(page
);
2813 EXPORT_SYMBOL(__test_set_page_writeback
);
2816 * Return true if any of the pages in the mapping are marked with the
2819 int mapping_tagged(struct address_space
*mapping
, int tag
)
2821 return radix_tree_tagged(&mapping
->page_tree
, tag
);
2823 EXPORT_SYMBOL(mapping_tagged
);
2826 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2827 * @page: The page to wait on.
2829 * This function determines if the given page is related to a backing device
2830 * that requires page contents to be held stable during writeback. If so, then
2831 * it will wait for any pending writeback to complete.
2833 void wait_for_stable_page(struct page
*page
)
2835 if (bdi_cap_stable_pages_required(inode_to_bdi(page
->mapping
->host
)))
2836 wait_on_page_writeback(page
);
2838 EXPORT_SYMBOL_GPL(wait_for_stable_page
);