4 * Copyright (C) 2002, Linus Torvalds.
5 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Contains functions related to writing back dirty pages at the
10 * 10Apr2002 Andrew Morton
14 #include <linux/kernel.h>
15 #include <linux/export.h>
16 #include <linux/spinlock.h>
19 #include <linux/swap.h>
20 #include <linux/slab.h>
21 #include <linux/pagemap.h>
22 #include <linux/writeback.h>
23 #include <linux/init.h>
24 #include <linux/backing-dev.h>
25 #include <linux/task_io_accounting_ops.h>
26 #include <linux/blkdev.h>
27 #include <linux/mpage.h>
28 #include <linux/rmap.h>
29 #include <linux/percpu.h>
30 #include <linux/notifier.h>
31 #include <linux/smp.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/syscalls.h>
35 #include <linux/buffer_head.h> /* __set_page_dirty_buffers */
36 #include <linux/pagevec.h>
37 #include <linux/timer.h>
38 #include <linux/sched/rt.h>
39 #include <linux/mm_inline.h>
40 #include <trace/events/writeback.h>
45 * Sleep at most 200ms at a time in balance_dirty_pages().
47 #define MAX_PAUSE max(HZ/5, 1)
50 * Try to keep balance_dirty_pages() call intervals higher than this many pages
51 * by raising pause time to max_pause when falls below it.
53 #define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
56 * Estimate write bandwidth at 200ms intervals.
58 #define BANDWIDTH_INTERVAL max(HZ/5, 1)
60 #define RATELIMIT_CALC_SHIFT 10
63 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
64 * will look to see if it needs to force writeback or throttling.
66 static long ratelimit_pages
= 32;
68 /* The following parameters are exported via /proc/sys/vm */
71 * Start background writeback (via writeback threads) at this percentage
73 int dirty_background_ratio
= 10;
76 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
77 * dirty_background_ratio * the amount of dirtyable memory
79 unsigned long dirty_background_bytes
;
82 * free highmem will not be subtracted from the total free memory
83 * for calculating free ratios if vm_highmem_is_dirtyable is true
85 int vm_highmem_is_dirtyable
;
88 * The generator of dirty data starts writeback at this percentage
90 int vm_dirty_ratio
= 20;
93 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
94 * vm_dirty_ratio * the amount of dirtyable memory
96 unsigned long vm_dirty_bytes
;
99 * The interval between `kupdate'-style writebacks
101 unsigned int dirty_writeback_interval
= 5 * 100; /* centiseconds */
103 EXPORT_SYMBOL_GPL(dirty_writeback_interval
);
106 * The longest time for which data is allowed to remain dirty
108 unsigned int dirty_expire_interval
= 30 * 100; /* centiseconds */
111 * Flag that makes the machine dump writes/reads and block dirtyings.
116 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
117 * a full sync is triggered after this time elapses without any disk activity.
121 EXPORT_SYMBOL(laptop_mode
);
123 /* End of sysctl-exported parameters */
125 struct wb_domain global_wb_domain
;
127 /* consolidated parameters for balance_dirty_pages() and its subroutines */
128 struct dirty_throttle_control
{
129 #ifdef CONFIG_CGROUP_WRITEBACK
130 struct wb_domain
*dom
;
131 struct dirty_throttle_control
*gdtc
; /* only set in memcg dtc's */
133 struct bdi_writeback
*wb
;
134 struct fprop_local_percpu
*wb_completions
;
136 unsigned long avail
; /* dirtyable */
137 unsigned long dirty
; /* file_dirty + write + nfs */
138 unsigned long thresh
; /* dirty threshold */
139 unsigned long bg_thresh
; /* dirty background threshold */
141 unsigned long wb_dirty
; /* per-wb counterparts */
142 unsigned long wb_thresh
;
143 unsigned long wb_bg_thresh
;
145 unsigned long pos_ratio
;
149 * Length of period for aging writeout fractions of bdis. This is an
150 * arbitrarily chosen number. The longer the period, the slower fractions will
151 * reflect changes in current writeout rate.
153 #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
155 #ifdef CONFIG_CGROUP_WRITEBACK
157 #define GDTC_INIT(__wb) .wb = (__wb), \
158 .dom = &global_wb_domain, \
159 .wb_completions = &(__wb)->completions
161 #define GDTC_INIT_NO_WB .dom = &global_wb_domain
163 #define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
164 .dom = mem_cgroup_wb_domain(__wb), \
165 .wb_completions = &(__wb)->memcg_completions, \
168 static bool mdtc_valid(struct dirty_throttle_control
*dtc
)
173 static struct wb_domain
*dtc_dom(struct dirty_throttle_control
*dtc
)
178 static struct dirty_throttle_control
*mdtc_gdtc(struct dirty_throttle_control
*mdtc
)
183 static struct fprop_local_percpu
*wb_memcg_completions(struct bdi_writeback
*wb
)
185 return &wb
->memcg_completions
;
188 static void wb_min_max_ratio(struct bdi_writeback
*wb
,
189 unsigned long *minp
, unsigned long *maxp
)
191 unsigned long this_bw
= wb
->avg_write_bandwidth
;
192 unsigned long tot_bw
= atomic_long_read(&wb
->bdi
->tot_write_bandwidth
);
193 unsigned long long min
= wb
->bdi
->min_ratio
;
194 unsigned long long max
= wb
->bdi
->max_ratio
;
197 * @wb may already be clean by the time control reaches here and
198 * the total may not include its bw.
200 if (this_bw
< tot_bw
) {
215 #else /* CONFIG_CGROUP_WRITEBACK */
217 #define GDTC_INIT(__wb) .wb = (__wb), \
218 .wb_completions = &(__wb)->completions
219 #define GDTC_INIT_NO_WB
220 #define MDTC_INIT(__wb, __gdtc)
222 static bool mdtc_valid(struct dirty_throttle_control
*dtc
)
227 static struct wb_domain
*dtc_dom(struct dirty_throttle_control
*dtc
)
229 return &global_wb_domain
;
232 static struct dirty_throttle_control
*mdtc_gdtc(struct dirty_throttle_control
*mdtc
)
237 static struct fprop_local_percpu
*wb_memcg_completions(struct bdi_writeback
*wb
)
242 static void wb_min_max_ratio(struct bdi_writeback
*wb
,
243 unsigned long *minp
, unsigned long *maxp
)
245 *minp
= wb
->bdi
->min_ratio
;
246 *maxp
= wb
->bdi
->max_ratio
;
249 #endif /* CONFIG_CGROUP_WRITEBACK */
252 * In a memory zone, there is a certain amount of pages we consider
253 * available for the page cache, which is essentially the number of
254 * free and reclaimable pages, minus some zone reserves to protect
255 * lowmem and the ability to uphold the zone's watermarks without
256 * requiring writeback.
258 * This number of dirtyable pages is the base value of which the
259 * user-configurable dirty ratio is the effictive number of pages that
260 * are allowed to be actually dirtied. Per individual zone, or
261 * globally by using the sum of dirtyable pages over all zones.
263 * Because the user is allowed to specify the dirty limit globally as
264 * absolute number of bytes, calculating the per-zone dirty limit can
265 * require translating the configured limit into a percentage of
266 * global dirtyable memory first.
270 * zone_dirtyable_memory - number of dirtyable pages in a zone
273 * Returns the zone's number of pages potentially available for dirty
274 * page cache. This is the base value for the per-zone dirty limits.
276 static unsigned long zone_dirtyable_memory(struct zone
*zone
)
278 unsigned long nr_pages
;
280 nr_pages
= zone_page_state(zone
, NR_FREE_PAGES
);
281 nr_pages
-= min(nr_pages
, zone
->dirty_balance_reserve
);
283 nr_pages
+= zone_page_state(zone
, NR_INACTIVE_FILE
);
284 nr_pages
+= zone_page_state(zone
, NR_ACTIVE_FILE
);
289 static unsigned long highmem_dirtyable_memory(unsigned long total
)
291 #ifdef CONFIG_HIGHMEM
295 for_each_node_state(node
, N_HIGH_MEMORY
) {
296 struct zone
*z
= &NODE_DATA(node
)->node_zones
[ZONE_HIGHMEM
];
298 x
+= zone_dirtyable_memory(z
);
301 * Unreclaimable memory (kernel memory or anonymous memory
302 * without swap) can bring down the dirtyable pages below
303 * the zone's dirty balance reserve and the above calculation
304 * will underflow. However we still want to add in nodes
305 * which are below threshold (negative values) to get a more
306 * accurate calculation but make sure that the total never
313 * Make sure that the number of highmem pages is never larger
314 * than the number of the total dirtyable memory. This can only
315 * occur in very strange VM situations but we want to make sure
316 * that this does not occur.
318 return min(x
, total
);
325 * global_dirtyable_memory - number of globally dirtyable pages
327 * Returns the global number of pages potentially available for dirty
328 * page cache. This is the base value for the global dirty limits.
330 static unsigned long global_dirtyable_memory(void)
334 x
= global_page_state(NR_FREE_PAGES
);
335 x
-= min(x
, dirty_balance_reserve
);
337 x
+= global_page_state(NR_INACTIVE_FILE
);
338 x
+= global_page_state(NR_ACTIVE_FILE
);
340 if (!vm_highmem_is_dirtyable
)
341 x
-= highmem_dirtyable_memory(x
);
343 return x
+ 1; /* Ensure that we never return 0 */
347 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
348 * @dtc: dirty_throttle_control of interest
350 * Calculate @dtc->thresh and ->bg_thresh considering
351 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
352 * must ensure that @dtc->avail is set before calling this function. The
353 * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
356 static void domain_dirty_limits(struct dirty_throttle_control
*dtc
)
358 const unsigned long available_memory
= dtc
->avail
;
359 struct dirty_throttle_control
*gdtc
= mdtc_gdtc(dtc
);
360 unsigned long bytes
= vm_dirty_bytes
;
361 unsigned long bg_bytes
= dirty_background_bytes
;
362 unsigned long ratio
= vm_dirty_ratio
;
363 unsigned long bg_ratio
= dirty_background_ratio
;
364 unsigned long thresh
;
365 unsigned long bg_thresh
;
366 struct task_struct
*tsk
;
368 /* gdtc is !NULL iff @dtc is for memcg domain */
370 unsigned long global_avail
= gdtc
->avail
;
373 * The byte settings can't be applied directly to memcg
374 * domains. Convert them to ratios by scaling against
375 * globally available memory.
