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 <trace/events/writeback.h>
40 * Sleep at most 200ms at a time in balance_dirty_pages().
42 #define MAX_PAUSE max(HZ/5, 1)
45 * Estimate write bandwidth at 200ms intervals.
47 #define BANDWIDTH_INTERVAL max(HZ/5, 1)
49 #define RATELIMIT_CALC_SHIFT 10
52 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
53 * will look to see if it needs to force writeback or throttling.
55 static long ratelimit_pages
= 32;
57 /* The following parameters are exported via /proc/sys/vm */
60 * Start background writeback (via writeback threads) at this percentage
62 int dirty_background_ratio
= 10;
65 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
66 * dirty_background_ratio * the amount of dirtyable memory
68 unsigned long dirty_background_bytes
;
71 * free highmem will not be subtracted from the total free memory
72 * for calculating free ratios if vm_highmem_is_dirtyable is true
74 int vm_highmem_is_dirtyable
;
77 * The generator of dirty data starts writeback at this percentage
79 int vm_dirty_ratio
= 20;
82 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
83 * vm_dirty_ratio * the amount of dirtyable memory
85 unsigned long vm_dirty_bytes
;
88 * The interval between `kupdate'-style writebacks
90 unsigned int dirty_writeback_interval
= 5 * 100; /* centiseconds */
93 * The longest time for which data is allowed to remain dirty
95 unsigned int dirty_expire_interval
= 30 * 100; /* centiseconds */
98 * Flag that makes the machine dump writes/reads and block dirtyings.
103 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
104 * a full sync is triggered after this time elapses without any disk activity.
108 EXPORT_SYMBOL(laptop_mode
);
110 /* End of sysctl-exported parameters */
112 unsigned long global_dirty_limit
;
115 * Scale the writeback cache size proportional to the relative writeout speeds.
117 * We do this by keeping a floating proportion between BDIs, based on page
118 * writeback completions [end_page_writeback()]. Those devices that write out
119 * pages fastest will get the larger share, while the slower will get a smaller
122 * We use page writeout completions because we are interested in getting rid of
123 * dirty pages. Having them written out is the primary goal.
125 * We introduce a concept of time, a period over which we measure these events,
126 * because demand can/will vary over time. The length of this period itself is
127 * measured in page writeback completions.
130 static struct prop_descriptor vm_completions
;
133 * Work out the current dirty-memory clamping and background writeout
136 * The main aim here is to lower them aggressively if there is a lot of mapped
137 * memory around. To avoid stressing page reclaim with lots of unreclaimable
138 * pages. It is better to clamp down on writers than to start swapping, and
139 * performing lots of scanning.
141 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
143 * We don't permit the clamping level to fall below 5% - that is getting rather
146 * We make sure that the background writeout level is below the adjusted
151 * In a memory zone, there is a certain amount of pages we consider
152 * available for the page cache, which is essentially the number of
153 * free and reclaimable pages, minus some zone reserves to protect
154 * lowmem and the ability to uphold the zone's watermarks without
155 * requiring writeback.
157 * This number of dirtyable pages is the base value of which the
158 * user-configurable dirty ratio is the effictive number of pages that
159 * are allowed to be actually dirtied. Per individual zone, or
160 * globally by using the sum of dirtyable pages over all zones.
162 * Because the user is allowed to specify the dirty limit globally as
163 * absolute number of bytes, calculating the per-zone dirty limit can
164 * require translating the configured limit into a percentage of
165 * global dirtyable memory first.
168 static unsigned long highmem_dirtyable_memory(unsigned long total
)
170 #ifdef CONFIG_HIGHMEM
174 for_each_node_state(node
, N_HIGH_MEMORY
) {
176 &NODE_DATA(node
)->node_zones
[ZONE_HIGHMEM
];
178 x
+= zone_page_state(z
, NR_FREE_PAGES
) +
179 zone_reclaimable_pages(z
) - z
->dirty_balance_reserve
;
182 * Make sure that the number of highmem pages is never larger
183 * than the number of the total dirtyable memory. This can only
184 * occur in very strange VM situations but we want to make sure
185 * that this does not occur.
187 return min(x
, total
);
194 * global_dirtyable_memory - number of globally dirtyable pages
196 * Returns the global number of pages potentially available for dirty
197 * page cache. This is the base value for the global dirty limits.
199 unsigned long global_dirtyable_memory(void)
203 x
= global_page_state(NR_FREE_PAGES
) + global_reclaimable_pages() -
204 dirty_balance_reserve
;
206 if (!vm_highmem_is_dirtyable
)
207 x
-= highmem_dirtyable_memory(x
);
209 return x
+ 1; /* Ensure that we never return 0 */
213 * global_dirty_limits - background-writeback and dirty-throttling thresholds
215 * Calculate the dirty thresholds based on sysctl parameters
216 * - vm.dirty_background_ratio or vm.dirty_background_bytes
217 * - vm.dirty_ratio or vm.dirty_bytes
218 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
221 void global_dirty_limits(unsigned long *pbackground
, unsigned long *pdirty
)
223 unsigned long background
;
225 unsigned long uninitialized_var(available_memory
);
226 struct task_struct
*tsk
;
228 if (!vm_dirty_bytes
|| !dirty_background_bytes
)
229 available_memory
= global_dirtyable_memory();
232 dirty
= DIV_ROUND_UP(vm_dirty_bytes
, PAGE_SIZE
);
234 dirty
= (vm_dirty_ratio
* available_memory
) / 100;
236 if (dirty_background_bytes
)
237 background
= DIV_ROUND_UP(dirty_background_bytes
, PAGE_SIZE
);
239 background
= (dirty_background_ratio
* available_memory
) / 100;
241 if (background
>= dirty
)
242 background
= dirty
/ 2;
244 if (tsk
->flags
& PF_LESS_THROTTLE
|| rt_task(tsk
)) {
245 background
+= background
/ 4;
248 *pbackground
= background
;
250 trace_global_dirty_state(background
, dirty
);
254 * zone_dirtyable_memory - number of dirtyable pages in a zone
257 * Returns the zone's number of pages potentially available for dirty
258 * page cache. This is the base value for the per-zone dirty limits.
260 static unsigned long zone_dirtyable_memory(struct zone
*zone
)
263 * The effective global number of dirtyable pages may exclude
264 * highmem as a big-picture measure to keep the ratio between
265 * dirty memory and lowmem reasonable.
267 * But this function is purely about the individual zone and a
268 * highmem zone can hold its share of dirty pages, so we don't
269 * care about vm_highmem_is_dirtyable here.
271 return zone_page_state(zone
, NR_FREE_PAGES
) +
272 zone_reclaimable_pages(zone
) -
273 zone
->dirty_balance_reserve
;
277 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
280 * Returns the maximum number of dirty pages allowed in a zone, based
281 * on the zone's dirtyable memory.
283 static unsigned long zone_dirty_limit(struct zone
*zone
)
285 unsigned long zone_memory
= zone_dirtyable_memory(zone
);
286 struct task_struct
*tsk
= current
;
290 dirty
= DIV_ROUND_UP(vm_dirty_bytes
, PAGE_SIZE
) *
291 zone_memory
/ global_dirtyable_memory();
293 dirty
= vm_dirty_ratio
* zone_memory
/ 100;
295 if (tsk
->flags
& PF_LESS_THROTTLE
|| rt_task(tsk
))
302 * zone_dirty_ok - tells whether a zone is within its dirty limits
303 * @zone: the zone to check
305 * Returns %true when the dirty pages in @zone are within the zone's
306 * dirty limit, %false if the limit is exceeded.
