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 unsigned long global_dirty_limit
;
128 * Scale the writeback cache size proportional to the relative writeout speeds.
130 * We do this by keeping a floating proportion between BDIs, based on page
131 * writeback completions [end_page_writeback()]. Those devices that write out
132 * pages fastest will get the larger share, while the slower will get a smaller
135 * We use page writeout completions because we are interested in getting rid of
136 * dirty pages. Having them written out is the primary goal.
138 * We introduce a concept of time, a period over which we measure these events,
139 * because demand can/will vary over time. The length of this period itself is
140 * measured in page writeback completions.
143 static struct fprop_global writeout_completions
;
145 static void writeout_period(unsigned long t
);
146 /* Timer for aging of writeout_completions */
147 static struct timer_list writeout_period_timer
=
148 TIMER_DEFERRED_INITIALIZER(writeout_period
, 0, 0);
149 static unsigned long writeout_period_time
= 0;
152 * Length of period for aging writeout fractions of bdis. This is an
153 * arbitrarily chosen number. The longer the period, the slower fractions will
154 * reflect changes in current writeout rate.
156 #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
158 #ifdef CONFIG_CGROUP_WRITEBACK
160 static void wb_min_max_ratio(struct bdi_writeback
*wb
,
161 unsigned long *minp
, unsigned long *maxp
)
163 unsigned long this_bw
= wb
->avg_write_bandwidth
;
164 unsigned long tot_bw
= atomic_long_read(&wb
->bdi
->tot_write_bandwidth
);
165 unsigned long long min
= wb
->bdi
->min_ratio
;
166 unsigned long long max
= wb
->bdi
->max_ratio
;
169 * @wb may already be clean by the time control reaches here and
170 * the total may not include its bw.
172 if (this_bw
< tot_bw
) {
187 #else /* CONFIG_CGROUP_WRITEBACK */
189 static void wb_min_max_ratio(struct bdi_writeback
*wb
,
190 unsigned long *minp
, unsigned long *maxp
)
192 *minp
= wb
->bdi
->min_ratio
;
193 *maxp
= wb
->bdi
->max_ratio
;
196 #endif /* CONFIG_CGROUP_WRITEBACK */
199 * In a memory zone, there is a certain amount of pages we consider
200 * available for the page cache, which is essentially the number of
201 * free and reclaimable pages, minus some zone reserves to protect
202 * lowmem and the ability to uphold the zone's watermarks without
203 * requiring writeback.
205 * This number of dirtyable pages is the base value of which the
206 * user-configurable dirty ratio is the effictive number of pages that
207 * are allowed to be actually dirtied. Per individual zone, or
208 * globally by using the sum of dirtyable pages over all zones.
210 * Because the user is allowed to specify the dirty limit globally as
211 * absolute number of bytes, calculating the per-zone dirty limit can
212 * require translating the configured limit into a percentage of
213 * global dirtyable memory first.
217 * zone_dirtyable_memory - number of dirtyable pages in a zone
220 * Returns the zone's number of pages potentially available for dirty
221 * page cache. This is the base value for the per-zone dirty limits.
223 static unsigned long zone_dirtyable_memory(struct zone
*zone
)
225 unsigned long nr_pages
;
227 nr_pages
= zone_page_state(zone
, NR_FREE_PAGES
);
228 nr_pages
-= min(nr_pages
, zone
->dirty_balance_reserve
);
230 nr_pages
+= zone_page_state(zone
, NR_INACTIVE_FILE
);
231 nr_pages
+= zone_page_state(zone
, NR_ACTIVE_FILE
);
236 static unsigned long highmem_dirtyable_memory(unsigned long total
)
238 #ifdef CONFIG_HIGHMEM
242 for_each_node_state(node
, N_HIGH_MEMORY
) {
243 struct zone
*z
= &NODE_DATA(node
)->node_zones
[ZONE_HIGHMEM
];
245 x
+= zone_dirtyable_memory(z
);
248 * Unreclaimable memory (kernel memory or anonymous memory
249 * without swap) can bring down the dirtyable pages below
250 * the zone's dirty balance reserve and the above calculation
251 * will underflow. However we still want to add in nodes
252 * which are below threshold (negative values) to get a more
253 * accurate calculation but make sure that the total never
260 * Make sure that the number of highmem pages is never larger
261 * than the number of the total dirtyable memory. This can only
262 * occur in very strange VM situations but we want to make sure
263 * that this does not occur.
265 return min(x
, total
);
272 * global_dirtyable_memory - number of globally dirtyable pages
274 * Returns the global number of pages potentially available for dirty
275 * page cache. This is the base value for the global dirty limits.
277 static unsigned long global_dirtyable_memory(void)
281 x
= global_page_state(NR_FREE_PAGES
);
282 x
-= min(x
, dirty_balance_reserve
);
284 x
+= global_page_state(NR_INACTIVE_FILE
);
285 x
+= global_page_state(NR_ACTIVE_FILE
);
287 if (!vm_highmem_is_dirtyable
)
288 x
-= highmem_dirtyable_memory(x
);
290 return x
+ 1; /* Ensure that we never return 0 */
294 * global_dirty_limits - background-writeback and dirty-throttling thresholds
296 * Calculate the dirty thresholds based on sysctl parameters
297 * - vm.dirty_background_ratio or vm.dirty_background_bytes
298 * - vm.dirty_ratio or vm.dirty_bytes
299 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
302 void global_dirty_limits(unsigned long *pbackground
, unsigned long *pdirty
)
304 const unsigned long available_memory
= global_dirtyable_memory();
305 unsigned long background
;
307 struct task_struct
*tsk
;
310 dirty
= DIV_ROUND_UP(vm_dirty_bytes
, PAGE_SIZE
);
312 dirty
= (vm_dirty_ratio
* available_memory
) / 100;
314 if (dirty_background_bytes
)
315 background
= DIV_ROUND_UP(dirty_background_bytes
, PAGE_SIZE
);
317 background
= (dirty_background_ratio
* available_memory
) / 100;
319 if (background
>= dirty
)
320 background
= dirty
/ 2;
322 if (tsk
->flags
& PF_LESS_THROTTLE
|| rt_task(tsk
)) {
323 background
+= background
/ 4;
326 *pbackground
= background
;
328 trace_global_dirty_state(background
, dirty
);
332 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
335 * Returns the maximum number of dirty pages allowed in a zone, based
336 * on the zone's dirtyable memory.
338 static unsigned long zone_dirty_limit(struct zone
*zone
)
340 unsigned long zone_memory
= zone_dirtyable_memory(zone
);
341 struct task_struct
*tsk
= current
;
345 dirty
= DIV_ROUND_UP(vm_dirty_bytes
, PAGE_SIZE
) *
346 zone_memory
/ global_dirtyable_memory();
348 dirty
= vm_dirty_ratio
* zone_memory
/ 100;
350 if (tsk
->flags
& PF_LESS_THROTTLE
|| rt_task(tsk
))
357 * zone_dirty_ok - tells whether a zone is within its dirty limits
358 * @zone: the zone to check
360 * Returns %true when the dirty pages in @zone are within the zone's
361 * dirty limit, %false if the limit is exceeded.
363 bool zone_dirty_ok(struct zone
*zone
)
365 unsigned long limit
= zone_dirty_limit(zone
);
367 return zone_page_state(zone
, NR_FILE_DIRTY
) +
368 zone_page_state(zone
, NR_UNSTABLE_NFS
) +
369 zone_page_state(zone
, NR_WRITEBACK
) <= limit
;
372 int dirty_background_ratio_handler(struct ctl_table
*table
, int write
,
373 void __user
*buffer
, size_t *lenp
,
378 ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
379 if (ret
== 0 && write
)
380 dirty_background_bytes
= 0;
384 int dirty_background_bytes_handler(struct ctl_table
*table
, int write
,
385 void __user
*buffer
, size_t *lenp
,
390 ret
= proc_doulongvec_minmax(table
, write
, buffer
, lenp
, ppos
);
391 if (ret
== 0 && write
)
392 dirty_background_ratio
= 0;
396 int dirty_ratio_handler(struct ctl_table
*table
, int write
,
397 void __user
*buffer
, size_t *lenp
,
400 int old_ratio
= vm_dirty_ratio
;
403 ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
404 if (ret
== 0 && write
&& vm_dirty_ratio
!= old_ratio
) {
405 writeback_set_ratelimit();
411 int dirty_bytes_handler(struct ctl_table
*table
, int write
,
412 void __user
*buffer
, size_t *lenp
,
415 unsigned long old_bytes
= vm_dirty_bytes
;
418 ret
= proc_doulongvec_minmax(table
, write
, buffer
, lenp
, ppos
);
419 if (ret
== 0 && write
&& vm_dirty_bytes
!= old_bytes
) {
420 writeback_set_ratelimit();
426 static unsigned long wp_next_time(unsigned long cur_time
)
428 cur_time
+= VM_COMPLETIONS_PERIOD_LEN
;
429 /* 0 has a special meaning... */
436 * Increment the BDI's writeout completion count and the global writeout
437 * completion count. Called from test_clear_page_writeback().