378 ratio
= min(DIV_ROUND_UP(bytes
, PAGE_SIZE
) * 100 /
379 global_avail
, 100UL);
381 bg_ratio
= min(DIV_ROUND_UP(bg_bytes
, PAGE_SIZE
) * 100 /
382 global_avail
, 100UL);
383 bytes
= bg_bytes
= 0;
387 thresh
= DIV_ROUND_UP(bytes
, PAGE_SIZE
);
389 thresh
= (ratio
* available_memory
) / 100;
392 bg_thresh
= DIV_ROUND_UP(bg_bytes
, PAGE_SIZE
);
394 bg_thresh
= (bg_ratio
* available_memory
) / 100;
396 if (bg_thresh
>= thresh
)
397 bg_thresh
= thresh
/ 2;
399 if (tsk
->flags
& PF_LESS_THROTTLE
|| rt_task(tsk
)) {
400 bg_thresh
+= bg_thresh
/ 4;
401 thresh
+= thresh
/ 4;
403 dtc
->thresh
= thresh
;
404 dtc
->bg_thresh
= bg_thresh
;
406 /* we should eventually report the domain in the TP */
408 trace_global_dirty_state(bg_thresh
, thresh
);
412 * global_dirty_limits - background-writeback and dirty-throttling thresholds
413 * @pbackground: out parameter for bg_thresh
414 * @pdirty: out parameter for thresh
416 * Calculate bg_thresh and thresh for global_wb_domain. See
417 * domain_dirty_limits() for details.
419 void global_dirty_limits(unsigned long *pbackground
, unsigned long *pdirty
)
421 struct dirty_throttle_control gdtc
= { GDTC_INIT_NO_WB
};
423 gdtc
.avail
= global_dirtyable_memory();
424 domain_dirty_limits(&gdtc
);
426 *pbackground
= gdtc
.bg_thresh
;
427 *pdirty
= gdtc
.thresh
;
431 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
434 * Returns the maximum number of dirty pages allowed in a zone, based
435 * on the zone's dirtyable memory.
437 static unsigned long zone_dirty_limit(struct zone
*zone
)
439 unsigned long zone_memory
= zone_dirtyable_memory(zone
);
440 struct task_struct
*tsk
= current
;
444 dirty
= DIV_ROUND_UP(vm_dirty_bytes
, PAGE_SIZE
) *
445 zone_memory
/ global_dirtyable_memory();
447 dirty
= vm_dirty_ratio
* zone_memory
/ 100;
449 if (tsk
->flags
& PF_LESS_THROTTLE
|| rt_task(tsk
))
456 * zone_dirty_ok - tells whether a zone is within its dirty limits
457 * @zone: the zone to check
459 * Returns %true when the dirty pages in @zone are within the zone's
460 * dirty limit, %false if the limit is exceeded.
462 bool zone_dirty_ok(struct zone
*zone
)
464 unsigned long limit
= zone_dirty_limit(zone
);
466 return zone_page_state(zone
, NR_FILE_DIRTY
) +
467 zone_page_state(zone
, NR_UNSTABLE_NFS
) +
468 zone_page_state(zone
, NR_WRITEBACK
) <= limit
;
471 int dirty_background_ratio_handler(struct ctl_table
*table
, int write
,
472 void __user
*buffer
, size_t *lenp
,
477 ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
478 if (ret
== 0 && write
)
479 dirty_background_bytes
= 0;
483 int dirty_background_bytes_handler(struct ctl_table
*table
, int write
,
484 void __user
*buffer
, size_t *lenp
,
489 ret
= proc_doulongvec_minmax(table
, write
, buffer
, lenp
, ppos
);
490 if (ret
== 0 && write
)
491 dirty_background_ratio
= 0;
495 int dirty_ratio_handler(struct ctl_table
*table
, int write
,
496 void __user
*buffer
, size_t *lenp
,
499 int old_ratio
= vm_dirty_ratio
;
502 ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
503 if (ret
== 0 && write
&& vm_dirty_ratio
!= old_ratio
) {
504 writeback_set_ratelimit();
510 int dirty_bytes_handler(struct ctl_table
*table
, int write
,
511 void __user
*buffer
, size_t *lenp
,
514 unsigned long old_bytes
= vm_dirty_bytes
;
517 ret
= proc_doulongvec_minmax(table
, write
, buffer
, lenp
, ppos
);
518 if (ret
== 0 && write
&& vm_dirty_bytes
!= old_bytes
) {
519 writeback_set_ratelimit();
525 static unsigned long wp_next_time(unsigned long cur_time
)
527 cur_time
+= VM_COMPLETIONS_PERIOD_LEN
;
528 /* 0 has a special meaning... */
534 static void wb_domain_writeout_inc(struct wb_domain
*dom
,
535 struct fprop_local_percpu
*completions
,
536 unsigned int max_prop_frac
)
538 __fprop_inc_percpu_max(&dom
->completions
, completions
,
540 /* First event after period switching was turned off? */
541 if (!unlikely(dom
->period_time
)) {
543 * We can race with other __bdi_writeout_inc calls here but
544 * it does not cause any harm since the resulting time when
545 * timer will fire and what is in writeout_period_time will be
548 dom
->period_time
= wp_next_time(jiffies
);
549 mod_timer(&dom
->period_timer
, dom
->period_time
);
554 * Increment @wb's writeout completion count and the global writeout
555 * completion count. Called from test_clear_page_writeback().
557 static inline void __wb_writeout_inc(struct bdi_writeback
*wb
)
559 struct wb_domain
*cgdom
;
561 __inc_wb_stat(wb
, WB_WRITTEN
);
562 wb_domain_writeout_inc(&global_wb_domain
, &wb
->completions
,
563 wb
->bdi
->max_prop_frac
);
565 cgdom
= mem_cgroup_wb_domain(wb
);
567 wb_domain_writeout_inc(cgdom
, wb_memcg_completions(wb
),
568 wb
->bdi
->max_prop_frac
);
571 void wb_writeout_inc(struct bdi_writeback
*wb
)
575 local_irq_save(flags
);
576 __wb_writeout_inc(wb
);
577 local_irq_restore(flags
);
579 EXPORT_SYMBOL_GPL(wb_writeout_inc
);
582 * On idle system, we can be called long after we scheduled because we use
583 * deferred timers so count with missed periods.
585 static void writeout_period(unsigned long t
)
587 struct wb_domain
*dom
= (void *)t
;
588 int miss_periods
= (jiffies
- dom
->period_time
) /
589 VM_COMPLETIONS_PERIOD_LEN
;
591 if (fprop_new_period(&dom
->completions
, miss_periods
+ 1)) {
592 dom
->period_time
= wp_next_time(dom
->period_time
+
593 miss_periods
* VM_COMPLETIONS_PERIOD_LEN
);
594 mod_timer(&dom
->period_timer
, dom
->period_time
);
597 * Aging has zeroed all fractions. Stop wasting CPU on period
600 dom
->period_time
= 0;
604 int wb_domain_init(struct wb_domain
*dom
, gfp_t gfp
)
606 memset(dom
, 0, sizeof(*dom
));
608 spin_lock_init(&dom
->lock
);
610 init_timer_deferrable(&dom
->period_timer
);
611 dom
->period_timer
.function
= writeout_period
;
612 dom
->period_timer
.data
= (unsigned long)dom
;
614 dom
->dirty_limit_tstamp
= jiffies
;
616 return fprop_global_init(&dom
->completions
, gfp
);
619 #ifdef CONFIG_CGROUP_WRITEBACK
620 void wb_domain_exit(struct wb_domain
*dom
)
622 del_timer_sync(&dom
->period_timer
);
623 fprop_global_destroy(&dom
->completions
);
628 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
629 * registered backing devices, which, for obvious reasons, can not
632 static unsigned int bdi_min_ratio
;
634 int bdi_set_min_ratio(struct backing_dev_info
*bdi
, unsigned int min_ratio
)
638 spin_lock_bh(&bdi_lock
);
639 if (min_ratio
> bdi
->max_ratio
) {
642 min_ratio
-= bdi
->min_ratio
;
643 if (bdi_min_ratio
+ min_ratio
< 100) {
644 bdi_min_ratio
+= min_ratio
;
645 bdi
->min_ratio
+= min_ratio
;
650 spin_unlock_bh(&bdi_lock
);
655 int bdi_set_max_ratio(struct backing_dev_info
*bdi
, unsigned max_ratio
)
662 spin_lock_bh(&bdi_lock
);
663 if (bdi
->min_ratio
> max_ratio
) {
666 bdi
->max_ratio
= max_ratio
;
667 bdi
->max_prop_frac
= (FPROP_FRAC_BASE
* max_ratio
) / 100;
669 spin_unlock_bh(&bdi_lock
);
673 EXPORT_SYMBOL(bdi_set_max_ratio
);
675 static unsigned long dirty_freerun_ceiling(unsigned long thresh
,
676 unsigned long bg_thresh
)
678 return (thresh
+ bg_thresh
) / 2;
681 static unsigned long hard_dirty_limit(struct wb_domain
*dom
,
682 unsigned long thresh
)
684 return max(thresh
, dom
->dirty_limit
);
688 * Memory which can be further allocated to a memcg domain is capped by
689 * system-wide clean memory excluding the amount being used in the domain.
691 static void mdtc_calc_avail(struct dirty_throttle_control
*mdtc
,
692 unsigned long filepages
, unsigned long headroom
)
694 struct dirty_throttle_control
*gdtc
= mdtc_gdtc(mdtc
);
695 unsigned long clean
= filepages
- min(filepages
, mdtc
->dirty
);
696 unsigned long global_clean
= gdtc
->avail
- min(gdtc
->avail
, gdtc
->dirty
);
697 unsigned long other_clean
= global_clean
- min(global_clean
, clean
);
699 mdtc
->avail
= filepages
+ min(headroom
, other_clean
);
703 * __wb_calc_thresh - @wb's share of dirty throttling threshold
704 * @dtc: dirty_throttle_context of interest
706 * Returns @wb's dirty limit in pages. The term "dirty" in the context of
707 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
709 * Note that balance_dirty_pages() will only seriously take it as a hard limit
710 * when sleeping max_pause per page is not enough to keep the dirty pages under
711 * control. For example, when the device is completely stalled due to some error
712 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
713 * In the other normal situations, it acts more gently by throttling the tasks
714 * more (rather than completely block them) when the wb dirty pages go high.
716 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
717 * - starving fast devices
718 * - piling up dirty pages (that will take long time to sync) on slow devices
720 * The wb's share of dirty limit will be adapting to its throughput and
721 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
723 static unsigned long __wb_calc_thresh(struct dirty_throttle_control
*dtc
)
725 struct wb_domain
*dom
= dtc_dom(dtc
);
726 unsigned long thresh
= dtc
->thresh
;
728 long numerator
, denominator
;
729 unsigned long wb_min_ratio
, wb_max_ratio
;
732 * Calculate this BDI's share of the thresh ratio.