308 bool zone_dirty_ok(struct zone
*zone
)
310 unsigned long limit
= zone_dirty_limit(zone
);
312 return zone_page_state(zone
, NR_FILE_DIRTY
) +
313 zone_page_state(zone
, NR_UNSTABLE_NFS
) +
314 zone_page_state(zone
, NR_WRITEBACK
) <= limit
;
318 * couple the period to the dirty_ratio:
320 * period/2 ~ roundup_pow_of_two(dirty limit)
322 static int calc_period_shift(void)
324 unsigned long dirty_total
;
327 dirty_total
= vm_dirty_bytes
/ PAGE_SIZE
;
329 dirty_total
= (vm_dirty_ratio
* global_dirtyable_memory()) /
331 return 2 + ilog2(dirty_total
- 1);
335 * update the period when the dirty threshold changes.
337 static void update_completion_period(void)
339 int shift
= calc_period_shift();
340 prop_change_shift(&vm_completions
, shift
);
342 writeback_set_ratelimit();
345 int dirty_background_ratio_handler(struct ctl_table
*table
, int write
,
346 void __user
*buffer
, size_t *lenp
,
351 ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
352 if (ret
== 0 && write
)
353 dirty_background_bytes
= 0;
357 int dirty_background_bytes_handler(struct ctl_table
*table
, int write
,
358 void __user
*buffer
, size_t *lenp
,
363 ret
= proc_doulongvec_minmax(table
, write
, buffer
, lenp
, ppos
);
364 if (ret
== 0 && write
)
365 dirty_background_ratio
= 0;
369 int dirty_ratio_handler(struct ctl_table
*table
, int write
,
370 void __user
*buffer
, size_t *lenp
,
373 int old_ratio
= vm_dirty_ratio
;
376 ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
377 if (ret
== 0 && write
&& vm_dirty_ratio
!= old_ratio
) {
378 update_completion_period();
384 int dirty_bytes_handler(struct ctl_table
*table
, int write
,
385 void __user
*buffer
, size_t *lenp
,
388 unsigned long old_bytes
= vm_dirty_bytes
;
391 ret
= proc_doulongvec_minmax(table
, write
, buffer
, lenp
, ppos
);
392 if (ret
== 0 && write
&& vm_dirty_bytes
!= old_bytes
) {
393 update_completion_period();
400 * Increment the BDI's writeout completion count and the global writeout
401 * completion count. Called from test_clear_page_writeback().
403 static inline void __bdi_writeout_inc(struct backing_dev_info
*bdi
)
405 __inc_bdi_stat(bdi
, BDI_WRITTEN
);
406 __prop_inc_percpu_max(&vm_completions
, &bdi
->completions
,
410 void bdi_writeout_inc(struct backing_dev_info
*bdi
)
414 local_irq_save(flags
);
415 __bdi_writeout_inc(bdi
);
416 local_irq_restore(flags
);
418 EXPORT_SYMBOL_GPL(bdi_writeout_inc
);
421 * Obtain an accurate fraction of the BDI's portion.
423 static void bdi_writeout_fraction(struct backing_dev_info
*bdi
,
424 long *numerator
, long *denominator
)
426 prop_fraction_percpu(&vm_completions
, &bdi
->completions
,
427 numerator
, denominator
);
431 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
432 * registered backing devices, which, for obvious reasons, can not
435 static unsigned int bdi_min_ratio
;
437 int bdi_set_min_ratio(struct backing_dev_info
*bdi
, unsigned int min_ratio
)
441 spin_lock_bh(&bdi_lock
);
442 if (min_ratio
> bdi
->max_ratio
) {
445 min_ratio
-= bdi
->min_ratio
;
446 if (bdi_min_ratio
+ min_ratio
< 100) {
447 bdi_min_ratio
+= min_ratio
;
448 bdi
->min_ratio
+= min_ratio
;
453 spin_unlock_bh(&bdi_lock
);
458 int bdi_set_max_ratio(struct backing_dev_info
*bdi
, unsigned max_ratio
)
465 spin_lock_bh(&bdi_lock
);
466 if (bdi
->min_ratio
> max_ratio
) {
469 bdi
->max_ratio
= max_ratio
;
470 bdi
->max_prop_frac
= (PROP_FRAC_BASE
* max_ratio
) / 100;
472 spin_unlock_bh(&bdi_lock
);
476 EXPORT_SYMBOL(bdi_set_max_ratio
);
478 static unsigned long dirty_freerun_ceiling(unsigned long thresh
,
479 unsigned long bg_thresh
)
481 return (thresh
+ bg_thresh
) / 2;
484 static unsigned long hard_dirty_limit(unsigned long thresh
)
486 return max(thresh
, global_dirty_limit
);
490 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
491 * @bdi: the backing_dev_info to query
492 * @dirty: global dirty limit in pages
494 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
495 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
497 * Note that balance_dirty_pages() will only seriously take it as a hard limit
498 * when sleeping max_pause per page is not enough to keep the dirty pages under
499 * control. For example, when the device is completely stalled due to some error
500 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
501 * In the other normal situations, it acts more gently by throttling the tasks
502 * more (rather than completely block them) when the bdi dirty pages go high.
504 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
505 * - starving fast devices
506 * - piling up dirty pages (that will take long time to sync) on slow devices
508 * The bdi's share of dirty limit will be adapting to its throughput and
509 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
511 unsigned long bdi_dirty_limit(struct backing_dev_info
*bdi
, unsigned long dirty
)
514 long numerator
, denominator
;
517 * Calculate this BDI's share of the dirty ratio.
519 bdi_writeout_fraction(bdi
, &numerator
, &denominator
);
521 bdi_dirty
= (dirty
* (100 - bdi_min_ratio
)) / 100;
522 bdi_dirty
*= numerator
;
523 do_div(bdi_dirty
, denominator
);
525 bdi_dirty
+= (dirty
* bdi
->min_ratio
) / 100;
526 if (bdi_dirty
> (dirty
* bdi
->max_ratio
) / 100)
527 bdi_dirty
= dirty
* bdi
->max_ratio
/ 100;
533 * Dirty position control.
535 * (o) global/bdi setpoints
537 * We want the dirty pages be balanced around the global/bdi setpoints.
538 * When the number of dirty pages is higher/lower than the setpoint, the
539 * dirty position control ratio (and hence task dirty ratelimit) will be
540 * decreased/increased to bring the dirty pages back to the setpoint.
542 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
544 * if (dirty < setpoint) scale up pos_ratio
545 * if (dirty > setpoint) scale down pos_ratio
547 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
548 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
550 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
552 * (o) global control line
556 * | |<===== global dirty control scope ======>|
564 * 1.0 ................................*
570 * 0 +------------.------------------.----------------------*------------->
571 * freerun^ setpoint^ limit^ dirty pages
573 * (o) bdi control line
581 * | * |<=========== span ============>|
582 * 1.0 .......................*
594 * 1/4 ...............................................* * * * * * * * * * * *
598 * 0 +----------------------.-------------------------------.------------->
599 * bdi_setpoint^ x_intercept^
601 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
602 * be smoothly throttled down to normal if it starts high in situations like
603 * - start writing to a slow SD card and a fast disk at the same time. The SD
604 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
605 * - the bdi dirty thresh drops quickly due to change of JBOD workload
607 static unsigned long bdi_position_ratio(struct backing_dev_info
*bdi
,
608 unsigned long thresh
,
609 unsigned long bg_thresh
,
611 unsigned long bdi_thresh
,
612 unsigned long bdi_dirty
)
614 unsigned long write_bw
= bdi
->avg_write_bandwidth
;
615 unsigned long freerun
= dirty_freerun_ceiling(thresh
, bg_thresh
);
616 unsigned long limit
= hard_dirty_limit(thresh
);
617 unsigned long x_intercept
;
618 unsigned long setpoint
; /* dirty pages' target balance point */
619 unsigned long bdi_setpoint
;
621 long long pos_ratio
; /* for scaling up/down the rate limit */
624 if (unlikely(dirty
>= limit
))
631 * f(dirty) := 1.0 + (----------------)
634 * it's a 3rd order polynomial that subjects to
636 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
637 * (2) f(setpoint) = 1.0 => the balance point
638 * (3) f(limit) = 0 => the hard limit
639 * (4) df/dx <= 0 => negative feedback control
640 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
641 * => fast response on large errors; small oscillation near setpoint
643 setpoint
= (freerun
+ limit
) / 2;
644 x
= div_s64((setpoint
- dirty
) << RATELIMIT_CALC_SHIFT
,
645 limit
- setpoint
+ 1);
647 pos_ratio
= pos_ratio
* x
>> RATELIMIT_CALC_SHIFT
;
648 pos_ratio
= pos_ratio
* x
>> RATELIMIT_CALC_SHIFT
;
649 pos_ratio
+= 1 << RATELIMIT_CALC_SHIFT
;
652 * We have computed basic pos_ratio above based on global situation. If
653 * the bdi is over/under its share of dirty pages, we want to scale
654 * pos_ratio further down/up. That is done by the following mechanism.