439 static inline void __wb_writeout_inc(struct bdi_writeback
*wb
)
441 __inc_wb_stat(wb
, WB_WRITTEN
);
442 __fprop_inc_percpu_max(&writeout_completions
, &wb
->completions
,
443 wb
->bdi
->max_prop_frac
);
444 /* First event after period switching was turned off? */
445 if (!unlikely(writeout_period_time
)) {
447 * We can race with other __bdi_writeout_inc calls here but
448 * it does not cause any harm since the resulting time when
449 * timer will fire and what is in writeout_period_time will be
452 writeout_period_time
= wp_next_time(jiffies
);
453 mod_timer(&writeout_period_timer
, writeout_period_time
);
457 void wb_writeout_inc(struct bdi_writeback
*wb
)
461 local_irq_save(flags
);
462 __wb_writeout_inc(wb
);
463 local_irq_restore(flags
);
465 EXPORT_SYMBOL_GPL(wb_writeout_inc
);
468 * Obtain an accurate fraction of the BDI's portion.
470 static void wb_writeout_fraction(struct bdi_writeback
*wb
,
471 long *numerator
, long *denominator
)
473 fprop_fraction_percpu(&writeout_completions
, &wb
->completions
,
474 numerator
, denominator
);
478 * On idle system, we can be called long after we scheduled because we use
479 * deferred timers so count with missed periods.
481 static void writeout_period(unsigned long t
)
483 int miss_periods
= (jiffies
- writeout_period_time
) /
484 VM_COMPLETIONS_PERIOD_LEN
;
486 if (fprop_new_period(&writeout_completions
, miss_periods
+ 1)) {
487 writeout_period_time
= wp_next_time(writeout_period_time
+
488 miss_periods
* VM_COMPLETIONS_PERIOD_LEN
);
489 mod_timer(&writeout_period_timer
, writeout_period_time
);
492 * Aging has zeroed all fractions. Stop wasting CPU on period
495 writeout_period_time
= 0;
500 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
501 * registered backing devices, which, for obvious reasons, can not
504 static unsigned int bdi_min_ratio
;
506 int bdi_set_min_ratio(struct backing_dev_info
*bdi
, unsigned int min_ratio
)
510 spin_lock_bh(&bdi_lock
);
511 if (min_ratio
> bdi
->max_ratio
) {
514 min_ratio
-= bdi
->min_ratio
;
515 if (bdi_min_ratio
+ min_ratio
< 100) {
516 bdi_min_ratio
+= min_ratio
;
517 bdi
->min_ratio
+= min_ratio
;
522 spin_unlock_bh(&bdi_lock
);
527 int bdi_set_max_ratio(struct backing_dev_info
*bdi
, unsigned max_ratio
)
534 spin_lock_bh(&bdi_lock
);
535 if (bdi
->min_ratio
> max_ratio
) {
538 bdi
->max_ratio
= max_ratio
;
539 bdi
->max_prop_frac
= (FPROP_FRAC_BASE
* max_ratio
) / 100;
541 spin_unlock_bh(&bdi_lock
);
545 EXPORT_SYMBOL(bdi_set_max_ratio
);
547 static unsigned long dirty_freerun_ceiling(unsigned long thresh
,
548 unsigned long bg_thresh
)
550 return (thresh
+ bg_thresh
) / 2;
553 static unsigned long hard_dirty_limit(unsigned long thresh
)
555 return max(thresh
, global_dirty_limit
);
559 * wb_calc_thresh - @wb's share of dirty throttling threshold
560 * @wb: bdi_writeback to query
561 * @dirty: global dirty limit in pages
563 * Returns @wb's dirty limit in pages. The term "dirty" in the context of
564 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
566 * Note that balance_dirty_pages() will only seriously take it as a hard limit
567 * when sleeping max_pause per page is not enough to keep the dirty pages under
568 * control. For example, when the device is completely stalled due to some error
569 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
570 * In the other normal situations, it acts more gently by throttling the tasks
571 * more (rather than completely block them) when the wb dirty pages go high.
573 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
574 * - starving fast devices
575 * - piling up dirty pages (that will take long time to sync) on slow devices
577 * The wb's share of dirty limit will be adapting to its throughput and
578 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
580 unsigned long wb_calc_thresh(struct bdi_writeback
*wb
, unsigned long thresh
)
583 long numerator
, denominator
;
584 unsigned long wb_min_ratio
, wb_max_ratio
;
587 * Calculate this BDI's share of the thresh ratio.
589 wb_writeout_fraction(wb
, &numerator
, &denominator
);
591 wb_thresh
= (thresh
* (100 - bdi_min_ratio
)) / 100;
592 wb_thresh
*= numerator
;
593 do_div(wb_thresh
, denominator
);
595 wb_min_max_ratio(wb
, &wb_min_ratio
, &wb_max_ratio
);
597 wb_thresh
+= (thresh
* wb_min_ratio
) / 100;
598 if (wb_thresh
> (thresh
* wb_max_ratio
) / 100)
599 wb_thresh
= thresh
* wb_max_ratio
/ 100;
606 * f(dirty) := 1.0 + (----------------)
609 * it's a 3rd order polynomial that subjects to
611 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
612 * (2) f(setpoint) = 1.0 => the balance point
613 * (3) f(limit) = 0 => the hard limit
614 * (4) df/dx <= 0 => negative feedback control
615 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
616 * => fast response on large errors; small oscillation near setpoint
618 static long long pos_ratio_polynom(unsigned long setpoint
,
625 x
= div64_s64(((s64
)setpoint
- (s64
)dirty
) << RATELIMIT_CALC_SHIFT
,
626 limit
- setpoint
+ 1);
628 pos_ratio
= pos_ratio
* x
>> RATELIMIT_CALC_SHIFT
;
629 pos_ratio
= pos_ratio
* x
>> RATELIMIT_CALC_SHIFT
;
630 pos_ratio
+= 1 << RATELIMIT_CALC_SHIFT
;
632 return clamp(pos_ratio
, 0LL, 2LL << RATELIMIT_CALC_SHIFT
);
636 * Dirty position control.
638 * (o) global/bdi setpoints
640 * We want the dirty pages be balanced around the global/wb setpoints.
641 * When the number of dirty pages is higher/lower than the setpoint, the
642 * dirty position control ratio (and hence task dirty ratelimit) will be
643 * decreased/increased to bring the dirty pages back to the setpoint.
645 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
647 * if (dirty < setpoint) scale up pos_ratio
648 * if (dirty > setpoint) scale down pos_ratio
650 * if (wb_dirty < wb_setpoint) scale up pos_ratio
651 * if (wb_dirty > wb_setpoint) scale down pos_ratio
653 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
655 * (o) global control line
659 * | |<===== global dirty control scope ======>|
667 * 1.0 ................................*
673 * 0 +------------.------------------.----------------------*------------->
674 * freerun^ setpoint^ limit^ dirty pages
676 * (o) wb control line
684 * | * |<=========== span ============>|
685 * 1.0 .......................*
697 * 1/4 ...............................................* * * * * * * * * * * *
701 * 0 +----------------------.-------------------------------.------------->
702 * wb_setpoint^ x_intercept^
704 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
705 * be smoothly throttled down to normal if it starts high in situations like
706 * - start writing to a slow SD card and a fast disk at the same time. The SD
707 * card's wb_dirty may rush to many times higher than wb_setpoint.
708 * - the wb dirty thresh drops quickly due to change of JBOD workload
710 static unsigned long wb_position_ratio(struct bdi_writeback
*wb
,
711 unsigned long thresh
,
712 unsigned long bg_thresh
,
714 unsigned long wb_thresh
,
715 unsigned long wb_dirty
)
717 unsigned long write_bw
= wb
->avg_write_bandwidth
;
718 unsigned long freerun
= dirty_freerun_ceiling(thresh
, bg_thresh
);
719 unsigned long limit
= hard_dirty_limit(thresh
);
720 unsigned long x_intercept
;
721 unsigned long setpoint
; /* dirty pages' target balance point */
722 unsigned long wb_setpoint
;
724 long long pos_ratio
; /* for scaling up/down the rate limit */
727 if (unlikely(dirty
>= limit
))
733 * See comment for pos_ratio_polynom().
735 setpoint
= (freerun
+ limit
) / 2;
736 pos_ratio
= pos_ratio_polynom(setpoint
, dirty
, limit
);
739 * The strictlimit feature is a tool preventing mistrusted filesystems
740 * from growing a large number of dirty pages before throttling. For
741 * such filesystems balance_dirty_pages always checks wb counters
742 * against wb limits. Even if global "nr_dirty" is under "freerun".
743 * This is especially important for fuse which sets bdi->max_ratio to
744 * 1% by default. Without strictlimit feature, fuse writeback may
745 * consume arbitrary amount of RAM because it is accounted in
746 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
748 * Here, in wb_position_ratio(), we calculate pos_ratio based on
749 * two values: wb_dirty and wb_thresh. Let's consider an example:
750 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
751 * limits are set by default to 10% and 20% (background and throttle).
752 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
753 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
754 * about ~6K pages (as the average of background and throttle wb
755 * limits). The 3rd order polynomial will provide positive feedback if
756 * wb_dirty is under wb_setpoint and vice versa.
758 * Note, that we cannot use global counters in these calculations
759 * because we want to throttle process writing to a strictlimit wb
760 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
761 * in the example above).