734 fprop_fraction_percpu(&dom
->completions
, dtc
->wb_completions
,
735 &numerator
, &denominator
);
737 wb_thresh
= (thresh
* (100 - bdi_min_ratio
)) / 100;
738 wb_thresh
*= numerator
;
739 do_div(wb_thresh
, denominator
);
741 wb_min_max_ratio(dtc
->wb
, &wb_min_ratio
, &wb_max_ratio
);
743 wb_thresh
+= (thresh
* wb_min_ratio
) / 100;
744 if (wb_thresh
> (thresh
* wb_max_ratio
) / 100)
745 wb_thresh
= thresh
* wb_max_ratio
/ 100;
750 unsigned long wb_calc_thresh(struct bdi_writeback
*wb
, unsigned long thresh
)
752 struct dirty_throttle_control gdtc
= { GDTC_INIT(wb
),
754 return __wb_calc_thresh(&gdtc
);
759 * f(dirty) := 1.0 + (----------------)
762 * it's a 3rd order polynomial that subjects to
764 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
765 * (2) f(setpoint) = 1.0 => the balance point
766 * (3) f(limit) = 0 => the hard limit
767 * (4) df/dx <= 0 => negative feedback control
768 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
769 * => fast response on large errors; small oscillation near setpoint
771 static long long pos_ratio_polynom(unsigned long setpoint
,
778 x
= div64_s64(((s64
)setpoint
- (s64
)dirty
) << RATELIMIT_CALC_SHIFT
,
779 (limit
- setpoint
) | 1);
781 pos_ratio
= pos_ratio
* x
>> RATELIMIT_CALC_SHIFT
;
782 pos_ratio
= pos_ratio
* x
>> RATELIMIT_CALC_SHIFT
;
783 pos_ratio
+= 1 << RATELIMIT_CALC_SHIFT
;
785 return clamp(pos_ratio
, 0LL, 2LL << RATELIMIT_CALC_SHIFT
);
789 * Dirty position control.
791 * (o) global/bdi setpoints
793 * We want the dirty pages be balanced around the global/wb setpoints.
794 * When the number of dirty pages is higher/lower than the setpoint, the
795 * dirty position control ratio (and hence task dirty ratelimit) will be
796 * decreased/increased to bring the dirty pages back to the setpoint.
798 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
800 * if (dirty < setpoint) scale up pos_ratio
801 * if (dirty > setpoint) scale down pos_ratio
803 * if (wb_dirty < wb_setpoint) scale up pos_ratio
804 * if (wb_dirty > wb_setpoint) scale down pos_ratio
806 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
808 * (o) global control line
812 * | |<===== global dirty control scope ======>|
820 * 1.0 ................................*
826 * 0 +------------.------------------.----------------------*------------->
827 * freerun^ setpoint^ limit^ dirty pages
829 * (o) wb control line
837 * | * |<=========== span ============>|
838 * 1.0 .......................*
850 * 1/4 ...............................................* * * * * * * * * * * *
854 * 0 +----------------------.-------------------------------.------------->
855 * wb_setpoint^ x_intercept^
857 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
858 * be smoothly throttled down to normal if it starts high in situations like
859 * - start writing to a slow SD card and a fast disk at the same time. The SD
860 * card's wb_dirty may rush to many times higher than wb_setpoint.
861 * - the wb dirty thresh drops quickly due to change of JBOD workload
863 static void wb_position_ratio(struct dirty_throttle_control
*dtc
)
865 struct bdi_writeback
*wb
= dtc
->wb
;
866 unsigned long write_bw
= wb
->avg_write_bandwidth
;
867 unsigned long freerun
= dirty_freerun_ceiling(dtc
->thresh
, dtc
->bg_thresh
);
868 unsigned long limit
= hard_dirty_limit(dtc_dom(dtc
), dtc
->thresh
);
869 unsigned long wb_thresh
= dtc
->wb_thresh
;
870 unsigned long x_intercept
;
871 unsigned long setpoint
; /* dirty pages' target balance point */
872 unsigned long wb_setpoint
;
874 long long pos_ratio
; /* for scaling up/down the rate limit */
879 if (unlikely(dtc
->dirty
>= limit
))
885 * See comment for pos_ratio_polynom().
887 setpoint
= (freerun
+ limit
) / 2;
888 pos_ratio
= pos_ratio_polynom(setpoint
, dtc
->dirty
, limit
);
891 * The strictlimit feature is a tool preventing mistrusted filesystems
892 * from growing a large number of dirty pages before throttling. For
893 * such filesystems balance_dirty_pages always checks wb counters
894 * against wb limits. Even if global "nr_dirty" is under "freerun".
895 * This is especially important for fuse which sets bdi->max_ratio to
896 * 1% by default. Without strictlimit feature, fuse writeback may
897 * consume arbitrary amount of RAM because it is accounted in
898 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
900 * Here, in wb_position_ratio(), we calculate pos_ratio based on
901 * two values: wb_dirty and wb_thresh. Let's consider an example:
902 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
903 * limits are set by default to 10% and 20% (background and throttle).
904 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
905 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
906 * about ~6K pages (as the average of background and throttle wb
907 * limits). The 3rd order polynomial will provide positive feedback if
908 * wb_dirty is under wb_setpoint and vice versa.
910 * Note, that we cannot use global counters in these calculations
911 * because we want to throttle process writing to a strictlimit wb
912 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
913 * in the example above).
915 if (unlikely(wb
->bdi
->capabilities
& BDI_CAP_STRICTLIMIT
)) {
916 long long wb_pos_ratio
;
918 if (dtc
->wb_dirty
< 8) {
919 dtc
->pos_ratio
= min_t(long long, pos_ratio
* 2,
920 2 << RATELIMIT_CALC_SHIFT
);
924 if (dtc
->wb_dirty
>= wb_thresh
)
927 wb_setpoint
= dirty_freerun_ceiling(wb_thresh
,
930 if (wb_setpoint
== 0 || wb_setpoint
== wb_thresh
)
933 wb_pos_ratio
= pos_ratio_polynom(wb_setpoint
, dtc
->wb_dirty
,
937 * Typically, for strictlimit case, wb_setpoint << setpoint
938 * and pos_ratio >> wb_pos_ratio. In the other words global
939 * state ("dirty") is not limiting factor and we have to
940 * make decision based on wb counters. But there is an
941 * important case when global pos_ratio should get precedence:
942 * global limits are exceeded (e.g. due to activities on other
943 * wb's) while given strictlimit wb is below limit.
945 * "pos_ratio * wb_pos_ratio" would work for the case above,
946 * but it would look too non-natural for the case of all
947 * activity in the system coming from a single strictlimit wb
948 * with bdi->max_ratio == 100%.
950 * Note that min() below somewhat changes the dynamics of the
951 * control system. Normally, pos_ratio value can be well over 3
952 * (when globally we are at freerun and wb is well below wb
953 * setpoint). Now the maximum pos_ratio in the same situation
954 * is 2. We might want to tweak this if we observe the control
955 * system is too slow to adapt.
957 dtc
->pos_ratio
= min(pos_ratio
, wb_pos_ratio
);
962 * We have computed basic pos_ratio above based on global situation. If
963 * the wb is over/under its share of dirty pages, we want to scale
964 * pos_ratio further down/up. That is done by the following mechanism.
970 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
972 * x_intercept - wb_dirty
973 * := --------------------------
974 * x_intercept - wb_setpoint
976 * The main wb control line is a linear function that subjects to
978 * (1) f(wb_setpoint) = 1.0
979 * (2) k = - 1 / (8 * write_bw) (in single wb case)
980 * or equally: x_intercept = wb_setpoint + 8 * write_bw
982 * For single wb case, the dirty pages are observed to fluctuate
983 * regularly within range
984 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
985 * for various filesystems, where (2) can yield in a reasonable 12.5%
986 * fluctuation range for pos_ratio.
988 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
989 * own size, so move the slope over accordingly and choose a slope that
990 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
992 if (unlikely(wb_thresh
> dtc
->thresh
))
993 wb_thresh
= dtc
->thresh
;
995 * It's very possible that wb_thresh is close to 0 not because the
996 * device is slow, but that it has remained inactive for long time.
997 * Honour such devices a reasonable good (hopefully IO efficient)
998 * threshold, so that the occasional writes won't be blocked and active
999 * writes can rampup the threshold quickly.
1001 wb_thresh
= max(wb_thresh
, (limit
- dtc
->dirty
) / 8);
1003 * scale global setpoint to wb's:
1004 * wb_setpoint = setpoint * wb_thresh / thresh
1006 x
= div_u64((u64
)wb_thresh
<< 16, dtc
->thresh
| 1);
1007 wb_setpoint
= setpoint
* (u64
)x
>> 16;
1009 * Use span=(8*write_bw) in single wb case as indicated by
1010 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1012 * wb_thresh thresh - wb_thresh
1013 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1016 span
= (dtc
->thresh
- wb_thresh
+ 8 * write_bw
) * (u64
)x
>> 16;
1017 x_intercept
= wb_setpoint
+ span
;
1019 if (dtc
->wb_dirty
< x_intercept
- span
/ 4) {
1020 pos_ratio
= div64_u64(pos_ratio
* (x_intercept
- dtc
->wb_dirty
),
1021 (x_intercept
- wb_setpoint
) | 1);
1026 * wb reserve area, safeguard against dirty pool underrun and disk idle
1027 * It may push the desired control point of global dirty pages higher
1030 x_intercept
= wb_thresh
/ 2;
1031 if (dtc
->wb_dirty
< x_intercept
) {
1032 if (dtc
->wb_dirty
> x_intercept
/ 8)
1033 pos_ratio
= div_u64(pos_ratio
* x_intercept
,
1039 dtc
->pos_ratio
= pos_ratio
;
1042 static void wb_update_write_bandwidth(struct bdi_writeback
*wb
,
1043 unsigned long elapsed
,
1044 unsigned long written
)
1046 const unsigned long period
= roundup_pow_of_two(3 * HZ
);
1047 unsigned long avg
= wb
->avg_write_bandwidth
;
1048 unsigned long old
= wb
->write_bandwidth
;
1052 * bw = written * HZ / elapsed
1054 * bw * elapsed + write_bandwidth * (period - elapsed)
1055 * write_bandwidth = ---------------------------------------------------
1058 * @written may have decreased due to account_page_redirty().
1059 * Avoid underflowing @bw calculation.
1061 bw
= written
- min(written
, wb
->written_stamp
);
1063 if (unlikely(elapsed
> period
)) {
1064 do_div(bw
, elapsed
);
1068 bw
+= (u64
)wb
->write_bandwidth
* (period
- elapsed
);
1069 bw
>>= ilog2(period
);
1072 * one more level of smoothing, for filtering out sudden spikes
1074 if (avg
> old
&& old
>= (unsigned long)bw
)
1075 avg
-= (avg
- old
) >> 3;
1077 if (avg
< old
&& old
<= (unsigned long)bw
)
1078 avg
+= (old
- avg
) >> 3;
1081 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1082 avg
= max(avg
, 1LU);
1083 if (wb_has_dirty_io(wb
)) {
1084 long delta
= avg
- wb
->avg_write_bandwidth
;
1085 WARN_ON_ONCE(atomic_long_add_return(delta
,
1086 &wb
->bdi
->tot_write_bandwidth
) <= 0);
1088 wb
->write_bandwidth
= bw
;
1089 wb
->avg_write_bandwidth
= avg
;
1092 static void update_dirty_limit(struct dirty_throttle_control
*dtc
)
1094 struct wb_domain
*dom
= dtc_dom(dtc
);
1095 unsigned long thresh
= dtc
->thresh
;
1096 unsigned long limit
= dom
->dirty_limit
;
1099 * Follow up in one step.