660 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
662 * x_intercept - bdi_dirty
663 * := --------------------------
664 * x_intercept - bdi_setpoint
666 * The main bdi control line is a linear function that subjects to
668 * (1) f(bdi_setpoint) = 1.0
669 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
670 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
672 * For single bdi case, the dirty pages are observed to fluctuate
673 * regularly within range
674 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
675 * for various filesystems, where (2) can yield in a reasonable 12.5%
676 * fluctuation range for pos_ratio.
678 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
679 * own size, so move the slope over accordingly and choose a slope that
680 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
682 if (unlikely(bdi_thresh
> thresh
))
685 * It's very possible that bdi_thresh is close to 0 not because the
686 * device is slow, but that it has remained inactive for long time.
687 * Honour such devices a reasonable good (hopefully IO efficient)
688 * threshold, so that the occasional writes won't be blocked and active
689 * writes can rampup the threshold quickly.
691 bdi_thresh
= max(bdi_thresh
, (limit
- dirty
) / 8);
693 * scale global setpoint to bdi's:
694 * bdi_setpoint = setpoint * bdi_thresh / thresh
696 x
= div_u64((u64
)bdi_thresh
<< 16, thresh
+ 1);
697 bdi_setpoint
= setpoint
* (u64
)x
>> 16;
699 * Use span=(8*write_bw) in single bdi case as indicated by
700 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
702 * bdi_thresh thresh - bdi_thresh
703 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
706 span
= (thresh
- bdi_thresh
+ 8 * write_bw
) * (u64
)x
>> 16;
707 x_intercept
= bdi_setpoint
+ span
;
709 if (bdi_dirty
< x_intercept
- span
/ 4) {
710 pos_ratio
= div_u64(pos_ratio
* (x_intercept
- bdi_dirty
),
711 x_intercept
- bdi_setpoint
+ 1);
716 * bdi reserve area, safeguard against dirty pool underrun and disk idle
717 * It may push the desired control point of global dirty pages higher
720 x_intercept
= bdi_thresh
/ 2;
721 if (bdi_dirty
< x_intercept
) {
722 if (bdi_dirty
> x_intercept
/ 8)
723 pos_ratio
= div_u64(pos_ratio
* x_intercept
, bdi_dirty
);
731 static void bdi_update_write_bandwidth(struct backing_dev_info
*bdi
,
732 unsigned long elapsed
,
733 unsigned long written
)
735 const unsigned long period
= roundup_pow_of_two(3 * HZ
);
736 unsigned long avg
= bdi
->avg_write_bandwidth
;
737 unsigned long old
= bdi
->write_bandwidth
;
741 * bw = written * HZ / elapsed
743 * bw * elapsed + write_bandwidth * (period - elapsed)
744 * write_bandwidth = ---------------------------------------------------
747 bw
= written
- bdi
->written_stamp
;
749 if (unlikely(elapsed
> period
)) {
754 bw
+= (u64
)bdi
->write_bandwidth
* (period
- elapsed
);
755 bw
>>= ilog2(period
);
758 * one more level of smoothing, for filtering out sudden spikes
760 if (avg
> old
&& old
>= (unsigned long)bw
)
761 avg
-= (avg
- old
) >> 3;
763 if (avg
< old
&& old
<= (unsigned long)bw
)
764 avg
+= (old
- avg
) >> 3;
767 bdi
->write_bandwidth
= bw
;
768 bdi
->avg_write_bandwidth
= avg
;
772 * The global dirtyable memory and dirty threshold could be suddenly knocked
773 * down by a large amount (eg. on the startup of KVM in a swapless system).
774 * This may throw the system into deep dirty exceeded state and throttle
775 * heavy/light dirtiers alike. To retain good responsiveness, maintain
776 * global_dirty_limit for tracking slowly down to the knocked down dirty
779 static void update_dirty_limit(unsigned long thresh
, unsigned long dirty
)
781 unsigned long limit
= global_dirty_limit
;
784 * Follow up in one step.
786 if (limit
< thresh
) {
792 * Follow down slowly. Use the higher one as the target, because thresh
793 * may drop below dirty. This is exactly the reason to introduce
794 * global_dirty_limit which is guaranteed to lie above the dirty pages.
796 thresh
= max(thresh
, dirty
);
797 if (limit
> thresh
) {
798 limit
-= (limit
- thresh
) >> 5;
803 global_dirty_limit
= limit
;
806 static void global_update_bandwidth(unsigned long thresh
,
810 static DEFINE_SPINLOCK(dirty_lock
);
811 static unsigned long update_time
;
814 * check locklessly first to optimize away locking for the most time
816 if (time_before(now
, update_time
+ BANDWIDTH_INTERVAL
))
819 spin_lock(&dirty_lock
);
820 if (time_after_eq(now
, update_time
+ BANDWIDTH_INTERVAL
)) {
821 update_dirty_limit(thresh
, dirty
);
824 spin_unlock(&dirty_lock
);
828 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
830 * Normal bdi tasks will be curbed at or below it in long term.
831 * Obviously it should be around (write_bw / N) when there are N dd tasks.
833 static void bdi_update_dirty_ratelimit(struct backing_dev_info
*bdi
,
834 unsigned long thresh
,
835 unsigned long bg_thresh
,
837 unsigned long bdi_thresh
,
838 unsigned long bdi_dirty
,
839 unsigned long dirtied
,
840 unsigned long elapsed
)
842 unsigned long freerun
= dirty_freerun_ceiling(thresh
, bg_thresh
);
843 unsigned long limit
= hard_dirty_limit(thresh
);
844 unsigned long setpoint
= (freerun
+ limit
) / 2;
845 unsigned long write_bw
= bdi
->avg_write_bandwidth
;
846 unsigned long dirty_ratelimit
= bdi
->dirty_ratelimit
;
847 unsigned long dirty_rate
;
848 unsigned long task_ratelimit
;
849 unsigned long balanced_dirty_ratelimit
;
850 unsigned long pos_ratio
;
855 * The dirty rate will match the writeout rate in long term, except
856 * when dirty pages are truncated by userspace or re-dirtied by FS.
858 dirty_rate
= (dirtied
- bdi
->dirtied_stamp
) * HZ
/ elapsed
;
860 pos_ratio
= bdi_position_ratio(bdi
, thresh
, bg_thresh
, dirty
,
861 bdi_thresh
, bdi_dirty
);
863 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
865 task_ratelimit
= (u64
)dirty_ratelimit
*
866 pos_ratio
>> RATELIMIT_CALC_SHIFT
;
867 task_ratelimit
++; /* it helps rampup dirty_ratelimit from tiny values */
870 * A linear estimation of the "balanced" throttle rate. The theory is,
871 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
872 * dirty_rate will be measured to be (N * task_ratelimit). So the below
873 * formula will yield the balanced rate limit (write_bw / N).