763 if (unlikely(wb
->bdi
->capabilities
& BDI_CAP_STRICTLIMIT
)) {
764 long long wb_pos_ratio
;
765 unsigned long wb_bg_thresh
;
768 return min_t(long long, pos_ratio
* 2,
769 2 << RATELIMIT_CALC_SHIFT
);
771 if (wb_dirty
>= wb_thresh
)
774 wb_bg_thresh
= div_u64((u64
)wb_thresh
* bg_thresh
, thresh
);
775 wb_setpoint
= dirty_freerun_ceiling(wb_thresh
, wb_bg_thresh
);
777 if (wb_setpoint
== 0 || wb_setpoint
== wb_thresh
)
780 wb_pos_ratio
= pos_ratio_polynom(wb_setpoint
, wb_dirty
,
784 * Typically, for strictlimit case, wb_setpoint << setpoint
785 * and pos_ratio >> wb_pos_ratio. In the other words global
786 * state ("dirty") is not limiting factor and we have to
787 * make decision based on wb counters. But there is an
788 * important case when global pos_ratio should get precedence:
789 * global limits are exceeded (e.g. due to activities on other
790 * wb's) while given strictlimit wb is below limit.
792 * "pos_ratio * wb_pos_ratio" would work for the case above,
793 * but it would look too non-natural for the case of all
794 * activity in the system coming from a single strictlimit wb
795 * with bdi->max_ratio == 100%.
797 * Note that min() below somewhat changes the dynamics of the
798 * control system. Normally, pos_ratio value can be well over 3
799 * (when globally we are at freerun and wb is well below wb
800 * setpoint). Now the maximum pos_ratio in the same situation
801 * is 2. We might want to tweak this if we observe the control
802 * system is too slow to adapt.
804 return min(pos_ratio
, wb_pos_ratio
);
808 * We have computed basic pos_ratio above based on global situation. If
809 * the wb is over/under its share of dirty pages, we want to scale
810 * pos_ratio further down/up. That is done by the following mechanism.
816 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
818 * x_intercept - wb_dirty
819 * := --------------------------
820 * x_intercept - wb_setpoint
822 * The main wb control line is a linear function that subjects to
824 * (1) f(wb_setpoint) = 1.0
825 * (2) k = - 1 / (8 * write_bw) (in single wb case)
826 * or equally: x_intercept = wb_setpoint + 8 * write_bw
828 * For single wb case, the dirty pages are observed to fluctuate
829 * regularly within range
830 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
831 * for various filesystems, where (2) can yield in a reasonable 12.5%
832 * fluctuation range for pos_ratio.
834 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
835 * own size, so move the slope over accordingly and choose a slope that
836 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
838 if (unlikely(wb_thresh
> thresh
))
841 * It's very possible that wb_thresh is close to 0 not because the
842 * device is slow, but that it has remained inactive for long time.
843 * Honour such devices a reasonable good (hopefully IO efficient)
844 * threshold, so that the occasional writes won't be blocked and active
845 * writes can rampup the threshold quickly.
847 wb_thresh
= max(wb_thresh
, (limit
- dirty
) / 8);
849 * scale global setpoint to wb's:
850 * wb_setpoint = setpoint * wb_thresh / thresh
852 x
= div_u64((u64
)wb_thresh
<< 16, thresh
+ 1);
853 wb_setpoint
= setpoint
* (u64
)x
>> 16;
855 * Use span=(8*write_bw) in single wb case as indicated by
856 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
858 * wb_thresh thresh - wb_thresh
859 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
862 span
= (thresh
- wb_thresh
+ 8 * write_bw
) * (u64
)x
>> 16;
863 x_intercept
= wb_setpoint
+ span
;
865 if (wb_dirty
< x_intercept
- span
/ 4) {
866 pos_ratio
= div64_u64(pos_ratio
* (x_intercept
- wb_dirty
),
867 x_intercept
- wb_setpoint
+ 1);
872 * wb reserve area, safeguard against dirty pool underrun and disk idle
873 * It may push the desired control point of global dirty pages higher
876 x_intercept
= wb_thresh
/ 2;
877 if (wb_dirty
< x_intercept
) {
878 if (wb_dirty
> x_intercept
/ 8)
879 pos_ratio
= div_u64(pos_ratio
* x_intercept
, wb_dirty
);
887 static void wb_update_write_bandwidth(struct bdi_writeback
*wb
,
888 unsigned long elapsed
,
889 unsigned long written
)
891 const unsigned long period
= roundup_pow_of_two(3 * HZ
);
892 unsigned long avg
= wb
->avg_write_bandwidth
;
893 unsigned long old
= wb
->write_bandwidth
;
897 * bw = written * HZ / elapsed
899 * bw * elapsed + write_bandwidth * (period - elapsed)
900 * write_bandwidth = ---------------------------------------------------
903 * @written may have decreased due to account_page_redirty().
904 * Avoid underflowing @bw calculation.
906 bw
= written
- min(written
, wb
->written_stamp
);
908 if (unlikely(elapsed
> period
)) {
913 bw
+= (u64
)wb
->write_bandwidth
* (period
- elapsed
);
914 bw
>>= ilog2(period
);
917 * one more level of smoothing, for filtering out sudden spikes
919 if (avg
> old
&& old
>= (unsigned long)bw
)
920 avg
-= (avg
- old
) >> 3;
922 if (avg
< old
&& old
<= (unsigned long)bw
)
923 avg
+= (old
- avg
) >> 3;
926 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
928 if (wb_has_dirty_io(wb
)) {
929 long delta
= avg
- wb
->avg_write_bandwidth
;
930 WARN_ON_ONCE(atomic_long_add_return(delta
,
931 &wb
->bdi
->tot_write_bandwidth
) <= 0);
933 wb
->write_bandwidth
= bw
;
934 wb
->avg_write_bandwidth
= avg
;
938 * The global dirtyable memory and dirty threshold could be suddenly knocked
939 * down by a large amount (eg. on the startup of KVM in a swapless system).
940 * This may throw the system into deep dirty exceeded state and throttle
941 * heavy/light dirtiers alike. To retain good responsiveness, maintain
942 * global_dirty_limit for tracking slowly down to the knocked down dirty
945 static void update_dirty_limit(unsigned long thresh
, unsigned long dirty
)
947 unsigned long limit
= global_dirty_limit
;
950 * Follow up in one step.
952 if (limit
< thresh
) {
958 * Follow down slowly. Use the higher one as the target, because thresh
959 * may drop below dirty. This is exactly the reason to introduce
960 * global_dirty_limit which is guaranteed to lie above the dirty pages.
962 thresh
= max(thresh
, dirty
);
963 if (limit
> thresh
) {
964 limit
-= (limit
- thresh
) >> 5;
969 global_dirty_limit
= limit
;
972 static void global_update_bandwidth(unsigned long thresh
,
976 static DEFINE_SPINLOCK(dirty_lock
);
977 static unsigned long update_time
= INITIAL_JIFFIES
;
980 * check locklessly first to optimize away locking for the most time
982 if (time_before(now
, update_time
+ BANDWIDTH_INTERVAL
))
985 spin_lock(&dirty_lock
);
986 if (time_after_eq(now
, update_time
+ BANDWIDTH_INTERVAL
)) {
987 update_dirty_limit(thresh
, dirty
);
990 spin_unlock(&dirty_lock
);
994 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
996 * Normal wb tasks will be curbed at or below it in long term.
997 * Obviously it should be around (write_bw / N) when there are N dd tasks.
999 static void wb_update_dirty_ratelimit(struct bdi_writeback
*wb
,
1000 unsigned long thresh
,
1001 unsigned long bg_thresh
,
1002 unsigned long dirty
,
1003 unsigned long wb_thresh
,
1004 unsigned long wb_dirty
,
1005 unsigned long dirtied
,
1006 unsigned long elapsed
)
1008 unsigned long freerun
= dirty_freerun_ceiling(thresh
, bg_thresh
);
1009 unsigned long limit
= hard_dirty_limit(thresh
);
1010 unsigned long setpoint
= (freerun
+ limit
) / 2;
1011 unsigned long write_bw
= wb
->avg_write_bandwidth
;
1012 unsigned long dirty_ratelimit
= wb
->dirty_ratelimit
;
1013 unsigned long dirty_rate
;
1014 unsigned long task_ratelimit
;
1015 unsigned long balanced_dirty_ratelimit
;
1016 unsigned long pos_ratio
;
1021 * The dirty rate will match the writeout rate in long term, except
1022 * when dirty pages are truncated by userspace or re-dirtied by FS.
1024 dirty_rate
= (dirtied
- wb
->dirtied_stamp
) * HZ
/ elapsed
;
1026 pos_ratio
= wb_position_ratio(wb
, thresh
, bg_thresh
, dirty
,
1027 wb_thresh
, wb_dirty
);
1029 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1031 task_ratelimit
= (u64
)dirty_ratelimit
*
1032 pos_ratio
>> RATELIMIT_CALC_SHIFT
;
1033 task_ratelimit
++; /* it helps rampup dirty_ratelimit from tiny values */
1036 * A linear estimation of the "balanced" throttle rate. The theory is,
1037 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1038 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1039 * formula will yield the balanced rate limit (write_bw / N).
1041 * Note that the expanded form is not a pure rate feedback:
1042 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1043 * but also takes pos_ratio into account:
1044 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1046 * (1) is not realistic because pos_ratio also takes part in balancing
1047 * the dirty rate. Consider the state
1048 * pos_ratio = 0.5 (3)
1049 * rate = 2 * (write_bw / N) (4)
1050 * If (1) is used, it will stuck in that state! Because each dd will
1052 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1054 * dirty_rate = N * task_ratelimit = write_bw (6)
1055 * put (6) into (1) we get
1056 * rate_(i+1) = rate_(i) (7)
1058 * So we end up using (2) to always keep
1059 * rate_(i+1) ~= (write_bw / N) (8)
1060 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1061 * pos_ratio is able to drive itself to 1.0, which is not only where
1062 * the dirty count meet the setpoint, but also where the slope of
1063 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1065 balanced_dirty_ratelimit
= div_u64((u64
)task_ratelimit
* write_bw
,
1068 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1070 if (unlikely(balanced_dirty_ratelimit
> write_bw
))
1071 balanced_dirty_ratelimit
= write_bw
;
1074 * We could safely do this and return immediately:
1076 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
1078 * However to get a more stable dirty_ratelimit, the below elaborated
1079 * code makes use of task_ratelimit to filter out singular points and
1080 * limit the step size.