1101 if (limit
< thresh
) {
1107 * Follow down slowly. Use the higher one as the target, because thresh
1108 * may drop below dirty. This is exactly the reason to introduce
1109 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1111 thresh
= max(thresh
, dtc
->dirty
);
1112 if (limit
> thresh
) {
1113 limit
-= (limit
- thresh
) >> 5;
1118 dom
->dirty_limit
= limit
;
1121 static void domain_update_bandwidth(struct dirty_throttle_control
*dtc
,
1124 struct wb_domain
*dom
= dtc_dom(dtc
);
1127 * check locklessly first to optimize away locking for the most time
1129 if (time_before(now
, dom
->dirty_limit_tstamp
+ BANDWIDTH_INTERVAL
))
1132 spin_lock(&dom
->lock
);
1133 if (time_after_eq(now
, dom
->dirty_limit_tstamp
+ BANDWIDTH_INTERVAL
)) {
1134 update_dirty_limit(dtc
);
1135 dom
->dirty_limit_tstamp
= now
;
1137 spin_unlock(&dom
->lock
);
1141 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1143 * Normal wb tasks will be curbed at or below it in long term.
1144 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1146 static void wb_update_dirty_ratelimit(struct dirty_throttle_control
*dtc
,
1147 unsigned long dirtied
,
1148 unsigned long elapsed
)
1150 struct bdi_writeback
*wb
= dtc
->wb
;
1151 unsigned long dirty
= dtc
->dirty
;
1152 unsigned long freerun
= dirty_freerun_ceiling(dtc
->thresh
, dtc
->bg_thresh
);
1153 unsigned long limit
= hard_dirty_limit(dtc_dom(dtc
), dtc
->thresh
);
1154 unsigned long setpoint
= (freerun
+ limit
) / 2;
1155 unsigned long write_bw
= wb
->avg_write_bandwidth
;
1156 unsigned long dirty_ratelimit
= wb
->dirty_ratelimit
;
1157 unsigned long dirty_rate
;
1158 unsigned long task_ratelimit
;
1159 unsigned long balanced_dirty_ratelimit
;
1164 * The dirty rate will match the writeout rate in long term, except
1165 * when dirty pages are truncated by userspace or re-dirtied by FS.
1167 dirty_rate
= (dirtied
- wb
->dirtied_stamp
) * HZ
/ elapsed
;
1170 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1172 task_ratelimit
= (u64
)dirty_ratelimit
*
1173 dtc
->pos_ratio
>> RATELIMIT_CALC_SHIFT
;
1174 task_ratelimit
++; /* it helps rampup dirty_ratelimit from tiny values */
1177 * A linear estimation of the "balanced" throttle rate. The theory is,
1178 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1179 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1180 * formula will yield the balanced rate limit (write_bw / N).
1182 * Note that the expanded form is not a pure rate feedback:
1183 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1184 * but also takes pos_ratio into account:
1185 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1187 * (1) is not realistic because pos_ratio also takes part in balancing
1188 * the dirty rate. Consider the state
1189 * pos_ratio = 0.5 (3)
1190 * rate = 2 * (write_bw / N) (4)
1191 * If (1) is used, it will stuck in that state! Because each dd will
1193 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1195 * dirty_rate = N * task_ratelimit = write_bw (6)
1196 * put (6) into (1) we get
1197 * rate_(i+1) = rate_(i) (7)
1199 * So we end up using (2) to always keep
1200 * rate_(i+1) ~= (write_bw / N) (8)
1201 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1202 * pos_ratio is able to drive itself to 1.0, which is not only where
1203 * the dirty count meet the setpoint, but also where the slope of
1204 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1206 balanced_dirty_ratelimit
= div_u64((u64
)task_ratelimit
* write_bw
,
1209 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1211 if (unlikely(balanced_dirty_ratelimit
> write_bw
))
1212 balanced_dirty_ratelimit
= write_bw
;
1215 * We could safely do this and return immediately:
1217 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1219 * However to get a more stable dirty_ratelimit, the below elaborated
1220 * code makes use of task_ratelimit to filter out singular points and
1221 * limit the step size.
1223 * The below code essentially only uses the relative value of
1225 * task_ratelimit - dirty_ratelimit
1226 * = (pos_ratio - 1) * dirty_ratelimit
1228 * which reflects the direction and size of dirty position error.
1232 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1233 * task_ratelimit is on the same side of dirty_ratelimit, too.
1235 * - dirty_ratelimit > balanced_dirty_ratelimit
1236 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1237 * lowering dirty_ratelimit will help meet both the position and rate
1238 * control targets. Otherwise, don't update dirty_ratelimit if it will
1239 * only help meet the rate target. After all, what the users ultimately
1240 * feel and care are stable dirty rate and small position error.
1242 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1243 * and filter out the singular points of balanced_dirty_ratelimit. Which
1244 * keeps jumping around randomly and can even leap far away at times
1245 * due to the small 200ms estimation period of dirty_rate (we want to
1246 * keep that period small to reduce time lags).
1251 * For strictlimit case, calculations above were based on wb counters
1252 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1253 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1254 * Hence, to calculate "step" properly, we have to use wb_dirty as
1255 * "dirty" and wb_setpoint as "setpoint".
1257 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1258 * it's possible that wb_thresh is close to zero due to inactivity
1259 * of backing device.
1261 if (unlikely(wb
->bdi
->capabilities
& BDI_CAP_STRICTLIMIT
)) {
1262 dirty
= dtc
->wb_dirty
;
1263 if (dtc
->wb_dirty
< 8)
1264 setpoint
= dtc
->wb_dirty
+ 1;
1266 setpoint
= (dtc
->wb_thresh
+ dtc
->wb_bg_thresh
) / 2;
1269 if (dirty
< setpoint
) {
1270 x
= min3(wb
->balanced_dirty_ratelimit
,
1271 balanced_dirty_ratelimit
, task_ratelimit
);
1272 if (dirty_ratelimit
< x
)
1273 step
= x
- dirty_ratelimit
;
1275 x
= max3(wb
->balanced_dirty_ratelimit
,
1276 balanced_dirty_ratelimit
, task_ratelimit
);
1277 if (dirty_ratelimit
> x
)
1278 step
= dirty_ratelimit
- x
;
1282 * Don't pursue 100% rate matching. It's impossible since the balanced
1283 * rate itself is constantly fluctuating. So decrease the track speed
1284 * when it gets close to the target. Helps eliminate pointless tremors.
1286 step
>>= dirty_ratelimit
/ (2 * step
+ 1);
1288 * Limit the tracking speed to avoid overshooting.
1290 step
= (step
+ 7) / 8;
1292 if (dirty_ratelimit
< balanced_dirty_ratelimit
)
1293 dirty_ratelimit
+= step
;
1295 dirty_ratelimit
-= step
;
1297 wb
->dirty_ratelimit
= max(dirty_ratelimit
, 1UL);
1298 wb
->balanced_dirty_ratelimit
= balanced_dirty_ratelimit
;
1300 trace_bdi_dirty_ratelimit(wb
, dirty_rate
, task_ratelimit
);
1303 static void __wb_update_bandwidth(struct dirty_throttle_control
*gdtc
,
1304 struct dirty_throttle_control
*mdtc
,
1305 unsigned long start_time
,
1306 bool update_ratelimit
)
1308 struct bdi_writeback
*wb
= gdtc
->wb
;
1309 unsigned long now
= jiffies
;
1310 unsigned long elapsed
= now
- wb
->bw_time_stamp
;
1311 unsigned long dirtied
;
1312 unsigned long written
;
1314 lockdep_assert_held(&wb
->list_lock
);
1317 * rate-limit, only update once every 200ms.
1319 if (elapsed
< BANDWIDTH_INTERVAL
)
1322 dirtied
= percpu_counter_read(&wb
->stat
[WB_DIRTIED
]);
1323 written
= percpu_counter_read(&wb
->stat
[WB_WRITTEN
]);
1326 * Skip quiet periods when disk bandwidth is under-utilized.
1327 * (at least 1s idle time between two flusher runs)
1329 if (elapsed
> HZ
&& time_before(wb
->bw_time_stamp
, start_time
))
1332 if (update_ratelimit
) {
1333 domain_update_bandwidth(gdtc
, now
);
1334 wb_update_dirty_ratelimit(gdtc
, dirtied
, elapsed
);
1337 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1338 * compiler has no way to figure that out. Help it.
1340 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK
) && mdtc
) {
1341 domain_update_bandwidth(mdtc
, now
);
1342 wb_update_dirty_ratelimit(mdtc
, dirtied
, elapsed
);
1345 wb_update_write_bandwidth(wb
, elapsed
, written
);
1348 wb
->dirtied_stamp
= dirtied
;
1349 wb
->written_stamp
= written
;
1350 wb
->bw_time_stamp
= now
;
1353 void wb_update_bandwidth(struct bdi_writeback
*wb
, unsigned long start_time
)
1355 struct dirty_throttle_control gdtc
= { GDTC_INIT(wb
) };
1357 __wb_update_bandwidth(&gdtc
, NULL
, start_time
, false);
1361 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1362 * will look to see if it needs to start dirty throttling.
1364 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1365 * global_page_state() too often. So scale it near-sqrt to the safety margin
1366 * (the number of pages we may dirty without exceeding the dirty limits).
1368 static unsigned long dirty_poll_interval(unsigned long dirty
,
1369 unsigned long thresh
)
1372 return 1UL << (ilog2(thresh
- dirty
) >> 1);
1377 static unsigned long wb_max_pause(struct bdi_writeback
*wb
,
1378 unsigned long wb_dirty
)
1380 unsigned long bw
= wb
->avg_write_bandwidth
;
1384 * Limit pause time for small memory systems. If sleeping for too long
1385 * time, a small pool of dirty/writeback pages may go empty and disk go
1388 * 8 serves as the safety ratio.
1390 t
= wb_dirty
/ (1 + bw
/ roundup_pow_of_two(1 + HZ
/ 8));
1393 return min_t(unsigned long, t
, MAX_PAUSE
);
1396 static long wb_min_pause(struct bdi_writeback
*wb
,
1398 unsigned long task_ratelimit
,
1399 unsigned long dirty_ratelimit
,
1400 int *nr_dirtied_pause
)
1402 long hi
= ilog2(wb
->avg_write_bandwidth
);
1403 long lo
= ilog2(wb
->dirty_ratelimit
);
1404 long t
; /* target pause */
1405 long pause
; /* estimated next pause */
1406 int pages
; /* target nr_dirtied_pause */
1408 /* target for 10ms pause on 1-dd case */
1409 t
= max(1, HZ
/ 100);
1412 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1415 * (N * 10ms) on 2^N concurrent tasks.