875 * Note that the expanded form is not a pure rate feedback:
876 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
877 * but also takes pos_ratio into account:
878 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
880 * (1) is not realistic because pos_ratio also takes part in balancing
881 * the dirty rate. Consider the state
882 * pos_ratio = 0.5 (3)
883 * rate = 2 * (write_bw / N) (4)
884 * If (1) is used, it will stuck in that state! Because each dd will
886 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
888 * dirty_rate = N * task_ratelimit = write_bw (6)
889 * put (6) into (1) we get
890 * rate_(i+1) = rate_(i) (7)
892 * So we end up using (2) to always keep
893 * rate_(i+1) ~= (write_bw / N) (8)
894 * regardless of the value of pos_ratio. As long as (8) is satisfied,
895 * pos_ratio is able to drive itself to 1.0, which is not only where
896 * the dirty count meet the setpoint, but also where the slope of
897 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
899 balanced_dirty_ratelimit
= div_u64((u64
)task_ratelimit
* write_bw
,
903 * We could safely do this and return immediately:
905 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
907 * However to get a more stable dirty_ratelimit, the below elaborated
908 * code makes use of task_ratelimit to filter out sigular points and
909 * limit the step size.
911 * The below code essentially only uses the relative value of
913 * task_ratelimit - dirty_ratelimit
914 * = (pos_ratio - 1) * dirty_ratelimit
916 * which reflects the direction and size of dirty position error.
920 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
921 * task_ratelimit is on the same side of dirty_ratelimit, too.
923 * - dirty_ratelimit > balanced_dirty_ratelimit
924 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
925 * lowering dirty_ratelimit will help meet both the position and rate
926 * control targets. Otherwise, don't update dirty_ratelimit if it will
927 * only help meet the rate target. After all, what the users ultimately
928 * feel and care are stable dirty rate and small position error.
930 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
931 * and filter out the sigular points of balanced_dirty_ratelimit. Which
932 * keeps jumping around randomly and can even leap far away at times
933 * due to the small 200ms estimation period of dirty_rate (we want to
934 * keep that period small to reduce time lags).
937 if (dirty
< setpoint
) {
938 x
= min(bdi
->balanced_dirty_ratelimit
,
939 min(balanced_dirty_ratelimit
, task_ratelimit
));
940 if (dirty_ratelimit
< x
)
941 step
= x
- dirty_ratelimit
;
943 x
= max(bdi
->balanced_dirty_ratelimit
,
944 max(balanced_dirty_ratelimit
, task_ratelimit
));
945 if (dirty_ratelimit
> x
)
946 step
= dirty_ratelimit
- x
;
950 * Don't pursue 100% rate matching. It's impossible since the balanced
951 * rate itself is constantly fluctuating. So decrease the track speed
952 * when it gets close to the target. Helps eliminate pointless tremors.
954 step
>>= dirty_ratelimit
/ (2 * step
+ 1);
956 * Limit the tracking speed to avoid overshooting.
958 step
= (step
+ 7) / 8;
960 if (dirty_ratelimit
< balanced_dirty_ratelimit
)
961 dirty_ratelimit
+= step
;
963 dirty_ratelimit
-= step
;
965 bdi
->dirty_ratelimit
= max(dirty_ratelimit
, 1UL);
966 bdi
->balanced_dirty_ratelimit
= balanced_dirty_ratelimit
;
968 trace_bdi_dirty_ratelimit(bdi
, dirty_rate
, task_ratelimit
);
971 void __bdi_update_bandwidth(struct backing_dev_info
*bdi
,
972 unsigned long thresh
,
973 unsigned long bg_thresh
,
975 unsigned long bdi_thresh
,
976 unsigned long bdi_dirty
,
977 unsigned long start_time
)
979 unsigned long now
= jiffies
;
980 unsigned long elapsed
= now
- bdi
->bw_time_stamp
;
981 unsigned long dirtied
;
982 unsigned long written
;
985 * rate-limit, only update once every 200ms.
987 if (elapsed
< BANDWIDTH_INTERVAL
)
990 dirtied
= percpu_counter_read(&bdi
->bdi_stat
[BDI_DIRTIED
]);
991 written
= percpu_counter_read(&bdi
->bdi_stat
[BDI_WRITTEN
]);
994 * Skip quiet periods when disk bandwidth is under-utilized.
995 * (at least 1s idle time between two flusher runs)
997 if (elapsed
> HZ
&& time_before(bdi
->bw_time_stamp
, start_time
))
1001 global_update_bandwidth(thresh
, dirty
, now
);
1002 bdi_update_dirty_ratelimit(bdi
, thresh
, bg_thresh
, dirty
,
1003 bdi_thresh
, bdi_dirty
,
1006 bdi_update_write_bandwidth(bdi
, elapsed
, written
);
1009 bdi
->dirtied_stamp
= dirtied
;
1010 bdi
->written_stamp
= written
;
1011 bdi
->bw_time_stamp
= now
;
1014 static void bdi_update_bandwidth(struct backing_dev_info
*bdi
,
1015 unsigned long thresh
,
1016 unsigned long bg_thresh
,
1017 unsigned long dirty
,
1018 unsigned long bdi_thresh
,
1019 unsigned long bdi_dirty
,
1020 unsigned long start_time
)
1022 if (time_is_after_eq_jiffies(bdi
->bw_time_stamp
+ BANDWIDTH_INTERVAL
))
1024 spin_lock(&bdi
->wb
.list_lock
);
1025 __bdi_update_bandwidth(bdi
, thresh
, bg_thresh
, dirty
,
1026 bdi_thresh
, bdi_dirty
, start_time
);
1027 spin_unlock(&bdi
->wb
.list_lock
);
1031 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
1032 * will look to see if it needs to start dirty throttling.
1034 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1035 * global_page_state() too often. So scale it near-sqrt to the safety margin
1036 * (the number of pages we may dirty without exceeding the dirty limits).
1038 static unsigned long dirty_poll_interval(unsigned long dirty
,
1039 unsigned long thresh
)
1042 return 1UL << (ilog2(thresh
- dirty
) >> 1);
1047 static unsigned long bdi_max_pause(struct backing_dev_info
*bdi
,
1048 unsigned long bdi_dirty
)
1050 unsigned long bw
= bdi
->avg_write_bandwidth
;
1051 unsigned long hi
= ilog2(bw
);
1052 unsigned long lo
= ilog2(bdi
->dirty_ratelimit
);
1055 /* target for 20ms max pause on 1-dd case */
1059 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1062 * (N * 20ms) on 2^N concurrent tasks.
1065 t
+= (hi
- lo
) * (20 * HZ
) / 1024;
1068 * Limit pause time for small memory systems. If sleeping for too long
1069 * time, a small pool of dirty/writeback pages may go empty and disk go
1072 * 8 serves as the safety ratio.
1074 t
= min(t
, bdi_dirty
* HZ
/ (8 * bw
+ 1));
1077 * The pause time will be settled within range (max_pause/4, max_pause).
1078 * Apply a minimal value of 4 to get a non-zero max_pause/4.
1080 return clamp_val(t
, 4, MAX_PAUSE
);
1084 * balance_dirty_pages() must be called by processes which are generating dirty
1085 * data. It looks at the number of dirty pages in the machine and will force
1086 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1087 * If we're over `background_thresh' then the writeback threads are woken to
1088 * perform some writeout.