1082 * The below code essentially only uses the relative value of
1084 * task_ratelimit - dirty_ratelimit
1085 * = (pos_ratio - 1) * dirty_ratelimit
1087 * which reflects the direction and size of dirty position error.
1091 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1092 * task_ratelimit is on the same side of dirty_ratelimit, too.
1094 * - dirty_ratelimit > balanced_dirty_ratelimit
1095 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1096 * lowering dirty_ratelimit will help meet both the position and rate
1097 * control targets. Otherwise, don't update dirty_ratelimit if it will
1098 * only help meet the rate target. After all, what the users ultimately
1099 * feel and care are stable dirty rate and small position error.
1101 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1102 * and filter out the singular points of balanced_dirty_ratelimit. Which
1103 * keeps jumping around randomly and can even leap far away at times
1104 * due to the small 200ms estimation period of dirty_rate (we want to
1105 * keep that period small to reduce time lags).
1110 * For strictlimit case, calculations above were based on wb counters
1111 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1112 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1113 * Hence, to calculate "step" properly, we have to use wb_dirty as
1114 * "dirty" and wb_setpoint as "setpoint".
1116 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1117 * it's possible that wb_thresh is close to zero due to inactivity
1118 * of backing device (see the implementation of wb_calc_thresh()).
1120 if (unlikely(wb
->bdi
->capabilities
& BDI_CAP_STRICTLIMIT
)) {
1123 setpoint
= wb_dirty
+ 1;
1125 setpoint
= (wb_thresh
+
1126 wb_calc_thresh(wb
, bg_thresh
)) / 2;
1129 if (dirty
< setpoint
) {
1130 x
= min3(wb
->balanced_dirty_ratelimit
,
1131 balanced_dirty_ratelimit
, task_ratelimit
);
1132 if (dirty_ratelimit
< x
)
1133 step
= x
- dirty_ratelimit
;
1135 x
= max3(wb
->balanced_dirty_ratelimit
,
1136 balanced_dirty_ratelimit
, task_ratelimit
);
1137 if (dirty_ratelimit
> x
)
1138 step
= dirty_ratelimit
- x
;
1142 * Don't pursue 100% rate matching. It's impossible since the balanced
1143 * rate itself is constantly fluctuating. So decrease the track speed
1144 * when it gets close to the target. Helps eliminate pointless tremors.
1146 step
>>= dirty_ratelimit
/ (2 * step
+ 1);
1148 * Limit the tracking speed to avoid overshooting.
1150 step
= (step
+ 7) / 8;
1152 if (dirty_ratelimit
< balanced_dirty_ratelimit
)
1153 dirty_ratelimit
+= step
;
1155 dirty_ratelimit
-= step
;
1157 wb
->dirty_ratelimit
= max(dirty_ratelimit
, 1UL);
1158 wb
->balanced_dirty_ratelimit
= balanced_dirty_ratelimit
;
1160 trace_bdi_dirty_ratelimit(wb
->bdi
, dirty_rate
, task_ratelimit
);
1163 void __wb_update_bandwidth(struct bdi_writeback
*wb
,
1164 unsigned long thresh
,
1165 unsigned long bg_thresh
,
1166 unsigned long dirty
,
1167 unsigned long wb_thresh
,
1168 unsigned long wb_dirty
,
1169 unsigned long start_time
)
1171 unsigned long now
= jiffies
;
1172 unsigned long elapsed
= now
- wb
->bw_time_stamp
;
1173 unsigned long dirtied
;
1174 unsigned long written
;
1177 * rate-limit, only update once every 200ms.
1179 if (elapsed
< BANDWIDTH_INTERVAL
)
1182 dirtied
= percpu_counter_read(&wb
->stat
[WB_DIRTIED
]);
1183 written
= percpu_counter_read(&wb
->stat
[WB_WRITTEN
]);
1186 * Skip quiet periods when disk bandwidth is under-utilized.
1187 * (at least 1s idle time between two flusher runs)
1189 if (elapsed
> HZ
&& time_before(wb
->bw_time_stamp
, start_time
))
1193 global_update_bandwidth(thresh
, dirty
, now
);
1194 wb_update_dirty_ratelimit(wb
, thresh
, bg_thresh
, dirty
,
1195 wb_thresh
, wb_dirty
,
1198 wb_update_write_bandwidth(wb
, elapsed
, written
);
1201 wb
->dirtied_stamp
= dirtied
;
1202 wb
->written_stamp
= written
;
1203 wb
->bw_time_stamp
= now
;
1206 static void wb_update_bandwidth(struct bdi_writeback
*wb
,
1207 unsigned long thresh
,
1208 unsigned long bg_thresh
,
1209 unsigned long dirty
,
1210 unsigned long wb_thresh
,
1211 unsigned long wb_dirty
,
1212 unsigned long start_time
)
1214 if (time_is_after_eq_jiffies(wb
->bw_time_stamp
+ BANDWIDTH_INTERVAL
))
1216 spin_lock(&wb
->list_lock
);
1217 __wb_update_bandwidth(wb
, thresh
, bg_thresh
, dirty
,
1218 wb_thresh
, wb_dirty
, start_time
);
1219 spin_unlock(&wb
->list_lock
);
1223 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1224 * will look to see if it needs to start dirty throttling.
1226 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1227 * global_page_state() too often. So scale it near-sqrt to the safety margin
1228 * (the number of pages we may dirty without exceeding the dirty limits).
1230 static unsigned long dirty_poll_interval(unsigned long dirty
,
1231 unsigned long thresh
)
1234 return 1UL << (ilog2(thresh
- dirty
) >> 1);
1239 static unsigned long wb_max_pause(struct bdi_writeback
*wb
,
1240 unsigned long wb_dirty
)
1242 unsigned long bw
= wb
->avg_write_bandwidth
;
1246 * Limit pause time for small memory systems. If sleeping for too long
1247 * time, a small pool of dirty/writeback pages may go empty and disk go
1250 * 8 serves as the safety ratio.
1252 t
= wb_dirty
/ (1 + bw
/ roundup_pow_of_two(1 + HZ
/ 8));
1255 return min_t(unsigned long, t
, MAX_PAUSE
);
1258 static long wb_min_pause(struct bdi_writeback
*wb
,
1260 unsigned long task_ratelimit
,
1261 unsigned long dirty_ratelimit
,
1262 int *nr_dirtied_pause
)
1264 long hi
= ilog2(wb
->avg_write_bandwidth
);
1265 long lo
= ilog2(wb
->dirty_ratelimit
);
1266 long t
; /* target pause */
1267 long pause
; /* estimated next pause */
1268 int pages
; /* target nr_dirtied_pause */
1270 /* target for 10ms pause on 1-dd case */
1271 t
= max(1, HZ
/ 100);
1274 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1277 * (N * 10ms) on 2^N concurrent tasks.
1280 t
+= (hi
- lo
) * (10 * HZ
) / 1024;
1283 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1284 * on the much more stable dirty_ratelimit. However the next pause time
1285 * will be computed based on task_ratelimit and the two rate limits may
1286 * depart considerably at some time. Especially if task_ratelimit goes
1287 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1288 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1289 * result task_ratelimit won't be executed faithfully, which could
1290 * eventually bring down dirty_ratelimit.
1292 * We apply two rules to fix it up:
1293 * 1) try to estimate the next pause time and if necessary, use a lower
1294 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1295 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1296 * 2) limit the target pause time to max_pause/2, so that the normal
1297 * small fluctuations of task_ratelimit won't trigger rule (1) and
1298 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1300 t
= min(t
, 1 + max_pause
/ 2);
1301 pages
= dirty_ratelimit
* t
/ roundup_pow_of_two(HZ
);
1304 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1305 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1306 * When the 16 consecutive reads are often interrupted by some dirty
1307 * throttling pause during the async writes, cfq will go into idles
1308 * (deadline is fine). So push nr_dirtied_pause as high as possible
1309 * until reaches DIRTY_POLL_THRESH=32 pages.
1311 if (pages
< DIRTY_POLL_THRESH
) {
1313 pages
= dirty_ratelimit
* t
/ roundup_pow_of_two(HZ
);
1314 if (pages
> DIRTY_POLL_THRESH
) {
1315 pages
= DIRTY_POLL_THRESH
;
1316 t
= HZ
* DIRTY_POLL_THRESH
/ dirty_ratelimit
;
1320 pause
= HZ
* pages
/ (task_ratelimit
+ 1);
1321 if (pause
> max_pause
) {
1323 pages
= task_ratelimit
* t
/ roundup_pow_of_two(HZ
);
1326 *nr_dirtied_pause
= pages
;
1328 * The minimal pause time will normally be half the target pause time.
1330 return pages
>= DIRTY_POLL_THRESH
? 1 + t
/ 2 : t
;
1333 static inline void wb_dirty_limits(struct bdi_writeback
*wb
,
1334 unsigned long dirty_thresh
,
1335 unsigned long background_thresh
,
1336 unsigned long *wb_dirty
,
1337 unsigned long *wb_thresh
,
1338 unsigned long *wb_bg_thresh
)
1340 unsigned long wb_reclaimable
;
1343 * wb_thresh is not treated as some limiting factor as
1344 * dirty_thresh, due to reasons
1345 * - in JBOD setup, wb_thresh can fluctuate a lot
1346 * - in a system with HDD and USB key, the USB key may somehow
1347 * go into state (wb_dirty >> wb_thresh) either because
1348 * wb_dirty starts high, or because wb_thresh drops low.