1418 t
+= (hi
- lo
) * (10 * HZ
) / 1024;
1421 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1422 * on the much more stable dirty_ratelimit. However the next pause time
1423 * will be computed based on task_ratelimit and the two rate limits may
1424 * depart considerably at some time. Especially if task_ratelimit goes
1425 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1426 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1427 * result task_ratelimit won't be executed faithfully, which could
1428 * eventually bring down dirty_ratelimit.
1430 * We apply two rules to fix it up:
1431 * 1) try to estimate the next pause time and if necessary, use a lower
1432 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1433 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1434 * 2) limit the target pause time to max_pause/2, so that the normal
1435 * small fluctuations of task_ratelimit won't trigger rule (1) and
1436 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1438 t
= min(t
, 1 + max_pause
/ 2);
1439 pages
= dirty_ratelimit
* t
/ roundup_pow_of_two(HZ
);
1442 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1443 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1444 * When the 16 consecutive reads are often interrupted by some dirty
1445 * throttling pause during the async writes, cfq will go into idles
1446 * (deadline is fine). So push nr_dirtied_pause as high as possible
1447 * until reaches DIRTY_POLL_THRESH=32 pages.
1449 if (pages
< DIRTY_POLL_THRESH
) {
1451 pages
= dirty_ratelimit
* t
/ roundup_pow_of_two(HZ
);
1452 if (pages
> DIRTY_POLL_THRESH
) {
1453 pages
= DIRTY_POLL_THRESH
;
1454 t
= HZ
* DIRTY_POLL_THRESH
/ dirty_ratelimit
;
1458 pause
= HZ
* pages
/ (task_ratelimit
+ 1);
1459 if (pause
> max_pause
) {
1461 pages
= task_ratelimit
* t
/ roundup_pow_of_two(HZ
);
1464 *nr_dirtied_pause
= pages
;
1466 * The minimal pause time will normally be half the target pause time.
1468 return pages
>= DIRTY_POLL_THRESH
? 1 + t
/ 2 : t
;
1471 static inline void wb_dirty_limits(struct dirty_throttle_control
*dtc
)
1473 struct bdi_writeback
*wb
= dtc
->wb
;
1474 unsigned long wb_reclaimable
;
1477 * wb_thresh is not treated as some limiting factor as
1478 * dirty_thresh, due to reasons
1479 * - in JBOD setup, wb_thresh can fluctuate a lot
1480 * - in a system with HDD and USB key, the USB key may somehow
1481 * go into state (wb_dirty >> wb_thresh) either because
1482 * wb_dirty starts high, or because wb_thresh drops low.
1483 * In this case we don't want to hard throttle the USB key
1484 * dirtiers for 100 seconds until wb_dirty drops under
1485 * wb_thresh. Instead the auxiliary wb control line in
1486 * wb_position_ratio() will let the dirtier task progress
1487 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1489 dtc
->wb_thresh
= __wb_calc_thresh(dtc
);
1490 dtc
->wb_bg_thresh
= dtc
->thresh
?
1491 div_u64((u64
)dtc
->wb_thresh
* dtc
->bg_thresh
, dtc
->thresh
) : 0;
1494 * In order to avoid the stacked BDI deadlock we need
1495 * to ensure we accurately count the 'dirty' pages when
1496 * the threshold is low.
1498 * Otherwise it would be possible to get thresh+n pages
1499 * reported dirty, even though there are thresh-m pages
1500 * actually dirty; with m+n sitting in the percpu
1503 if (dtc
->wb_thresh
< 2 * wb_stat_error(wb
)) {
1504 wb_reclaimable
= wb_stat_sum(wb
, WB_RECLAIMABLE
);
1505 dtc
->wb_dirty
= wb_reclaimable
+ wb_stat_sum(wb
, WB_WRITEBACK
);
1507 wb_reclaimable
= wb_stat(wb
, WB_RECLAIMABLE
);
1508 dtc
->wb_dirty
= wb_reclaimable
+ wb_stat(wb
, WB_WRITEBACK
);
1513 * balance_dirty_pages() must be called by processes which are generating dirty
1514 * data. It looks at the number of dirty pages in the machine and will force
1515 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1516 * If we're over `background_thresh' then the writeback threads are woken to
1517 * perform some writeout.
1519 static void balance_dirty_pages(struct address_space
*mapping
,
1520 struct bdi_writeback
*wb
,
1521 unsigned long pages_dirtied
)
1523 struct dirty_throttle_control gdtc_stor
= { GDTC_INIT(wb
) };
1524 struct dirty_throttle_control mdtc_stor
= { MDTC_INIT(wb
, &gdtc_stor
) };
1525 struct dirty_throttle_control
* const gdtc
= &gdtc_stor
;
1526 struct dirty_throttle_control
* const mdtc
= mdtc_valid(&mdtc_stor
) ?
1528 struct dirty_throttle_control
*sdtc
;
1529 unsigned long nr_reclaimable
; /* = file_dirty + unstable_nfs */
1534 int nr_dirtied_pause
;
1535 bool dirty_exceeded
= false;
1536 unsigned long task_ratelimit
;
1537 unsigned long dirty_ratelimit
;
1538 struct backing_dev_info
*bdi
= wb
->bdi
;
1539 bool strictlimit
= bdi
->capabilities
& BDI_CAP_STRICTLIMIT
;
1540 unsigned long start_time
= jiffies
;
1543 unsigned long now
= jiffies
;
1544 unsigned long dirty
, thresh
, bg_thresh
;
1545 unsigned long m_dirty
, m_thresh
, m_bg_thresh
;
1548 * Unstable writes are a feature of certain networked
1549 * filesystems (i.e. NFS) in which data may have been
1550 * written to the server's write cache, but has not yet
1551 * been flushed to permanent storage.
1553 nr_reclaimable
= global_page_state(NR_FILE_DIRTY
) +
1554 global_page_state(NR_UNSTABLE_NFS
);
1555 gdtc
->avail
= global_dirtyable_memory();
1556 gdtc
->dirty
= nr_reclaimable
+ global_page_state(NR_WRITEBACK
);
1558 domain_dirty_limits(gdtc
);
1560 if (unlikely(strictlimit
)) {
1561 wb_dirty_limits(gdtc
);
1563 dirty
= gdtc
->wb_dirty
;
1564 thresh
= gdtc
->wb_thresh
;
1565 bg_thresh
= gdtc
->wb_bg_thresh
;
1567 dirty
= gdtc
->dirty
;
1568 thresh
= gdtc
->thresh
;
1569 bg_thresh
= gdtc
->bg_thresh
;
1573 unsigned long filepages
, headroom
, writeback
;
1576 * If @wb belongs to !root memcg, repeat the same
1577 * basic calculations for the memcg domain.
1579 mem_cgroup_wb_stats(wb
, &filepages
, &headroom
,
1580 &mdtc
->dirty
, &writeback
);
1581 mdtc
->dirty
+= writeback
;
1582 mdtc_calc_avail(mdtc
, filepages
, headroom
);
1584 domain_dirty_limits(mdtc
);
1586 if (unlikely(strictlimit
)) {
1587 wb_dirty_limits(mdtc
);
1588 m_dirty
= mdtc
->wb_dirty
;
1589 m_thresh
= mdtc
->wb_thresh
;
1590 m_bg_thresh
= mdtc
->wb_bg_thresh
;
1592 m_dirty
= mdtc
->dirty
;
1593 m_thresh
= mdtc
->thresh
;
1594 m_bg_thresh
= mdtc
->bg_thresh
;
1599 * Throttle it only when the background writeback cannot
1600 * catch-up. This avoids (excessively) small writeouts
1601 * when the wb limits are ramping up in case of !strictlimit.
1603 * In strictlimit case make decision based on the wb counters
1604 * and limits. Small writeouts when the wb limits are ramping
1605 * up are the price we consciously pay for strictlimit-ing.
1607 * If memcg domain is in effect, @dirty should be under
1608 * both global and memcg freerun ceilings.
1610 if (dirty
<= dirty_freerun_ceiling(thresh
, bg_thresh
) &&
1612 m_dirty
<= dirty_freerun_ceiling(m_thresh
, m_bg_thresh
))) {
1613 unsigned long intv
= dirty_poll_interval(dirty
, thresh
);
1614 unsigned long m_intv
= ULONG_MAX
;
1616 current
->dirty_paused_when
= now
;
1617 current
->nr_dirtied
= 0;
1619 m_intv
= dirty_poll_interval(m_dirty
, m_thresh
);
1620 current
->nr_dirtied_pause
= min(intv
, m_intv
);
1624 if (unlikely(!writeback_in_progress(wb
)))
1625 wb_start_background_writeback(wb
);
1628 * Calculate global domain's pos_ratio and select the
1629 * global dtc by default.
1632 wb_dirty_limits(gdtc
);
1634 dirty_exceeded
= (gdtc
->wb_dirty
> gdtc
->wb_thresh
) &&
1635 ((gdtc
->dirty
> gdtc
->thresh
) || strictlimit
);
1637 wb_position_ratio(gdtc
);
1642 * If memcg domain is in effect, calculate its
1643 * pos_ratio. @wb should satisfy constraints from
1644 * both global and memcg domains. Choose the one
1645 * w/ lower pos_ratio.
1648 wb_dirty_limits(mdtc
);
1650 dirty_exceeded
|= (mdtc
->wb_dirty
> mdtc
->wb_thresh
) &&
1651 ((mdtc
->dirty
> mdtc
->thresh
) || strictlimit
);
1653 wb_position_ratio(mdtc
);
1654 if (mdtc
->pos_ratio
< gdtc
->pos_ratio
)
1658 if (dirty_exceeded
&& !wb
->dirty_exceeded
)
1659 wb
->dirty_exceeded
= 1;
1661 if (time_is_before_jiffies(wb
->bw_time_stamp
+
1662 BANDWIDTH_INTERVAL
)) {
1663 spin_lock(&wb
->list_lock
);
1664 __wb_update_bandwidth(gdtc
, mdtc
, start_time
, true);
1665 spin_unlock(&wb
->list_lock
);
1668 /* throttle according to the chosen dtc */
1669 dirty_ratelimit
= wb
->dirty_ratelimit
;
1670 task_ratelimit
= ((u64
)dirty_ratelimit
* sdtc
->pos_ratio
) >>
1671 RATELIMIT_CALC_SHIFT
;
1672 max_pause
= wb_max_pause(wb
, sdtc
->wb_dirty
);
1673 min_pause
= wb_min_pause(wb
, max_pause
,
1674 task_ratelimit
, dirty_ratelimit
,
1677 if (unlikely(task_ratelimit
== 0)) {
1682 period
= HZ
* pages_dirtied
/ task_ratelimit
;
1684 if (current
->dirty_paused_when
)
1685 pause
-= now
- current
->dirty_paused_when
;
1687 * For less than 1s think time (ext3/4 may block the dirtier
1688 * for up to 800ms from time to time on 1-HDD; so does xfs,
1689 * however at much less frequency), try to compensate it in
1690 * future periods by updating the virtual time; otherwise just
1691 * do a reset, as it may be a light dirtier.