1090 static void balance_dirty_pages(struct address_space
*mapping
,
1091 unsigned long pages_dirtied
)
1093 unsigned long nr_reclaimable
; /* = file_dirty + unstable_nfs */
1094 unsigned long bdi_reclaimable
;
1095 unsigned long nr_dirty
; /* = file_dirty + writeback + unstable_nfs */
1096 unsigned long bdi_dirty
;
1097 unsigned long freerun
;
1098 unsigned long background_thresh
;
1099 unsigned long dirty_thresh
;
1100 unsigned long bdi_thresh
;
1102 long uninitialized_var(max_pause
);
1103 bool dirty_exceeded
= false;
1104 unsigned long task_ratelimit
;
1105 unsigned long uninitialized_var(dirty_ratelimit
);
1106 unsigned long pos_ratio
;
1107 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
1108 unsigned long start_time
= jiffies
;
1112 * Unstable writes are a feature of certain networked
1113 * filesystems (i.e. NFS) in which data may have been
1114 * written to the server's write cache, but has not yet
1115 * been flushed to permanent storage.
1117 nr_reclaimable
= global_page_state(NR_FILE_DIRTY
) +
1118 global_page_state(NR_UNSTABLE_NFS
);
1119 nr_dirty
= nr_reclaimable
+ global_page_state(NR_WRITEBACK
);
1121 global_dirty_limits(&background_thresh
, &dirty_thresh
);
1124 * Throttle it only when the background writeback cannot
1125 * catch-up. This avoids (excessively) small writeouts
1126 * when the bdi limits are ramping up.
1128 freerun
= dirty_freerun_ceiling(dirty_thresh
,
1130 if (nr_dirty
<= freerun
)
1133 if (unlikely(!writeback_in_progress(bdi
)))
1134 bdi_start_background_writeback(bdi
);
1137 * bdi_thresh is not treated as some limiting factor as
1138 * dirty_thresh, due to reasons
1139 * - in JBOD setup, bdi_thresh can fluctuate a lot
1140 * - in a system with HDD and USB key, the USB key may somehow
1141 * go into state (bdi_dirty >> bdi_thresh) either because
1142 * bdi_dirty starts high, or because bdi_thresh drops low.
1143 * In this case we don't want to hard throttle the USB key
1144 * dirtiers for 100 seconds until bdi_dirty drops under
1145 * bdi_thresh. Instead the auxiliary bdi control line in
1146 * bdi_position_ratio() will let the dirtier task progress
1147 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1149 bdi_thresh
= bdi_dirty_limit(bdi
, dirty_thresh
);
1152 * In order to avoid the stacked BDI deadlock we need
1153 * to ensure we accurately count the 'dirty' pages when
1154 * the threshold is low.
1156 * Otherwise it would be possible to get thresh+n pages
1157 * reported dirty, even though there are thresh-m pages
1158 * actually dirty; with m+n sitting in the percpu
1161 if (bdi_thresh
< 2 * bdi_stat_error(bdi
)) {
1162 bdi_reclaimable
= bdi_stat_sum(bdi
, BDI_RECLAIMABLE
);
1163 bdi_dirty
= bdi_reclaimable
+
1164 bdi_stat_sum(bdi
, BDI_WRITEBACK
);
1166 bdi_reclaimable
= bdi_stat(bdi
, BDI_RECLAIMABLE
);
1167 bdi_dirty
= bdi_reclaimable
+
1168 bdi_stat(bdi
, BDI_WRITEBACK
);
1171 dirty_exceeded
= (bdi_dirty
> bdi_thresh
) ||
1172 (nr_dirty
> dirty_thresh
);
1173 if (dirty_exceeded
&& !bdi
->dirty_exceeded
)
1174 bdi
->dirty_exceeded
= 1;
1176 bdi_update_bandwidth(bdi
, dirty_thresh
, background_thresh
,
1177 nr_dirty
, bdi_thresh
, bdi_dirty
,
1180 max_pause
= bdi_max_pause(bdi
, bdi_dirty
);
1182 dirty_ratelimit
= bdi
->dirty_ratelimit
;
1183 pos_ratio
= bdi_position_ratio(bdi
, dirty_thresh
,
1184 background_thresh
, nr_dirty
,
1185 bdi_thresh
, bdi_dirty
);
1186 task_ratelimit
= ((u64
)dirty_ratelimit
* pos_ratio
) >>
1187 RATELIMIT_CALC_SHIFT
;
1188 if (unlikely(task_ratelimit
== 0)) {
1192 pause
= HZ
* pages_dirtied
/ task_ratelimit
;
1193 if (unlikely(pause
<= 0)) {
1194 trace_balance_dirty_pages(bdi
,
1205 pause
= 1; /* avoid resetting nr_dirtied_pause below */
1208 pause
= min(pause
, max_pause
);
1211 trace_balance_dirty_pages(bdi
,
1222 __set_current_state(TASK_KILLABLE
);
1223 io_schedule_timeout(pause
);
1226 * This is typically equal to (nr_dirty < dirty_thresh) and can
1227 * also keep "1000+ dd on a slow USB stick" under control.
1233 * In the case of an unresponding NFS server and the NFS dirty
1234 * pages exceeds dirty_thresh, give the other good bdi's a pipe
1235 * to go through, so that tasks on them still remain responsive.
1237 * In theory 1 page is enough to keep the comsumer-producer
1238 * pipe going: the flusher cleans 1 page => the task dirties 1
1239 * more page. However bdi_dirty has accounting errors. So use
1240 * the larger and more IO friendly bdi_stat_error.
1242 if (bdi_dirty
<= bdi_stat_error(bdi
))
1245 if (fatal_signal_pending(current
))
1249 if (!dirty_exceeded
&& bdi
->dirty_exceeded
)
1250 bdi
->dirty_exceeded
= 0;
1252 current
->nr_dirtied
= 0;
1253 if (pause
== 0) { /* in freerun area */
1254 current
->nr_dirtied_pause
=
1255 dirty_poll_interval(nr_dirty
, dirty_thresh
);
1256 } else if (pause
<= max_pause
/ 4 &&
1257 pages_dirtied
>= current
->nr_dirtied_pause
) {
1258 current
->nr_dirtied_pause
= clamp_val(
1259 dirty_ratelimit
* (max_pause
/ 2) / HZ
,
1260 pages_dirtied
+ pages_dirtied
/ 8,
1262 } else if (pause
>= max_pause
) {
1263 current
->nr_dirtied_pause
= 1 | clamp_val(
1264 dirty_ratelimit
* (max_pause
/ 2) / HZ
,
1266 pages_dirtied
- pages_dirtied
/ 8);
1269 if (writeback_in_progress(bdi
))
1273 * In laptop mode, we wait until hitting the higher threshold before
1274 * starting background writeout, and then write out all the way down
1275 * to the lower threshold. So slow writers cause minimal disk activity.
1277 * In normal mode, we start background writeout at the lower
1278 * background_thresh, to keep the amount of dirty memory low.
1283 if (nr_reclaimable
> background_thresh
)
1284 bdi_start_background_writeback(bdi
);
1287 void set_page_dirty_balance(struct page
*page
, int page_mkwrite
)
1289 if (set_page_dirty(page
) || page_mkwrite
) {
1290 struct address_space
*mapping
= page_mapping(page
);
1293 balance_dirty_pages_ratelimited(mapping
);
1297 static DEFINE_PER_CPU(int, bdp_ratelimits
);
1300 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
1301 * @mapping: address_space which was dirtied
1302 * @nr_pages_dirtied: number of pages which the caller has just dirtied
1304 * Processes which are dirtying memory should call in here once for each page
1305 * which was newly dirtied. The function will periodically check the system's
1306 * dirty state and will initiate writeback if needed.
1308 * On really big machines, get_writeback_state is expensive, so try to avoid
1309 * calling it too often (ratelimiting). But once we're over the dirty memory
1310 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1311 * from overshooting the limit by (ratelimit_pages) each.