1349 * In this case we don't want to hard throttle the USB key
1350 * dirtiers for 100 seconds until wb_dirty drops under
1351 * wb_thresh. Instead the auxiliary wb control line in
1352 * wb_position_ratio() will let the dirtier task progress
1353 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
1355 *wb_thresh
= wb_calc_thresh(wb
, dirty_thresh
);
1358 *wb_bg_thresh
= dirty_thresh
? div_u64((u64
)*wb_thresh
*
1363 * In order to avoid the stacked BDI deadlock we need
1364 * to ensure we accurately count the 'dirty' pages when
1365 * the threshold is low.
1367 * Otherwise it would be possible to get thresh+n pages
1368 * reported dirty, even though there are thresh-m pages
1369 * actually dirty; with m+n sitting in the percpu
1372 if (*wb_thresh
< 2 * wb_stat_error(wb
)) {
1373 wb_reclaimable
= wb_stat_sum(wb
, WB_RECLAIMABLE
);
1374 *wb_dirty
= wb_reclaimable
+ wb_stat_sum(wb
, WB_WRITEBACK
);
1376 wb_reclaimable
= wb_stat(wb
, WB_RECLAIMABLE
);
1377 *wb_dirty
= wb_reclaimable
+ wb_stat(wb
, WB_WRITEBACK
);
1382 * balance_dirty_pages() must be called by processes which are generating dirty
1383 * data. It looks at the number of dirty pages in the machine and will force
1384 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1385 * If we're over `background_thresh' then the writeback threads are woken to
1386 * perform some writeout.
1388 static void balance_dirty_pages(struct address_space
*mapping
,
1389 struct bdi_writeback
*wb
,
1390 unsigned long pages_dirtied
)
1392 unsigned long nr_reclaimable
; /* = file_dirty + unstable_nfs */
1393 unsigned long nr_dirty
; /* = file_dirty + writeback + unstable_nfs */
1394 unsigned long background_thresh
;
1395 unsigned long dirty_thresh
;
1400 int nr_dirtied_pause
;
1401 bool dirty_exceeded
= false;
1402 unsigned long task_ratelimit
;
1403 unsigned long dirty_ratelimit
;
1404 unsigned long pos_ratio
;
1405 struct backing_dev_info
*bdi
= wb
->bdi
;
1406 bool strictlimit
= bdi
->capabilities
& BDI_CAP_STRICTLIMIT
;
1407 unsigned long start_time
= jiffies
;
1410 unsigned long now
= jiffies
;
1411 unsigned long uninitialized_var(wb_thresh
);
1412 unsigned long thresh
;
1413 unsigned long uninitialized_var(wb_dirty
);
1414 unsigned long dirty
;
1415 unsigned long bg_thresh
;
1418 * Unstable writes are a feature of certain networked
1419 * filesystems (i.e. NFS) in which data may have been
1420 * written to the server's write cache, but has not yet
1421 * been flushed to permanent storage.
1423 nr_reclaimable
= global_page_state(NR_FILE_DIRTY
) +
1424 global_page_state(NR_UNSTABLE_NFS
);
1425 nr_dirty
= nr_reclaimable
+ global_page_state(NR_WRITEBACK
);
1427 global_dirty_limits(&background_thresh
, &dirty_thresh
);
1429 if (unlikely(strictlimit
)) {
1430 wb_dirty_limits(wb
, dirty_thresh
, background_thresh
,
1431 &wb_dirty
, &wb_thresh
, &bg_thresh
);
1437 thresh
= dirty_thresh
;
1438 bg_thresh
= background_thresh
;
1442 * Throttle it only when the background writeback cannot
1443 * catch-up. This avoids (excessively) small writeouts
1444 * when the wb limits are ramping up in case of !strictlimit.
1446 * In strictlimit case make decision based on the wb counters
1447 * and limits. Small writeouts when the wb limits are ramping
1448 * up are the price we consciously pay for strictlimit-ing.
1450 if (dirty
<= dirty_freerun_ceiling(thresh
, bg_thresh
)) {
1451 current
->dirty_paused_when
= now
;
1452 current
->nr_dirtied
= 0;
1453 current
->nr_dirtied_pause
=
1454 dirty_poll_interval(dirty
, thresh
);
1458 if (unlikely(!writeback_in_progress(wb
)))
1459 wb_start_background_writeback(wb
);
1462 wb_dirty_limits(wb
, dirty_thresh
, background_thresh
,
1463 &wb_dirty
, &wb_thresh
, NULL
);
1465 dirty_exceeded
= (wb_dirty
> wb_thresh
) &&
1466 ((nr_dirty
> dirty_thresh
) || strictlimit
);
1467 if (dirty_exceeded
&& !wb
->dirty_exceeded
)
1468 wb
->dirty_exceeded
= 1;
1470 wb_update_bandwidth(wb
, dirty_thresh
, background_thresh
,
1471 nr_dirty
, wb_thresh
, wb_dirty
, start_time
);
1473 dirty_ratelimit
= wb
->dirty_ratelimit
;
1474 pos_ratio
= wb_position_ratio(wb
, dirty_thresh
,
1475 background_thresh
, nr_dirty
,
1476 wb_thresh
, wb_dirty
);
1477 task_ratelimit
= ((u64
)dirty_ratelimit
* pos_ratio
) >>
1478 RATELIMIT_CALC_SHIFT
;
1479 max_pause
= wb_max_pause(wb
, wb_dirty
);
1480 min_pause
= wb_min_pause(wb
, max_pause
,
1481 task_ratelimit
, dirty_ratelimit
,
1484 if (unlikely(task_ratelimit
== 0)) {
1489 period
= HZ
* pages_dirtied
/ task_ratelimit
;
1491 if (current
->dirty_paused_when
)
1492 pause
-= now
- current
->dirty_paused_when
;
1494 * For less than 1s think time (ext3/4 may block the dirtier
1495 * for up to 800ms from time to time on 1-HDD; so does xfs,
1496 * however at much less frequency), try to compensate it in
1497 * future periods by updating the virtual time; otherwise just
1498 * do a reset, as it may be a light dirtier.
1500 if (pause
< min_pause
) {
1501 trace_balance_dirty_pages(bdi
,
1514 current
->dirty_paused_when
= now
;
1515 current
->nr_dirtied
= 0;
1516 } else if (period
) {
1517 current
->dirty_paused_when
+= period
;
1518 current
->nr_dirtied
= 0;
1519 } else if (current
->nr_dirtied_pause
<= pages_dirtied
)
1520 current
->nr_dirtied_pause
+= pages_dirtied
;
1523 if (unlikely(pause
> max_pause
)) {
1524 /* for occasional dropped task_ratelimit */
1525 now
+= min(pause
- max_pause
, max_pause
);
1530 trace_balance_dirty_pages(bdi
,
1542 __set_current_state(TASK_KILLABLE
);
1543 io_schedule_timeout(pause
);
1545 current
->dirty_paused_when
= now
+ pause
;
1546 current
->nr_dirtied
= 0;
1547 current
->nr_dirtied_pause
= nr_dirtied_pause
;
1550 * This is typically equal to (nr_dirty < dirty_thresh) and can
1551 * also keep "1000+ dd on a slow USB stick" under control.
1557 * In the case of an unresponding NFS server and the NFS dirty
1558 * pages exceeds dirty_thresh, give the other good wb's a pipe
1559 * to go through, so that tasks on them still remain responsive.
1561 * In theory 1 page is enough to keep the comsumer-producer
1562 * pipe going: the flusher cleans 1 page => the task dirties 1
1563 * more page. However wb_dirty has accounting errors. So use
1564 * the larger and more IO friendly wb_stat_error.
1566 if (wb_dirty
<= wb_stat_error(wb
))
1569 if (fatal_signal_pending(current
))
1573 if (!dirty_exceeded
&& wb
->dirty_exceeded
)
1574 wb
->dirty_exceeded
= 0;
1576 if (writeback_in_progress(wb
))
1580 * In laptop mode, we wait until hitting the higher threshold before
1581 * starting background writeout, and then write out all the way down
1582 * to the lower threshold. So slow writers cause minimal disk activity.
1584 * In normal mode, we start background writeout at the lower
1585 * background_thresh, to keep the amount of dirty memory low.
1590 if (nr_reclaimable
> background_thresh
)
1591 wb_start_background_writeback(wb
);
1594 static DEFINE_PER_CPU(int, bdp_ratelimits
);
1597 * Normal tasks are throttled by
1599 * dirty tsk->nr_dirtied_pause pages;
1600 * take a snap in balance_dirty_pages();
1602 * However there is a worst case. If every task exit immediately when dirtied
1603 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1604 * called to throttle the page dirties. The solution is to save the not yet
1605 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1606 * randomly into the running tasks. This works well for the above worst case,
1607 * as the new task will pick up and accumulate the old task's leaked dirty
1608 * count and eventually get throttled.
1610 DEFINE_PER_CPU(int, dirty_throttle_leaks
) = 0;
1613 * balance_dirty_pages_ratelimited - balance dirty memory state
1614 * @mapping: address_space which was dirtied
1616 * Processes which are dirtying memory should call in here once for each page
1617 * which was newly dirtied. The function will periodically check the system's
1618 * dirty state and will initiate writeback if needed.