1693 if (pause
< min_pause
) {
1694 trace_balance_dirty_pages(wb
,
1707 current
->dirty_paused_when
= now
;
1708 current
->nr_dirtied
= 0;
1709 } else if (period
) {
1710 current
->dirty_paused_when
+= period
;
1711 current
->nr_dirtied
= 0;
1712 } else if (current
->nr_dirtied_pause
<= pages_dirtied
)
1713 current
->nr_dirtied_pause
+= pages_dirtied
;
1716 if (unlikely(pause
> max_pause
)) {
1717 /* for occasional dropped task_ratelimit */
1718 now
+= min(pause
- max_pause
, max_pause
);
1723 trace_balance_dirty_pages(wb
,
1735 __set_current_state(TASK_KILLABLE
);
1736 io_schedule_timeout(pause
);
1738 current
->dirty_paused_when
= now
+ pause
;
1739 current
->nr_dirtied
= 0;
1740 current
->nr_dirtied_pause
= nr_dirtied_pause
;
1743 * This is typically equal to (dirty < thresh) and can also
1744 * keep "1000+ dd on a slow USB stick" under control.
1750 * In the case of an unresponding NFS server and the NFS dirty
1751 * pages exceeds dirty_thresh, give the other good wb's a pipe
1752 * to go through, so that tasks on them still remain responsive.
1754 * In theory 1 page is enough to keep the comsumer-producer
1755 * pipe going: the flusher cleans 1 page => the task dirties 1
1756 * more page. However wb_dirty has accounting errors. So use
1757 * the larger and more IO friendly wb_stat_error.
1759 if (sdtc
->wb_dirty
<= wb_stat_error(wb
))
1762 if (fatal_signal_pending(current
))
1766 if (!dirty_exceeded
&& wb
->dirty_exceeded
)
1767 wb
->dirty_exceeded
= 0;
1769 if (writeback_in_progress(wb
))
1773 * In laptop mode, we wait until hitting the higher threshold before
1774 * starting background writeout, and then write out all the way down
1775 * to the lower threshold. So slow writers cause minimal disk activity.
1777 * In normal mode, we start background writeout at the lower
1778 * background_thresh, to keep the amount of dirty memory low.
1783 if (nr_reclaimable
> gdtc
->bg_thresh
)
1784 wb_start_background_writeback(wb
);
1787 static DEFINE_PER_CPU(int, bdp_ratelimits
);
1790 * Normal tasks are throttled by
1792 * dirty tsk->nr_dirtied_pause pages;
1793 * take a snap in balance_dirty_pages();
1795 * However there is a worst case. If every task exit immediately when dirtied
1796 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1797 * called to throttle the page dirties. The solution is to save the not yet
1798 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1799 * randomly into the running tasks. This works well for the above worst case,
1800 * as the new task will pick up and accumulate the old task's leaked dirty
1801 * count and eventually get throttled.
1803 DEFINE_PER_CPU(int, dirty_throttle_leaks
) = 0;
1806 * balance_dirty_pages_ratelimited - balance dirty memory state
1807 * @mapping: address_space which was dirtied
1809 * Processes which are dirtying memory should call in here once for each page
1810 * which was newly dirtied. The function will periodically check the system's
1811 * dirty state and will initiate writeback if needed.
1813 * On really big machines, get_writeback_state is expensive, so try to avoid
1814 * calling it too often (ratelimiting). But once we're over the dirty memory
1815 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1816 * from overshooting the limit by (ratelimit_pages) each.
1818 void balance_dirty_pages_ratelimited(struct address_space
*mapping
)
1820 struct inode
*inode
= mapping
->host
;
1821 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
1822 struct bdi_writeback
*wb
= NULL
;
1826 if (!bdi_cap_account_dirty(bdi
))
1829 if (inode_cgwb_enabled(inode
))
1830 wb
= wb_get_create_current(bdi
, GFP_KERNEL
);
1834 ratelimit
= current
->nr_dirtied_pause
;
1835 if (wb
->dirty_exceeded
)
1836 ratelimit
= min(ratelimit
, 32 >> (PAGE_SHIFT
- 10));
1840 * This prevents one CPU to accumulate too many dirtied pages without
1841 * calling into balance_dirty_pages(), which can happen when there are
1842 * 1000+ tasks, all of them start dirtying pages at exactly the same
1843 * time, hence all honoured too large initial task->nr_dirtied_pause.
1845 p
= this_cpu_ptr(&bdp_ratelimits
);
1846 if (unlikely(current
->nr_dirtied
>= ratelimit
))
1848 else if (unlikely(*p
>= ratelimit_pages
)) {
1853 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1854 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1855 * the dirty throttling and livelock other long-run dirtiers.
1857 p
= this_cpu_ptr(&dirty_throttle_leaks
);
1858 if (*p
> 0 && current
->nr_dirtied
< ratelimit
) {
1859 unsigned long nr_pages_dirtied
;
1860 nr_pages_dirtied
= min(*p
, ratelimit
- current
->nr_dirtied
);
1861 *p
-= nr_pages_dirtied
;
1862 current
->nr_dirtied
+= nr_pages_dirtied
;
1866 if (unlikely(current
->nr_dirtied
>= ratelimit
))
1867 balance_dirty_pages(mapping
, wb
, current
->nr_dirtied
);
1871 EXPORT_SYMBOL(balance_dirty_pages_ratelimited
);
1874 * wb_over_bg_thresh - does @wb need to be written back?
1875 * @wb: bdi_writeback of interest
1877 * Determines whether background writeback should keep writing @wb or it's
1878 * clean enough. Returns %true if writeback should continue.
1880 bool wb_over_bg_thresh(struct bdi_writeback
*wb
)
1882 struct dirty_throttle_control gdtc_stor
= { GDTC_INIT(wb
) };
1883 struct dirty_throttle_control mdtc_stor
= { MDTC_INIT(wb
, &gdtc_stor
) };
1884 struct dirty_throttle_control
* const gdtc
= &gdtc_stor
;
1885 struct dirty_throttle_control
* const mdtc
= mdtc_valid(&mdtc_stor
) ?
1889 * Similar to balance_dirty_pages() but ignores pages being written
1890 * as we're trying to decide whether to put more under writeback.
1892 gdtc
->avail
= global_dirtyable_memory();
1893 gdtc
->dirty
= global_page_state(NR_FILE_DIRTY
) +
1894 global_page_state(NR_UNSTABLE_NFS
);
1895 domain_dirty_limits(gdtc
);
1897 if (gdtc
->dirty
> gdtc
->bg_thresh
)
1900 if (wb_stat(wb
, WB_RECLAIMABLE
) > __wb_calc_thresh(gdtc
))
1904 unsigned long filepages
, headroom
, writeback
;
1906 mem_cgroup_wb_stats(wb
, &filepages
, &headroom
, &mdtc
->dirty
,
1908 mdtc_calc_avail(mdtc
, filepages
, headroom
);
1909 domain_dirty_limits(mdtc
); /* ditto, ignore writeback */
1911 if (mdtc
->dirty
> mdtc
->bg_thresh
)
1914 if (wb_stat(wb
, WB_RECLAIMABLE
) > __wb_calc_thresh(mdtc
))
1921 void throttle_vm_writeout(gfp_t gfp_mask
)
1923 unsigned long background_thresh
;
1924 unsigned long dirty_thresh
;
1927 global_dirty_limits(&background_thresh
, &dirty_thresh
);
1928 dirty_thresh
= hard_dirty_limit(&global_wb_domain
, dirty_thresh
);
1931 * Boost the allowable dirty threshold a bit for page
1932 * allocators so they don't get DoS'ed by heavy writers
1934 dirty_thresh
+= dirty_thresh
/ 10; /* wheeee... */
1936 if (global_page_state(NR_UNSTABLE_NFS
) +
1937 global_page_state(NR_WRITEBACK
) <= dirty_thresh
)
1939 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1942 * The caller might hold locks which can prevent IO completion
1943 * or progress in the filesystem. So we cannot just sit here
1944 * waiting for IO to complete.
1946 if ((gfp_mask
& (__GFP_FS
|__GFP_IO
)) != (__GFP_FS
|__GFP_IO
))
1952 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1954 int dirty_writeback_centisecs_handler(struct ctl_table
*table
, int write
,
1955 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1957 proc_dointvec(table
, write
, buffer
, length
, ppos
);
1962 void laptop_mode_timer_fn(unsigned long data
)
1964 struct request_queue
*q
= (struct request_queue
*)data
;
1965 int nr_pages
= global_page_state(NR_FILE_DIRTY
) +
1966 global_page_state(NR_UNSTABLE_NFS
);
1967 struct bdi_writeback
*wb
;
1970 * We want to write everything out, not just down to the dirty
1973 if (!bdi_has_dirty_io(&q
->backing_dev_info
))
1977 list_for_each_entry_rcu(wb
, &q
->backing_dev_info
.wb_list
, bdi_node
)
1978 if (wb_has_dirty_io(wb
))
1979 wb_start_writeback(wb
, nr_pages
, true,
1980 WB_REASON_LAPTOP_TIMER
);
1985 * We've spun up the disk and we're in laptop mode: schedule writeback
1986 * of all dirty data a few seconds from now. If the flush is already scheduled
1987 * then push it back - the user is still using the disk.
1989 void laptop_io_completion(struct backing_dev_info
*info
)
1991 mod_timer(&info
->laptop_mode_wb_timer
, jiffies
+ laptop_mode
);
1995 * We're in laptop mode and we've just synced. The sync's writes will have
1996 * caused another writeback to be scheduled by laptop_io_completion.
1997 * Nothing needs to be written back anymore, so we unschedule the writeback.
1999 void laptop_sync_completion(void)
2001 struct backing_dev_info
*bdi
;
2005 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
)
2006 del_timer(&bdi
->laptop_mode_wb_timer
);
2013 * If ratelimit_pages is too high then we can get into dirty-data overload
2014 * if a large number of processes all perform writes at the same time.
2015 * If it is too low then SMP machines will call the (expensive)
2016 * get_writeback_state too often.