1313 void balance_dirty_pages_ratelimited_nr(struct address_space
*mapping
,
1314 unsigned long nr_pages_dirtied
)
1316 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
1320 if (!bdi_cap_account_dirty(bdi
))
1323 ratelimit
= current
->nr_dirtied_pause
;
1324 if (bdi
->dirty_exceeded
)
1325 ratelimit
= min(ratelimit
, 32 >> (PAGE_SHIFT
- 10));
1327 current
->nr_dirtied
+= nr_pages_dirtied
;
1331 * This prevents one CPU to accumulate too many dirtied pages without
1332 * calling into balance_dirty_pages(), which can happen when there are
1333 * 1000+ tasks, all of them start dirtying pages at exactly the same
1334 * time, hence all honoured too large initial task->nr_dirtied_pause.
1336 p
= &__get_cpu_var(bdp_ratelimits
);
1337 if (unlikely(current
->nr_dirtied
>= ratelimit
))
1340 *p
+= nr_pages_dirtied
;
1341 if (unlikely(*p
>= ratelimit_pages
)) {
1348 if (unlikely(current
->nr_dirtied
>= ratelimit
))
1349 balance_dirty_pages(mapping
, current
->nr_dirtied
);
1351 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr
);
1353 void throttle_vm_writeout(gfp_t gfp_mask
)
1355 unsigned long background_thresh
;
1356 unsigned long dirty_thresh
;
1359 global_dirty_limits(&background_thresh
, &dirty_thresh
);
1362 * Boost the allowable dirty threshold a bit for page
1363 * allocators so they don't get DoS'ed by heavy writers
1365 dirty_thresh
+= dirty_thresh
/ 10; /* wheeee... */
1367 if (global_page_state(NR_UNSTABLE_NFS
) +
1368 global_page_state(NR_WRITEBACK
) <= dirty_thresh
)
1370 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1373 * The caller might hold locks which can prevent IO completion
1374 * or progress in the filesystem. So we cannot just sit here
1375 * waiting for IO to complete.
1377 if ((gfp_mask
& (__GFP_FS
|__GFP_IO
)) != (__GFP_FS
|__GFP_IO
))
1383 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1385 int dirty_writeback_centisecs_handler(ctl_table
*table
, int write
,
1386 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1388 proc_dointvec(table
, write
, buffer
, length
, ppos
);
1389 bdi_arm_supers_timer();
1394 void laptop_mode_timer_fn(unsigned long data
)
1396 struct request_queue
*q
= (struct request_queue
*)data
;
1397 int nr_pages
= global_page_state(NR_FILE_DIRTY
) +
1398 global_page_state(NR_UNSTABLE_NFS
);
1401 * We want to write everything out, not just down to the dirty
1404 if (bdi_has_dirty_io(&q
->backing_dev_info
))
1405 bdi_start_writeback(&q
->backing_dev_info
, nr_pages
,
1406 WB_REASON_LAPTOP_TIMER
);
1410 * We've spun up the disk and we're in laptop mode: schedule writeback
1411 * of all dirty data a few seconds from now. If the flush is already scheduled
1412 * then push it back - the user is still using the disk.
1414 void laptop_io_completion(struct backing_dev_info
*info
)
1416 mod_timer(&info
->laptop_mode_wb_timer
, jiffies
+ laptop_mode
);
1420 * We're in laptop mode and we've just synced. The sync's writes will have
1421 * caused another writeback to be scheduled by laptop_io_completion.
1422 * Nothing needs to be written back anymore, so we unschedule the writeback.
1424 void laptop_sync_completion(void)
1426 struct backing_dev_info
*bdi
;
1430 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
)
1431 del_timer(&bdi
->laptop_mode_wb_timer
);
1438 * If ratelimit_pages is too high then we can get into dirty-data overload
1439 * if a large number of processes all perform writes at the same time.
1440 * If it is too low then SMP machines will call the (expensive)
1441 * get_writeback_state too often.
1443 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1444 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1448 void writeback_set_ratelimit(void)
1450 unsigned long background_thresh
;
1451 unsigned long dirty_thresh
;
1452 global_dirty_limits(&background_thresh
, &dirty_thresh
);
1453 ratelimit_pages
= dirty_thresh
/ (num_online_cpus() * 32);
1454 if (ratelimit_pages
< 16)
1455 ratelimit_pages
= 16;
1458 static int __cpuinit
1459 ratelimit_handler(struct notifier_block
*self
, unsigned long u
, void *v
)
1461 writeback_set_ratelimit();
1465 static struct notifier_block __cpuinitdata ratelimit_nb
= {
1466 .notifier_call
= ratelimit_handler
,
1471 * Called early on to tune the page writeback dirty limits.
1473 * We used to scale dirty pages according to how total memory
1474 * related to pages that could be allocated for buffers (by
1475 * comparing nr_free_buffer_pages() to vm_total_pages.
1477 * However, that was when we used "dirty_ratio" to scale with
1478 * all memory, and we don't do that any more. "dirty_ratio"
1479 * is now applied to total non-HIGHPAGE memory (by subtracting
1480 * totalhigh_pages from vm_total_pages), and as such we can't
1481 * get into the old insane situation any more where we had
1482 * large amounts of dirty pages compared to a small amount of
1483 * non-HIGHMEM memory.
1485 * But we might still want to scale the dirty_ratio by how
1486 * much memory the box has..
1488 void __init
page_writeback_init(void)
1492 writeback_set_ratelimit();
1493 register_cpu_notifier(&ratelimit_nb
);
1495 shift
= calc_period_shift();
1496 prop_descriptor_init(&vm_completions
, shift
);
1500 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1501 * @mapping: address space structure to write
1502 * @start: starting page index
1503 * @end: ending page index (inclusive)
1505 * This function scans the page range from @start to @end (inclusive) and tags
1506 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1507 * that write_cache_pages (or whoever calls this function) will then use
1508 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1509 * used to avoid livelocking of writeback by a process steadily creating new
1510 * dirty pages in the file (thus it is important for this function to be quick
1511 * so that it can tag pages faster than a dirtying process can create them).
1514 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1516 void tag_pages_for_writeback(struct address_space
*mapping
,
1517 pgoff_t start
, pgoff_t end
)
1519 #define WRITEBACK_TAG_BATCH 4096
1520 unsigned long tagged
;
1523 spin_lock_irq(&mapping
->tree_lock
);
1524 tagged
= radix_tree_range_tag_if_tagged(&mapping
->page_tree
,
1525 &start
, end
, WRITEBACK_TAG_BATCH
,
1526 PAGECACHE_TAG_DIRTY
, PAGECACHE_TAG_TOWRITE
);
1527 spin_unlock_irq(&mapping
->tree_lock
);
1528 WARN_ON_ONCE(tagged
> WRITEBACK_TAG_BATCH
);
1530 /* We check 'start' to handle wrapping when end == ~0UL */
1531 } while (tagged
>= WRITEBACK_TAG_BATCH
&& start
);
1533 EXPORT_SYMBOL(tag_pages_for_writeback
);
1536 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1537 * @mapping: address space structure to write
1538 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1539 * @writepage: function called for each page
1540 * @data: data passed to writepage function
1542 * If a page is already under I/O, write_cache_pages() skips it, even
1543 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1544 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1545 * and msync() need to guarantee that all the data which was dirty at the time
1546 * the call was made get new I/O started against them. If wbc->sync_mode is
1547 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1548 * existing IO to complete.
1550 * To avoid livelocks (when other process dirties new pages), we first tag
1551 * pages which should be written back with TOWRITE tag and only then start
1552 * writing them. For data-integrity sync we have to be careful so that we do
1553 * not miss some pages (e.g., because some other process has cleared TOWRITE
1554 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1555 * by the process clearing the DIRTY tag (and submitting the page for IO).