1620 * On really big machines, get_writeback_state is expensive, so try to avoid
1621 * calling it too often (ratelimiting). But once we're over the dirty memory
1622 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1623 * from overshooting the limit by (ratelimit_pages) each.
1625 void balance_dirty_pages_ratelimited(struct address_space
*mapping
)
1627 struct inode
*inode
= mapping
->host
;
1628 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
1629 struct bdi_writeback
*wb
= NULL
;
1633 if (!bdi_cap_account_dirty(bdi
))
1636 if (inode_cgwb_enabled(inode
))
1637 wb
= wb_get_create_current(bdi
, GFP_KERNEL
);
1641 ratelimit
= current
->nr_dirtied_pause
;
1642 if (wb
->dirty_exceeded
)
1643 ratelimit
= min(ratelimit
, 32 >> (PAGE_SHIFT
- 10));
1647 * This prevents one CPU to accumulate too many dirtied pages without
1648 * calling into balance_dirty_pages(), which can happen when there are
1649 * 1000+ tasks, all of them start dirtying pages at exactly the same
1650 * time, hence all honoured too large initial task->nr_dirtied_pause.
1652 p
= this_cpu_ptr(&bdp_ratelimits
);
1653 if (unlikely(current
->nr_dirtied
>= ratelimit
))
1655 else if (unlikely(*p
>= ratelimit_pages
)) {
1660 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1661 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1662 * the dirty throttling and livelock other long-run dirtiers.
1664 p
= this_cpu_ptr(&dirty_throttle_leaks
);
1665 if (*p
> 0 && current
->nr_dirtied
< ratelimit
) {
1666 unsigned long nr_pages_dirtied
;
1667 nr_pages_dirtied
= min(*p
, ratelimit
- current
->nr_dirtied
);
1668 *p
-= nr_pages_dirtied
;
1669 current
->nr_dirtied
+= nr_pages_dirtied
;
1673 if (unlikely(current
->nr_dirtied
>= ratelimit
))
1674 balance_dirty_pages(mapping
, wb
, current
->nr_dirtied
);
1678 EXPORT_SYMBOL(balance_dirty_pages_ratelimited
);
1680 void throttle_vm_writeout(gfp_t gfp_mask
)
1682 unsigned long background_thresh
;
1683 unsigned long dirty_thresh
;
1686 global_dirty_limits(&background_thresh
, &dirty_thresh
);
1687 dirty_thresh
= hard_dirty_limit(dirty_thresh
);
1690 * Boost the allowable dirty threshold a bit for page
1691 * allocators so they don't get DoS'ed by heavy writers
1693 dirty_thresh
+= dirty_thresh
/ 10; /* wheeee... */
1695 if (global_page_state(NR_UNSTABLE_NFS
) +
1696 global_page_state(NR_WRITEBACK
) <= dirty_thresh
)
1698 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1701 * The caller might hold locks which can prevent IO completion
1702 * or progress in the filesystem. So we cannot just sit here
1703 * waiting for IO to complete.
1705 if ((gfp_mask
& (__GFP_FS
|__GFP_IO
)) != (__GFP_FS
|__GFP_IO
))
1711 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1713 int dirty_writeback_centisecs_handler(struct ctl_table
*table
, int write
,
1714 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1716 proc_dointvec(table
, write
, buffer
, length
, ppos
);
1721 void laptop_mode_timer_fn(unsigned long data
)
1723 struct request_queue
*q
= (struct request_queue
*)data
;
1724 int nr_pages
= global_page_state(NR_FILE_DIRTY
) +
1725 global_page_state(NR_UNSTABLE_NFS
);
1726 struct bdi_writeback
*wb
;
1727 struct wb_iter iter
;
1730 * We want to write everything out, not just down to the dirty
1733 if (!bdi_has_dirty_io(&q
->backing_dev_info
))
1736 bdi_for_each_wb(wb
, &q
->backing_dev_info
, &iter
, 0)
1737 if (wb_has_dirty_io(wb
))
1738 wb_start_writeback(wb
, nr_pages
, true,
1739 WB_REASON_LAPTOP_TIMER
);
1743 * We've spun up the disk and we're in laptop mode: schedule writeback
1744 * of all dirty data a few seconds from now. If the flush is already scheduled
1745 * then push it back - the user is still using the disk.
1747 void laptop_io_completion(struct backing_dev_info
*info
)
1749 mod_timer(&info
->laptop_mode_wb_timer
, jiffies
+ laptop_mode
);
1753 * We're in laptop mode and we've just synced. The sync's writes will have
1754 * caused another writeback to be scheduled by laptop_io_completion.
1755 * Nothing needs to be written back anymore, so we unschedule the writeback.
1757 void laptop_sync_completion(void)
1759 struct backing_dev_info
*bdi
;
1763 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
)
1764 del_timer(&bdi
->laptop_mode_wb_timer
);
1771 * If ratelimit_pages is too high then we can get into dirty-data overload
1772 * if a large number of processes all perform writes at the same time.
1773 * If it is too low then SMP machines will call the (expensive)
1774 * get_writeback_state too often.
1776 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1777 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1781 void writeback_set_ratelimit(void)
1783 unsigned long background_thresh
;
1784 unsigned long dirty_thresh
;
1785 global_dirty_limits(&background_thresh
, &dirty_thresh
);
1786 global_dirty_limit
= dirty_thresh
;
1787 ratelimit_pages
= dirty_thresh
/ (num_online_cpus() * 32);
1788 if (ratelimit_pages
< 16)
1789 ratelimit_pages
= 16;
1793 ratelimit_handler(struct notifier_block
*self
, unsigned long action
,
1797 switch (action
& ~CPU_TASKS_FROZEN
) {
1800 writeback_set_ratelimit();
1807 static struct notifier_block ratelimit_nb
= {
1808 .notifier_call
= ratelimit_handler
,
1813 * Called early on to tune the page writeback dirty limits.
1815 * We used to scale dirty pages according to how total memory
1816 * related to pages that could be allocated for buffers (by
1817 * comparing nr_free_buffer_pages() to vm_total_pages.
1819 * However, that was when we used "dirty_ratio" to scale with
1820 * all memory, and we don't do that any more. "dirty_ratio"
1821 * is now applied to total non-HIGHPAGE memory (by subtracting
1822 * totalhigh_pages from vm_total_pages), and as such we can't
1823 * get into the old insane situation any more where we had
1824 * large amounts of dirty pages compared to a small amount of
1825 * non-HIGHMEM memory.
1827 * But we might still want to scale the dirty_ratio by how
1828 * much memory the box has..
1830 void __init
page_writeback_init(void)
1832 writeback_set_ratelimit();
1833 register_cpu_notifier(&ratelimit_nb
);
1835 fprop_global_init(&writeout_completions
, GFP_KERNEL
);
1839 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1840 * @mapping: address space structure to write
1841 * @start: starting page index
1842 * @end: ending page index (inclusive)
1844 * This function scans the page range from @start to @end (inclusive) and tags
1845 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1846 * that write_cache_pages (or whoever calls this function) will then use
1847 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1848 * used to avoid livelocking of writeback by a process steadily creating new
1849 * dirty pages in the file (thus it is important for this function to be quick
1850 * so that it can tag pages faster than a dirtying process can create them).
1853 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1855 void tag_pages_for_writeback(struct address_space
*mapping
,
1856 pgoff_t start
, pgoff_t end
)
1858 #define WRITEBACK_TAG_BATCH 4096
1859 unsigned long tagged
;
1862 spin_lock_irq(&mapping
->tree_lock
);
1863 tagged
= radix_tree_range_tag_if_tagged(&mapping
->page_tree
,
1864 &start
, end
, WRITEBACK_TAG_BATCH
,
1865 PAGECACHE_TAG_DIRTY
, PAGECACHE_TAG_TOWRITE
);
1866 spin_unlock_irq(&mapping
->tree_lock
);
1867 WARN_ON_ONCE(tagged
> WRITEBACK_TAG_BATCH
);
1869 /* We check 'start' to handle wrapping when end == ~0UL */
1870 } while (tagged
>= WRITEBACK_TAG_BATCH
&& start
);
1872 EXPORT_SYMBOL(tag_pages_for_writeback
);
1875 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1876 * @mapping: address space structure to write
1877 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1878 * @writepage: function called for each page
1879 * @data: data passed to writepage function
1881 * If a page is already under I/O, write_cache_pages() skips it, even
1882 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1883 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1884 * and msync() need to guarantee that all the data which was dirty at the time
1885 * the call was made get new I/O started against them. If wbc->sync_mode is
1886 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1887 * existing IO to complete.
1889 * To avoid livelocks (when other process dirties new pages), we first tag
1890 * pages which should be written back with TOWRITE tag and only then start
1891 * writing them. For data-integrity sync we have to be careful so that we do
1892 * not miss some pages (e.g., because some other process has cleared TOWRITE
1893 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1894 * by the process clearing the DIRTY tag (and submitting the page for IO).