2018 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2019 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2023 void writeback_set_ratelimit(void)
2025 struct wb_domain
*dom
= &global_wb_domain
;
2026 unsigned long background_thresh
;
2027 unsigned long dirty_thresh
;
2029 global_dirty_limits(&background_thresh
, &dirty_thresh
);
2030 dom
->dirty_limit
= dirty_thresh
;
2031 ratelimit_pages
= dirty_thresh
/ (num_online_cpus() * 32);
2032 if (ratelimit_pages
< 16)
2033 ratelimit_pages
= 16;
2037 ratelimit_handler(struct notifier_block
*self
, unsigned long action
,
2041 switch (action
& ~CPU_TASKS_FROZEN
) {
2044 writeback_set_ratelimit();
2051 static struct notifier_block ratelimit_nb
= {
2052 .notifier_call
= ratelimit_handler
,
2057 * Called early on to tune the page writeback dirty limits.
2059 * We used to scale dirty pages according to how total memory
2060 * related to pages that could be allocated for buffers (by
2061 * comparing nr_free_buffer_pages() to vm_total_pages.
2063 * However, that was when we used "dirty_ratio" to scale with
2064 * all memory, and we don't do that any more. "dirty_ratio"
2065 * is now applied to total non-HIGHPAGE memory (by subtracting
2066 * totalhigh_pages from vm_total_pages), and as such we can't
2067 * get into the old insane situation any more where we had
2068 * large amounts of dirty pages compared to a small amount of
2069 * non-HIGHMEM memory.
2071 * But we might still want to scale the dirty_ratio by how
2072 * much memory the box has..
2074 void __init
page_writeback_init(void)
2076 BUG_ON(wb_domain_init(&global_wb_domain
, GFP_KERNEL
));
2078 writeback_set_ratelimit();
2079 register_cpu_notifier(&ratelimit_nb
);
2083 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2084 * @mapping: address space structure to write
2085 * @start: starting page index
2086 * @end: ending page index (inclusive)
2088 * This function scans the page range from @start to @end (inclusive) and tags
2089 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2090 * that write_cache_pages (or whoever calls this function) will then use
2091 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2092 * used to avoid livelocking of writeback by a process steadily creating new
2093 * dirty pages in the file (thus it is important for this function to be quick
2094 * so that it can tag pages faster than a dirtying process can create them).
2097 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
2099 void tag_pages_for_writeback(struct address_space
*mapping
,
2100 pgoff_t start
, pgoff_t end
)
2102 #define WRITEBACK_TAG_BATCH 4096
2103 unsigned long tagged
;
2106 spin_lock_irq(&mapping
->tree_lock
);
2107 tagged
= radix_tree_range_tag_if_tagged(&mapping
->page_tree
,
2108 &start
, end
, WRITEBACK_TAG_BATCH
,
2109 PAGECACHE_TAG_DIRTY
, PAGECACHE_TAG_TOWRITE
);
2110 spin_unlock_irq(&mapping
->tree_lock
);
2111 WARN_ON_ONCE(tagged
> WRITEBACK_TAG_BATCH
);
2113 /* We check 'start' to handle wrapping when end == ~0UL */
2114 } while (tagged
>= WRITEBACK_TAG_BATCH
&& start
);
2116 EXPORT_SYMBOL(tag_pages_for_writeback
);
2119 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2120 * @mapping: address space structure to write
2121 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2122 * @writepage: function called for each page
2123 * @data: data passed to writepage function
2125 * If a page is already under I/O, write_cache_pages() skips it, even
2126 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2127 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2128 * and msync() need to guarantee that all the data which was dirty at the time
2129 * the call was made get new I/O started against them. If wbc->sync_mode is
2130 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2131 * existing IO to complete.
2133 * To avoid livelocks (when other process dirties new pages), we first tag
2134 * pages which should be written back with TOWRITE tag and only then start
2135 * writing them. For data-integrity sync we have to be careful so that we do
2136 * not miss some pages (e.g., because some other process has cleared TOWRITE
2137 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2138 * by the process clearing the DIRTY tag (and submitting the page for IO).
2140 int write_cache_pages(struct address_space
*mapping
,
2141 struct writeback_control
*wbc
, writepage_t writepage
,
2146 struct pagevec pvec
;
2148 pgoff_t
uninitialized_var(writeback_index
);
2150 pgoff_t end
; /* Inclusive */
2153 int range_whole
= 0;
2156 pagevec_init(&pvec
, 0);
2157 if (wbc
->range_cyclic
) {
2158 writeback_index
= mapping
->writeback_index
; /* prev offset */
2159 index
= writeback_index
;
2166 index
= wbc
->range_start
>> PAGE_CACHE_SHIFT
;
2167 end
= wbc
->range_end
>> PAGE_CACHE_SHIFT
;
2168 if (wbc
->range_start
== 0 && wbc
->range_end
== LLONG_MAX
)
2170 cycled
= 1; /* ignore range_cyclic tests */
2172 if (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
)
2173 tag
= PAGECACHE_TAG_TOWRITE
;
2175 tag
= PAGECACHE_TAG_DIRTY
;
2177 if (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
)
2178 tag_pages_for_writeback(mapping
, index
, end
);
2180 while (!done
&& (index
<= end
)) {
2183 nr_pages
= pagevec_lookup_tag(&pvec
, mapping
, &index
, tag
,
2184 min(end
- index
, (pgoff_t
)PAGEVEC_SIZE
-1) + 1);
2188 for (i
= 0; i
< nr_pages
; i
++) {
2189 struct page
*page
= pvec
.pages
[i
];
2192 * At this point, the page may be truncated or
2193 * invalidated (changing page->mapping to NULL), or
2194 * even swizzled back from swapper_space to tmpfs file
2195 * mapping. However, page->index will not change
2196 * because we have a reference on the page.
2198 if (page
->index
> end
) {
2200 * can't be range_cyclic (1st pass) because
2201 * end == -1 in that case.
2207 done_index
= page
->index
;
2212 * Page truncated or invalidated. We can freely skip it
2213 * then, even for data integrity operations: the page
2214 * has disappeared concurrently, so there could be no
2215 * real expectation of this data interity operation
2216 * even if there is now a new, dirty page at the same
2217 * pagecache address.
2219 if (unlikely(page
->mapping
!= mapping
)) {
2225 if (!PageDirty(page
)) {
2226 /* someone wrote it for us */
2227 goto continue_unlock
;
2230 if (PageWriteback(page
)) {
2231 if (wbc
->sync_mode
!= WB_SYNC_NONE
)
2232 wait_on_page_writeback(page
);
2234 goto continue_unlock
;
2237 BUG_ON(PageWriteback(page
));
2238 if (!clear_page_dirty_for_io(page
))
2239 goto continue_unlock
;
2241 trace_wbc_writepage(wbc
, inode_to_bdi(mapping
->host
));
2242 ret
= (*writepage
)(page
, wbc
, data
);
2243 if (unlikely(ret
)) {
2244 if (ret
== AOP_WRITEPAGE_ACTIVATE
) {
2249 * done_index is set past this page,
2250 * so media errors will not choke
2251 * background writeout for the entire
2252 * file. This has consequences for
2253 * range_cyclic semantics (ie. it may
2254 * not be suitable for data integrity
2257 done_index
= page
->index
+ 1;
2264 * We stop writing back only if we are not doing
2265 * integrity sync. In case of integrity sync we have to
2266 * keep going until we have written all the pages
2267 * we tagged for writeback prior to entering this loop.
2269 if (--wbc
->nr_to_write
<= 0 &&
2270 wbc
->sync_mode
== WB_SYNC_NONE
) {
2275 pagevec_release(&pvec
);
2278 if (!cycled
&& !done
) {
2281 * We hit the last page and there is more work to be done: wrap
2282 * back to the start of the file
2286 end
= writeback_index
- 1;
2289 if (wbc
->range_cyclic
|| (range_whole
&& wbc
->nr_to_write
> 0))
2290 mapping
->writeback_index
= done_index
;
2294 EXPORT_SYMBOL(write_cache_pages
);
2297 * Function used by generic_writepages to call the real writepage
2298 * function and set the mapping flags on error
2300 static int __writepage(struct page
*page
, struct writeback_control
*wbc
,
2303 struct address_space
*mapping
= data
;
2304 int ret
= mapping
->a_ops
->writepage(page
, wbc
);
2305 mapping_set_error(mapping
, ret
);
2310 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2311 * @mapping: address space structure to write
2312 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2314 * This is a library function, which implements the writepages()
2315 * address_space_operation.
2317 int generic_writepages(struct address_space
*mapping
,
2318 struct writeback_control
*wbc
)
2320 struct blk_plug plug
;
2323 /* deal with chardevs and other special file */
2324 if (!mapping
->a_ops
->writepage
)
2327 blk_start_plug(&plug
);
2328 ret
= write_cache_pages(mapping
, wbc
, __writepage
, mapping
);
2329 blk_finish_plug(&plug
);
2333 EXPORT_SYMBOL(generic_writepages
);
2335 int do_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
2339 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
);
2349 * write_one_page - write out a single page and optionally wait on I/O
2350 * @page: the page to write
2351 * @wait: if true, wait on writeout
2353 * The page must be locked by the caller and will be unlocked upon return.
2355 * write_one_page() returns a negative error code if I/O failed.
2357 int write_one_page(struct page
*page
, int wait
)
2359 struct address_space
*mapping
= page
->mapping
;
2361 struct writeback_control wbc
= {
2362 .sync_mode
= WB_SYNC_ALL
,
2366 BUG_ON(!PageLocked(page
));
2369 wait_on_page_writeback(page
);
2371 if (clear_page_dirty_for_io(page
)) {
2372 page_cache_get(page
);
2373 ret
= mapping
->a_ops
->writepage(page
, &wbc
);
2374 if (ret
== 0 && wait
) {
2375 wait_on_page_writeback(page
);
2376 if (PageError(page
))
2379 page_cache_release(page
);
2385 EXPORT_SYMBOL(write_one_page
);
2388 * For address_spaces which do not use buffers nor write back.
2390 int __set_page_dirty_no_writeback(struct page
*page
)
2392 if (!PageDirty(page
))
2393 return !TestSetPageDirty(page
);
2398 * Helper function for set_page_dirty family.
2400 * Caller must hold mem_cgroup_begin_page_stat().
2402 * NOTE: This relies on being atomic wrt interrupts.
2404 void account_page_dirtied(struct page
*page
, struct address_space
*mapping
,
2405 struct mem_cgroup
*memcg
)
2407 struct inode
*inode
= mapping
->host
;
2409 trace_writeback_dirty_page(page
, mapping
);
2411 if (mapping_cap_account_dirty(mapping
)) {
2412 struct bdi_writeback
*wb
;
2414 inode_attach_wb(inode
, page
);
2415 wb
= inode_to_wb(inode
);
2417 mem_cgroup_inc_page_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
2418 __inc_zone_page_state(page
, NR_FILE_DIRTY
);
2419 __inc_zone_page_state(page
, NR_DIRTIED
);
2420 __inc_wb_stat(wb
, WB_RECLAIMABLE
);
2421 __inc_wb_stat(wb
, WB_DIRTIED
);
2422 task_io_account_write(PAGE_CACHE_SIZE
);
2423 current
->nr_dirtied
++;
2424 this_cpu_inc(bdp_ratelimits
);
2427 EXPORT_SYMBOL(account_page_dirtied
);
2430 * Helper function for deaccounting dirty page without writeback.