1557 int write_cache_pages(struct address_space
*mapping
,
1558 struct writeback_control
*wbc
, writepage_t writepage
,
1563 struct pagevec pvec
;
1565 pgoff_t
uninitialized_var(writeback_index
);
1567 pgoff_t end
; /* Inclusive */
1570 int range_whole
= 0;
1573 pagevec_init(&pvec
, 0);
1574 if (wbc
->range_cyclic
) {
1575 writeback_index
= mapping
->writeback_index
; /* prev offset */
1576 index
= writeback_index
;
1583 index
= wbc
->range_start
>> PAGE_CACHE_SHIFT
;
1584 end
= wbc
->range_end
>> PAGE_CACHE_SHIFT
;
1585 if (wbc
->range_start
== 0 && wbc
->range_end
== LLONG_MAX
)
1587 cycled
= 1; /* ignore range_cyclic tests */
1589 if (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
)
1590 tag
= PAGECACHE_TAG_TOWRITE
;
1592 tag
= PAGECACHE_TAG_DIRTY
;
1594 if (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
)
1595 tag_pages_for_writeback(mapping
, index
, end
);
1597 while (!done
&& (index
<= end
)) {
1600 nr_pages
= pagevec_lookup_tag(&pvec
, mapping
, &index
, tag
,
1601 min(end
- index
, (pgoff_t
)PAGEVEC_SIZE
-1) + 1);
1605 for (i
= 0; i
< nr_pages
; i
++) {
1606 struct page
*page
= pvec
.pages
[i
];
1609 * At this point, the page may be truncated or
1610 * invalidated (changing page->mapping to NULL), or
1611 * even swizzled back from swapper_space to tmpfs file
1612 * mapping. However, page->index will not change
1613 * because we have a reference on the page.
1615 if (page
->index
> end
) {
1617 * can't be range_cyclic (1st pass) because
1618 * end == -1 in that case.
1624 done_index
= page
->index
;
1629 * Page truncated or invalidated. We can freely skip it
1630 * then, even for data integrity operations: the page
1631 * has disappeared concurrently, so there could be no
1632 * real expectation of this data interity operation
1633 * even if there is now a new, dirty page at the same
1634 * pagecache address.
1636 if (unlikely(page
->mapping
!= mapping
)) {
1642 if (!PageDirty(page
)) {
1643 /* someone wrote it for us */
1644 goto continue_unlock
;
1647 if (PageWriteback(page
)) {
1648 if (wbc
->sync_mode
!= WB_SYNC_NONE
)
1649 wait_on_page_writeback(page
);
1651 goto continue_unlock
;
1654 BUG_ON(PageWriteback(page
));
1655 if (!clear_page_dirty_for_io(page
))
1656 goto continue_unlock
;
1658 trace_wbc_writepage(wbc
, mapping
->backing_dev_info
);
1659 ret
= (*writepage
)(page
, wbc
, data
);
1660 if (unlikely(ret
)) {
1661 if (ret
== AOP_WRITEPAGE_ACTIVATE
) {
1666 * done_index is set past this page,
1667 * so media errors will not choke
1668 * background writeout for the entire
1669 * file. This has consequences for
1670 * range_cyclic semantics (ie. it may
1671 * not be suitable for data integrity
1674 done_index
= page
->index
+ 1;
1681 * We stop writing back only if we are not doing
1682 * integrity sync. In case of integrity sync we have to
1683 * keep going until we have written all the pages
1684 * we tagged for writeback prior to entering this loop.
1686 if (--wbc
->nr_to_write
<= 0 &&
1687 wbc
->sync_mode
== WB_SYNC_NONE
) {
1692 pagevec_release(&pvec
);
1695 if (!cycled
&& !done
) {
1698 * We hit the last page and there is more work to be done: wrap
1699 * back to the start of the file
1703 end
= writeback_index
- 1;
1706 if (wbc
->range_cyclic
|| (range_whole
&& wbc
->nr_to_write
> 0))
1707 mapping
->writeback_index
= done_index
;
1711 EXPORT_SYMBOL(write_cache_pages
);
1714 * Function used by generic_writepages to call the real writepage
1715 * function and set the mapping flags on error
1717 static int __writepage(struct page
*page
, struct writeback_control
*wbc
,
1720 struct address_space
*mapping
= data
;
1721 int ret
= mapping
->a_ops
->writepage(page
, wbc
);
1722 mapping_set_error(mapping
, ret
);
1727 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1728 * @mapping: address space structure to write
1729 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1731 * This is a library function, which implements the writepages()
1732 * address_space_operation.
1734 int generic_writepages(struct address_space
*mapping
,
1735 struct writeback_control
*wbc
)
1737 struct blk_plug plug
;
1740 /* deal with chardevs and other special file */
1741 if (!mapping
->a_ops
->writepage
)
1744 blk_start_plug(&plug
);
1745 ret
= write_cache_pages(mapping
, wbc
, __writepage
, mapping
);
1746 blk_finish_plug(&plug
);
1750 EXPORT_SYMBOL(generic_writepages
);
1752 int do_writepages(struct address_space
*mapping
, struct writeback_control
*wbc
)
1756 if (wbc
->nr_to_write
<= 0)
1758 if (mapping
->a_ops
->writepages
)
1759 ret
= mapping
->a_ops
->writepages(mapping
, wbc
);
1761 ret
= generic_writepages(mapping
, wbc
);
1766 * write_one_page - write out a single page and optionally wait on I/O
1767 * @page: the page to write
1768 * @wait: if true, wait on writeout
1770 * The page must be locked by the caller and will be unlocked upon return.
1772 * write_one_page() returns a negative error code if I/O failed.
1774 int write_one_page(struct page
*page
, int wait
)
1776 struct address_space
*mapping
= page
->mapping
;
1778 struct writeback_control wbc
= {
1779 .sync_mode
= WB_SYNC_ALL
,
1783 BUG_ON(!PageLocked(page
));
1786 wait_on_page_writeback(page
);
1788 if (clear_page_dirty_for_io(page
)) {
1789 page_cache_get(page
);
1790 ret
= mapping
->a_ops
->writepage(page
, &wbc
);
1791 if (ret
== 0 && wait
) {
1792 wait_on_page_writeback(page
);
1793 if (PageError(page
))
1796 page_cache_release(page
);
1802 EXPORT_SYMBOL(write_one_page
);
1805 * For address_spaces which do not use buffers nor write back.
1807 int __set_page_dirty_no_writeback(struct page
*page
)
1809 if (!PageDirty(page
))
1810 return !TestSetPageDirty(page
);
1815 * Helper function for set_page_dirty family.
1816 * NOTE: This relies on being atomic wrt interrupts.
1818 void account_page_dirtied(struct page
*page
, struct address_space
*mapping
)
1820 if (mapping_cap_account_dirty(mapping
)) {
1821 __inc_zone_page_state(page
, NR_FILE_DIRTY
);
1822 __inc_zone_page_state(page
, NR_DIRTIED
);
1823 __inc_bdi_stat(mapping
->backing_dev_info
, BDI_RECLAIMABLE
);
1824 __inc_bdi_stat(mapping
->backing_dev_info
, BDI_DIRTIED
);
1825 task_io_account_write(PAGE_CACHE_SIZE
);
1828 EXPORT_SYMBOL(account_page_dirtied
);
1831 * Helper function for set_page_writeback family.
1832 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1835 void account_page_writeback(struct page
*page
)
1837 inc_zone_page_state(page
, NR_WRITEBACK
);
1839 EXPORT_SYMBOL(account_page_writeback
);
1842 * For address_spaces which do not use buffers. Just tag the page as dirty in
1845 * This is also used when a single buffer is being dirtied: we want to set the
1846 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1847 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1849 * Most callers have locked the page, which pins the address_space in memory.
1850 * But zap_pte_range() does not lock the page, however in that case the
1851 * mapping is pinned by the vma's ->vm_file reference.
1853 * We take care to handle the case where the page was truncated from the
1854 * mapping by re-checking page_mapping() inside tree_lock.