1896 int write_cache_pages(struct address_space
*mapping
,
1897 struct writeback_control
*wbc
, writepage_t writepage
,
1902 struct pagevec pvec
;
1904 pgoff_t
uninitialized_var(writeback_index
);
1906 pgoff_t end
; /* Inclusive */
1909 int range_whole
= 0;
1912 pagevec_init(&pvec
, 0);
1913 if (wbc
->range_cyclic
) {
1914 writeback_index
= mapping
->writeback_index
; /* prev offset */
1915 index
= writeback_index
;
1922 index
= wbc
->range_start
>> PAGE_CACHE_SHIFT
;
1923 end
= wbc
->range_end
>> PAGE_CACHE_SHIFT
;
1924 if (wbc
->range_start
== 0 && wbc
->range_end
== LLONG_MAX
)
1926 cycled
= 1; /* ignore range_cyclic tests */
1928 if (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
)
1929 tag
= PAGECACHE_TAG_TOWRITE
;
1931 tag
= PAGECACHE_TAG_DIRTY
;
1933 if (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
)
1934 tag_pages_for_writeback(mapping
, index
, end
);
1936 while (!done
&& (index
<= end
)) {
1939 nr_pages
= pagevec_lookup_tag(&pvec
, mapping
, &index
, tag
,
1940 min(end
- index
, (pgoff_t
)PAGEVEC_SIZE
-1) + 1);
1944 for (i
= 0; i
< nr_pages
; i
++) {
1945 struct page
*page
= pvec
.pages
[i
];
1948 * At this point, the page may be truncated or
1949 * invalidated (changing page->mapping to NULL), or
1950 * even swizzled back from swapper_space to tmpfs file
1951 * mapping. However, page->index will not change
1952 * because we have a reference on the page.
1954 if (page
->index
> end
) {
1956 * can't be range_cyclic (1st pass) because
1957 * end == -1 in that case.
1963 done_index
= page
->index
;
1968 * Page truncated or invalidated. We can freely skip it
1969 * then, even for data integrity operations: the page
1970 * has disappeared concurrently, so there could be no
1971 * real expectation of this data interity operation
1972 * even if there is now a new, dirty page at the same
1973 * pagecache address.
1975 if (unlikely(page
->mapping
!= mapping
)) {
1981 if (!PageDirty(page
)) {
1982 /* someone wrote it for us */
1983 goto continue_unlock
;
1986 if (PageWriteback(page
)) {
1987 if (wbc
->sync_mode
!= WB_SYNC_NONE
)
1988 wait_on_page_writeback(page
);
1990 goto continue_unlock
;
1993 BUG_ON(PageWriteback(page
));
1994 if (!clear_page_dirty_for_io(page
))
1995 goto continue_unlock
;
1997 trace_wbc_writepage(wbc
, inode_to_bdi(mapping
->host
));
1998 ret
= (*writepage
)(page
, wbc
, data
);
1999 if (unlikely(ret
)) {
2000 if (ret
== AOP_WRITEPAGE_ACTIVATE
) {
2005 * done_index is set past this page,
2006 * so media errors will not choke
2007 * background writeout for the entire
2008 * file. This has consequences for
2009 * range_cyclic semantics (ie. it may
2010 * not be suitable for data integrity
2013 done_index
= page
->index
+ 1;
2020 * We stop writing back only if we are not doing
2021 * integrity sync. In case of integrity sync we have to
2022 * keep going until we have written all the pages
2023 * we tagged for writeback prior to entering this loop.
2025 if (--wbc
->nr_to_write
<= 0 &&
2026 wbc
->sync_mode
== WB_SYNC_NONE
) {
2031 pagevec_release(&pvec
);
2034 if (!cycled
&& !done
) {
2037 * We hit the last page and there is more work to be done: wrap
2038 * back to the start of the file
2042 end
= writeback_index
- 1;
2045 if (wbc
->range_cyclic
|| (range_whole
&& wbc
->nr_to_write
> 0))
2046 mapping
->writeback_index
= done_index
;
2050 EXPORT_SYMBOL(write_cache_pages
);
2053 * Function used by generic_writepages to call the real writepage
2054 * function and set the mapping flags on error
2056 static int __writepage(struct page
*page
, struct writeback_control
*wbc
,
2059 struct address_space
*mapping
= data
;
2060 int ret
= mapping
->a_ops
->writepage(page
, wbc
);
2061 mapping_set_error(mapping
, ret
);
2066 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2067 * @mapping: address space structure to write
2068 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2070 * This is a library function, which implements the writepages()
2071 * address_space_operation.
2073 int generic_writepages(struct address_space
*mapping
,
2074 struct writeback_control
*wbc
)
2076 struct blk_plug plug
;
2079 /* deal with chardevs and other special file */
2080 if (!mapping
->a_ops
->writepage
)
2083 blk_start_plug(&plug
);
2084 ret
= write_cache_pages(mapping
, wbc
, __writepage
, mapping
);
2085 blk_finish_plug(&plug
);
2089 EXPORT_SYMBOL(generic_writepages
);
2091 int do_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
2095 if (wbc
->nr_to_write
<= 0)
2097 if (mapping
->a_ops
->writepages
)
2098 ret
= mapping
->a_ops
->writepages(mapping
, wbc
);
2100 ret
= generic_writepages(mapping
, wbc
);
2105 * write_one_page - write out a single page and optionally wait on I/O
2106 * @page: the page to write
2107 * @wait: if true, wait on writeout
2109 * The page must be locked by the caller and will be unlocked upon return.
2111 * write_one_page() returns a negative error code if I/O failed.
2113 int write_one_page(struct page
*page
, int wait
)
2115 struct address_space
*mapping
= page
->mapping
;
2117 struct writeback_control wbc
= {
2118 .sync_mode
= WB_SYNC_ALL
,
2122 BUG_ON(!PageLocked(page
));
2125 wait_on_page_writeback(page
);
2127 if (clear_page_dirty_for_io(page
)) {
2128 page_cache_get(page
);
2129 ret
= mapping
->a_ops
->writepage(page
, &wbc
);
2130 if (ret
== 0 && wait
) {
2131 wait_on_page_writeback(page
);
2132 if (PageError(page
))
2135 page_cache_release(page
);
2141 EXPORT_SYMBOL(write_one_page
);
2144 * For address_spaces which do not use buffers nor write back.
2146 int __set_page_dirty_no_writeback(struct page
*page
)
2148 if (!PageDirty(page
))
2149 return !TestSetPageDirty(page
);
2154 * Helper function for set_page_dirty family.
2156 * Caller must hold mem_cgroup_begin_page_stat().
2158 * NOTE: This relies on being atomic wrt interrupts.
2160 void account_page_dirtied(struct page
*page
, struct address_space
*mapping
,
2161 struct mem_cgroup
*memcg
)
2163 struct inode
*inode
= mapping
->host
;
2165 trace_writeback_dirty_page(page
, mapping
);
2167 if (mapping_cap_account_dirty(mapping
)) {
2168 struct bdi_writeback
*wb
;
2170 inode_attach_wb(inode
, page
);
2171 wb
= inode_to_wb(inode
);
2173 mem_cgroup_inc_page_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
2174 __inc_zone_page_state(page
, NR_FILE_DIRTY
);
2175 __inc_zone_page_state(page
, NR_DIRTIED
);
2176 __inc_wb_stat(wb
, WB_RECLAIMABLE
);
2177 __inc_wb_stat(wb
, WB_DIRTIED
);
2178 task_io_account_write(PAGE_CACHE_SIZE
);
2179 current
->nr_dirtied
++;
2180 this_cpu_inc(bdp_ratelimits
);
2183 EXPORT_SYMBOL(account_page_dirtied
);
2186 * Helper function for deaccounting dirty page without writeback.
2188 * Caller must hold mem_cgroup_begin_page_stat().
2190 void account_page_cleaned(struct page
*page
, struct address_space
*mapping
,
2191 struct mem_cgroup
*memcg
)
2193 if (mapping_cap_account_dirty(mapping
)) {
2194 mem_cgroup_dec_page_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
2195 dec_zone_page_state(page
, NR_FILE_DIRTY
);
2196 dec_wb_stat(inode_to_wb(mapping
->host
), WB_RECLAIMABLE
);
2197 task_io_account_cancelled_write(PAGE_CACHE_SIZE
);
2202 * For address_spaces which do not use buffers. Just tag the page as dirty in
2205 * This is also used when a single buffer is being dirtied: we want to set the
2206 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2207 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2209 * The caller must ensure this doesn't race with truncation. Most will simply
2210 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2211 * the pte lock held, which also locks out truncation.
2213 int __set_page_dirty_nobuffers(struct page
*page
)
2215 struct mem_cgroup
*memcg
;
2217 memcg
= mem_cgroup_begin_page_stat(page
);
2218 if (!TestSetPageDirty(page
)) {
2219 struct address_space
*mapping
= page_mapping(page
);
2220 unsigned long flags
;
2223 mem_cgroup_end_page_stat(memcg
);
2227 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
2228 BUG_ON(page_mapping(page
) != mapping
);
2229 WARN_ON_ONCE(!PagePrivate(page
) && !PageUptodate(page
));
2230 account_page_dirtied(page
, mapping
, memcg
);
2231 radix_tree_tag_set(&mapping
->page_tree
, page_index(page
),
2232 PAGECACHE_TAG_DIRTY
);
2233 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
2234 mem_cgroup_end_page_stat(memcg
);
2236 if (mapping
->host
) {
2237 /* !PageAnon && !swapper_space */
2238 __mark_inode_dirty(mapping
->host
, I_DIRTY_PAGES
);
2242 mem_cgroup_end_page_stat(memcg
);
2245 EXPORT_SYMBOL(__set_page_dirty_nobuffers
);
2248 * Call this whenever redirtying a page, to de-account the dirty counters
2249 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2250 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2251 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2254 void account_page_redirty(struct page
*page
)
2256 struct address_space
*mapping
= page
->mapping
;
2258 if (mapping
&& mapping_cap_account_dirty(mapping
)) {
2259 struct bdi_writeback
*wb
= inode_to_wb(mapping
->host
);
2261 current
->nr_dirtied
--;
2262 dec_zone_page_state(page
, NR_DIRTIED
);
2263 dec_wb_stat(wb
, WB_DIRTIED
);
2266 EXPORT_SYMBOL(account_page_redirty
);
2269 * When a writepage implementation decides that it doesn't want to write this
2270 * page for some reason, it should redirty the locked page via
2271 * redirty_page_for_writepage() and it should then unlock the page and return 0
2273 int redirty_page_for_writepage(struct writeback_control
*wbc
, struct page
*page
)
2277 wbc
->pages_skipped
++;
2278 ret
= __set_page_dirty_nobuffers(page
);
2279 account_page_redirty(page
);
2282 EXPORT_SYMBOL(redirty_page_for_writepage
);
2287 * For pages with a mapping this should be done under the page lock
2288 * for the benefit of asynchronous memory errors who prefer a consistent
2289 * dirty state. This rule can be broken in some special cases,
2290 * but should be better not to.