2432 * Caller must hold mem_cgroup_begin_page_stat().
2434 void account_page_cleaned(struct page
*page
, struct address_space
*mapping
,
2435 struct mem_cgroup
*memcg
, struct bdi_writeback
*wb
)
2437 if (mapping_cap_account_dirty(mapping
)) {
2438 mem_cgroup_dec_page_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
2439 dec_zone_page_state(page
, NR_FILE_DIRTY
);
2440 dec_wb_stat(wb
, WB_RECLAIMABLE
);
2441 task_io_account_cancelled_write(PAGE_CACHE_SIZE
);
2446 * For address_spaces which do not use buffers. Just tag the page as dirty in
2449 * This is also used when a single buffer is being dirtied: we want to set the
2450 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2451 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2453 * The caller must ensure this doesn't race with truncation. Most will simply
2454 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2455 * the pte lock held, which also locks out truncation.
2457 int __set_page_dirty_nobuffers(struct page
*page
)
2459 struct mem_cgroup
*memcg
;
2461 memcg
= mem_cgroup_begin_page_stat(page
);
2462 if (!TestSetPageDirty(page
)) {
2463 struct address_space
*mapping
= page_mapping(page
);
2464 unsigned long flags
;
2467 mem_cgroup_end_page_stat(memcg
);
2471 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
2472 BUG_ON(page_mapping(page
) != mapping
);
2473 WARN_ON_ONCE(!PagePrivate(page
) && !PageUptodate(page
));
2474 account_page_dirtied(page
, mapping
, memcg
);
2475 radix_tree_tag_set(&mapping
->page_tree
, page_index(page
),
2476 PAGECACHE_TAG_DIRTY
);
2477 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
2478 mem_cgroup_end_page_stat(memcg
);
2480 if (mapping
->host
) {
2481 /* !PageAnon && !swapper_space */
2482 __mark_inode_dirty(mapping
->host
, I_DIRTY_PAGES
);
2486 mem_cgroup_end_page_stat(memcg
);
2489 EXPORT_SYMBOL(__set_page_dirty_nobuffers
);
2492 * Call this whenever redirtying a page, to de-account the dirty counters
2493 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2494 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2495 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2498 void account_page_redirty(struct page
*page
)
2500 struct address_space
*mapping
= page
->mapping
;
2502 if (mapping
&& mapping_cap_account_dirty(mapping
)) {
2503 struct inode
*inode
= mapping
->host
;
2504 struct bdi_writeback
*wb
;
2507 wb
= unlocked_inode_to_wb_begin(inode
, &locked
);
2508 current
->nr_dirtied
--;
2509 dec_zone_page_state(page
, NR_DIRTIED
);
2510 dec_wb_stat(wb
, WB_DIRTIED
);
2511 unlocked_inode_to_wb_end(inode
, locked
);
2514 EXPORT_SYMBOL(account_page_redirty
);
2517 * When a writepage implementation decides that it doesn't want to write this
2518 * page for some reason, it should redirty the locked page via
2519 * redirty_page_for_writepage() and it should then unlock the page and return 0
2521 int redirty_page_for_writepage(struct writeback_control
*wbc
, struct page
*page
)
2525 wbc
->pages_skipped
++;
2526 ret
= __set_page_dirty_nobuffers(page
);
2527 account_page_redirty(page
);
2530 EXPORT_SYMBOL(redirty_page_for_writepage
);
2535 * For pages with a mapping this should be done under the page lock
2536 * for the benefit of asynchronous memory errors who prefer a consistent
2537 * dirty state. This rule can be broken in some special cases,
2538 * but should be better not to.
2540 * If the mapping doesn't provide a set_page_dirty a_op, then
2541 * just fall through and assume that it wants buffer_heads.
2543 int set_page_dirty(struct page
*page
)
2545 struct address_space
*mapping
= page_mapping(page
);
2547 if (likely(mapping
)) {
2548 int (*spd
)(struct page
*) = mapping
->a_ops
->set_page_dirty
;
2550 * readahead/lru_deactivate_page could remain
2551 * PG_readahead/PG_reclaim due to race with end_page_writeback
2552 * About readahead, if the page is written, the flags would be
2553 * reset. So no problem.
2554 * About lru_deactivate_page, if the page is redirty, the flag
2555 * will be reset. So no problem. but if the page is used by readahead
2556 * it will confuse readahead and make it restart the size rampup
2557 * process. But it's a trivial problem.
2559 if (PageReclaim(page
))
2560 ClearPageReclaim(page
);
2563 spd
= __set_page_dirty_buffers
;
2565 return (*spd
)(page
);
2567 if (!PageDirty(page
)) {
2568 if (!TestSetPageDirty(page
))
2573 EXPORT_SYMBOL(set_page_dirty
);
2576 * set_page_dirty() is racy if the caller has no reference against
2577 * page->mapping->host, and if the page is unlocked. This is because another
2578 * CPU could truncate the page off the mapping and then free the mapping.
2580 * Usually, the page _is_ locked, or the caller is a user-space process which
2581 * holds a reference on the inode by having an open file.
2583 * In other cases, the page should be locked before running set_page_dirty().
2585 int set_page_dirty_lock(struct page
*page
)
2590 ret
= set_page_dirty(page
);
2594 EXPORT_SYMBOL(set_page_dirty_lock
);
2597 * This cancels just the dirty bit on the kernel page itself, it does NOT
2598 * actually remove dirty bits on any mmap's that may be around. It also
2599 * leaves the page tagged dirty, so any sync activity will still find it on
2600 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2601 * look at the dirty bits in the VM.
2603 * Doing this should *normally* only ever be done when a page is truncated,
2604 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2605 * this when it notices that somebody has cleaned out all the buffers on a
2606 * page without actually doing it through the VM. Can you say "ext3 is
2607 * horribly ugly"? Thought you could.
2609 void cancel_dirty_page(struct page
*page
)
2611 struct address_space
*mapping
= page_mapping(page
);
2613 if (mapping_cap_account_dirty(mapping
)) {
2614 struct inode
*inode
= mapping
->host
;
2615 struct bdi_writeback
*wb
;
2616 struct mem_cgroup
*memcg
;
2619 memcg
= mem_cgroup_begin_page_stat(page
);
2620 wb
= unlocked_inode_to_wb_begin(inode
, &locked
);
2622 if (TestClearPageDirty(page
))
2623 account_page_cleaned(page
, mapping
, memcg
, wb
);
2625 unlocked_inode_to_wb_end(inode
, locked
);
2626 mem_cgroup_end_page_stat(memcg
);
2628 ClearPageDirty(page
);
2631 EXPORT_SYMBOL(cancel_dirty_page
);
2634 * Clear a page's dirty flag, while caring for dirty memory accounting.
2635 * Returns true if the page was previously dirty.
2637 * This is for preparing to put the page under writeout. We leave the page
2638 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2639 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2640 * implementation will run either set_page_writeback() or set_page_dirty(),
2641 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2644 * This incoherency between the page's dirty flag and radix-tree tag is
2645 * unfortunate, but it only exists while the page is locked.
2647 int clear_page_dirty_for_io(struct page
*page
)
2649 struct address_space
*mapping
= page_mapping(page
);
2652 BUG_ON(!PageLocked(page
));
2654 if (mapping
&& mapping_cap_account_dirty(mapping
)) {
2655 struct inode
*inode
= mapping
->host
;
2656 struct bdi_writeback
*wb
;
2657 struct mem_cgroup
*memcg
;
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 memcg
= mem_cgroup_begin_page_stat(page
);
2696 wb
= unlocked_inode_to_wb_begin(inode
, &locked
);
2697 if (TestClearPageDirty(page
)) {
2698 mem_cgroup_dec_page_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
2699 dec_zone_page_state(page
, NR_FILE_DIRTY
);
2700 dec_wb_stat(wb
, WB_RECLAIMABLE
);
2703 unlocked_inode_to_wb_end(inode
, locked
);
2704 mem_cgroup_end_page_stat(memcg
);
2707 return TestClearPageDirty(page
);
2709 EXPORT_SYMBOL(clear_page_dirty_for_io
);
2711 int test_clear_page_writeback(struct page
*page
)
2713 struct address_space
*mapping
= page_mapping(page
);
2714 struct mem_cgroup
*memcg
;
2717 memcg
= mem_cgroup_begin_page_stat(page
);
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
);
2736 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
2738 ret
= TestClearPageWriteback(page
);
2741 mem_cgroup_dec_page_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
2742 dec_zone_page_state(page
, NR_WRITEBACK
);
2743 inc_zone_page_state(page
, NR_WRITTEN
);
2745 mem_cgroup_end_page_stat(memcg
);
2749 int __test_set_page_writeback(struct page
*page
, bool keep_write
)
2751 struct address_space
*mapping
= page_mapping(page
);
2752 struct mem_cgroup
*memcg
;
2755 memcg
= mem_cgroup_begin_page_stat(page
);
2757 struct inode
*inode
= mapping
->host
;
2758 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
2759 unsigned long flags
;
2761 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
2762 ret
= TestSetPageWriteback(page
);
2764 radix_tree_tag_set(&mapping
->page_tree
,
2766 PAGECACHE_TAG_WRITEBACK
);
2767 if (bdi_cap_account_writeback(bdi
))
2768 __inc_wb_stat(inode_to_wb(inode
), WB_WRITEBACK
);
2770 if (!PageDirty(page
))
2771 radix_tree_tag_clear(&mapping
->page_tree
,
2773 PAGECACHE_TAG_DIRTY
);
2775 radix_tree_tag_clear(&mapping
->page_tree
,
2777 PAGECACHE_TAG_TOWRITE
);
2778 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
2780 ret
= TestSetPageWriteback(page
);
2783 mem_cgroup_inc_page_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
2784 inc_zone_page_state(page
, NR_WRITEBACK
);
2786 mem_cgroup_end_page_stat(memcg
);
2790 EXPORT_SYMBOL(__test_set_page_writeback
);
2793 * Return true if any of the pages in the mapping are marked with the
2796 int mapping_tagged(struct address_space
*mapping
, int tag
)
2798 return radix_tree_tagged(&mapping
->page_tree
, tag
);
2800 EXPORT_SYMBOL(mapping_tagged
);
2803 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2804 * @page: The page to wait on.
2806 * This function determines if the given page is related to a backing device
2807 * that requires page contents to be held stable during writeback. If so, then
2808 * it will wait for any pending writeback to complete.
2810 void wait_for_stable_page(struct page
*page
)
2812 if (bdi_cap_stable_pages_required(inode_to_bdi(page
->mapping
->host
)))
2813 wait_on_page_writeback(page
);
2815 EXPORT_SYMBOL_GPL(wait_for_stable_page
);