1856 int __set_page_dirty_nobuffers(struct page
*page
)
1858 if (!TestSetPageDirty(page
)) {
1859 struct address_space
*mapping
= page_mapping(page
);
1860 struct address_space
*mapping2
;
1865 spin_lock_irq(&mapping
->tree_lock
);
1866 mapping2
= page_mapping(page
);
1867 if (mapping2
) { /* Race with truncate? */
1868 BUG_ON(mapping2
!= mapping
);
1869 WARN_ON_ONCE(!PagePrivate(page
) && !PageUptodate(page
));
1870 account_page_dirtied(page
, mapping
);
1871 radix_tree_tag_set(&mapping
->page_tree
,
1872 page_index(page
), PAGECACHE_TAG_DIRTY
);
1874 spin_unlock_irq(&mapping
->tree_lock
);
1875 if (mapping
->host
) {
1876 /* !PageAnon && !swapper_space */
1877 __mark_inode_dirty(mapping
->host
, I_DIRTY_PAGES
);
1883 EXPORT_SYMBOL(__set_page_dirty_nobuffers
);
1886 * When a writepage implementation decides that it doesn't want to write this
1887 * page for some reason, it should redirty the locked page via
1888 * redirty_page_for_writepage() and it should then unlock the page and return 0
1890 int redirty_page_for_writepage(struct writeback_control
*wbc
, struct page
*page
)
1892 wbc
->pages_skipped
++;
1893 return __set_page_dirty_nobuffers(page
);
1895 EXPORT_SYMBOL(redirty_page_for_writepage
);
1900 * For pages with a mapping this should be done under the page lock
1901 * for the benefit of asynchronous memory errors who prefer a consistent
1902 * dirty state. This rule can be broken in some special cases,
1903 * but should be better not to.
1905 * If the mapping doesn't provide a set_page_dirty a_op, then
1906 * just fall through and assume that it wants buffer_heads.
1908 int set_page_dirty(struct page
*page
)
1910 struct address_space
*mapping
= page_mapping(page
);
1912 if (likely(mapping
)) {
1913 int (*spd
)(struct page
*) = mapping
->a_ops
->set_page_dirty
;
1915 * readahead/lru_deactivate_page could remain
1916 * PG_readahead/PG_reclaim due to race with end_page_writeback
1917 * About readahead, if the page is written, the flags would be
1918 * reset. So no problem.
1919 * About lru_deactivate_page, if the page is redirty, the flag
1920 * will be reset. So no problem. but if the page is used by readahead
1921 * it will confuse readahead and make it restart the size rampup
1922 * process. But it's a trivial problem.
1924 ClearPageReclaim(page
);
1927 spd
= __set_page_dirty_buffers
;
1929 return (*spd
)(page
);
1931 if (!PageDirty(page
)) {
1932 if (!TestSetPageDirty(page
))
1937 EXPORT_SYMBOL(set_page_dirty
);
1940 * set_page_dirty() is racy if the caller has no reference against
1941 * page->mapping->host, and if the page is unlocked. This is because another
1942 * CPU could truncate the page off the mapping and then free the mapping.
1944 * Usually, the page _is_ locked, or the caller is a user-space process which
1945 * holds a reference on the inode by having an open file.
1947 * In other cases, the page should be locked before running set_page_dirty().
1949 int set_page_dirty_lock(struct page
*page
)
1954 ret
= set_page_dirty(page
);
1958 EXPORT_SYMBOL(set_page_dirty_lock
);
1961 * Clear a page's dirty flag, while caring for dirty memory accounting.
1962 * Returns true if the page was previously dirty.
1964 * This is for preparing to put the page under writeout. We leave the page
1965 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1966 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1967 * implementation will run either set_page_writeback() or set_page_dirty(),
1968 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1971 * This incoherency between the page's dirty flag and radix-tree tag is
1972 * unfortunate, but it only exists while the page is locked.
1974 int clear_page_dirty_for_io(struct page
*page
)
1976 struct address_space
*mapping
= page_mapping(page
);
1978 BUG_ON(!PageLocked(page
));
1980 if (mapping
&& mapping_cap_account_dirty(mapping
)) {
1982 * Yes, Virginia, this is indeed insane.
1984 * We use this sequence to make sure that
1985 * (a) we account for dirty stats properly
1986 * (b) we tell the low-level filesystem to
1987 * mark the whole page dirty if it was
1988 * dirty in a pagetable. Only to then
1989 * (c) clean the page again and return 1 to
1990 * cause the writeback.
1992 * This way we avoid all nasty races with the
1993 * dirty bit in multiple places and clearing
1994 * them concurrently from different threads.
1996 * Note! Normally the "set_page_dirty(page)"
1997 * has no effect on the actual dirty bit - since
1998 * that will already usually be set. But we
1999 * need the side effects, and it can help us
2002 * We basically use the page "master dirty bit"
2003 * as a serialization point for all the different
2004 * threads doing their things.
2006 if (page_mkclean(page
))
2007 set_page_dirty(page
);
2009 * We carefully synchronise fault handlers against
2010 * installing a dirty pte and marking the page dirty
2011 * at this point. We do this by having them hold the
2012 * page lock at some point after installing their
2013 * pte, but before marking the page dirty.
2014 * Pages are always locked coming in here, so we get
2015 * the desired exclusion. See mm/memory.c:do_wp_page()
2016 * for more comments.
2018 if (TestClearPageDirty(page
)) {
2019 dec_zone_page_state(page
, NR_FILE_DIRTY
);
2020 dec_bdi_stat(mapping
->backing_dev_info
,
2026 return TestClearPageDirty(page
);
2028 EXPORT_SYMBOL(clear_page_dirty_for_io
);
2030 int test_clear_page_writeback(struct page
*page
)
2032 struct address_space
*mapping
= page_mapping(page
);
2036 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
2037 unsigned long flags
;
2039 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
2040 ret
= TestClearPageWriteback(page
);
2042 radix_tree_tag_clear(&mapping
->page_tree
,
2044 PAGECACHE_TAG_WRITEBACK
);
2045 if (bdi_cap_account_writeback(bdi
)) {
2046 __dec_bdi_stat(bdi
, BDI_WRITEBACK
);
2047 __bdi_writeout_inc(bdi
);
2050 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
2052 ret
= TestClearPageWriteback(page
);
2055 dec_zone_page_state(page
, NR_WRITEBACK
);
2056 inc_zone_page_state(page
, NR_WRITTEN
);
2061 int test_set_page_writeback(struct page
*page
)
2063 struct address_space
*mapping
= page_mapping(page
);
2067 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
2068 unsigned long flags
;
2070 spin_lock_irqsave(&mapping
->tree_lock
, flags
);
2071 ret
= TestSetPageWriteback(page
);
2073 radix_tree_tag_set(&mapping
->page_tree
,
2075 PAGECACHE_TAG_WRITEBACK
);
2076 if (bdi_cap_account_writeback(bdi
))
2077 __inc_bdi_stat(bdi
, BDI_WRITEBACK
);
2079 if (!PageDirty(page
))
2080 radix_tree_tag_clear(&mapping
->page_tree
,
2082 PAGECACHE_TAG_DIRTY
);
2083 radix_tree_tag_clear(&mapping
->page_tree
,
2085 PAGECACHE_TAG_TOWRITE
);
2086 spin_unlock_irqrestore(&mapping
->tree_lock
, flags
);
2088 ret
= TestSetPageWriteback(page
);
2091 account_page_writeback(page
);
2095 EXPORT_SYMBOL(test_set_page_writeback
);
2098 * Return true if any of the pages in the mapping are marked with the
2101 int mapping_tagged(struct address_space
*mapping
, int tag
)
2103 return radix_tree_tagged(&mapping
->page_tree
, tag
);
2105 EXPORT_SYMBOL(mapping_tagged
);