2292 * If the mapping doesn't provide a set_page_dirty a_op, then
2293 * just fall through and assume that it wants buffer_heads.
2295 int set_page_dirty(struct page
*page
)
2297 struct address_space
*mapping
= page_mapping(page
);
2299 if (likely(mapping
)) {
2300 int (*spd
)(struct page
*) = mapping
->a_ops
->set_page_dirty
;
2302 * readahead/lru_deactivate_page could remain
2303 * PG_readahead/PG_reclaim due to race with end_page_writeback
2304 * About readahead, if the page is written, the flags would be
2305 * reset. So no problem.
2306 * About lru_deactivate_page, if the page is redirty, the flag
2307 * will be reset. So no problem. but if the page is used by readahead
2308 * it will confuse readahead and make it restart the size rampup
2309 * process. But it's a trivial problem.
2311 if (PageReclaim(page
))
2312 ClearPageReclaim(page
);
2315 spd
= __set_page_dirty_buffers
;
2317 return (*spd
)(page
);
2319 if (!PageDirty(page
)) {
2320 if (!TestSetPageDirty(page
))
2325 EXPORT_SYMBOL(set_page_dirty
);
2328 * set_page_dirty() is racy if the caller has no reference against
2329 * page->mapping->host, and if the page is unlocked. This is because another
2330 * CPU could truncate the page off the mapping and then free the mapping.
2332 * Usually, the page _is_ locked, or the caller is a user-space process which
2333 * holds a reference on the inode by having an open file.
2335 * In other cases, the page should be locked before running set_page_dirty().
2337 int set_page_dirty_lock(struct page
*page
)
2342 ret
= set_page_dirty(page
);
2346 EXPORT_SYMBOL(set_page_dirty_lock
);
2349 * This cancels just the dirty bit on the kernel page itself, it does NOT
2350 * actually remove dirty bits on any mmap's that may be around. It also
2351 * leaves the page tagged dirty, so any sync activity will still find it on
2352 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2353 * look at the dirty bits in the VM.
2355 * Doing this should *normally* only ever be done when a page is truncated,
2356 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2357 * this when it notices that somebody has cleaned out all the buffers on a
2358 * page without actually doing it through the VM. Can you say "ext3 is
2359 * horribly ugly"? Thought you could.
2361 void cancel_dirty_page(struct page
*page
)
2363 struct address_space
*mapping
= page_mapping(page
);
2365 if (mapping_cap_account_dirty(mapping
)) {
2366 struct mem_cgroup
*memcg
;
2368 memcg
= mem_cgroup_begin_page_stat(page
);
2370 if (TestClearPageDirty(page
))
2371 account_page_cleaned(page
, mapping
, memcg
);
2373 mem_cgroup_end_page_stat(memcg
);
2375 ClearPageDirty(page
);
2378 EXPORT_SYMBOL(cancel_dirty_page
);
2381 * Clear a page's dirty flag, while caring for dirty memory accounting.
2382 * Returns true if the page was previously dirty.
2384 * This is for preparing to put the page under writeout. We leave the page
2385 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2386 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2387 * implementation will run either set_page_writeback() or set_page_dirty(),
2388 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2391 * This incoherency between the page's dirty flag and radix-tree tag is
2392 * unfortunate, but it only exists while the page is locked.
2394 int clear_page_dirty_for_io(struct page
*page
)
2396 struct address_space
*mapping
= page_mapping(page
);
2397 struct mem_cgroup
*memcg
;
2400 BUG_ON(!PageLocked(page
));
2402 if (mapping
&& mapping_cap_account_dirty(mapping
)) {
2404 * Yes, Virginia, this is indeed insane.
2406 * We use this sequence to make sure that
2407 * (a) we account for dirty stats properly
2408 * (b) we tell the low-level filesystem to
2409 * mark the whole page dirty if it was
2410 * dirty in a pagetable. Only to then
2411 * (c) clean the page again and return 1 to
2412 * cause the writeback.
2414 * This way we avoid all nasty races with the
2415 * dirty bit in multiple places and clearing
2416 * them concurrently from different threads.
2418 * Note! Normally the "set_page_dirty(page)"
2419 * has no effect on the actual dirty bit - since
2420 * that will already usually be set. But we
2421 * need the side effects, and it can help us
2424 * We basically use the page "master dirty bit"
2425 * as a serialization point for all the different
2426 * threads doing their things.
2428 if (page_mkclean(page
))
2429 set_page_dirty(page
);
2431 * We carefully synchronise fault handlers against
2432 * installing a dirty pte and marking the page dirty
2433 * at this point. We do this by having them hold the
2434 * page lock while dirtying the page, and pages are
2435 * always locked coming in here, so we get the desired
2438 memcg
= mem_cgroup_begin_page_stat(page
);
2439 if (TestClearPageDirty(page
)) {
2440 mem_cgroup_dec_page_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
2441 dec_zone_page_state(page
, NR_FILE_DIRTY
);
2442 dec_wb_stat(inode_to_wb(mapping
->host
), WB_RECLAIMABLE
);
2445 mem_cgroup_end_page_stat(memcg
);
2448 return TestClearPageDirty(page
);
2450 EXPORT_SYMBOL(clear_page_dirty_for_io
);
2452 int test_clear_page_writeback(struct page
*page
)
2454 struct address_space
*mapping
= page_mapping(page
);
2455 struct mem_cgroup
*memcg
;
2458 memcg
= mem_cgroup_begin_page_stat(page
);
2460 struct inode
*inode
= mapping
->host
;
2461 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
2462 unsigned long flags
;
2464 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
2465 ret
= TestClearPageWriteback(page
);
2467 radix_tree_tag_clear(&mapping
->page_tree
,
2469 PAGECACHE_TAG_WRITEBACK
);
2470 if (bdi_cap_account_writeback(bdi
)) {
2471 struct bdi_writeback
*wb
= inode_to_wb(inode
);
2473 __dec_wb_stat(wb
, WB_WRITEBACK
);
2474 __wb_writeout_inc(wb
);
2477 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
2479 ret
= TestClearPageWriteback(page
);
2482 mem_cgroup_dec_page_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
2483 dec_zone_page_state(page
, NR_WRITEBACK
);
2484 inc_zone_page_state(page
, NR_WRITTEN
);
2486 mem_cgroup_end_page_stat(memcg
);
2490 int __test_set_page_writeback(struct page
*page
, bool keep_write
)
2492 struct address_space
*mapping
= page_mapping(page
);
2493 struct mem_cgroup
*memcg
;
2496 memcg
= mem_cgroup_begin_page_stat(page
);
2498 struct inode
*inode
= mapping
->host
;
2499 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
2500 unsigned long flags
;
2502 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
2503 ret
= TestSetPageWriteback(page
);
2505 radix_tree_tag_set(&mapping
->page_tree
,
2507 PAGECACHE_TAG_WRITEBACK
);
2508 if (bdi_cap_account_writeback(bdi
))
2509 __inc_wb_stat(inode_to_wb(inode
), WB_WRITEBACK
);
2511 if (!PageDirty(page
))
2512 radix_tree_tag_clear(&mapping
->page_tree
,
2514 PAGECACHE_TAG_DIRTY
);
2516 radix_tree_tag_clear(&mapping
->page_tree
,
2518 PAGECACHE_TAG_TOWRITE
);
2519 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
2521 ret
= TestSetPageWriteback(page
);
2524 mem_cgroup_inc_page_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
2525 inc_zone_page_state(page
, NR_WRITEBACK
);
2527 mem_cgroup_end_page_stat(memcg
);
2531 EXPORT_SYMBOL(__test_set_page_writeback
);
2534 * Return true if any of the pages in the mapping are marked with the
2537 int mapping_tagged(struct address_space
*mapping
, int tag
)
2539 return radix_tree_tagged(&mapping
->page_tree
, tag
);
2541 EXPORT_SYMBOL(mapping_tagged
);
2544 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2545 * @page: The page to wait on.
2547 * This function determines if the given page is related to a backing device
2548 * that requires page contents to be held stable during writeback. If so, then
2549 * it will wait for any pending writeback to complete.
2551 void wait_for_stable_page(struct page
*page
)
2553 if (bdi_cap_stable_pages_required(inode_to_bdi(page
->mapping
->host
)))
2554 wait_on_page_writeback(page
);
2556 EXPORT_SYMBOL_GPL(wait_for_stable_page
);